app/vmselect/promql: do not take into account the previous point before time window in square brackets for min_over_time, max_over_time, rollup_first and rollup_last functions

This makes the behaviour for these functions similar to Prometheus when processing broken time series with irregular data points
like `gitlab_runner_jobs`. See https://gitlab.com/gitlab-org/gitlab-exporter/issues/50 for details.

.github/workflows/github-pages.yml

@@ -0,0 +1,30 @@
+name: github-pages
+on:
+  push:
+    paths:
+      - 'docs/*.md'
+      - 'README.md'
+    branches:
+      - master
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@master
+      - name: publish
+        shell: bash
+        env:
+          TOKEN: ${{secrets.CI_TOKEN}}
+        run: |
+          git clone https://vika:${TOKEN}@github.com/VictoriaMetrics/VictoriaMetrics.github.io.git gpages
+          cp docs/*.md gpages
+          cp README.md gpages
+          cd gpages
+          git config --local user.email "info@victoriametrics.com"
+          git config --local user.name "Vika"
+          git add "*.md"
+          git commit -m "update github pages"
+          remote_repo="https://vika:${TOKEN}@github.com/VictoriaMetrics/VictoriaMetrics.github.io.git"
+          git push "${remote_repo}"
+          cd ..
+          rm -rf gpages

.github/workflows/main.yml

@@ -0,0 +1,51 @@
+name: main
+on:
+  push:
+    paths-ignore:
+      - 'docs/**'
+      - '**.md'
+  pull_request:
+    paths-ignore:
+      - 'docs/**'
+      - '**.md'
+jobs:
+  build:
+    name: Build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Setup Go
+        uses: actions/setup-go@v1
+        with:
+          go-version: 1.13
+        id: go
+      - name: Code checkout
+        uses: actions/checkout@v1
+      - name: Dependencies
+        env:
+          GO111MODULE: off
+        run: |
+          go get -v golang.org/x/lint/golint
+          go get -u github.com/kisielk/errcheck
+      - name: Build
+        env:
+          GO111MODULE: on
+        run: |
+          export PATH=$PATH:$(go env GOPATH)/bin # temporary fix. See https://github.com/actions/setup-go/issues/14
+          make check-all
+          git diff --exit-code
+          make test-full
+          make test-pure
+          make test-full-386
+          make victoria-metrics
+          make victoria-metrics-pure
+          make victoria-metrics-arm
+          make victoria-metrics-arm64
+          make vmutils
+          GOOS=freebsd go build -mod=vendor ./app/victoria-metrics
+          GOOS=darwin go build -mod=vendor ./app/victoria-metrics
+      - name: Publish coverage
+        uses: codecov/codecov-action@v1.0.4
+        with:
+          token: ${{secrets.CODECOV_TOKEN}}
+          file: ./coverage.txt
+

.github/workflows/wiki.yml

@@ -0,0 +1,29 @@
+name: wiki
+on:
+  push:
+    paths:
+      - 'docs/*.md'
+    branches:
+      - master
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@master
+      - name: publish
+        shell: bash
+        env:
+          TOKEN: ${{secrets.CI_TOKEN}}
+        run: |
+          cd docs
+          git clone https://vika:${TOKEN}@github.com/VictoriaMetrics/VictoriaMetrics.wiki.git wiki
+          find ./ -name '*.md' -exec cp -prv '{}' 'wiki' ';'
+          cd wiki
+          git config --local user.email "info@victoriametrics.com"
+          git config --local user.name "Vika"
+          git add "*.md"
+          git commit -m "update wiki pages"
+          remote_repo="https://vika:${TOKEN}@github.com/VictoriaMetrics/VictoriaMetrics.wiki.git"
+          git push "${remote_repo}"
+          cd ..
+          rm -rf wiki

.gitignore

@@ -1,3 +1,4 @@
+/tmp
 /tags
 /pkg
 *.pprof

.travis.yml

@@ -1,26 +0,0 @@
-language: go
-
-go:
-  - 1.12.x
-
-install: make
-
-env:
-  - GO111MODULE=on
-
-before_install:
-  - GO111MODULE=off go get -v golang.org/x/lint/golint
-  - GO111MODULE=off go get -u github.com/kisielk/errcheck
-
-script:
-  - make check_all
-  - git diff --exit-code
-  - make test-full
-  - make test-pure
-  - make victoria-metrics
-  - make victoria-metrics-pure
-  - make victoria-metrics-arm
-  - make victoria-metrics-arm64
-
-after_success:
-  - bash <(curl -s https://codecov.io/bash)

LICENSE

@@ -175,7 +175,7 @@
 
    END OF TERMS AND CONDITIONS
 
-   Copyright 2019 VictoriaMetrics, Inc.
+   Copyright 2019-2020 VictoriaMetrics, Inc.
 
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.

Makefile

@@ -1,7 +1,7 @@
 PKG_PREFIX := github.com/VictoriaMetrics/VictoriaMetrics
 
 BUILDINFO_TAG ?= $(shell echo $$(git describe --long --all | tr '/' '-')$$( \
-	      git diff-index --quiet HEAD -- || echo '-dirty-'$$(git diff-index -u HEAD | sha1sum | grep -oP '^.{8}')))
+	      git diff-index --quiet HEAD -- || echo '-dirty-'$$(git diff-index -u HEAD | openssl sha1 | cut -c 10-17)))
 
 PKG_TAG ?= $(shell git tag -l --points-at HEAD)
 ifeq ($(PKG_TAG),)
@@ -19,12 +19,36 @@ include deployment/*/Makefile
 clean:
 	rm -rf bin/*
 
-publish: publish-victoria-metrics
+publish: \
+	publish-victoria-metrics \
+	publish-vmbackup \
+	publish-vmrestore
 
-package: package-victoria-metrics
+package: \
+	package-victoria-metrics \
+	package-vmbackup \
+	package-vmrestore
 
-release: victoria-metrics-prod
-	cd bin && tar czf victoria-metrics-$(PKG_TAG).tar.gz victoria-metrics-prod
+vmutils: \
+	vmbackup \
+	vmrestore
+
+release: \
+	release-victoria-metrics \
+	release-vmutils
+
+release-victoria-metrics: victoria-metrics-prod
+	cd bin && tar czf victoria-metrics-$(PKG_TAG).tar.gz victoria-metrics-prod && \
+		sha256sum victoria-metrics-$(PKG_TAG).tar.gz > victoria-metrics-$(PKG_TAG)_checksums.txt
+
+release-vmutils: \
+	vmbackup-prod \
+	vmrestore-prod
+	cd bin && tar czf vmutils-$(PKG_TAG).tar.gz vmbackup-prod vmrestore-prod && \
+		sha256sum vmutils-$(PKG_TAG).tar.gz > vmutils-$(PKG_TAG)_checksums.txt
+
+pprof-cpu:
+	go tool pprof -trim_path=github.com/VictoriaMetrics/VictoriaMetrics@ $(PPROF_FILE)
 
 fmt:
 	GO111MODULE=on gofmt -l -w -s ./lib
@@ -39,18 +63,20 @@ lint: install-golint
 	golint app/...
 
 install-golint:
-	which golint || GO111MODULE=off go get -u github.com/golang/lint/golint
+	which golint || GO111MODULE=off go get -u golang.org/x/lint/golint
 
 errcheck: install-errcheck
 	errcheck -exclude=errcheck_excludes.txt ./lib/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vminsert/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vmselect/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vmstorage/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmbackup/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmrestore/...
 
 install-errcheck:
 	which errcheck || GO111MODULE=off go get -u github.com/kisielk/errcheck
 
-check_all: fmt vet lint errcheck golangci-lint
+check-all: fmt vet lint errcheck golangci-lint
 
 test:
 	GO111MODULE=on go test -tags=integration -mod=vendor ./lib/... ./app/...
@@ -61,6 +87,9 @@ test-pure:
 test-full:
 	GO111MODULE=on go test -tags=integration -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
 
+test-full-386:
+	GO111MODULE=on GOARCH=386 go test -tags=integration -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
+
 benchmark:
 	GO111MODULE=on go test -mod=vendor -bench=. ./lib/...
 	GO111MODULE=on go test -mod=vendor -bench=. ./app/...
@@ -75,6 +104,12 @@ vendor-update:
 	GO111MODULE=on go mod tidy
 	GO111MODULE=on go mod vendor
 
+app-local:
+	CGO_ENABLED=1 GO111MODULE=on go build $(RACE) -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/$(APP_NAME)$(RACE) $(PKG_PREFIX)/app/$(APP_NAME)
+
+app-local-pure:
+	CGO_ENABLED=0 GO111MODULE=on go build $(RACE) -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/$(APP_NAME)-pure$(RACE) $(PKG_PREFIX)/app/$(APP_NAME)
+
 quicktemplate-gen: install-qtc
 	qtc
 
@@ -83,7 +118,7 @@ install-qtc:
 
 
 golangci-lint: install-golangci-lint
-	golangci-lint run --exclude '(SA4003|SA1019):' -D errcheck
+	golangci-lint run --exclude '(SA4003|SA1019):' -D errcheck -D structcheck
 
 install-golangci-lint:
 	which golangci-lint || GO111MODULE=off go get -u github.com/golangci/golangci-lint/cmd/golangci-lint

README.md

@@ -1,50 +1,65 @@
 [![Latest Release](https://img.shields.io/github/release/VictoriaMetrics/VictoriaMetrics.svg?style=flat-square)](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/latest)
+[![Docker Pulls](https://img.shields.io/docker/pulls/victoriametrics/victoria-metrics.svg?maxAge=604800)](https://hub.docker.com/r/victoriametrics/victoria-metrics)
 [![Slack](https://img.shields.io/badge/join%20slack-%23victoriametrics-brightgreen.svg)](http://slack.victoriametrics.com/)
 [![GitHub license](https://img.shields.io/github/license/VictoriaMetrics/VictoriaMetrics.svg)](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/LICENSE)
 [![Go Report](https://goreportcard.com/badge/github.com/VictoriaMetrics/VictoriaMetrics)](https://goreportcard.com/report/github.com/VictoriaMetrics/VictoriaMetrics)
-[![Build Status](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics.svg?branch=master)](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics)
+[![Build Status](https://github.com/VictoriaMetrics/VictoriaMetrics/workflows/main/badge.svg)](https://github.com/VictoriaMetrics/VictoriaMetrics/actions)
 [![codecov](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics/branch/master/graph/badge.svg)](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics)
 
 <img alt="Victoria Metrics" src="logo.png">
 
-## Single-node VictoriaMetrics
+## VictoriaMetrics
 
 VictoriaMetrics is fast, cost-effective and scalable time-series database. It can be used as long-term remote storage for Prometheus.
 It is available in [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases),
 [docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) and
-in [source code](https://github.com/VictoriaMetrics/VictoriaMetrics).
+in [source code](https://github.com/VictoriaMetrics/VictoriaMetrics). Just download VictoriaMetrics and see [how to start it](#how-to-start-victoriametrics).
 
 Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
 
 
+## Case studies
+
+* [Wix.com](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/CaseStudies#wixcom)
+* [Wedos.com](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/CaseStudies#wedoscom)
+* [Dreamteam](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/CaseStudies#dreamteam)
+
+
 ## Prominent features
 
 * Supports [Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/), so it can be used as Prometheus drop-in replacement in Grafana.
-  Additionally, VictoriaMetrics extends PromQL with opt-in [useful features](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/ExtendedPromQL).
-* Global query view. Multiple Prometheus instances may write data into VictoriaMetrics. Later this data may be used in a single query.
+  VictoriaMetrics implements [MetricsQL](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/ExtendedPromQL) query language, which is inspired by PromQL.
+* Supports global query view. Multiple Prometheus instances may write data into VictoriaMetrics. Later this data may be used in a single query.
 * High performance and good scalability for both [inserts](https://medium.com/@valyala/high-cardinality-tsdb-benchmarks-victoriametrics-vs-timescaledb-vs-influxdb-13e6ee64dd6b)
   and [selects](https://medium.com/@valyala/when-size-matters-benchmarking-victoriametrics-vs-timescale-and-influxdb-6035811952d4).
   [Outperforms InfluxDB and TimescaleDB by up to 20x](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
 * [Uses 10x less RAM than InfluxDB](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) when working with millions of unique time series (aka high cardinality).
+* Optimized for time series with high churn rate. Think about [prometheus-operator](https://github.com/coreos/prometheus-operator) metrics from frequent deployments in Kubernetes.
 * High data compression, so [up to 70x more data points](https://medium.com/@valyala/when-size-matters-benchmarking-victoriametrics-vs-timescale-and-influxdb-6035811952d4)
   may be crammed into limited storage comparing to TimescaleDB.
 * Optimized for storage with high-latency IO and low IOPS (HDD and network storage in AWS, Google Cloud, Microsoft Azure, etc). See [graphs from these benchmarks](https://medium.com/@valyala/high-cardinality-tsdb-benchmarks-victoriametrics-vs-timescaledb-vs-influxdb-13e6ee64dd6b).
-* A single-node VictoriaMetrics may substitute moderately sized clusters built with competing solutions such as Thanos, Uber M3, Cortex, InfluxDB or TimescaleDB.
-  See [vertical scalability benchmarks](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae)
-  and [comparing Thanos to VictoriaMetrics cluster](https://medium.com/@valyala/comparing-thanos-to-victoriametrics-cluster-b193bea1683).
+* A single-node VictoriaMetrics may substitute moderately sized clusters built with competing solutions such as Thanos, M3DB, Cortex, InfluxDB or TimescaleDB.
+  See [vertical scalability benchmarks](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae),
+  [comparing Thanos to VictoriaMetrics cluster](https://medium.com/@valyala/comparing-thanos-to-victoriametrics-cluster-b193bea1683)
+  and [Remote Write Storage Wars](https://promcon.io/2019-munich/talks/remote-write-storage-wars/) talk
+  from [PromCon 2019](https://promcon.io/2019-munich/talks/remote-write-storage-wars/).
 * Easy operation:
-  * VictoriaMetrics consists of a single executable without external dependencies.
+  * VictoriaMetrics consists of a single [small executable](https://medium.com/@valyala/stripping-dependency-bloat-in-victoriametrics-docker-image-983fb5912b0d) without external dependencies.
   * All the configuration is done via explicit command-line flags with reasonable defaults.
   * All the data is stored in a single directory pointed by `-storageDataPath` flag.
-  * Easy backups from [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
-* Storage is protected from corruption on unclean shutdown (i.e. hardware reset or `kill -9`) thanks to [the storage architecture](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
-* Supports metrics' ingestion and backfilling via the following protocols:
+  * Easy and fast backups from [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
+  to S3 or GCS with [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md) / [vmrestore](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md).
+  See [this article](https://medium.com/@valyala/speeding-up-backups-for-big-time-series-databases-533c1a927883) for more details.
+* Storage is protected from corruption on unclean shutdown (i.e. OOM, hardware reset or `kill -9`) thanks to [the storage architecture](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
+* Supports metrics' ingestion and [backfilling](#backfilling) via the following protocols:
   * [Prometheus remote write API](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write)
-  * [InfluxDB line protocol](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/)
-  * [Graphite plaintext protocol](https://graphite.readthedocs.io/en/latest/feeding-carbon.html) with [tags](https://graphite.readthedocs.io/en/latest/tags.html#carbon)
+  * [InfluxDB line protocol](#how-to-send-data-from-influxdb-compatible-agents-such-as-telegraf)
+  * [Graphite plaintext protocol](#how-to-send-data-from-graphite-compatible-agents-such-as-statsd) with [tags](https://graphite.readthedocs.io/en/latest/tags.html#carbon)
     if `-graphiteListenAddr` is set.
-  * [OpenTSDB put message](http://opentsdb.net/docs/build/html/api_telnet/put.html) if `-opentsdbListenAddr` is set.
-* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars and industrial telemetry.
+  * [OpenTSDB put message](#sending-data-via-telnet-put-protocol) if `-opentsdbListenAddr` is set.
+  * [HTTP OpenTSDB /api/put requests](#sending-opentsdb-data-via-http-apiput-requests) if `-opentsdbHTTPListenAddr` is set.
+  * [/api/v1/import](#how-to-import-time-series-data)
+* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars, industrial telemetry, financial data and various Enterprise workloads.
 * Has open source [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
 
 
@@ -62,6 +77,7 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [How to send data from Graphite-compatible agents such as StatsD?](#how-to-send-data-from-graphite-compatible-agents-such-as-statsd)
   - [Querying Graphite data](#querying-graphite-data)
   - [How to send data from OpenTSDB-compatible agents?](#how-to-send-data-from-opentsdb-compatible-agents)
+  - [Prometheus querying API usage](#prometheus-querying-api-usage)
   - [How to build from sources](#how-to-build-from-sources)
     - [Development build](#development-build)
     - [Production build](#production-build)
@@ -74,6 +90,7 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [How to work with snapshots?](#how-to-work-with-snapshots)
   - [How to delete time series?](#how-to-delete-time-series)
   - [How to export time series?](#how-to-export-time-series)
+  - [How to import time series data?](#how-to-import-time-series-data)
   - [Federation](#federation)
   - [Capacity planning](#capacity-planning)
   - [High availability](#high-availability)
@@ -86,6 +103,9 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [Tuning](#tuning)
   - [Monitoring](#monitoring)
   - [Troubleshooting](#troubleshooting)
+  - [Backfilling](#backfilling)
+  - [Profiling](#profiling)
+- [Integrations](#integrations)
 - [Roadmap](#roadmap)
 - [Contacts](#contacts)
 - [Community and contributions](#community-and-contributions)
@@ -104,11 +124,12 @@ or [docker image](https://hub.docker.com/r/victoriametrics/victoria-metrics/) wi
 
 The following command-line flags are used the most:
 
-* `-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory.
-* `-retentionPeriod` - retention period in months for the data. Older data is automatically deleted.
+* `-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory. Default path is `victoria-metrics-data` in current working directory.
+* `-retentionPeriod` - retention period in months for the data. Older data is automatically deleted. Default period is 1 month.
 * `-httpListenAddr` - TCP address to listen to for http requests. By default, it listens port `8428` on all the network interfaces.
 * `-graphiteListenAddr` - TCP and UDP address to listen to for Graphite data. By default, it is disabled.
-* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data. By default, it is disabled.
+* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data over telnet protocol. By default, it is disabled.
+* `-opentsdbHTTPListenAddr` - TCP address to listen to for HTTP OpenTSDB data over `/api/put`. By default, it is disabled.
 
 Pass `-help` to see all the available flags with description and default values.
 
@@ -117,14 +138,13 @@ It is recommended setting up [monitoring](#monitoring) for VictoriaMetrics.
 
 ### Prometheus setup
 
-Add the following lines to Prometheus config file (it is usually located at `/etc/prometheus/prometheus.yml`):
+Prometheus must be configured with [remote_write](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write) 
+in order to send data to VictoriaMetrics. Add the following lines 
+to Prometheus config file (it is usually located at `/etc/prometheus/prometheus.yml`):
 
 ```yml
 remote_write:
   - url: http://<victoriametrics-addr>:8428/api/v1/write
-    queue_config:
-      max_samples_per_send: 10000
-      max_shards: 100
 ```
 
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
@@ -151,8 +171,24 @@ This instructs Prometheus to add `datacenter=dc-123` label to each time series s
 The label name may be arbitrary - `datacenter` is just an example. The label value must be unique
 across Prometheus instances, so those time series may be filtered and grouped by this label.
 
+For highly loaded Prometheus instances (400k+ samples per second)
+the following tuning may be applied:
+```
+remote_write:
+  - url: http://<victoriametrics-addr>:8428/api/v1/write
+    queue_config:
+      max_samples_per_send: 10000
+      capacity: 20000
+      max_shards: 30
+```
 
-It is recommended upgrading Prometheus to [v2.10.0](https://github.com/prometheus/prometheus/releases) or newer,
+Using remote write increases memory usage for Prometheus up to ~25%
+and depends on the shape of data. If you are experiencing issues with
+too high memory consumption try to lower `max_samples_per_send` 
+and `capacity` params (keep in mind that these two params are tightly connected).
+Read more about tuning remote write for Prometheus [here](https://prometheus.io/docs/practices/remote_write).
+
+It is recommended upgrading Prometheus to [v2.12.0](https://github.com/prometheus/prometheus/releases) or newer,
 since the previous versions may have issues with `remote_write`.
 
 
@@ -167,7 +203,7 @@ http://<victoriametrics-addr>:8428
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
 
 Then build graphs with the created datasource using [Prometheus query language](https://prometheus.io/docs/prometheus/latest/querying/basics/).
-VictoriaMetrics supports native PromQL and [extends it with useful features](ExtendedPromQL).
+VictoriaMetrics supports native PromQL and [extends it with useful features](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/ExtendedPromQL).
 
 
 ### How to upgrade VictoriaMetrics?
@@ -182,6 +218,9 @@ Follow the following steps during the upgrade:
 2) Wait until the process stops. This can take a few seconds.
 3) Start the upgraded VictoriaMetrics.
 
+Prometheus doesn't drop data during VictoriaMetrics restart.
+See [this article](https://grafana.com/blog/2019/03/25/whats-new-in-prometheus-2.8-wal-based-remote-write/) for details.
+
 
 ### How to apply new config to VictoriaMetrics?
 
@@ -191,6 +230,9 @@ VictoriaMetrics must be restarted for applying new config:
 2) Wait until the process stops. This can take a few seconds.
 3) Start VictoriaMetrics with the new config.
 
+Prometheus doesn't drop data during VictoriaMetrics restart.
+See [this article](https://grafana.com/blog/2019/03/25/whats-new-in-prometheus-2.8-wal-based-remote-write/) for details.
+
 
 ### How to send data from InfluxDB-compatible agents such as [Telegraf](https://www.influxdata.com/time-series-platform/telegraf/)?
 
@@ -205,10 +247,11 @@ For instance, put the following lines into `Telegraf` config, so it sends data t
 Do not forget substituting `<victoriametrics-addr>` with the real address where VictoriaMetrics runs.
 
 VictoriaMetrics maps Influx data using the following rules:
-* [`db` query arg](https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint) is mapped into `db` label value.
+* [`db` query arg](https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint) is mapped into `db` label value
+  unless `db` tag exists in the Influx line.
 * Field names are mapped to time series names prefixed with `{measurement}{separator}` value,
   where `{separator}` equals to `_` by default. It can be changed with `-influxMeasurementFieldSeparator` command-line flag.
-  See also `-influxSkipSingleField` command-line flag.
+  See also `-influxSkipSingleField` command-line flag. If `{measurement}` is empty, then time series names correspond to field names.
 * Field values are mapped to time series values.
 * Tags are mapped to Prometheus labels as-is.
 
@@ -221,8 +264,8 @@ foo,tag1=value1,tag2=value2 field1=12,field2=40
 is converted into the following Prometheus data points:
 
 ```
-foo.field1{tag1="value1", tag2="value2"} 12
-foo.field2{tag1="value1", tag2="value2"} 40
+foo_field1{tag1="value1", tag2="value2"} 12
+foo_field2{tag1="value1", tag2="value2"} 40
 ```
 
 Example for writing data with [Influx line protocol](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/)
@@ -236,16 +279,19 @@ An arbitrary number of lines delimited by '\n' may be sent in a single request.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match={__name__=~"measurement_.*"}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
 
 ```
-{"metric":{"__name__":"measurement.field1","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560272508147]}
-{"metric":{"__name__":"measurement.field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
+{"metric":{"__name__":"measurement_field1","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560272508147]}
+{"metric":{"__name__":"measurement_field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
 ```
 
+Note that Influx line protocol expects [timestamps in *nanoseconds* by default](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/#timestamp),
+while VictoriaMetrics stores them with *milliseconds* precision.
+
 
 ### How to send data from Graphite-compatible agents such as [StatsD](https://github.com/etsy/statsd)?
 
@@ -271,7 +317,7 @@ An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match=foo.bar.baz'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -284,15 +330,20 @@ The `/api/v1/export` endpoint should return the following response:
 ### Querying Graphite data
 
 Data sent to VictoriaMetrics via `Graphite plaintext protocol` may be read either via
-[Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/)
+[Prometheus querying API](#prometheus-querying-api-usage)
 or via [go-graphite/carbonapi](https://github.com/go-graphite/carbonapi/blob/master/cmd/carbonapi/carbonapi.example.prometheus.yaml).
 
 
 
 ### How to send data from OpenTSDB-compatible agents?
 
+VictoriaMetrics supports [telnet put protocol](http://opentsdb.net/docs/build/html/api_telnet/put.html)
+and [HTTP /api/put requests](http://opentsdb.net/docs/build/html/api_http/put.html) for ingesting OpenTSDB data.
+
+#### Sending data via `telnet put` protocol
+
 1) Enable OpenTSDB receiver in VictoriaMetrics by setting `-opentsdbListenAddr` command line flag. For instance,
-the following command will enable OpenTSDB receiver in VictoriaMetrics on TCP and UDP port `4242`:
+the following command enables OpenTSDB receiver in VictoriaMetrics on TCP and UDP port `4242`:
 
 ```
 /path/to/victoria-metrics-prod -opentsdbListenAddr=:4242
@@ -311,7 +362,7 @@ An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match=foo.bar.baz'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -321,6 +372,69 @@ The `/api/v1/export` endpoint should return the following response:
 ```
 
 
+#### Sending OpenTSDB data via HTTP `/api/put` requests
+
+1) Enable HTTP server for OpenTSDB `/api/put` requests by setting `-opentsdbHTTPListenAddr` command line flag. For instance,
+the following command enables OpenTSDB HTTP server on port `4242`:
+
+```
+/path/to/victoria-metrics-prod -opentsdbHTTPListenAddr=:4242
+```
+
+2) Send data to the given address from OpenTSDB-compatible agents.
+
+Example for writing a single data point:
+
+```
+curl -H 'Content-Type: application/json' -d '{"metric":"x.y.z","value":45.34,"tags":{"t1":"v1","t2":"v2"}}' http://localhost:4242/api/put
+```
+
+Example for writing multiple data points in a single request:
+
+```
+curl -H 'Content-Type: application/json' -d '[{"metric":"foo","value":45.34},{"metric":"bar","value":43}]' http://localhost:4242/api/put
+```
+
+After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
+
+```
+curl -G 'http://localhost:8428/api/v1/export' -d 'match[]=x.y.z' -d 'match[]=foo' -d 'match[]=bar'
+```
+
+The `/api/v1/export` endpoint should return the following response:
+
+```
+{"metric":{"__name__":"foo"},"values":[45.34],"timestamps":[1566464846000]}
+{"metric":{"__name__":"bar"},"values":[43],"timestamps":[1566464846000]}
+{"metric":{"__name__":"x.y.z","t1":"v1","t2":"v2"},"values":[45.34],"timestamps":[1566464763000]}
+```
+
+
+### Prometheus querying API usage
+
+VictoriaMetrics supports the following handlers from [Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/):
+
+* [/api/v1/query](https://prometheus.io/docs/prometheus/latest/querying/api/#instant-queries)
+* [/api/v1/query_range](https://prometheus.io/docs/prometheus/latest/querying/api/#range-queries)
+* [/api/v1/series](https://prometheus.io/docs/prometheus/latest/querying/api/#finding-series-by-label-matchers)
+* [/api/v1/labels](https://prometheus.io/docs/prometheus/latest/querying/api/#getting-label-names)
+* [/api/v1/label/.../values](https://prometheus.io/docs/prometheus/latest/querying/api/#querying-label-values)
+
+These handlers can be queried from Prometheus-compatible clients such as Grafana or curl.
+
+VictoriaMetrics accepts additional args for `/api/v1/labels` and `/api/v1/label/.../values` handlers.
+See [this feature request](https://github.com/prometheus/prometheus/issues/6178) for details:
+
+* Any number [time series selectors](https://prometheus.io/docs/prometheus/latest/querying/basics/#time-series-selectors) via `match[]` query arg.
+* Optional `start` and `end` query args for limiting the time range for the selected labels or label values.
+
+Additionally VictoriaMetrics provides the following handlers:
+
+* `/api/v1/series/count` - it returns the total number of time series in the database. Note that this handler scans all the inverted index,
+  so it can be slow if the database contains tens of millions of time series.
+* `/api/v1/labels/count` - it returns a list of `label: values_count` entries. It can be used for determining labels with the maximum number of values.
+
+
 ### How to build from sources
 
 We recommend using either [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) or
@@ -403,8 +517,8 @@ The page will return the following JSON response:
 ```
 
 Snapshots are created under `<-storageDataPath>/snapshots` directory, where `<-storageDataPath>`
-is the command-line flag value. Snapshots can be archived to backup storage via `cp -L`, `rsync -L`, `scp -r`
-or any similar tool that follows symlinks during copying.
+is the command-line flag value. Snapshots can be archived to backup storage at any time
+with [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md).
 
 The `http://<victoriametrics-addr>:8428/snapshot/list` page contains the list of available snapshots.
 
@@ -415,9 +529,9 @@ Navigate to `http://<victoriametrics-addr>:8428/snapshot/delete_all` in order to
 
 Steps for restoring from a snapshot:
 1. Stop VictoriaMetrics with `kill -INT`.
-2. Remove the entire contents of the directory pointed by `-storageDataPath` command-line flag.
-3. Copy snapshot contents to the directory pointed by `-storageDataPath`.
-4. Start VictoriaMetrics.
+2. Restore snapshot contents from backup with [vmrestore](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md)
+   to the directory pointed by `-storageDataPath`.
+3. Start VictoriaMetrics.
 
 
 ### How to delete time series?
@@ -427,12 +541,16 @@ where `<timeseries_selector_for_delete>` may contain any [time series selector](
 for metrics to delete. After that all the time series matching the given selector are deleted. Storage space for
 the deleted time series isn't freed instantly - it is freed during subsequent merges of data files.
 
+It is recommended verifying which metrics will be deleted with the call to `http://<victoria-metrics-addr>:8428/api/v1/series?match[]=<timeseries_selector_for_delete>`
+before actually deleting the metrics.
+
 
 ### How to export time series?
 
 Send a request to `http://<victoriametrics-addr>:8428/api/v1/export?match[]=<timeseries_selector_for_export>`,
 where `<timeseries_selector_for_export>` may contain any [time series selector](https://prometheus.io/docs/prometheus/latest/querying/basics/#time-series-selectors)
-for metrics to export. The response would contain all the data for the selected time series in [JSON streaming format](https://en.wikipedia.org/wiki/JSON_streaming#Line-delimited_JSON).
+for metrics to export. Use `{__name__!=""}` selector for fetching all the time series.
+The response would contain all the data for the selected time series in [JSON streaming format](https://en.wikipedia.org/wiki/JSON_streaming#Line-delimited_JSON).
 Each JSON line would contain data for a single time series. An example output:
 
 ```
@@ -443,6 +561,52 @@ Each JSON line would contain data for a single time series. An example output:
 Optional `start` and `end` args may be added to the request in order to limit the time frame for the exported data. These args may contain either
 unix timestamp in seconds or [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) values.
 
+Pass `Accept-Encoding: gzip` HTTP header in the request to `/api/v1/export` in order to reduce network bandwidth during exporing big amounts
+of time series data. This enables gzip compression for the exported data. Example for exporting gzipped data:
+
+```
+curl -H 'Accept-Encoding: gzip' http://localhost:8428/api/v1/export -d 'match[]={__name__!=""}' > data.jsonl.gz
+```
+
+The maximum duration for each request to `/api/v1/export` is limited by `-search.maxExportDuration` command-line flag.
+
+Exported data can be imported via POST'ing it to [/api/v1/import](#how-to-import-time-series-data).
+
+
+### How to import time series data?
+
+Time series data can be imported via any supported ingestion protocol:
+
+* [Prometheus remote_write API](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write)
+* [Influx line protocol](#how-to-send-data-from-influxdb-compatible-agents-such-as-telegraf)
+* [Graphite plaintext protocol](#how-to-send-data-from-graphite-compatible-agents-such-as-statsd)
+* [OpenTSDB telnet put protocol](#sending-data-via-telnet-put-protocol)
+* [OpenTSDB http /api/put](#sending-opentsdb-data-via-http-apiput-requests)
+* `/api/v1/import` http POST handler, which accepts data from [/api/v1/export](#how-to-export-time-series).
+
+The most efficient protocol for importing data into VictoriaMetrics is `/api/v1/import`. Example for importing data obtained via `/api/v1/export`:
+
+```
+# Export the data from <source-victoriametrics>:
+curl http://source-victoriametrics:8428/api/v1/export -d 'match={__name__!=""}' > exported_data.jsonl
+
+# Import the data to <destination-victoriametrics>:
+curl -X POST http://destination-victoriametrics:8428/api/v1/import -T exported_data.jsonl
+```
+
+Pass `Content-Encoding: gzip` HTTP request header to `/api/v1/import` for importing gzipped data:
+
+```
+# Export gzipped data from <source-victoriametrics>:
+curl -H 'Accept-Encoding: gzip' http://source-victoriametrics:8428/api/v1/export -d 'match={__name__!=""}' > exported_data.jsonl.gz
+
+# Import gzipped data to <destination-victoriametrics>:
+curl -X POST -H 'Content-Encoding: gzip' http://destination-victoriametrics:8428/api/v1/import -T exported_data.jsonl.gz
+```
+
+Each request to `/api/v1/import` can load up to a single vCPU core on VictoriaMetrics. Import speed can be improved by splitting the original file into smaller parts
+and importing them concurrently. Note that the original file must be split on newlines.
+
 
 ### Federation
 
@@ -451,7 +615,7 @@ at `http://<victoriametrics-addr>:8428/federate?match[]=<timeseries_selector_for
 
 Optional `start` and `end` args may be added to the request in order to scrape the last point for each selected time series on the `[start ... end]` interval.
 `start` and `end` may contain either unix timestamp in seconds or [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) values. By default, the last point
-on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but can be overridden.
+on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but it can be overridden.
 For instance, `/federate?match[]=up&max_lookback=1h` would return last points on the `[now - 1h ... now]` interval. This may be useful for time series federation
 with scrape intervals exceeding `5m`.
 
@@ -463,11 +627,11 @@ A rough estimation of the required resources for ingestion path:
 * RAM size: less than 1KB per active time series. So, ~1GB of RAM is required for 1M active time series.
   Time series is considered active if new data points have been added to it recently or if it has been recently queried.
   The number of active time series may be obtained from `vm_cache_entries{type="storage/hour_metric_ids"}` metric
-  exproted on the `/metrics` page.
+  exported on the `/metrics` page.
   VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited by `-memory.allowedPercent` flag.
 
 * CPU cores: a CPU core per 300K inserted data points per second. So, ~4 CPU cores are required for processing
-  the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data.
+  the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data or for time series with high number of labels.
   See [this article](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) for details.
   If you see lower numbers per CPU core, then it is likely active time series info doesn't fit caches,
   so you need more RAM for lowering CPU usage.
@@ -524,7 +688,7 @@ kill -HUP `pidof prometheus`
 
 
 If you have Prometheus HA pairs with replicas `r1` and `r2` in each pair, then configure each `r1`
-to write data to `<victoriametrics-addr-1`, while each `r2` should write data to `victoriametrics-addr-2`.
+to write data to `victoriametrics-addr-1`, while each `r2` should write data to `victoriametrics-addr-2`.
 
 
 ### Multiple retentions
@@ -544,7 +708,7 @@ There is no downsampling support at the moment, but:
 - VictoriaMetrics has good compression for on-disk data. See [this article](https://medium.com/@valyala/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932)
   for details.
 
-These properties reduce the need in downsampling. We plan to implement downsampling in the future.
+These properties reduce the need of downsampling. We plan to implement downsampling in the future.
 See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/36) for details.
 
 
@@ -566,8 +730,10 @@ horizontally scalable long-term remote storage for really large Prometheus deplo
 
 ### Alerting
 
-VictoriaMetrics doesn't support rule evaluation and alerting yet, so these actions must be performed either
-on [Prometheus side](https://prometheus.io/docs/alerting/overview/) or on [Grafana side](https://grafana.com/docs/alerting/rules/).
+VictoriaMetrics doesn't support rule evaluation and alerting yet, so these actions can be performed at the following places:
+* At Prometheus - see [the corresponding docs](https://prometheus.io/docs/alerting/overview/).
+* At Promxy - see [the corresponding docs](https://github.com/jacksontj/promxy/blob/master/README.md#how-do-i-use-alertingrecording-rules-in-promxy).
+* At Grafana - see [the corresponding docs](https://grafana.com/docs/alerting/rules/).
 
 
 ### Security
@@ -587,27 +753,33 @@ For example, substitute `-graphiteListenAddr=:2003` with `-graphiteListenAddr=<i
 
 ### Tuning
 
-* There is no need in VictoriaMetrics tuning since it uses reasonable defaults for command-line flags,
+* There is no need for VictoriaMetrics tuning since it uses reasonable defaults for command-line flags,
   which are automatically adjusted for the available CPU and RAM resources.
-* There is no need in Operating System tuning since VictoriaMetrics is optimized for default OS settings.
+* There is no need for Operating System tuning since VictoriaMetrics is optimized for default OS settings.
   The only option is increasing the limit on [the number of open files in the OS](https://medium.com/@muhammadtriwibowo/set-permanently-ulimit-n-open-files-in-ubuntu-4d61064429a),
   so Prometheus instances could establish more connections to VictoriaMetrics.
+* The recommended filesystem is `ext4`, the recommended persistent storage is [persistent HDD-based disk on GCP](https://cloud.google.com/compute/docs/disks/#pdspecs),
+  since it is protected from hardware failures via internal replication and it can be [resized on the fly](https://cloud.google.com/compute/docs/disks/add-persistent-disk#resize_pd).
+  If you plan to store more than 1TB of data on `ext4` partition or plan extending it to more than 16TB,
+  then the following options are recommended to pass to `mkfs.ext4`:
+
+```
+mkfs.ext4 ... -O 64bit,huge_file,extent -T huge
+```
 
 
 ### Monitoring
 
 VictoriaMetrics exports internal metrics in Prometheus format on the `/metrics` page.
 Add this page to Prometheus' scrape config in order to collect VictoriaMetrics metrics.
-There is [an official Grafana dashboard for single-node VictoriaMetrics](https://grafana.com/dashboards/10229).
+There are officials Grafana dashboards for [single-node VictoriaMetrics](https://grafana.com/dashboards/10229) and [clustered VictoriaMetrics](https://grafana.com/grafana/dashboards/11176).
 
 The most interesting metrics are:
 
 * `vm_cache_entries{type="storage/hour_metric_ids"}` - the number of time series with new data points during the last hour
   aka active time series.
-* `vm_rows{type="indexdb"}` - the number of rows in inverted index. Each label in each unique time series adds a single
-  row into the inverted index. An approximate number of time series in the database may be calculated as
-  `vm_rows{type="indexdb"} / (avg_labels_per_series + 1)`, where `avg_labels_per_series` is the average number of labels
-  per each time series.
+* `rate(vm_new_timeseries_created_total[5m])` - time series churn rate.
+* `vm_rows{type="indexdb"}` - the number of rows in inverted index. High value for this number usually mean high churn rate for time series.
 * Sum of `vm_rows{type="storage/big"}` and `vm_rows{type="storage/small"}` - total number of `(timestamp, value)` data points
   in the database.
 * Sum of all the `vm_cache_size_bytes` metrics - the total size of all the caches in the database.
@@ -618,6 +790,9 @@ The most interesting metrics are:
 
 ### Troubleshooting
 
+* It is recommended to use default command-line flag values (i.e. don't set them explicitly) until the need
+  of tweaking these flag values arises.
+
 * If VictoriaMetrics works slowly and eats more than a CPU core per 100K ingested data points per second,
   then it is likely you have too many active time series for the current amount of RAM.
   It is recommended increasing the amount of RAM on the node with VictoriaMetrics in order to improve
@@ -630,10 +805,48 @@ The most interesting metrics are:
   has at least 20% of free space comparing to disk size.
 
 * If VictoriaMetrics doesn't work because of certain parts are corrupted due to disk errors,
-  then just remove directoreis with broken parts. This will recover VictoriaMetrics at the cost
+  then just remove directories with broken parts. This will recover VictoriaMetrics at the cost
   of data loss stored in the broken parts. In the future, `vmrecover` tool will be created
   for automatic recovering from such errors.
 
+
+### Backfilling
+
+Make sure that configured `-retentionPeriod` covers timestamps for the backfilled data.
+
+It is recommended disabling query cache with `-search.disableCache` command-line flag when writing
+historical data with timestamps from the past, since the cache assumes that the data is written with
+the current timestamps. Query cache can be enabled after the backfilling is complete.
+
+
+### Profiling
+
+VictoriaMetrics provides handlers for collecting the following [Go profiles](https://blog.golang.org/profiling-go-programs):
+
+- Memory profile. It can be collected with the following command:
+```
+curl -s http://<victoria-metrics-host>:8428/debug/pprof/heap > mem.pprof
+```
+
+- CPU profile. It can be collected with the following command:
+```
+curl -s http://<victoria-metrics-host>:8428/debug/pprof/profile > cpu.pprof
+```
+
+The command for collecting CPU profile waits for 30 seconds before returning.
+
+The collected profiles may be analyzed with [go tool pprof](https://github.com/google/pprof).
+
+
+## Integrations
+
+* [netdata](https://github.com/netdata/netdata) can push data into VictoriaMetrics via `Prometheus remote_write API`.
+  See [these docs](https://github.com/netdata/netdata#integrations).
+* [go-graphite/carbonapi](https://github.com/go-graphite/carbonapi) can use VictoriaMetrics as time series backend.
+  See [this example](/blob/master/cmd/carbonapi/carbonapi.example.prometheus.yaml).
+* [Ansible role for installing VictoriaMetrics](https://github.com/dreamteam-gg/ansible-victoriametrics-role).
+
+
 ## Roadmap
 
 - [ ] Replication [#118](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/118)
@@ -643,7 +856,7 @@ The most interesting metrics are:
 - [ ] CLI tool for data migration, re-balancing and adding/removing nodes [#103](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/103)
 
 
-The discussion happens [here](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/129). Feel free to comment any item or add own one.
+The discussion happens [here](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/129). Feel free to comment on any item or add you own one.
 
 
 ## Contacts
@@ -656,8 +869,8 @@ Contact us with any questions regarding VictoriaMetrics at [info@victoriametrics
 Feel free asking any questions regarding VictoriaMetrics:
 
 - [slack](http://slack.victoriametrics.com/)
-- [telergam-en](https://t.me/VictoriaMetrics_en)
-- [telergam-ru](https://t.me/VictoriaMetrics_ru1)
+- [telegram-en](https://t.me/VictoriaMetrics_en)
+- [telegram-ru](https://t.me/VictoriaMetrics_ru1)
 - [google groups](https://groups.google.com/forum/#!forum/victorametrics-users)
 
 
@@ -686,7 +899,7 @@ Report bugs and propose new features [here](https://github.com/VictoriaMetrics/V
 
 ## Victoria Metrics Logo
 
-[Zip](VM_logo.zip) contains three folders with different image orientation (main color and inverted version).
+[Zip](VM_logo.zip) contains three folders with different image orientations (main color and inverted version).
 
 Files included in each folder:
 

app/victoria-metrics/Makefile

@@ -1,14 +1,49 @@
 # All these commands must run from repository root.
 
 victoria-metrics:
-	GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics ./app/victoria-metrics
+	APP_NAME=victoria-metrics $(MAKE) app-local
 
 victoria-metrics-prod:
 	APP_NAME=victoria-metrics $(MAKE) app-via-docker
 
+victoria-metrics-pure-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-pure
+
+victoria-metrics-amd64-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-amd64
+
+victoria-metrics-arm-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-arm
+
+victoria-metrics-arm64-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-arm64
+
+victoria-metrics-ppc64le-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-ppc64le
+
+victoria-metrics-386-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-386
+
 package-victoria-metrics:
-	APP_NAME=victoria-metrics \
-	$(MAKE) package-via-docker
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker
+
+package-victoria-metrics-pure:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-pure
+
+package-victoria-metrics-amd64:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-amd64
+
+package-victoria-metrics-arm:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-arm
+
+package-victoria-metrics-arm64:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-arm64
+
+package-victoria-metrics-ppc64le:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-ppc64le
+
+package-victoria-metrics-386:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-386
 
 publish-victoria-metrics:
 	APP_NAME=victoria-metrics $(MAKE) publish-via-docker
@@ -20,23 +55,23 @@ run-victoria-metrics:
 	ARGS='-graphiteListenAddr=:2003 -opentsdbListenAddr=:4242 -retentionPeriod=12 -search.maxUniqueTimeseries=1000000 -search.maxQueryDuration=10m' \
 	$(MAKE) run-via-docker
 
+victoria-metrics-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-amd64 ./app/victoria-metrics
+
 victoria-metrics-arm:
 	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm ./app/victoria-metrics
 
-victoria-metrics-arm-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-arm' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm' $(MAKE) app-via-docker
-
 victoria-metrics-arm64:
 	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm64 ./app/victoria-metrics
 
-victoria-metrics-arm64-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-arm64' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm64' $(MAKE) app-via-docker
+victoria-metrics-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-ppc64le ./app/victoria-metrics
+
+victoria-metrics-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-386 ./app/victoria-metrics
 
 victoria-metrics-pure:
-	GO111MODULE=on CGO_ENABLED=0 go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-pure ./app/victoria-metrics
-
-victoria-metrics-pure-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-pure' DOCKER_OPTS='--env CGO_ENABLED=0' $(MAKE) app-via-docker
+	APP_NAME=victoria-metrics $(MAKE) app-local-pure
 
 ### Packaging as DEB - amd64
 victoria-metrics-package-deb: victoria-metrics-prod

app/victoria-metrics/deployment/Dockerfile

@@ -1,5 +1,8 @@
+ARG certs_image
+FROM $certs_image AS certs
 FROM scratch
-COPY --from=local/certs:1.0.2 /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
-COPY bin/victoria-metrics-prod .
+COPY --from=certs /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
+ARG src_binary
+COPY $src_binary ./victoria-metrics-prod
 EXPOSE 8428
 ENTRYPOINT ["/victoria-metrics-prod"]

app/victoria-metrics/main.go

@@ -9,6 +9,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/procutil"
@@ -20,7 +21,7 @@ func main() {
 	flag.Parse()
 	buildinfo.Init()
 	logger.Init()
-	logger.Infof("starting VictoraMetrics at %q...", *httpListenAddr)
+	logger.Infof("starting VictoriaMetrics at %q...", *httpListenAddr)
 	startTime := time.Now()
 	vmstorage.Init()
 	vmselect.Init()
@@ -43,6 +44,8 @@ func main() {
 	vmstorage.Stop()
 	vmselect.Stop()
 
+	fs.MustStopDirRemover()
+
 	logger.Infof("the VictoriaMetrics has been stopped in %s", time.Since(startTime))
 }
 

app/victoria-metrics/main_test.go

@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"flag"
 	"fmt"
+	"io"
 	"io/ioutil"
 	"log"
 	"net"
@@ -18,26 +19,31 @@ import (
 	"testing"
 	"time"
 
+	testutil "github.com/VictoriaMetrics/VictoriaMetrics/app/victoria-metrics/test"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
 const (
-	testFixturesDir        = "testdata"
-	testStorageSuffix      = "vm-test-storage"
-	testHTTPListenAddr     = ":7654"
-	testStatsDListenAddr   = ":2003"
-	testOpenTSDBListenAddr = ":4242"
-	testLogLevel           = "INFO"
+	testFixturesDir            = "testdata"
+	testStorageSuffix          = "vm-test-storage"
+	testHTTPListenAddr         = ":7654"
+	testStatsDListenAddr       = ":2003"
+	testOpenTSDBListenAddr     = ":4242"
+	testOpenTSDBHTTPListenAddr = ":4243"
+	testLogLevel               = "INFO"
 )
 
 const (
-	testReadHTTPPath   = "http://127.0.0.1" + testHTTPListenAddr
-	testWriteHTTPPath  = "http://127.0.0.1" + testHTTPListenAddr + "/write"
-	testHealthHTTPPath = "http://127.0.0.1" + testHTTPListenAddr + "/health"
+	testReadHTTPPath          = "http://127.0.0.1" + testHTTPListenAddr
+	testWriteHTTPPath         = "http://127.0.0.1" + testHTTPListenAddr + "/write"
+	testOpenTSDBWriteHTTPPath = "http://127.0.0.1" + testOpenTSDBHTTPListenAddr + "/api/put"
+	testPromWriteHTTPPath     = "http://127.0.0.1" + testHTTPListenAddr + "/api/v1/write"
+	testHealthHTTPPath        = "http://127.0.0.1" + testHTTPListenAddr + "/health"
 )
 
 const (
@@ -50,18 +56,69 @@ var (
 )
 
 type test struct {
-	Name   string `json:"name"`
-	Data   string `json:"data"`
-	Query  string `json:"query"`
-	Result []Row  `json:"result"`
+	Name             string     `json:"name"`
+	Data             []string   `json:"data"`
+	Query            []string   `json:"query"`
+	ResultMetrics    []Metric   `json:"result_metrics"`
+	ResultSeries     Series     `json:"result_series"`
+	ResultQuery      Query      `json:"result_query"`
+	ResultQueryRange QueryRange `json:"result_query_range"`
+	Issue            string     `json:"issue"`
 }
 
-type Row struct {
+type Metric struct {
 	Metric     map[string]string `json:"metric"`
 	Values     []float64         `json:"values"`
 	Timestamps []int64           `json:"timestamps"`
 }
 
+func (r *Metric) UnmarshalJSON(b []byte) error {
+	type plain Metric
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
+type Series struct {
+	Status string              `json:"status"`
+	Data   []map[string]string `json:"data"`
+}
+type Query struct {
+	Status string    `json:"status"`
+	Data   QueryData `json:"data"`
+}
+type QueryData struct {
+	ResultType string            `json:"resultType"`
+	Result     []QueryDataResult `json:"result"`
+}
+
+type QueryDataResult struct {
+	Metric map[string]string `json:"metric"`
+	Value  []interface{}     `json:"value"`
+}
+
+func (r *QueryDataResult) UnmarshalJSON(b []byte) error {
+	type plain QueryDataResult
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
+type QueryRange struct {
+	Status string         `json:"status"`
+	Data   QueryRangeData `json:"data"`
+}
+type QueryRangeData struct {
+	ResultType string                 `json:"resultType"`
+	Result     []QueryRangeDataResult `json:"result"`
+}
+
+type QueryRangeDataResult struct {
+	Metric map[string]string `json:"metric"`
+	Values [][]interface{}   `json:"values"`
+}
+
+func (r *QueryRangeDataResult) UnmarshalJSON(b []byte) error {
+	type plain QueryRangeDataResult
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
 func TestMain(m *testing.M) {
 	setUp()
 	code := m.Run()
@@ -92,7 +149,7 @@ func setUp() {
 
 func processFlags() {
 	flag.Parse()
-	for _, fs := range []struct {
+	for _, fv := range []struct {
 		flag  string
 		value string
 	}{
@@ -101,10 +158,11 @@ func processFlags() {
 		{flag: "graphiteListenAddr", value: testStatsDListenAddr},
 		{flag: "opentsdbListenAddr", value: testOpenTSDBListenAddr},
 		{flag: "loggerLevel", value: testLogLevel},
+		{flag: "opentsdbHTTPListenAddr", value: testOpenTSDBHTTPListenAddr},
 	} {
 		// panics if flag doesn't exist
-		if err := flag.Lookup(fs.flag).Value.Set(fs.value); err != nil {
-			log.Fatalf("unable to set %q with value %q, err: %v", fs.flag, fs.value, err)
+		if err := flag.Lookup(fv.flag).Value.Set(fv.value); err != nil {
+			log.Fatalf("unable to set %q with value %q, err: %v", fv.flag, fv.value, err)
 		}
 	}
 }
@@ -121,67 +179,125 @@ func waitFor(timeout time.Duration, f func() bool) error {
 }
 
 func tearDown() {
+	if err := httpserver.Stop(*httpListenAddr); err != nil {
+		log.Printf("cannot stop the webservice: %s", err)
+	}
 	vminsert.Stop()
 	vmstorage.Stop()
 	vmselect.Stop()
-	if err := httpserver.Stop(*httpListenAddr); err != nil {
-		log.Fatalf("cannot stop the webservice: %s", err)
-	}
-	os.RemoveAll(storagePath)
+	fs.MustRemoveAll(storagePath)
 }
 
 func TestWriteRead(t *testing.T) {
 	t.Run("write", testWrite)
 	time.Sleep(1 * time.Second)
 	vmstorage.Stop()
-
 	// open storage after stop in write
 	vmstorage.InitWithoutMetrics()
 	t.Run("read", testRead)
 }
 
 func testWrite(t *testing.T) {
+	t.Run("prometheus", func(t *testing.T) {
+		for _, test := range readIn("prometheus", t, insertionTime) {
+			s := newSuite(t)
+			r := testutil.WriteRequest{}
+			s.noError(json.Unmarshal([]byte(strings.Join(test.Data, "\n")), &r.Timeseries))
+			data, err := testutil.Compress(r)
+			s.greaterThan(len(r.Timeseries), 0)
+			if err != nil {
+				t.Errorf("error compressing %v %s", r, err)
+				t.Fail()
+			}
+			httpWrite(t, testPromWriteHTTPPath, bytes.NewBuffer(data))
+		}
+	})
+
 	t.Run("influxdb", func(t *testing.T) {
-		for _, test := range readIn("influxdb", t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
+		for _, x := range readIn("influxdb", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				httpWrite(t, testWriteHTTPPath, test.Data)
+				httpWrite(t, testWriteHTTPPath, bytes.NewBufferString(strings.Join(test.Data, "\n")))
 			})
 		}
 	})
 	t.Run("graphite", func(t *testing.T) {
-		for _, test := range readIn("graphite", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+		for _, x := range readIn("graphite", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				tcpWrite(t, "127.0.0.1"+testStatsDListenAddr, test.Data)
+				tcpWrite(t, "127.0.0.1"+testStatsDListenAddr, strings.Join(test.Data, "\n"))
 			})
 		}
 	})
 	t.Run("opentsdb", func(t *testing.T) {
-		for _, test := range readIn("opentsdb", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+		for _, x := range readIn("opentsdb", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				tcpWrite(t, "127.0.0.1"+testOpenTSDBListenAddr, test.Data)
+				tcpWrite(t, "127.0.0.1"+testOpenTSDBListenAddr, strings.Join(test.Data, "\n"))
+			})
+		}
+	})
+	t.Run("opentsdbhttp", func(t *testing.T) {
+		for _, x := range readIn("opentsdbhttp", t, insertionTime) {
+			test := x
+			t.Run(test.Name, func(t *testing.T) {
+				t.Parallel()
+				logger.Infof("writing %s", test.Data)
+				httpWrite(t, testOpenTSDBWriteHTTPPath, bytes.NewBufferString(strings.Join(test.Data, "\n")))
 			})
 		}
 	})
 }
 
 func testRead(t *testing.T) {
-	for _, engine := range []string{"graphite", "opentsdb", "influxdb"} {
+	for _, engine := range []string{"prometheus", "graphite", "opentsdb", "influxdb", "opentsdbhttp"} {
 		t.Run(engine, func(t *testing.T) {
-			for _, test := range readIn(engine, t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
-				test := test
+			for _, x := range readIn(engine, t, insertionTime) {
+				test := x
 				t.Run(test.Name, func(t *testing.T) {
 					t.Parallel()
-					rowContains(t, httpRead(t, testReadHTTPPath, test.Query), test.Result)
+					for _, q := range test.Query {
+						q = testutil.PopulateTimeTplString(q, insertionTime)
+						if test.Issue != "" {
+							test.Issue = "Regression in " + test.Issue
+						}
+						switch true {
+						case strings.HasPrefix(q, "/api/v1/export"):
+							if err := checkMetricsResult(httpReadMetrics(t, testReadHTTPPath, q), test.ResultMetrics); err != nil {
+								t.Fatalf("Export. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/series"):
+							s := Series{}
+							httpReadStruct(t, testReadHTTPPath, q, &s)
+							if err := checkSeriesResult(s, test.ResultSeries); err != nil {
+								t.Fatalf("Series. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/query_range"):
+							queryResult := QueryRange{}
+							httpReadStruct(t, testReadHTTPPath, q, &queryResult)
+							if err := checkQueryRangeResult(queryResult, test.ResultQueryRange); err != nil {
+								t.Fatalf("Query Range. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/query"):
+							queryResult := Query{}
+							httpReadStruct(t, testReadHTTPPath, q, &queryResult)
+							if err := checkQueryResult(queryResult, test.ResultQuery); err != nil {
+								t.Fatalf("Query. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						default:
+							t.Fatalf("unsupported read query %s", q)
+						}
+					}
 				})
 			}
 		})
 	}
 }
 
-func readIn(readFor string, t *testing.T, timeStr string) []test {
+func readIn(readFor string, t *testing.T, insertTime time.Time) []test {
 	t.Helper()
 	s := newSuite(t)
 	var tt []test
@@ -193,7 +309,9 @@ func readIn(readFor string, t *testing.T, timeStr string) []test {
 		s.noError(err)
 		item := test{}
 		s.noError(json.Unmarshal(b, &item))
-		item.Data = strings.Replace(item.Data, "{TIME}", timeStr, 1)
+		for i := range item.Data {
+			item.Data[i] = testutil.PopulateTimeTplString(item.Data[i], insertTime)
+		}
 		tt = append(tt, item)
 		return nil
 	}))
@@ -203,10 +321,10 @@ func readIn(readFor string, t *testing.T, timeStr string) []test {
 	return tt
 }
 
-func httpWrite(t *testing.T, address string, data string) {
+func httpWrite(t *testing.T, address string, r io.Reader) {
 	t.Helper()
 	s := newSuite(t)
-	resp, err := http.Post(address, "", bytes.NewBufferString(data))
+	resp, err := http.Post(address, "", r)
 	s.noError(err)
 	s.noError(resp.Body.Close())
 	s.equalInt(resp.StatusCode, 204)
@@ -223,35 +341,122 @@ func tcpWrite(t *testing.T, address string, data string) {
 	s.equalInt(n, len(data))
 }
 
-func httpRead(t *testing.T, address, query string) []Row {
+func httpReadMetrics(t *testing.T, address, query string) []Metric {
 	t.Helper()
 	s := newSuite(t)
 	resp, err := http.Get(address + query)
 	s.noError(err)
 	defer resp.Body.Close()
 	s.equalInt(resp.StatusCode, 200)
-	var rows []Row
+	var rows []Metric
 	for dec := json.NewDecoder(resp.Body); dec.More(); {
-		var row Row
+		var row Metric
 		s.noError(dec.Decode(&row))
 		rows = append(rows, row)
 	}
 	return rows
 }
-
-func rowContains(t *testing.T, rows, contains []Row) {
+func httpReadStruct(t *testing.T, address, query string, dst interface{}) {
 	t.Helper()
-	for _, r := range rows {
-		contains = removeIfFound(r, contains)
-	}
-	if len(contains) > 0 {
-		t.Fatalf("result rows %+v not found in %+v", contains, rows)
-	}
+	s := newSuite(t)
+	resp, err := http.Get(address + query)
+	s.noError(err)
+	defer resp.Body.Close()
+	s.equalInt(resp.StatusCode, 200)
+	s.noError(json.NewDecoder(resp.Body).Decode(dst))
 }
 
-func removeIfFound(r Row, contains []Row) []Row {
+func checkMetricsResult(got, want []Metric) error {
+	for _, r := range append([]Metric(nil), got...) {
+		want = removeIfFoundMetrics(r, want)
+	}
+	if len(want) > 0 {
+		return fmt.Errorf("exptected metrics %+v not found in %+v", want, got)
+	}
+	return nil
+}
+
+func removeIfFoundMetrics(r Metric, contains []Metric) []Metric {
+	for i, item := range contains {
+		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Values, item.Values) &&
+			reflect.DeepEqual(r.Timestamps, item.Timestamps) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkSeriesResult(got, want Series) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	wantData := append([]map[string]string(nil), want.Data...)
+	for _, r := range got.Data {
+		wantData = removeIfFoundSeries(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected seria(s) %+v not found in %+v", wantData, got.Data)
+	}
+	return nil
+}
+
+func removeIfFoundSeries(r map[string]string, contains []map[string]string) []map[string]string {
+	for i, item := range contains {
+		if reflect.DeepEqual(r, item) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkQueryResult(got, want Query) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	if got.Data.ResultType != want.Data.ResultType {
+		return fmt.Errorf("result type mismatch %q - %q", want.Data.ResultType, got.Data.ResultType)
+	}
+	wantData := append([]QueryDataResult(nil), want.Data.Result...)
+	for _, r := range got.Data.Result {
+		wantData = removeIfFoundQueryData(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected query result %+v not found in %+v", wantData, got.Data.Result)
+	}
+	return nil
+}
+
+func removeIfFoundQueryData(r QueryDataResult, contains []QueryDataResult) []QueryDataResult {
+	for i, item := range contains {
+		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Value[0], item.Value[0]) && reflect.DeepEqual(r.Value[1], item.Value[1]) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkQueryRangeResult(got, want QueryRange) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	if got.Data.ResultType != want.Data.ResultType {
+		return fmt.Errorf("result type mismatch %q - %q", want.Data.ResultType, got.Data.ResultType)
+	}
+	wantData := append([]QueryRangeDataResult(nil), want.Data.Result...)
+	for _, r := range got.Data.Result {
+		wantData = removeIfFoundQueryRangeData(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected query range result %+v not found in %+v", wantData, got.Data.Result)
+	}
+	return nil
+}
+
+func removeIfFoundQueryRangeData(r QueryRangeDataResult, contains []QueryRangeDataResult) []QueryRangeDataResult {
 	for i, item := range contains {
-		// todo check time
 		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Values, item.Values) {
 			contains[i] = contains[len(contains)-1]
 			return contains[:len(contains)-1]
@@ -279,3 +484,11 @@ func (s *suite) equalInt(a, b int) {
 		s.t.FailNow()
 	}
 }
+
+func (s *suite) greaterThan(a, b int) {
+	s.t.Helper()
+	if a <= b {
+		s.t.Errorf("%d less or equal then %d", a, b)
+		s.t.FailNow()
+	}
+}

app/victoria-metrics/test/parser.go

@@ -0,0 +1,52 @@
+package test
+
+import (
+	"fmt"
+	"log"
+	"regexp"
+	"strings"
+	"time"
+)
+
+var (
+	parseTimeExpRegex = regexp.MustCompile(`"?{TIME[^}]*}"?`)
+	extractRegex      = regexp.MustCompile(`"?{([^}]*)}"?`)
+)
+
+// PopulateTimeTplString substitutes {TIME_*} with t in s and returns the result.
+func PopulateTimeTplString(s string, t time.Time) string {
+	return string(PopulateTimeTpl([]byte(s), t))
+}
+
+// PopulateTimeTpl substitutes {TIME_*} with tGlobal in b and returns the result.
+func PopulateTimeTpl(b []byte, tGlobal time.Time) []byte {
+	return parseTimeExpRegex.ReplaceAllFunc(b, func(repl []byte) []byte {
+		t := tGlobal
+		repl = extractRegex.FindSubmatch(repl)[1]
+		parts := strings.SplitN(string(repl), "-", 2)
+		if len(parts) == 2 {
+			duration, err := time.ParseDuration(strings.TrimSpace(parts[1]))
+			if err != nil {
+				log.Fatalf("error %s parsing duration %s in %s", err, parts[1], repl)
+			}
+			t = t.Add(-duration)
+		}
+		switch strings.TrimSpace(parts[0]) {
+		case `TIME_S`:
+			return []byte(fmt.Sprintf("%d", t.Unix()))
+		case `TIME_MSZ`:
+			return []byte(fmt.Sprintf("%d", t.Unix()*1e3))
+		case `TIME_MS`:
+			return []byte(fmt.Sprintf("%d", timeToMillis(t)))
+		case `TIME_NS`:
+			return []byte(fmt.Sprintf("%d", t.UnixNano()))
+		default:
+			log.Fatalf("unknown time pattern %s in %s", parts[0], repl)
+		}
+		return repl
+	})
+}
+
+func timeToMillis(t time.Time) int64 {
+	return t.UnixNano() / 1e6
+}

app/victoria-metrics/test/parser_test.go

@@ -0,0 +1,24 @@
+package test
+
+import (
+	"testing"
+	"time"
+)
+
+func TestPopulateTimeTplString(t *testing.T) {
+	now, err := time.Parse(time.RFC3339, "2006-01-02T15:04:05Z")
+	if err != nil {
+		t.Fatalf("unexpected error when parsing time: %s", err)
+	}
+	f := func(s, resultExpected string) {
+		t.Helper()
+		result := PopulateTimeTplString(s, now)
+		if result != resultExpected {
+			t.Fatalf("unexpected result; got %q; want %q", result, resultExpected)
+		}
+	}
+	f("", "")
+	f("{TIME_S}", "1136214245")
+	f("now: {TIME_S}, past 30s: {TIME_MS-30s}, now: {TIME_S}", "now: 1136214245, past 30s: 1136214215000, now: 1136214245")
+	f("now: {TIME_MS}, past 30m: {TIME_MSZ-30m}, past 2h: {TIME_NS-2h}", "now: 1136214245000, past 30m: 1136212445000, past 2h: 1136207045000000000")
+}

app/victoria-metrics/test/prom_types.go

@@ -0,0 +1,338 @@
+// +build integration
+
+// Source https://github.com/prometheus/prometheus/blob/master/prompb/remote.pb.go . Code is copy pasted and cleaned up
+package test
+
+import (
+	"encoding/binary"
+	"math"
+	"math/bits"
+)
+
+type WriteRequest struct {
+	Timeseries []TimeSeries `protobuf:"bytes,1,rep,name=timeseries,proto3" json:"timeseries"`
+}
+
+func (m *WriteRequest) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Timeseries) > 0 {
+		for _, e := range m.Timeseries {
+			l = e.Size()
+			n += 1 + l + sovRemote(uint64(l))
+		}
+	}
+	return n
+}
+func sovRemote(x uint64) (n int) {
+	return (bits.Len64(x|1) + 6) / 7
+}
+
+func (m *WriteRequest) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *WriteRequest) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *WriteRequest) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	if len(m.Timeseries) > 0 {
+		for iNdEx := len(m.Timeseries) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Timeseries[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintRemote(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func encodeVarintRemote(dAtA []byte, offset int, v uint64) int {
+	offset -= sovRemote(v)
+	base := offset
+	for v >= 1<<7 {
+		dAtA[offset] = uint8(v&0x7f | 0x80)
+		v >>= 7
+		offset++
+	}
+	dAtA[offset] = uint8(v)
+	return base
+}
+
+type Sample struct {
+	Value     float64 `protobuf:"fixed64,1,opt,name=value,proto3" json:"value,omitempty"`
+	Timestamp int64   `protobuf:"varint,2,opt,name=timestamp,proto3" json:"timestamp,omitempty"`
+}
+
+func (m *Sample) Reset() { *m = Sample{} }
+
+// TimeSeries represents samples and labels for a single time series.
+type TimeSeries struct {
+	Labels  []Label  `protobuf:"bytes,1,rep,name=labels,proto3" json:"labels"`
+	Samples []Sample `protobuf:"bytes,2,rep,name=samples,proto3" json:"samples"`
+}
+
+func (m *TimeSeries) Reset() { *m = TimeSeries{} }
+
+type Label struct {
+	Name  string `protobuf:"bytes,1,opt,name=name,proto3" json:"name,omitempty"`
+	Value string `protobuf:"bytes,2,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *Label) Reset() { *m = Label{} }
+
+type Labels struct {
+	Labels []Label `protobuf:"bytes,1,rep,name=labels,proto3" json:"labels"`
+}
+
+func (m *Labels) Reset() { *m = Labels{} }
+
+func (m *Sample) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *Sample) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *Sample) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	if m.Timestamp != 0 {
+		i = encodeVarintTypes(dAtA, i, uint64(m.Timestamp))
+		i--
+		dAtA[i] = 0x10
+	}
+	if m.Value != 0 {
+		i -= 8
+		binary.LittleEndian.PutUint64(dAtA[i:], uint64(math.Float64bits(float64(m.Value))))
+		i--
+		dAtA[i] = 0x9
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *TimeSeries) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *TimeSeries) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *TimeSeries) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	if len(m.Samples) > 0 {
+		for iNdEx := len(m.Samples) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Samples[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintTypes(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0x12
+		}
+	}
+	if len(m.Labels) > 0 {
+		for iNdEx := len(m.Labels) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Labels[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintTypes(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *Label) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *Label) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *Label) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if len(m.Value) > 0 {
+		i -= len(m.Value)
+		copy(dAtA[i:], m.Value)
+		i = encodeVarintTypes(dAtA, i, uint64(len(m.Value)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if len(m.Name) > 0 {
+		i -= len(m.Name)
+		copy(dAtA[i:], m.Name)
+		i = encodeVarintTypes(dAtA, i, uint64(len(m.Name)))
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *Labels) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *Labels) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *Labels) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	if len(m.Labels) > 0 {
+		for iNdEx := len(m.Labels) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Labels[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintTypes(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func encodeVarintTypes(dAtA []byte, offset int, v uint64) int {
+	offset -= sovTypes(v)
+	base := offset
+	for v >= 1<<7 {
+		dAtA[offset] = uint8(v&0x7f | 0x80)
+		v >>= 7
+		offset++
+	}
+	dAtA[offset] = uint8(v)
+	return base
+}
+
+func (m *Sample) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	if m.Value != 0 {
+		n += 9
+	}
+	if m.Timestamp != 0 {
+		n += 1 + sovTypes(uint64(m.Timestamp))
+	}
+	return n
+}
+
+func (m *TimeSeries) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Labels) > 0 {
+		for _, e := range m.Labels {
+			l = e.Size()
+			n += 1 + l + sovTypes(uint64(l))
+		}
+	}
+	if len(m.Samples) > 0 {
+		for _, e := range m.Samples {
+			l = e.Size()
+			n += 1 + l + sovTypes(uint64(l))
+		}
+	}
+	return n
+}
+
+func (m *Label) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	l = len(m.Name)
+	if l > 0 {
+		n += 1 + l + sovTypes(uint64(l))
+	}
+	l = len(m.Value)
+	if l > 0 {
+		n += 1 + l + sovTypes(uint64(l))
+	}
+	return n
+}
+
+func (m *Labels) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Labels) > 0 {
+		for _, e := range m.Labels {
+			l = e.Size()
+			n += 1 + l + sovTypes(uint64(l))
+		}
+	}
+	return n
+}
+
+func sovTypes(x uint64) (n int) {
+	return (bits.Len64(x|1) + 6) / 7
+}

app/victoria-metrics/test/prom_writter.go

@@ -0,0 +1,13 @@
+// +build integration
+
+package test
+
+import "github.com/golang/snappy"
+
+func Compress(wr WriteRequest) ([]byte, error) {
+	data, err := wr.Marshal()
+	if err != nil {
+		return nil, err
+	}
+	return snappy.Encode(nil, data), nil
+}

app/victoria-metrics/testdata/graphite/basic.json

@@ -1,8 +1,8 @@
 {
   "name": "basic_insertion",
-  "data": "graphite.foo.bar.baz;tag1=value1;tag2=value2 123 {TIME}",
-  "query": "/api/v1/export?match={__name__!=\"\"}",
-  "result": [
-    {"metric":{"__name__":"graphite.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123]}
+  "data": ["graphite.foo.bar.baz;tag1=value1;tag2=value2 123 {TIME_S}"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"graphite.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123], "timestamps": ["{TIME_MSZ}"]}
   ]
 }

app/victoria-metrics/testdata/graphite/comparison-not-inf-not-nan.json

@@ -0,0 +1,16 @@
+{
+  "name": "comparison-not-inf-not-nan",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150",
+  "data": [
+    "not_nan_not_inf;item=x 1 {TIME_S-1m}",
+    "not_nan_not_inf;item=x 1 {TIME_S-2m}",
+    "not_nan_not_inf;item=y 3 {TIME_S-1m}",
+    "not_nan_not_inf;item=y 1 {TIME_S-2m}"],
+  "query": ["/api/v1/query_range?query=1/(not_nan_not_inf-1)!=inf!=nan&start={TIME_S-3m}&end={TIME_S}&step=60"],
+  "result_query_range": {
+    "status":"success",
+    "data":{"resultType":"matrix",
+      "result":[
+	      {"metric":{"item":"y"},"values":[["{TIME_S-1m}","0.5"],["{TIME_S}","0.5"]]}
+      ]}}
+}

app/victoria-metrics/testdata/graphite/max_lookback_set.json

@@ -0,0 +1,24 @@
+{
+  "name": "max_lookback_set",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/209",
+  "data": [
+    "max_lookback_set 1 {TIME_S-30s}",
+    "max_lookback_set 2 {TIME_S-60s}",
+    "max_lookback_set 3 {TIME_S-120s}",
+    "max_lookback_set 4 {TIME_S-150s}"
+  ],
+  "query": ["/api/v1/query_range?query=max_lookback_set&start={TIME_S-150s}&end={TIME_S}&step=10s&max_lookback=1s"],
+  "result_query_range": {
+    "status":"success",
+    "data":{"resultType":"matrix",
+      "result":[{"metric":{"__name__":"max_lookback_set"},"values":[
+	      ["{TIME_S-150s}","4"],
+	      ["{TIME_S-140s}","4"],
+	      ["{TIME_S-120s}","3"],
+	      ["{TIME_S-110s}","3"],
+	      ["{TIME_S-60s}","2"],
+	      ["{TIME_S-50s}","2"],
+	      ["{TIME_S-30s}","1"],
+	      ["{TIME_S-20s}","1"]
+      ]}]}}
+}

app/victoria-metrics/testdata/graphite/max_lookback_unset.json

@@ -0,0 +1,32 @@
+{
+  "name": "max_lookback_unset",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/209",
+  "data": [
+    "max_lookback_unset 1 {TIME_S-30s}",
+    "max_lookback_unset 2 {TIME_S-60s}",
+    "max_lookback_unset 3 {TIME_S-120s}",
+    "max_lookback_unset 4 {TIME_S-150s}"
+  ],
+  "query": ["/api/v1/query_range?query=max_lookback_unset&start={TIME_S-150s}&end={TIME_S}&step=10s"],
+  "result_query_range": {
+    "status":"success",
+    "data":{"resultType":"matrix",
+      "result":[{"metric":{"__name__":"max_lookback_unset"},"values":[
+	      ["{TIME_S-150s}","4"],
+	      ["{TIME_S-140s}","4"],
+	      ["{TIME_S-130s}","4"],
+	      ["{TIME_S-120s}","3"],
+	      ["{TIME_S-110s}","3"],
+	      ["{TIME_S-100s}","3"],
+	      ["{TIME_S-90s}","3"],
+	      ["{TIME_S-80s}","3"],
+	      ["{TIME_S-70s}","3"],
+	      ["{TIME_S-60s}","2"],
+	      ["{TIME_S-50s}","2"],
+	      ["{TIME_S-40s}","2"],
+	      ["{TIME_S-30s}","1"],
+	      ["{TIME_S-20s}","1"],
+	      ["{TIME_S-10s}","1"],
+	      ["{TIME_S}","1"]
+      ]}]}}
+}

app/victoria-metrics/testdata/graphite/not-nan-as-missing-data.json

@@ -0,0 +1,18 @@
+{
+  "name": "not-nan-as-missing-data",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/153",
+  "data": [
+    "not_nan_as_missing_data;item=x 2 {TIME_S-2m}",
+    "not_nan_as_missing_data;item=x 1 {TIME_S-1m}",
+    "not_nan_as_missing_data;item=y 4 {TIME_S-2m}",
+    "not_nan_as_missing_data;item=y 3 {TIME_S-1m}"
+  ],
+  "query": ["/api/v1/query_range?query=not_nan_as_missing_data>1&start={TIME_S-2m}&end={TIME_S}&step=60"],
+  "result_query_range": {
+    "status":"success",
+    "data":{"resultType":"matrix",
+      "result":[
+	      {"metric":{"__name__":"not_nan_as_missing_data","item":"x"},"values":[["{TIME_S-2m}","2"]]},
+	      {"metric":{"__name__":"not_nan_as_missing_data","item":"y"},"values":[["{TIME_S-2m}","4"],["{TIME_S-1m}","3"],["{TIME_S}","3"]]}
+      ]}}
+}

app/victoria-metrics/testdata/graphite/subquery-aggregation.json

@@ -0,0 +1,14 @@
+{
+  "name": "subquery-aggregation",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/184",
+  "data": [
+    "forms_daily_count;item=x 1 {TIME_S-1m}",
+    "forms_daily_count;item=x 2 {TIME_S-2m}",
+    "forms_daily_count;item=y 3 {TIME_S-1m}",
+    "forms_daily_count;item=y 4 {TIME_S-2m}"],
+  "query": ["/api/v1/query?query=min%20by%20(item)%20(min_over_time(forms_daily_count[10m:1m]))&time={TIME_S-1m}"],
+  "result_query": {
+    "status":"success",
+    "data":{"resultType":"vector","result":[{"metric":{"item":"x"},"value":["{TIME_S-1m}","1"]},{"metric":{"item":"y"},"value":["{TIME_S-1m}","3"]}]}
+  }
+}

app/victoria-metrics/testdata/influxdb/basic.json

@@ -1,9 +1,9 @@
 {
   "name": "basic_insertion",
-  "data": "measurement,tag1=value1,tag2=value2 field1=1.23,field2=123",
-  "query": "/api/v1/export?match={__name__!=\"\"}",
-  "result": [
-    {"metric":{"__name__":"measurement_field2","tag1":"value1","tag2":"value2"},"values":[123]},
-    {"metric":{"__name__":"measurement_field1","tag1":"value1","tag2":"value2"},"values":[1.23]}
+  "data": ["measurement,tag1=value1,tag2=value2 field1=1.23,field2=123 {TIME_NS}"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"measurement_field2","tag1":"value1","tag2":"value2"},"values":[123], "timestamps": ["{TIME_MS}"]},
+    {"metric":{"__name__":"measurement_field1","tag1":"value1","tag2":"value2"},"values":[1.23], "timestamps": ["{TIME_MS}"]}
   ]
 }

app/victoria-metrics/testdata/opentsdb/basic.json

@@ -1,8 +1,8 @@
 {
   "name": "basic_insertion",
-  "data": "put openstdb.foo.bar.baz {TIME} 123 tag1=value1 tag2=value2",
-  "query": "/api/v1/export?match={__name__!=\"\"}",
-  "result": [
-    {"metric":{"__name__":"openstdb.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123]}
+  "data": ["put openstdb.foo.bar.baz {TIME_S} 123 tag1=value1 tag2=value2"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"openstdb.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123], "timestamps": ["{TIME_MSZ}"]}
   ]
 }

app/victoria-metrics/testdata/opentsdbhttp/basic.json

@@ -0,0 +1,8 @@
+{
+  "name": "basic_insertion",
+  "data": ["{\"metric\": \"opentsdbhttp.foo\", \"value\": 1001, \"timestamp\": {TIME_S}, \"tags\": {\"bar\":\"baz\", \"x\": \"y\"}}"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"opentsdbhttp.foo","bar":"baz","x":"y"},"values":[1001], "timestamps": ["{TIME_MSZ}"]}
+  ]
+}

app/victoria-metrics/testdata/opentsdbhttp/multi_line.json

@@ -0,0 +1,9 @@
+{
+  "name": "multiline",
+  "data": ["[{\"metric\": \"opentsdbhttp.multiline1\", \"value\": 1001, \"timestamp\": \"{TIME_S}\", \"tags\": {\"bar\":\"baz\", \"x\": \"y\"}}, {\"metric\": \"opentsdbhttp.multiline2\", \"value\": 1002, \"timestamp\": {TIME_S}}]"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"opentsdbhttp.multiline1","bar":"baz","x":"y"},"values":[1001], "timestamps": ["{TIME_MSZ}"]},
+    {"metric":{"__name__":"opentsdbhttp.multiline2"},"values":[1002], "timestamps": ["{TIME_MSZ}"]}
+  ]
+}

app/victoria-metrics/testdata/prometheus/basic.json

@@ -0,0 +1,8 @@
+{
+  "name": "basic_insertion",
+  "data": ["[{\"labels\":[{\"name\":\"__name__\",\"value\":\"prometheus.bar\"},{\"name\":\"baz\",\"value\":\"qux\"}],\"samples\":[{\"value\":100000,\"timestamp\":\"{TIME_MS}\"}]}]"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"prometheus.bar","baz":"qux"},"values":[100000], "timestamps": ["{TIME_MS}"]}
+  ]
+}

app/victoria-metrics/testdata/prometheus/case-sensitive-regex.json

@@ -0,0 +1,10 @@
+{
+  "name": "case-sensitive-regex",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/161",
+  "data": ["[{\"labels\":[{\"name\":\"__name__\",\"value\":\"prometheus.sensitiveRegex\"},{\"name\":\"label\",\"value\":\"sensitiveRegex\"}],\"samples\":[{\"value\":2,\"timestamp\":\"{TIME_MS}\"}]},{\"labels\":[{\"name\":\"__name__\",\"value\":\"prometheus.sensitiveRegex\"},{\"name\":\"label\",\"value\":\"SensitiveRegex\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]}]"],
+  "query": ["/api/v1/export?match={label=~'(?i)sensitiveregex'}"],
+  "result_metrics": [
+    {"metric":{"__name__":"prometheus.sensitiveRegex","label":"sensitiveRegex"},"values":[2], "timestamps": ["{TIME_MS}"]},
+    {"metric":{"__name__":"prometheus.sensitiveRegex","label":"SensitiveRegex"},"values":[1], "timestamps": ["{TIME_MS}"]}
+  ]
+}

app/victoria-metrics/testdata/prometheus/duplicate-label.json

@@ -0,0 +1,9 @@
+{
+  "name": "duplicate_label",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/172",
+  "data": ["[{\"labels\":[{\"name\":\"__name__\",\"value\":\"prometheus.duplicate_label\"},{\"name\":\"duplicate\",\"value\":\"label\"},{\"name\":\"duplicate\",\"value\":\"label\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]}]"],
+  "query": ["/api/v1/export?match={__name__!=''}"],
+  "result_metrics": [
+    {"metric":{"__name__":"prometheus.duplicate_label","duplicate":"label"},"values":[1], "timestamps": ["{TIME_MS}"]}
+  ]
+}

app/victoria-metrics/testdata/prometheus/match-series.json

@@ -0,0 +1,15 @@
+{
+  "name": "match_series",
+  "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/155",
+  "data": ["[{\"labels\":[{\"name\":\"__name__\",\"value\":\"MatchSeries\"},{\"name\":\"db\",\"value\":\"TenMinute\"},{\"name\":\"TurbineType\",\"value\":\"V112\"},{\"name\":\"Park\",\"value\":\"1\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]},{\"labels\":[{\"name\":\"__name__\",\"value\":\"MatchSeries\"},{\"name\":\"db\",\"value\":\"TenMinute\"},{\"name\":\"TurbineType\",\"value\":\"V112\"},{\"name\":\"Park\",\"value\":\"2\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]},{\"labels\":[{\"name\":\"__name__\",\"value\":\"MatchSeries\"},{\"name\":\"db\",\"value\":\"TenMinute\"},{\"name\":\"TurbineType\",\"value\":\"V112\"},{\"name\":\"Park\",\"value\":\"3\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]},{\"labels\":[{\"name\":\"__name__\",\"value\":\"MatchSeries\"},{\"name\":\"db\",\"value\":\"TenMinute\"},{\"name\":\"TurbineType\",\"value\":\"V112\"},{\"name\":\"Park\",\"value\":\"4\"}],\"samples\":[{\"value\":1,\"timestamp\":\"{TIME_MS}\"}]}]"],
+  "query": ["/api/v1/series?match[]={__name__='MatchSeries'}", "/api/v1/series?match[]={__name__=~'MatchSeries.*'}"],
+  "result_series": {
+    "status": "success",
+    "data": [
+      {"__name__":"MatchSeries","db":"TenMinute","Park":"1","TurbineType":"V112"},
+      {"__name__":"MatchSeries","db":"TenMinute","Park":"2","TurbineType":"V112"},
+      {"__name__":"MatchSeries","db":"TenMinute","Park":"3","TurbineType":"V112"},
+      {"__name__":"MatchSeries","db":"TenMinute","Park":"4","TurbineType":"V112"}
+    ]
+  }
+}

app/vmbackup/Makefile

@@ -0,0 +1,67 @@
+# All these commands must run from repository root.
+
+vmbackup:
+	APP_NAME=vmbackup $(MAKE) app-local
+
+vmbackup-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker
+
+vmbackup-pure-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-pure
+
+vmbackup-amd64-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-amd64
+
+vmbackup-arm-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-arm
+
+vmbackup-arm64-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-arm64
+
+vmbackup-ppc64le-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-ppc64le
+
+vmbackup-386-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-386
+
+package-vmbackup:
+	APP_NAME=vmbackup $(MAKE) package-via-docker
+
+package-vmbackup-pure:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-pure
+
+package-vmbackup-amd64:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-amd64
+
+package-vmbackup-arm:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-arm
+
+package-vmbackup-arm64:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-arm64
+
+package-vmbackup-ppc64le:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-ppc64le
+
+package-vmbackup-386:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-386
+
+publish-vmbackup:
+	APP_NAME=vmbackup $(MAKE) publish-via-docker
+
+vmbackup-pure:
+	APP_NAME=vmbackup $(MAKE) app-local-pure
+
+vmbackup-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-amd64 ./app/vmbackup
+
+vmbackup-arm:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-arm ./app/vmbackup
+
+vmbackup-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-arm64 ./app/vmbackup
+
+vmbackup-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-ppc64le ./app/vmbackup
+
+vmbackup-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-386 ./app/vmbackup

app/vmbackup/README.md

@@ -0,0 +1,181 @@
+## vmbackup
+
+`vmbackup` creates VictoriaMetrics data backups from [instant snapshots](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots).
+
+Supported storage systems for backups:
+
+* [GCS](https://cloud.google.com/storage/). Example: `gcs://<bucket>/<path/to/backup>`
+* [S3](https://aws.amazon.com/s3/). Example: `s3://<bucket>/<path/to/backup>`
+* Any S3-compatible storage such as [MinIO](https://github.com/minio/minio), [Ceph](https://docs.ceph.com/docs/mimic/radosgw/s3/) or [Swift](https://www.swiftstack.com/docs/admin/middleware/s3_middleware.html). See `-customS3Endpoint` command-line flag.
+* Local filesystem. Example: `fs://</absolute/path/to/backup>`
+
+Incremental backups and full backups are supported. Incremental backups are created automatically if the destination path already contains data from the previous backup.
+Full backups can be sped up with `-origin` pointing to already existing backup on the same remote storage. In this case `vmbackup` makes server-side copy for the shared
+data between the existing backup and new backup. This saves time and costs on data transfer.
+
+Backup process can be interrupted at any time. It is automatically resumed from the interruption point when restarting `vmbackup` with the same args.
+
+Backed up data can be restored with [vmrestore](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md).
+
+See [this article](https://medium.com/@valyala/speeding-up-backups-for-big-time-series-databases-533c1a927883) for more details.
+
+
+### Use cases
+
+#### Regular backups
+
+Regular backup can be performed with the following command:
+
+```
+vmbackup -storageDataPath=</path/to/victoria-metrics-data> -snapshotName=<local-snapshot> -dst=gcs://<bucket>/<path/to/new/backup>
+```
+
+* `</path/to/victoria-metrics-data>` - path to VictoriaMetrics data pointed by `-storageDataPath` command-line flag in single-node VictoriaMetrics or in cluster `vmstorage`.
+  There is no need to stop VictoriaMetrics for creating backups, since they are performed from immutable [instant snapshots](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots).
+* `<local-snapshot>` is the snapshot to backup. See [how to create instant snapshots](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots).
+* `<bucket>` is already existing name for [GCS bucket](https://cloud.google.com/storage/docs/creating-buckets).
+* `<path/to/new/backup>` is the destination path where new backup will be placed.
+
+
+#### Regular backups with server-side copy from existing backup
+
+If the destination GCS bucket already contains the previous backup at `-origin` path, then new backup can be sped up
+with the following command:
+
+```
+vmbackup -storageDataPath=</path/to/victoria-metrics-data> -snapshotName=<local-snapshot> -dst=gcs://<bucket>/<path/to/new/backup> -origin=gcs://<bucket>/<path/to/existing/backup>
+```
+
+This saves time and network bandwidth costs by performing server-side copy for the shared data from the `-origin` to `-dst`.
+
+
+#### Incremental backups
+
+Incremental backups are performed if `-dst` points to already existing backup. In this case only new data is uploaded to remote storage.
+This saves time and network bandwidth costs when working with big backups:
+
+```
+vmbackup -storageDataPath=</path/to/victoria-metrics-data> -snapshotName=<local-snapshot> -dst=gcs://<bucket>/<path/to/existing/backup>
+```
+
+
+#### Smart backups
+
+Smart backups mean storing full daily backups into `YYYYMMDD` folders and creating incremental hourly backup into `latest` folder:
+
+* Run the following command every hour:
+
+```
+vmbackup -snapshotName=<latest-snapshot> -dst=gcs://<bucket>/latest
+```
+
+Where `<latest-snapshot>` is the latest [snapshot](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots).
+The command will upload only changed data to `gcs://<bucket>/latest`.
+
+* Run the following command once a day:
+
+```
+vmbackup -snapshotName=<daily-snapshot> -dst=gcs://<bucket>/<YYYYMMDD> -origin=gcs://<bucket>/latest
+```
+
+Where `<daily-snapshot>` is the snapshot for the last day `<YYYYMMDD>`.
+
+
+This apporach saves network bandwidth costs on hourly backups (since they are incremental) and allows recovering data from either the last hour (`latest` backup)
+or from any day (`YYYYMMDD` backups). Note that hourly backup shouldn't run when creating daily backup.
+
+Do not forget removing old snapshots and backups when they are no longer needed for saving storage costs.
+
+
+### How does it work?
+
+The backup algorithm is the following:
+
+1. Collect information about files in the `-snapshotName`, in the `-dst` and in the `-origin`.
+2. Determine files in `-dst`, which are missing in `-snapshotName`, and delete them. These are usually small files, which are already merged into bigger files in the snapshot.
+3. Determine files from `-snapshotName`, which are missing in `-dst`. These are usually small new files and bigger merged files.
+4. Determine files from step 3, which exist in the `-origin`, and perform server-side copy of these files from `-origin` to `-dst`.
+   This are usually the biggest and the oldest files, which are shared between backups.
+5. Upload the remaining files from setp 3 from `-snapshotName` to `-dst`.
+
+The algorithm splits source files into 100MB chunks in the backup. Each chunk is stored as a separate file in the backup.
+Such splitting minimizes the amounts of data to re-transfer after temporary errors.
+
+`vmbackup` relies on [instant snapshot](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282) properties:
+
+- All the files in the snapshot are immutable.
+- Old files are periodically merged into new files.
+- Smaller files have higher probability to be merged.
+- Consecutive snapshots share many identical files.
+
+These properties allow performing fast and cheap incremental backups and server-side copying from `-origin` paths.
+See [this article](https://medium.com/@valyala/speeding-up-backups-for-big-time-series-databases-533c1a927883) for more details.
+`vmbackup` can work improperly or slowly when these properties are violated.
+
+
+### Troubleshooting
+
+* If the backup is slow, then try setting higher value for `-concurrency` flag. This will increase the number of concurrent workers that upload data to backup storage.
+* If `vmbackup` eats all the network bandwidth, then set `-maxBytesPerSecond` to the desired value.
+* If `vmbackup` has been interrupted due to temporary error, then just restart it with the same args. It will resume the backup process.
+
+
+### Advanced usage
+
+Run `vmbackup -help` in order to see all the available options:
+
+```
+  -concurrency int
+    	The number of concurrent workers. Higher concurrency may reduce backup duration (default 10)
+  -configFilePath string
+    	Path to file with S3 configs. Configs are loaded from default location if not set.
+    	See https://docs.aws.amazon.com/general/latest/gr/aws-security-credentials.html
+  -configProfile string
+    	Profile name for S3 configs (default "default")
+  -credsFilePath string
+    	Path to file with GCS or S3 credentials. Credentials are loaded from default locations if not set.
+    	See https://cloud.google.com/iam/docs/creating-managing-service-account-keys and https://docs.aws.amazon.com/general/latest/gr/aws-security-credentials.html
+  -customS3Endpoint string
+    	Custom S3 endpoint for use with S3-compatible storages (e.g. MinIO). S3 is used if not set
+  -dst string
+    	Where to put the backup on the remote storage. Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir
+    	-dst can point to the previous backup. In this case incremental backup is performed, i.e. only changed data is uploaded
+  -loggerLevel string
+    	Minimum level of errors to log. Possible values: INFO, ERROR, FATAL, PANIC (default "INFO")
+  -maxBytesPerSecond int
+    	The maximum upload speed. There is no limit if it is set to 0
+  -memory.allowedPercent float
+    	Allowed percent of system memory VictoriaMetrics caches may occupy (default 60)
+  -origin string
+    	Optional origin directory on the remote storage with old backup for server-side copying when performing full backup. This speeds up full backups
+  -snapshotName string
+    	Name for the snapshot to backup. See https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots
+  -storageDataPath string
+    	Path to VictoriaMetrics data. Must match -storageDataPath from VictoriaMetrics or vmstorage (default "victoria-metrics-data")
+  -version
+    	Show VictoriaMetrics version
+```
+
+
+### How to build from sources
+
+It is recommended using [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) - see `vmutils-*` archives there.
+
+
+#### Development build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make vmbackup` from the root folder of the repository.
+   It builds `vmbackup` binary and puts it into the `bin` folder.
+
+#### Production build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make vmbackup-prod` from the root folder of the repository.
+   It builds `vmbackup-prod` binary and puts it into the `bin` folder.
+
+#### Building docker images
+
+Run `make package-vmbackup`. It builds `victoriametrics/vmbackup:<PKG_TAG>` docker image locally.
+`<PKG_TAG>` is auto-generated image tag, which depends on source code in the repository.
+The `<PKG_TAG>` may be manually set via `PKG_TAG=foobar make package-vmbackup`.

app/vmbackup/deployment/Dockerfile

@@ -0,0 +1,7 @@
+ARG certs_image
+FROM $certs_image AS certs
+FROM scratch
+COPY --from=certs /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
+ARG src_binary
+COPY $src_binary ./vmbackup-prod
+ENTRYPOINT ["/vmbackup-prod"]

app/vmbackup/main.go

@@ -0,0 +1,114 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"os"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/actions"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/fslocal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+var (
+	storageDataPath = flag.String("storageDataPath", "victoria-metrics-data", "Path to VictoriaMetrics data. Must match -storageDataPath from VictoriaMetrics or vmstorage")
+	snapshotName    = flag.String("snapshotName", "", "Name for the snapshot to backup. See https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots")
+	dst             = flag.String("dst", "", "Where to put the backup on the remote storage. "+
+		"Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir\n"+
+		"-dst can point to the previous backup. In this case incremental backup is performed, i.e. only changed data is uploaded")
+	origin            = flag.String("origin", "", "Optional origin directory on the remote storage with old backup for server-side copying when performing full backup. This speeds up full backups")
+	concurrency       = flag.Int("concurrency", 10, "The number of concurrent workers. Higher concurrency may reduce backup duration")
+	maxBytesPerSecond = flag.Int("maxBytesPerSecond", 0, "The maximum upload speed. There is no limit if it is set to 0")
+)
+
+func main() {
+	flag.Usage = usage
+	flag.Parse()
+	buildinfo.Init()
+
+	srcFS, err := newSrcFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	dstFS, err := newDstFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	originFS, err := newOriginFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	a := &actions.Backup{
+		Concurrency: *concurrency,
+		Src:         srcFS,
+		Dst:         dstFS,
+		Origin:      originFS,
+	}
+	if err := a.Run(); err != nil {
+		logger.Fatalf("cannot create backup: %s", err)
+	}
+}
+
+func usage() {
+	const s = `
+vmbackup performs backups for VictoriaMetrics data from instant snapshots to gcs, s3
+or local filesystem. Backed up data can be restored with vmrestore.
+
+See the docs at https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md .
+`
+
+	f := flag.CommandLine.Output()
+	fmt.Fprintf(f, "%s\n", s)
+	flag.PrintDefaults()
+}
+
+func newSrcFS() (*fslocal.FS, error) {
+	if len(*snapshotName) == 0 {
+		return nil, fmt.Errorf("`-snapshotName` cannot be empty")
+	}
+	snapshotPath := *storageDataPath + "/snapshots/" + *snapshotName
+
+	// Verify the snapshot exists.
+	f, err := os.Open(snapshotPath)
+	if err != nil {
+		return nil, fmt.Errorf("cannot open snapshot at %q: %s", snapshotPath, err)
+	}
+	fi, err := f.Stat()
+	_ = f.Close()
+	if err != nil {
+		return nil, fmt.Errorf("cannot stat %q: %s", snapshotPath, err)
+	}
+	if !fi.IsDir() {
+		return nil, fmt.Errorf("snapshot %q must be a directory", snapshotPath)
+	}
+
+	fs := &fslocal.FS{
+		Dir:               snapshotPath,
+		MaxBytesPerSecond: *maxBytesPerSecond,
+	}
+	if err := fs.Init(); err != nil {
+		return nil, fmt.Errorf("cannot initialize fs: %s", err)
+	}
+	return fs, nil
+}
+
+func newDstFS() (common.RemoteFS, error) {
+	fs, err := actions.NewRemoteFS(*dst)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-dst`=%q: %s", *dst, err)
+	}
+	return fs, nil
+}
+
+func newOriginFS() (common.RemoteFS, error) {
+	if len(*origin) == 0 {
+		return nil, nil
+	}
+	fs, err := actions.NewRemoteFS(*origin)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-origin`=%q: %s", *origin, err)
+	}
+	return fs, nil
+}

app/vminsert/common/gzip_reader.go

@@ -0,0 +1,30 @@
+package common
+
+import (
+	"compress/gzip"
+	"io"
+	"sync"
+)
+
+// GetGzipReader returns new gzip reader from the pool.
+//
+// Return back the gzip reader when it no longer needed with PutGzipReader.
+func GetGzipReader(r io.Reader) (*gzip.Reader, error) {
+	v := gzipReaderPool.Get()
+	if v == nil {
+		return gzip.NewReader(r)
+	}
+	zr := v.(*gzip.Reader)
+	if err := zr.Reset(r); err != nil {
+		return nil, err
+	}
+	return zr, nil
+}
+
+// PutGzipReader returns back gzip reader obtained via GetGzipReader.
+func PutGzipReader(zr *gzip.Reader) {
+	_ = zr.Close()
+	gzipReaderPool.Put(zr)
+}
+
+var gzipReaderPool sync.Pool

app/vminsert/common/insert_ctx.go

@@ -2,9 +2,11 @@ package common
 
 import (
 	"fmt"
+	"net/http"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
@@ -45,7 +47,7 @@ func (ctx *InsertCtx) marshalMetricNameRaw(prefix []byte, labels []prompb.Label)
 	return metricNameRaw[:len(metricNameRaw):len(metricNameRaw)]
 }
 
-// WriteDataPoint writes (timestamp, value) with the given prefix and lables into ctx buffer.
+// WriteDataPoint writes (timestamp, value) with the given prefix and labels into ctx buffer.
 func (ctx *InsertCtx) WriteDataPoint(prefix []byte, labels []prompb.Label, timestamp int64, value float64) {
 	metricNameRaw := ctx.marshalMetricNameRaw(prefix, labels)
 	ctx.addRow(metricNameRaw, timestamp, value)
@@ -76,6 +78,26 @@ func (ctx *InsertCtx) addRow(metricNameRaw []byte, timestamp int64, value float6
 	mr.Value = value
 }
 
+// AddLabelBytes adds (name, value) label to ctx.Labels.
+//
+// name and value must exist until ctx.Labels is used.
+func (ctx *InsertCtx) AddLabelBytes(name, value []byte) {
+	labels := ctx.Labels
+	if cap(labels) > len(labels) {
+		labels = labels[:len(labels)+1]
+	} else {
+		labels = append(labels, prompb.Label{})
+	}
+	label := &labels[len(labels)-1]
+
+	// Do not copy name and value contents for performance reasons.
+	// This reduces GC overhead on the number of objects and allocations.
+	label.Name = name
+	label.Value = value
+
+	ctx.Labels = labels
+}
+
 // AddLabel adds (name, value) label to ctx.Labels.
 //
 // name and value must exist until ctx.Labels is used.
@@ -99,7 +121,10 @@ func (ctx *InsertCtx) AddLabel(name, value string) {
 // FlushBufs flushes buffered rows to the underlying storage.
 func (ctx *InsertCtx) FlushBufs() error {
 	if err := vmstorage.AddRows(ctx.mrs); err != nil {
-		return fmt.Errorf("cannot store metrics: %s", err)
+		return &httpserver.ErrorWithStatusCode{
+			Err:        fmt.Errorf("cannot store metrics: %s", err),
+			StatusCode: http.StatusServiceUnavailable,
+		}
 	}
 	return nil
 }

app/vminsert/common/lines_reader.go

@@ -20,6 +20,17 @@ const defaultBlockSize = 64 * 1024
 //
 // Returns (dstBuf, tailBuf).
 func ReadLinesBlock(r io.Reader, dstBuf, tailBuf []byte) ([]byte, []byte, error) {
+	return ReadLinesBlockExt(r, dstBuf, tailBuf, maxLineSize)
+}
+
+// ReadLinesBlockExt reads a block of lines delimited by '\n' from tailBuf and r into dstBuf.
+//
+// Trailing chars after the last newline are put into tailBuf.
+//
+// Returns (dstBuf, tailBuf).
+//
+// maxLineLen limits the maximum length of a single line.
+func ReadLinesBlockExt(r io.Reader, dstBuf, tailBuf []byte, maxLineLen int) ([]byte, []byte, error) {
 	if cap(dstBuf) < defaultBlockSize {
 		dstBuf = bytesutil.Resize(dstBuf, defaultBlockSize)
 	}
@@ -48,8 +59,8 @@ again:
 	nn := bytes.LastIndexByte(dstBuf[len(dstBuf)-n:], '\n')
 	if nn < 0 {
 		// Didn't found at least a single line.
-		if len(dstBuf) > maxLineSize {
-			return dstBuf, tailBuf, fmt.Errorf("too long line: more than %d bytes", maxLineSize)
+		if len(dstBuf) > maxLineLen {
+			return dstBuf, tailBuf, fmt.Errorf("too long line: more than %d bytes", maxLineLen)
 		}
 		if cap(dstBuf) < 2*len(dstBuf) {
 			// Increase dsbBuf capacity, so more data could be read into it.

app/vminsert/concurrencylimiter/concurrencylimiter.go

@@ -3,9 +3,11 @@ package concurrencylimiter
 import (
 	"flag"
 	"fmt"
+	"net/http"
 	"runtime"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -53,7 +55,10 @@ func Do(f func() error) error {
 	case <-t.C:
 		timerpool.Put(t)
 		concurrencyLimitTimeout.Inc()
-		return fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase -maxConcurrentInserts or reduce the load", cap(ch))
+		return &httpserver.ErrorWithStatusCode{
+			Err:        fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase -maxConcurrentInserts or reduce the load", cap(ch)),
+			StatusCode: http.StatusServiceUnavailable,
+		}
 	}
 }
 

app/vminsert/graphite/parser.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/fastjson/fastfloat"
 )
 
@@ -34,13 +36,8 @@ func (rs *Rows) Reset() {
 // See https://graphite.readthedocs.io/en/latest/feeding-carbon.html#the-plaintext-protocol
 //
 // s must be unchanged until rs is in use.
-func (rs *Rows) Unmarshal(s string) error {
-	var err error
-	rs.Rows, rs.tagsPool, err = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0])
-	if err != nil {
-		return err
-	}
-	return err
+func (rs *Rows) Unmarshal(s string) {
+	rs.Rows, rs.tagsPool = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0])
 }
 
 // Row is a single graphite row.
@@ -83,6 +80,9 @@ func (r *Row) unmarshal(s string, tagsPool []Tag) ([]Tag, error) {
 		tags := tagsPool[tagsStart:]
 		r.Tags = tags[:len(tags):len(tags)]
 	}
+	if len(r.Metric) == 0 {
+		return tagsPool, fmt.Errorf("metric cannot be empty")
+	}
 
 	n = strings.IndexByte(tail, ' ')
 	if n < 0 {
@@ -95,41 +95,46 @@ func (r *Row) unmarshal(s string, tagsPool []Tag) ([]Tag, error) {
 	return tagsPool, nil
 }
 
-func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
+func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag) {
 	for len(s) > 0 {
 		n := strings.IndexByte(s, '\n')
-		if n == 0 {
-			// Skip empty line
-			s = s[1:]
-			continue
-		}
-		if cap(dst) > len(dst) {
-			dst = dst[:len(dst)+1]
-		} else {
-			dst = append(dst, Row{})
-		}
-		r := &dst[len(dst)-1]
 		if n < 0 {
 			// The last line.
-			var err error
-			tagsPool, err = r.unmarshal(s, tagsPool)
-			if err != nil {
-				err = fmt.Errorf("cannot unmarshal Graphite line %q: %s", s, err)
-				return dst, tagsPool, err
-			}
-			return dst, tagsPool, nil
-		}
-		var err error
-		tagsPool, err = r.unmarshal(s[:n], tagsPool)
-		if err != nil {
-			err = fmt.Errorf("cannot unmarshal Graphite line %q: %s", s[:n], err)
-			return dst, tagsPool, err
+			return unmarshalRow(dst, s, tagsPool)
 		}
+		dst, tagsPool = unmarshalRow(dst, s[:n], tagsPool)
 		s = s[n+1:]
 	}
-	return dst, tagsPool, nil
+	return dst, tagsPool
 }
 
+func unmarshalRow(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag) {
+	if len(s) > 0 && s[len(s)-1] == '\r' {
+		s = s[:len(s)-1]
+	}
+	if len(s) == 0 {
+		// Skip empty line
+		return dst, tagsPool
+	}
+
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	var err error
+	tagsPool, err = r.unmarshal(s, tagsPool)
+	if err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal Graphite line %q: %s", s, err)
+		invalidLines.Inc()
+	}
+	return dst, tagsPool
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="graphite"}`)
+
 func unmarshalTags(dst []Tag, s string) ([]Tag, error) {
 	for {
 		if cap(dst) > len(dst) {
@@ -145,12 +150,20 @@ func unmarshalTags(dst []Tag, s string) ([]Tag, error) {
 			if err := tag.unmarshal(s); err != nil {
 				return dst[:len(dst)-1], err
 			}
+			if len(tag.Key) == 0 || len(tag.Value) == 0 {
+				// Skip empty tag
+				dst = dst[:len(dst)-1]
+			}
 			return dst, nil
 		}
 		if err := tag.unmarshal(s[:n]); err != nil {
 			return dst[:len(dst)-1], err
 		}
 		s = s[n+1:]
+		if len(tag.Key) == 0 || len(tag.Value) == 0 {
+			// Skip empty tag
+			dst = dst[:len(dst)-1]
+		}
 	}
 }
 
@@ -172,9 +185,6 @@ func (t *Tag) unmarshal(s string) error {
 		return fmt.Errorf("missing tag value for %q", s)
 	}
 	t.Key = s[:n]
-	if len(t.Key) == 0 {
-		return fmt.Errorf("tag key cannot be empty for %q", s)
-	}
 	t.Value = s[n+1:]
 	return nil
 }

app/vminsert/graphite/parser_test.go

@@ -9,45 +9,42 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	f := func(s string) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
 		}
 
 		// Try again
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
 		}
 	}
 
+	// Missing metric
+	f(" 123 455")
+
 	// Missing value
 	f("aaa")
 
-	// Invalid multiline
-	f("aaa\nbbb 123 34")
-
 	// missing tag
 	f("aa; 12 34")
 
 	// missing tag value
 	f("aa;bb 23 34")
-	f("aa;=dsd 234 45")
 }
 
 func TestRowsUnmarshalSuccess(t *testing.T) {
 	f := func(s string, rowsExpected *Rows) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
 
 		// Try unmarshaling again
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
@@ -60,7 +57,9 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 
 	// Empty line
 	f("", &Rows{})
+	f("\r", &Rows{})
 	f("\n\n", &Rows{})
+	f("\n\r\n", &Rows{})
 
 	// Single line
 	f("foobar -123.456 789", &Rows{
@@ -86,6 +85,15 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 		}},
 	})
 
+	// Timestamp bigger than 1<<31
+	f("aaa 1123 429496729600", &Rows{
+		Rows: []Row{{
+			Metric:    "aaa",
+			Value:     1123,
+			Timestamp: 429496729600,
+		}},
+	})
+
 	// Tags
 	f("foo;bar=baz 1 2", &Rows{
 		Rows: []Row{{
@@ -98,7 +106,8 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			Timestamp: 2,
 		}},
 	})
-	f("foo;bar=baz;aa=;x=y 1 2", &Rows{
+	// Empty tags
+	f("foo;bar=baz;aa=;x=y;=z 1 2", &Rows{
 		Rows: []Row{{
 			Metric: "foo",
 			Tags: []Tag{
@@ -106,10 +115,6 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 					Key:   "bar",
 					Value: "baz",
 				},
-				{
-					Key:   "aa",
-					Value: "",
-				},
 				{
 					Key:   "x",
 					Value: "y",
@@ -139,4 +144,20 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+
+	// Multi lines with invalid line
+	f("foo 0.3 2\naaa\nbar.baz 0.34 43\n", &Rows{
+		Rows: []Row{
+			{
+				Metric:    "foo",
+				Value:     0.3,
+				Timestamp: 2,
+			},
+			{
+				Metric:    "bar.baz",
+				Value:     0.34,
+				Timestamp: 43,
+			},
+		},
+	})
 }

app/vminsert/graphite/parser_timing_test.go

@@ -16,8 +16,9 @@ cpu.usage_irq 0.34432 1234556768
 	b.RunParallel(func(pb *testing.PB) {
 		var rows Rows
 		for pb.Next() {
-			if err := rows.Unmarshal(s); err != nil {
-				panic(fmt.Errorf("cannot unmarshal %q: %s", s, err))
+			rows.Unmarshal(s)
+			if len(rows.Rows) != 4 {
+				panic(fmt.Errorf("unexpected number of rows unmarshaled: got %d; want 4", len(rows.Rows)))
 			}
 		}
 	})

app/vminsert/graphite/request_handler.go

@@ -61,13 +61,13 @@ func (ctx *pushCtx) InsertRows() error {
 const flushTimeout = 3 * time.Second
 
 func (ctx *pushCtx) Read(r io.Reader) bool {
-	graphiteReadCalls.Inc()
+	readCalls.Inc()
 	if ctx.err != nil {
 		return false
 	}
 	if c, ok := r.(net.Conn); ok {
 		if err := c.SetReadDeadline(time.Now().Add(flushTimeout)); err != nil {
-			graphiteReadErrors.Inc()
+			readErrors.Inc()
 			ctx.err = fmt.Errorf("cannot set read deadline: %s", err)
 			return false
 		}
@@ -79,17 +79,13 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			ctx.err = nil
 		} else {
 			if ctx.err != io.EOF {
-				graphiteReadErrors.Inc()
+				readErrors.Inc()
 				ctx.err = fmt.Errorf("cannot read graphite plaintext protocol data: %s", ctx.err)
 			}
 			return false
 		}
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		graphiteUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal graphite plaintext protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
-	}
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
 
 	// Fill missing timestamps with the current timestamp rounded to seconds.
 	currentTimestamp := time.Now().Unix()
@@ -136,9 +132,8 @@ func (ctx *pushCtx) reset() {
 }
 
 var (
-	graphiteReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
-	graphiteReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
-	graphiteUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="graphite"}`)
+	readCalls  = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
+	readErrors = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/graphite/server.go

@@ -21,36 +21,62 @@ var (
 	writeErrorsUDP   = metrics.NewCounter(`vm_graphite_request_errors_total{name="write", net="udp"}`)
 )
 
-// Serve starts graphite server on the given addr.
-func Serve(addr string) {
+// Server accepts Graphite plaintext lines over TCP and UDP.
+type Server struct {
+	addr  string
+	lnTCP net.Listener
+	lnUDP net.PacketConn
+	wg    sync.WaitGroup
+}
+
+// MustStart starts graphite server on the given addr.
+//
+// MustStop must be called on the returned server when it is no longer needed.
+func MustStart(addr string) *Server {
 	logger.Infof("starting TCP Graphite server at %q", addr)
 	lnTCP, err := netutil.NewTCPListener("graphite", addr)
 	if err != nil {
 		logger.Fatalf("cannot start TCP Graphite server at %q: %s", addr, err)
 	}
-	listenerTCP = lnTCP
 
 	logger.Infof("starting UDP Graphite server at %q", addr)
 	lnUDP, err := net.ListenPacket("udp4", addr)
 	if err != nil {
 		logger.Fatalf("cannot start UDP Graphite server at %q: %s", addr, err)
 	}
-	listenerUDP = lnUDP
 
-	var wg sync.WaitGroup
-	wg.Add(1)
+	s := &Server{
+		addr:  addr,
+		lnTCP: lnTCP,
+		lnUDP: lnUDP,
+	}
+	s.wg.Add(1)
 	go func() {
-		defer wg.Done()
-		serveTCP(listenerTCP)
+		defer s.wg.Done()
+		serveTCP(lnTCP)
 		logger.Infof("stopped TCP Graphite server at %q", addr)
 	}()
-	wg.Add(1)
+	s.wg.Add(1)
 	go func() {
-		defer wg.Done()
-		serveUDP(listenerUDP)
+		defer s.wg.Done()
+		serveUDP(lnUDP)
 		logger.Infof("stopped UDP Graphite server at %q", addr)
 	}()
-	wg.Wait()
+	return s
+}
+
+// MustStop stops the server.
+func (s *Server) MustStop() {
+	logger.Infof("stopping TCP Graphite server at %q...", s.addr)
+	if err := s.lnTCP.Close(); err != nil {
+		logger.Errorf("cannot close TCP Graphite server: %s", err)
+	}
+	logger.Infof("stopping UDP Graphite server at %q...", s.addr)
+	if err := s.lnUDP.Close(); err != nil {
+		logger.Errorf("cannot close UDP Graphite server: %s", err)
+	}
+	s.wg.Wait()
+	logger.Infof("TCP and UDP Graphite servers at %q have been stopped", s.addr)
 }
 
 func serveTCP(ln net.Listener) {
@@ -59,6 +85,7 @@ func serveTCP(ln net.Listener) {
 		if err != nil {
 			if ne, ok := err.(net.Error); ok {
 				if ne.Temporary() {
+					logger.Errorf("graphite: temporary error when listening for TCP addr %q: %s", ln.Addr(), err)
 					time.Sleep(time.Second)
 					continue
 				}
@@ -97,6 +124,7 @@ func serveUDP(ln net.PacketConn) {
 					writeErrorsUDP.Inc()
 					if ne, ok := err.(net.Error); ok {
 						if ne.Temporary() {
+							logger.Errorf("graphite: temporary error when listening for UDP addr %q: %s", ln.LocalAddr(), err)
 							time.Sleep(time.Second)
 							continue
 						}
@@ -119,20 +147,3 @@ func serveUDP(ln net.PacketConn) {
 	}
 	wg.Wait()
 }
-
-var (
-	listenerTCP net.Listener
-	listenerUDP net.PacketConn
-)
-
-// Stop stops the server.
-func Stop() {
-	logger.Infof("stopping TCP Graphite server at %q...", listenerTCP.Addr())
-	if err := listenerTCP.Close(); err != nil {
-		logger.Errorf("cannot close TCP Graphite server: %s", err)
-	}
-	logger.Infof("stopping UDP Graphite server at %q...", listenerUDP.LocalAddr())
-	if err := listenerUDP.Close(); err != nil {
-		logger.Errorf("cannot close UDP Graphite server: %s", err)
-	}
-}

app/vminsert/influx/parser.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/fastjson/fastfloat"
 )
 
@@ -41,13 +43,8 @@ func (rs *Rows) Reset() {
 // See https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/
 //
 // s must be unchanged until rs is in use.
-func (rs *Rows) Unmarshal(s string) error {
-	var err error
-	rs.Rows, rs.tagsPool, rs.fieldsPool, err = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0], rs.fieldsPool[:0])
-	if err != nil {
-		return err
-	}
-	return err
+func (rs *Rows) Unmarshal(s string) {
+	rs.Rows, rs.tagsPool, rs.fieldsPool = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0], rs.fieldsPool[:0])
 }
 
 // Row is a single influx row.
@@ -65,9 +62,8 @@ func (r *Row) reset() {
 	r.Timestamp = 0
 }
 
-func (r *Row) unmarshal(s string, tagsPool []Tag, fieldsPool []Field) ([]Tag, []Field, error) {
+func (r *Row) unmarshal(s string, tagsPool []Tag, fieldsPool []Field, noEscapeChars bool) ([]Tag, []Field, error) {
 	r.reset()
-	noEscapeChars := strings.IndexByte(s, '\\') < 0
 	n := nextUnescapedChar(s, ' ', noEscapeChars)
 	if n < 0 {
 		return tagsPool, fieldsPool, fmt.Errorf("cannot find Whitespace I in %q", s)
@@ -89,9 +85,7 @@ func (r *Row) unmarshal(s string, tagsPool []Tag, fieldsPool []Field) ([]Tag, []
 		measurementTags = measurementTags[:n]
 	}
 	r.Measurement = unescapeTagValue(measurementTags, noEscapeChars)
-	if len(r.Measurement) == 0 {
-		return tagsPool, fieldsPool, fmt.Errorf("measurement cannot be empty. measurementTags=%q", s)
-	}
+	// Allow empty r.Measurement. In this case metric name is constructed directly from field keys.
 
 	// Parse fields
 	fieldsStart := len(fieldsPool)
@@ -141,9 +135,6 @@ func (tag *Tag) unmarshal(s string, noEscapeChars bool) error {
 		return fmt.Errorf("missing tag value for %q", s)
 	}
 	tag.Key = unescapeTagValue(s[:n], noEscapeChars)
-	if len(tag.Key) == 0 {
-		return fmt.Errorf("tag key cannot be empty")
-	}
 	tag.Value = unescapeTagValue(s[n+1:], noEscapeChars)
 	return nil
 }
@@ -177,41 +168,51 @@ func (f *Field) unmarshal(s string, noEscapeChars, hasQuotedFields bool) error {
 	return nil
 }
 
-func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]Row, []Tag, []Field, error) {
+func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]Row, []Tag, []Field) {
+	noEscapeChars := strings.IndexByte(s, '\\') < 0
 	for len(s) > 0 {
 		n := strings.IndexByte(s, '\n')
-		if n == 0 {
-			// Skip empty line
-			s = s[1:]
-			continue
-		}
-		if cap(dst) > len(dst) {
-			dst = dst[:len(dst)+1]
-		} else {
-			dst = append(dst, Row{})
-		}
-		r := &dst[len(dst)-1]
 		if n < 0 {
 			// The last line.
-			var err error
-			tagsPool, fieldsPool, err = r.unmarshal(s, tagsPool, fieldsPool)
-			if err != nil {
-				err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s, err)
-				return dst, tagsPool, fieldsPool, err
-			}
-			return dst, tagsPool, fieldsPool, nil
-		}
-		var err error
-		tagsPool, fieldsPool, err = r.unmarshal(s[:n], tagsPool, fieldsPool)
-		if err != nil {
-			err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s[:n], err)
-			return dst, tagsPool, fieldsPool, err
+			return unmarshalRow(dst, s, tagsPool, fieldsPool, noEscapeChars)
 		}
+		dst, tagsPool, fieldsPool = unmarshalRow(dst, s[:n], tagsPool, fieldsPool, noEscapeChars)
 		s = s[n+1:]
 	}
-	return dst, tagsPool, fieldsPool, nil
+	return dst, tagsPool, fieldsPool
 }
 
+func unmarshalRow(dst []Row, s string, tagsPool []Tag, fieldsPool []Field, noEscapeChars bool) ([]Row, []Tag, []Field) {
+	if len(s) > 0 && s[len(s)-1] == '\r' {
+		s = s[:len(s)-1]
+	}
+	if len(s) == 0 {
+		// Skip empty line
+		return dst, tagsPool, fieldsPool
+	}
+	if s[0] == '#' {
+		// Skip comment
+		return dst, tagsPool, fieldsPool
+	}
+
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	var err error
+	tagsPool, fieldsPool, err = r.unmarshal(s, tagsPool, fieldsPool, noEscapeChars)
+	if err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal Influx line %q: %s; skipping it", s, err)
+		invalidLines.Inc()
+	}
+	return dst, tagsPool, fieldsPool
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="influx"}`)
+
 func unmarshalTags(dst []Tag, s string, noEscapeChars bool) ([]Tag, error) {
 	for {
 		if cap(dst) > len(dst) {
@@ -223,14 +224,22 @@ func unmarshalTags(dst []Tag, s string, noEscapeChars bool) ([]Tag, error) {
 		n := nextUnescapedChar(s, ',', noEscapeChars)
 		if n < 0 {
 			if err := tag.unmarshal(s, noEscapeChars); err != nil {
-				return dst, err
+				return dst[:len(dst)-1], err
+			}
+			if len(tag.Key) == 0 || len(tag.Value) == 0 {
+				// Skip empty tag
+				dst = dst[:len(dst)-1]
 			}
 			return dst, nil
 		}
 		if err := tag.unmarshal(s[:n], noEscapeChars); err != nil {
-			return dst, err
+			return dst[:len(dst)-1], err
 		}
 		s = s[n+1:]
+		if len(tag.Key) == 0 || len(tag.Value) == 0 {
+			// Skip empty tag
+			dst = dst[:len(dst)-1]
+		}
 	}
 }
 

app/vminsert/influx/parser_test.go

@@ -74,19 +74,18 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	f := func(s string) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
 		}
 
 		// Try again
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
 		}
 	}
 
-	// Missing measurement
-	f(",foo=bar baz=123")
-
 	// No fields
 	f("foo")
 	f("foo,bar=baz 1234")
@@ -94,12 +93,8 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	// Missing tag value
 	f("foo,bar")
 	f("foo,bar baz")
-	f("foo,bar= baz")
 	f("foo,bar=123, 123")
 
-	// Missing tag name
-	f("foo,=bar baz=234")
-
 	// Missing field value
 	f("foo bar")
 	f("foo bar=")
@@ -122,17 +117,13 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	f := func(s string, rowsExpected *Rows) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
 
 		// Try unmarshaling again
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
@@ -146,6 +137,36 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	// Empty line
 	f("", &Rows{})
 	f("\n\n", &Rows{})
+	f("\n\r\n", &Rows{})
+
+	// Comment
+	f("\n# foobar\n", &Rows{})
+	f("#foobar baz", &Rows{})
+	f("#foobar baz\n#sss", &Rows{})
+
+	// Missing measurement
+	f(" baz=123", &Rows{
+		Rows: []Row{{
+			Measurement: "",
+			Fields: []Field{{
+				Key:   "baz",
+				Value: 123,
+			}},
+		}},
+	})
+	f(",foo=bar baz=123", &Rows{
+		Rows: []Row{{
+			Measurement: "",
+			Tags: []Tag{{
+				Key:   "foo",
+				Value: "bar",
+			}},
+			Fields: []Field{{
+				Key:   "baz",
+				Value: 123,
+			}},
+		}},
+	})
 
 	// Minimal line without tags and timestamp
 	f("foo bar=123", &Rows{
@@ -157,6 +178,15 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			}},
 		}},
 	})
+	f("# comment\nfoo bar=123\r\n#comment2 sdsf dsf", &Rows{
+		Rows: []Row{{
+			Measurement: "foo",
+			Fields: []Field{{
+				Key:   "bar",
+				Value: 123,
+			}},
+		}},
+	})
 	f("foo bar=123\n", &Rows{
 		Rows: []Row{{
 			Measurement: "foo",
@@ -216,7 +246,7 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Line with empty tag values
-	f("foo,tag1=xyz,tagN=,tag2=43as bar=123", &Rows{
+	f("foo,tag1=xyz,tagN=,tag2=43as,=xxx bar=123", &Rows{
 		Rows: []Row{{
 			Measurement: "foo",
 			Tags: []Tag{
@@ -224,10 +254,6 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 					Key:   "tag1",
 					Value: "xyz",
 				},
-				{
-					Key:   "tagN",
-					Value: "",
-				},
 				{
 					Key:   "tag2",
 					Value: "43as",
@@ -309,11 +335,11 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Escape chars
-	f(`fo\,bar\=baz,x\==\\a\,\=\q\  \\\a\=\,=4.34`, &Rows{
+	f(`fo\,bar\=baz,x\=\b=\\a\,\=\q\  \\\a\=\,=4.34`, &Rows{
 		Rows: []Row{{
 			Measurement: `fo,bar=baz`,
 			Tags: []Tag{{
-				Key:   `x=`,
+				Key:   `x=\b`,
 				Value: `\a,=\q `,
 			}},
 			Fields: []Field{{
@@ -322,6 +348,36 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			}},
 		}},
 	})
+	// Test case from https://community.librenms.org/t/integration-with-victoriametrics/9689
+	f("ports,foo=a,bar=et\\ +\\ V,baz=ype INDISCARDS=245333676,OUTDISCARDS=1798680", &Rows{
+		Rows: []Row{{
+			Measurement: "ports",
+			Tags: []Tag{
+				{
+					Key:   "foo",
+					Value: "a",
+				},
+				{
+					Key:   "bar",
+					Value: "et + V",
+				},
+				{
+					Key:   "baz",
+					Value: "ype",
+				},
+			},
+			Fields: []Field{
+				{
+					Key:   "INDISCARDS",
+					Value: 245333676,
+				},
+				{
+					Key:   "OUTDISCARDS",
+					Value: 1798680,
+				},
+			},
+		}},
+	})
 
 	// Multiple lines
 	f("foo,tag=xyz field=1.23 48934\n"+
@@ -348,6 +404,34 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+
+	// Multiple lines with invalid line in the middle.
+	f("foo,tag=xyz field=1.23 48934\n"+
+		"invalid line\n"+
+		"bar x=-1i\n\n", &Rows{
+		Rows: []Row{
+			{
+				Measurement: "foo",
+				Tags: []Tag{{
+					Key:   "tag",
+					Value: "xyz",
+				}},
+				Fields: []Field{{
+					Key:   "field",
+					Value: 1.23,
+				}},
+				Timestamp: 48934,
+			},
+			{
+				Measurement: "bar",
+				Fields: []Field{{
+					Key:   "x",
+					Value: -1,
+				}},
+			},
+		},
+	})
+
 	// No newline after the second line.
 	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/82
 	f("foo,tag=xyz field=1.23 48934\n"+
@@ -374,4 +458,24 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+
+	f("x,y=z,g=p:\\ \\ 5432\\,\\ gp\\ mon\\ [lol]\\ con10\\ cmd5\\ SELECT f=1", &Rows{
+		Rows: []Row{{
+			Measurement: "x",
+			Tags: []Tag{
+				{
+					Key:   "y",
+					Value: "z",
+				},
+				{
+					Key:   "g",
+					Value: "p:  5432, gp mon [lol] con10 cmd5 SELECT",
+				},
+			},
+			Fields: []Field{{
+				Key:   "f",
+				Value: 1,
+			}},
+		}},
+	})
 }

app/vminsert/influx/parser_timing_test.go

@@ -6,14 +6,19 @@ import (
 )
 
 func BenchmarkRowsUnmarshal(b *testing.B) {
-	s := `cpu usage_user=1.23,usage_system=4.34,usage_iowait=0.1112 1234556768`
+	s := `cpu usage_user=1.23,usage_system=4.34,usage_iowait=0.1112 1234556768
+cpu usage_user=1.23,usage_system=4.34,usage_iowait=0.1112 123455676344
+aaa usage_user=1.23,usage_system=4.34,usage_iowait=0.1112 123455676344
+bbb usage_user=1.23,usage_system=4.34,usage_iowait=0.1112 123455676344
+`
 	b.SetBytes(int64(len(s)))
 	b.ReportAllocs()
 	b.RunParallel(func(pb *testing.PB) {
 		var rows Rows
 		for pb.Next() {
-			if err := rows.Unmarshal(s); err != nil {
-				panic(fmt.Errorf("cannot unmarshal %q: %s", s, err))
+			rows.Unmarshal(s)
+			if len(rows.Rows) != 4 {
+				panic(fmt.Errorf("unexpected number of rows parsed; got %d; want 4", len(rows.Rows)))
 			}
 		}
 	})

app/vminsert/influx/request_handler.go

@@ -1,7 +1,6 @@
 package influx
 
 import (
-	"compress/gzip"
 	"flag"
 	"fmt"
 	"io"
@@ -37,15 +36,15 @@ func InsertHandler(req *http.Request) error {
 }
 
 func insertHandlerInternal(req *http.Request) error {
-	influxReadCalls.Inc()
+	readCalls.Inc()
 
 	r := req.Body
 	if req.Header.Get("Content-Encoding") == "gzip" {
-		zr, err := getGzipReader(r)
+		zr, err := common.GetGzipReader(r)
 		if err != nil {
 			return fmt.Errorf("cannot read gzipped influx line protocol data: %s", err)
 		}
-		defer putGzipReader(zr)
+		defer common.PutGzipReader(zr)
 		r = zr
 	}
 
@@ -83,7 +82,7 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	rows := ctx.Rows.Rows
 	rowsLen := 0
 	for i := range rows {
-		rowsLen += len(rows[i].Tags)
+		rowsLen += len(rows[i].Fields)
 	}
 	ic := &ctx.Common
 	ic.Reset(rowsLen)
@@ -91,15 +90,21 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	for i := range rows {
 		r := &rows[i]
 		ic.Labels = ic.Labels[:0]
-		ic.AddLabel("db", db)
+		hasDBLabel := false
 		for j := range r.Tags {
 			tag := &r.Tags[j]
+			if tag.Key == "db" {
+				hasDBLabel = true
+			}
 			ic.AddLabel(tag.Key, tag.Value)
 		}
+		if len(db) > 0 && !hasDBLabel {
+			ic.AddLabel("db", db)
+		}
 		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
 		ctx.metricGroupBuf = append(ctx.metricGroupBuf[:0], r.Measurement...)
 		skipFieldKey := len(r.Fields) == 1 && *skipSingleField
-		if !skipFieldKey {
+		if len(ctx.metricGroupBuf) > 0 && !skipFieldKey {
 			ctx.metricGroupBuf = append(ctx.metricGroupBuf, *measurementFieldSeparator...)
 		}
 		metricGroupPrefixLen := len(ctx.metricGroupBuf)
@@ -120,25 +125,6 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	return ic.FlushBufs()
 }
 
-func getGzipReader(r io.Reader) (*gzip.Reader, error) {
-	v := gzipReaderPool.Get()
-	if v == nil {
-		return gzip.NewReader(r)
-	}
-	zr := v.(*gzip.Reader)
-	if err := zr.Reset(r); err != nil {
-		return nil, err
-	}
-	return zr, nil
-}
-
-func putGzipReader(zr *gzip.Reader) {
-	_ = zr.Close()
-	gzipReaderPool.Put(zr)
-}
-
-var gzipReaderPool sync.Pool
-
 func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 	if ctx.err != nil {
 		return false
@@ -146,16 +132,12 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
 	if ctx.err != nil {
 		if ctx.err != io.EOF {
-			influxReadErrors.Inc()
+			readErrors.Inc()
 			ctx.err = fmt.Errorf("cannot read influx line protocol data: %s", ctx.err)
 		}
 		return false
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		influxUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal influx line protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
-	}
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
 
 	// Adjust timestamps according to tsMultiplier
 	currentTs := time.Now().UnixNano() / 1e6
@@ -184,9 +166,8 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 }
 
 var (
-	influxReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
-	influxReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
-	influxUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="influx"}`)
+	readCalls  = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
+	readErrors = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
 )
 
 type pushCtx struct {

app/vminsert/main.go

@@ -10,35 +10,56 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/graphite"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/influx"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdbhttp"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/prometheus"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/vmimport"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 )
 
 var (
-	graphiteListenAddr   = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
-	opentsdbListenAddr   = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB put messages. Usually :4242 must be set. Doesn't work if empty")
-	maxInsertRequestSize = flag.Int("maxInsertRequestSize", 32*1024*1024, "The maximum size of a single insert request in bytes")
+	graphiteListenAddr = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
+	opentsdbListenAddr = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB metrics. "+
+		"Telnet put messages and HTTP /api/put messages are simultaneously served on TCP port. "+
+		"Usually :4242 must be set. Doesn't work if empty")
+	opentsdbHTTPListenAddr = flag.String("opentsdbHTTPListenAddr", "", "TCP address to listen for OpentTSDB HTTP put requests. Usually :4242 must be set. Doesn't work if empty")
+	maxInsertRequestSize   = flag.Int("maxInsertRequestSize", 32*1024*1024, "The maximum size of a single insert request in bytes")
+	maxLabelsPerTimeseries = flag.Int("maxLabelsPerTimeseries", 30, "The maximum number of labels accepted per time series. Superflouos labels are dropped")
+)
+
+var (
+	graphiteServer     *graphite.Server
+	opentsdbServer     *opentsdb.Server
+	opentsdbhttpServer *opentsdbhttp.Server
 )
 
 // Init initializes vminsert.
 func Init() {
+	storage.SetMaxLabelsPerTimeseries(*maxLabelsPerTimeseries)
+
 	concurrencylimiter.Init()
 	if len(*graphiteListenAddr) > 0 {
-		go graphite.Serve(*graphiteListenAddr)
+		graphiteServer = graphite.MustStart(*graphiteListenAddr)
 	}
 	if len(*opentsdbListenAddr) > 0 {
-		go opentsdb.Serve(*opentsdbListenAddr)
+		opentsdbServer = opentsdb.MustStart(*opentsdbListenAddr, int64(*maxInsertRequestSize))
+	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttpServer = opentsdbhttp.MustStart(*opentsdbHTTPListenAddr, int64(*maxInsertRequestSize))
 	}
 }
 
 // Stop stops vminsert.
 func Stop() {
 	if len(*graphiteListenAddr) > 0 {
-		graphite.Stop()
+		graphiteServer.MustStop()
 	}
 	if len(*opentsdbListenAddr) > 0 {
-		opentsdb.Stop()
+		opentsdbServer.MustStop()
+	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttpServer.MustStop()
 	}
 }
 
@@ -55,6 +76,15 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		}
 		w.WriteHeader(http.StatusNoContent)
 		return true
+	case "/api/v1/import":
+		vmimportRequests.Inc()
+		if err := vmimport.InsertHandler(r); err != nil {
+			vmimportErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
 	case "/write", "/api/v2/write":
 		influxWriteRequests.Inc()
 		if err := influx.InsertHandler(r); err != nil {
@@ -80,6 +110,9 @@ var (
 	prometheusWriteRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/write", protocol="prometheus"}`)
 	prometheusWriteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/write", protocol="prometheus"}`)
 
+	vmimportRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/import", protocol="vm"}`)
+	vmimportErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/import", protocol="vm"}`)
+
 	influxWriteRequests = metrics.NewCounter(`vm_http_requests_total{path="/write", protocol="influx"}`)
 	influxWriteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/write", protocol="influx"}`)
 

app/vminsert/opentsdb/listener_switch.go

@@ -0,0 +1,159 @@
+package opentsdb
+
+import (
+	"io"
+	"net"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+// listenerSwitch listens for incoming connections and multiplexes them to OpenTSDB http or telnet listeners
+// depending on the first byte in the accepted connection.
+//
+// It is expected that both listeners - http and telnet consume incoming connections as soon as possible.
+type listenerSwitch struct {
+	ln net.Listener
+	wg sync.WaitGroup
+
+	telnetConnsCh chan net.Conn
+	httpConnsCh   chan net.Conn
+
+	closeLock sync.Mutex
+	closed    bool
+	acceptErr error
+	closeErr  error
+}
+
+func newListenerSwitch(ln net.Listener) *listenerSwitch {
+	ls := &listenerSwitch{
+		ln: ln,
+	}
+	ls.telnetConnsCh = make(chan net.Conn)
+	ls.httpConnsCh = make(chan net.Conn)
+	ls.wg.Add(1)
+	go func() {
+		ls.worker()
+		close(ls.telnetConnsCh)
+		close(ls.httpConnsCh)
+		ls.wg.Done()
+	}()
+	return ls
+}
+
+func (ls *listenerSwitch) stop() error {
+	var err error
+	ls.closeLock.Lock()
+	if !ls.closed {
+		err = ls.ln.Close()
+		ls.closeErr = err
+		ls.closed = true
+	}
+	ls.closeLock.Unlock()
+
+	if err == nil {
+		// Wait until worker detects the closed ls.ln and exits.
+		ls.wg.Wait()
+	}
+	return err
+}
+
+func (ls *listenerSwitch) worker() {
+	var buf [1]byte
+	for {
+		c, err := ls.ln.Accept()
+		if err != nil {
+			if ne, ok := err.(net.Error); ok && ne.Temporary() {
+				logger.Infof("listenerSwitch: temporary error at %q: %s; sleeping for a second...", ls.ln.Addr(), err)
+				time.Sleep(time.Second)
+				continue
+			}
+			ls.closeLock.Lock()
+			ls.acceptErr = err
+			ls.closeLock.Unlock()
+			return
+		}
+		if _, err := io.ReadFull(c, buf[:]); err != nil {
+			logger.Errorf("listenerSwitch: cannot read one byte from the underlying connection for %q: %s", ls.ln.Addr(), err)
+			_ = c.Close()
+			continue
+		}
+
+		// It is expected that both listeners - http and telnet consume incoming connections as soon as possible,
+		// so the below code shouldn't block for extended periods of time.
+		pc := &peekedConn{
+			Conn:      c,
+			firstChar: buf[0],
+		}
+		if buf[0] == 'p' {
+			// Assume the request starts with `put`.
+			ls.telnetConnsCh <- pc
+		} else {
+			// Assume the request starts with `POST`.
+			ls.httpConnsCh <- pc
+		}
+	}
+}
+
+type peekedConn struct {
+	net.Conn
+	firstChar     byte
+	firstCharRead bool
+}
+
+func (pc *peekedConn) Read(p []byte) (int, error) {
+	// It is assumed that the pc cannot be read from concurrent goroutines.
+	if pc.firstCharRead {
+		// Fast path - first char already read.
+		return pc.Conn.Read(p)
+	}
+
+	// Slow path - read the first char.
+	if len(p) == 0 {
+		return 0, nil
+	}
+	p[0] = pc.firstChar
+	pc.firstCharRead = true
+	n, err := pc.Conn.Read(p[1:])
+	return n + 1, err
+}
+
+func (ls *listenerSwitch) newTelnetListener() *chanListener {
+	return &chanListener{
+		ls: ls,
+		ch: ls.telnetConnsCh,
+	}
+}
+
+func (ls *listenerSwitch) newHTTPListener() *chanListener {
+	return &chanListener{
+		ls: ls,
+		ch: ls.httpConnsCh,
+	}
+}
+
+type chanListener struct {
+	ls *listenerSwitch
+	ch chan net.Conn
+}
+
+func (cl *chanListener) Accept() (net.Conn, error) {
+	c, ok := <-cl.ch
+	if ok {
+		return c, nil
+	}
+
+	cl.ls.closeLock.Lock()
+	err := cl.ls.acceptErr
+	cl.ls.closeLock.Unlock()
+	return nil, err
+}
+
+func (cl *chanListener) Close() error {
+	return cl.ls.stop()
+}
+
+func (cl *chanListener) Addr() net.Addr {
+	return cl.ls.ln.Addr()
+}

app/vminsert/opentsdb/parser.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/fastjson/fastfloat"
 )
 
@@ -34,13 +36,8 @@ func (rs *Rows) Reset() {
 // See http://opentsdb.net/docs/build/html/api_telnet/put.html
 //
 // s must be unchanged until rs is in use.
-func (rs *Rows) Unmarshal(s string) error {
-	var err error
-	rs.Rows, rs.tagsPool, err = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0])
-	if err != nil {
-		return err
-	}
-	return err
+func (rs *Rows) Unmarshal(s string) {
+	rs.Rows, rs.tagsPool = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0])
 }
 
 // Row is a single OpenTSDB row.
@@ -69,6 +66,9 @@ func (r *Row) unmarshal(s string, tagsPool []Tag) ([]Tag, error) {
 		return tagsPool, fmt.Errorf("cannot find whitespace between metric and timestamp in %q", s)
 	}
 	r.Metric = s[:n]
+	if len(r.Metric) == 0 {
+		return tagsPool, fmt.Errorf("metric cannot be empty")
+	}
 	tail := s[n+1:]
 	n = strings.IndexByte(tail, ' ')
 	if n < 0 {
@@ -92,41 +92,46 @@ func (r *Row) unmarshal(s string, tagsPool []Tag) ([]Tag, error) {
 	return tagsPool, nil
 }
 
-func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
+func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag) {
 	for len(s) > 0 {
 		n := strings.IndexByte(s, '\n')
-		if n == 0 {
-			// Skip empty line
-			s = s[1:]
-			continue
-		}
-		if cap(dst) > len(dst) {
-			dst = dst[:len(dst)+1]
-		} else {
-			dst = append(dst, Row{})
-		}
-		r := &dst[len(dst)-1]
 		if n < 0 {
 			// The last line.
-			var err error
-			tagsPool, err = r.unmarshal(s, tagsPool)
-			if err != nil {
-				err = fmt.Errorf("cannot unmarshal OpenTSDB line %q: %s", s, err)
-				return dst, tagsPool, err
-			}
-			return dst, tagsPool, nil
-		}
-		var err error
-		tagsPool, err = r.unmarshal(s[:n], tagsPool)
-		if err != nil {
-			err = fmt.Errorf("cannot unmarshal OpenTSDB line %q: %s", s[:n], err)
-			return dst, tagsPool, err
+			return unmarshalRow(dst, s, tagsPool)
 		}
+		dst, tagsPool = unmarshalRow(dst, s[:n], tagsPool)
 		s = s[n+1:]
 	}
-	return dst, tagsPool, nil
+	return dst, tagsPool
 }
 
+func unmarshalRow(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag) {
+	if len(s) > 0 && s[len(s)-1] == '\r' {
+		s = s[:len(s)-1]
+	}
+	if len(s) == 0 {
+		// Skip empty line
+		return dst, tagsPool
+	}
+
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	var err error
+	tagsPool, err = r.unmarshal(s, tagsPool)
+	if err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal OpenTSDB line %q: %s", s, err)
+		invalidLines.Inc()
+	}
+	return dst, tagsPool
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="opentsdb"}`)
+
 func unmarshalTags(dst []Tag, s string) ([]Tag, error) {
 	for {
 		if cap(dst) > len(dst) {
@@ -142,12 +147,20 @@ func unmarshalTags(dst []Tag, s string) ([]Tag, error) {
 			if err := tag.unmarshal(s); err != nil {
 				return dst[:len(dst)-1], err
 			}
+			if len(tag.Key) == 0 || len(tag.Value) == 0 {
+				// Skip empty tag
+				dst = dst[:len(dst)-1]
+			}
 			return dst, nil
 		}
 		if err := tag.unmarshal(s[:n]); err != nil {
 			return dst[:len(dst)-1], err
 		}
 		s = s[n+1:]
+		if len(tag.Key) == 0 || len(tag.Value) == 0 {
+			// Skip empty tag
+			dst = dst[:len(dst)-1]
+		}
 	}
 }
 
@@ -169,9 +182,6 @@ func (t *Tag) unmarshal(s string) error {
 		return fmt.Errorf("missing tag value for %q", s)
 	}
 	t.Key = s[:n]
-	if len(t.Key) == 0 {
-		return fmt.Errorf("tag key cannot be empty for %q", s)
-	}
 	t.Value = s[n+1:]
 	return nil
 }

app/vminsert/opentsdb/parser_test.go

@@ -9,19 +9,24 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	f := func(s string) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
 		}
 
 		// Try again
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
 		}
 	}
 
 	// Missing put prefix
 	f("xx")
 
+	// Missing metric
+	f("put  111 34")
+
 	// Missing timestamp
 	f("put aaa")
 
@@ -42,26 +47,19 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 
 	// Invalid tag
 	f("put aaa 123 4.5 foo")
-	f("put aaa 123 4.5 =")
-	f("put aaa 123 4.5 =foo")
-	f("put aaa 123 4.5 =foo a=b")
 }
 
 func TestRowsUnmarshalSuccess(t *testing.T) {
 	f := func(s string, rowsExpected *Rows) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
 
 		// Try unmarshaling again
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
@@ -74,7 +72,9 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 
 	// Empty line
 	f("", &Rows{})
+	f("\r", &Rows{})
 	f("\n\n", &Rows{})
+	f("\n\r\n", &Rows{})
 
 	// Single line
 	f("put foobar 789 -123.456 a=b", &Rows{
@@ -88,17 +88,13 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			}},
 		}},
 	})
-	// Empty tag value
-	f("put foobar 789 -123.456 a= b=c", &Rows{
+	// Empty tag
+	f("put foobar 789 -123.456 a= b=c =d", &Rows{
 		Rows: []Row{{
 			Metric:    "foobar",
 			Value:     -123.456,
 			Timestamp: 789,
 			Tags: []Tag{
-				{
-					Key:   "a",
-					Value: "",
-				},
 				{
 					Key:   "b",
 					Value: "c",
@@ -200,4 +196,27 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+	// Multi lines with invalid line
+	f("put foo 2 0.3 a=b\naaa bbb\nput bar.baz 43 0.34 a=b\n", &Rows{
+		Rows: []Row{
+			{
+				Metric:    "foo",
+				Value:     0.3,
+				Timestamp: 2,
+				Tags: []Tag{{
+					Key:   "a",
+					Value: "b",
+				}},
+			},
+			{
+				Metric:    "bar.baz",
+				Value:     0.34,
+				Timestamp: 43,
+				Tags: []Tag{{
+					Key:   "a",
+					Value: "b",
+				}},
+			},
+		},
+	})
 }

app/vminsert/opentsdb/parser_timing_test.go

@@ -6,18 +6,19 @@ import (
 )
 
 func BenchmarkRowsUnmarshal(b *testing.B) {
-	s := `cpu.usage_user 1234556768 1.23 a=b
-cpu.usage_system 1234556768 23.344 a=b
-cpu.usage_iowait 1234556769 3.3443 a=b
-cpu.usage_irq 1234556768 0.34432 a=b
+	s := `put cpu.usage_user 1234556768 1.23 a=b
+put cpu.usage_system 1234556768 23.344 a=b
+put cpu.usage_iowait 1234556769 3.3443 a=b
+put cpu.usage_irq 1234556768 0.34432 a=b
 `
 	b.SetBytes(int64(len(s)))
 	b.ReportAllocs()
 	b.RunParallel(func(pb *testing.PB) {
 		var rows Rows
 		for pb.Next() {
-			if err := rows.Unmarshal(s); err != nil {
-				panic(fmt.Errorf("cannot unmarshal %q: %s", s, err))
+			rows.Unmarshal(s)
+			if len(rows.Rows) != 4 {
+				panic(fmt.Errorf("unexpected number of parsed rows; got %d; want 4", len(rows.Rows)))
 			}
 		}
 	})

app/vminsert/opentsdb/request_handler.go

@@ -61,13 +61,13 @@ func (ctx *pushCtx) InsertRows() error {
 const flushTimeout = 3 * time.Second
 
 func (ctx *pushCtx) Read(r io.Reader) bool {
-	opentsdbReadCalls.Inc()
+	readCalls.Inc()
 	if ctx.err != nil {
 		return false
 	}
 	if c, ok := r.(net.Conn); ok {
 		if err := c.SetReadDeadline(time.Now().Add(flushTimeout)); err != nil {
-			opentsdbReadErrors.Inc()
+			readErrors.Inc()
 			ctx.err = fmt.Errorf("cannot set read deadline: %s", err)
 			return false
 		}
@@ -79,21 +79,27 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			ctx.err = nil
 		} else {
 			if ctx.err != io.EOF {
-				opentsdbReadErrors.Inc()
+				readErrors.Inc()
 				ctx.err = fmt.Errorf("cannot read OpenTSDB put protocol data: %s", ctx.err)
 			}
 			return false
 		}
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		opentsdbUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal OpenTSDB put protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
+
+	// Fill in missing timestamps
+	currentTimestamp := time.Now().Unix()
+	rows := ctx.Rows.Rows
+	for i := range rows {
+		r := &rows[i]
+		if r.Timestamp == 0 {
+			r.Timestamp = currentTimestamp
+		}
 	}
 
 	// Convert timestamps from seconds to milliseconds
-	for i := range ctx.Rows.Rows {
-		ctx.Rows.Rows[i].Timestamp *= 1e3
+	for i := range rows {
+		rows[i].Timestamp *= 1e3
 	}
 	return true
 }
@@ -125,9 +131,8 @@ func (ctx *pushCtx) reset() {
 }
 
 var (
-	opentsdbReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
-	opentsdbReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
-	opentsdbUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="opentsdb"}`)
+	readCalls  = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
+	readErrors = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/opentsdb/server.go

@@ -7,6 +7,7 @@ import (
 	"sync"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdbhttp"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
@@ -21,44 +22,91 @@ var (
 	writeErrorsUDP   = metrics.NewCounter(`vm_opentsdb_request_errors_total{name="write", net="udp"}`)
 )
 
-// Serve starts OpenTSDB collector on the given addr.
-func Serve(addr string) {
+// Server is a server for collecting OpenTSDB TCP and UDP metrics.
+//
+// It accepts simultaneously Telnet put requests and HTTP put requests over TCP.
+type Server struct {
+	addr       string
+	ls         *listenerSwitch
+	httpServer *opentsdbhttp.Server
+	lnUDP      net.PacketConn
+	wg         sync.WaitGroup
+}
+
+// MustStart starts OpenTSDB collector on the given addr.
+//
+// MustStop must be called on the returned server when it is no longer needed.
+func MustStart(addr string, maxRequestSize int64) *Server {
 	logger.Infof("starting TCP OpenTSDB collector at %q", addr)
 	lnTCP, err := netutil.NewTCPListener("opentsdb", addr)
 	if err != nil {
 		logger.Fatalf("cannot start TCP OpenTSDB collector at %q: %s", addr, err)
 	}
-	listenerTCP = lnTCP
+	ls := newListenerSwitch(lnTCP)
+	lnHTTP := ls.newHTTPListener()
+	lnTelnet := ls.newTelnetListener()
+	httpServer := opentsdbhttp.MustServe(lnHTTP, maxRequestSize)
 
 	logger.Infof("starting UDP OpenTSDB collector at %q", addr)
 	lnUDP, err := net.ListenPacket("udp4", addr)
 	if err != nil {
 		logger.Fatalf("cannot start UDP OpenTSDB collector at %q: %s", addr, err)
 	}
-	listenerUDP = lnUDP
 
-	var wg sync.WaitGroup
-	wg.Add(1)
+	s := &Server{
+		addr:       addr,
+		ls:         ls,
+		httpServer: httpServer,
+		lnUDP:      lnUDP,
+	}
+	s.wg.Add(1)
 	go func() {
-		defer wg.Done()
-		serveTCP(listenerTCP)
-		logger.Infof("stopped TCP OpenTSDB collector at %q", addr)
+		defer s.wg.Done()
+		serveTelnet(lnTelnet)
+		logger.Infof("stopped TCP telnet OpenTSDB server at %q", addr)
 	}()
-	wg.Add(1)
+	s.wg.Add(1)
 	go func() {
-		defer wg.Done()
-		serveUDP(listenerUDP)
-		logger.Infof("stopped UDP OpenTSDB collector at %q", addr)
+		defer s.wg.Done()
+		httpServer.Wait()
+		// Do not log when httpServer is stopped, since this is logged by the server itself.
 	}()
-	wg.Wait()
+	s.wg.Add(1)
+	go func() {
+		defer s.wg.Done()
+		serveUDP(lnUDP)
+		logger.Infof("stopped UDP OpenTSDB server at %q", addr)
+	}()
+	return s
 }
 
-func serveTCP(ln net.Listener) {
+// MustStop stops the server.
+func (s *Server) MustStop() {
+	// Stop HTTP server. Do not emit log message, since it is emitted by the httpServer.
+	s.httpServer.MustStop()
+
+	logger.Infof("stopping TCP telnet OpenTSDB server at %q...", s.addr)
+	if err := s.ls.stop(); err != nil {
+		logger.Errorf("cannot stop TCP telnet OpenTSDB server: %s", err)
+	}
+
+	logger.Infof("stopping UDP OpenTSDB server at %q...", s.addr)
+	if err := s.lnUDP.Close(); err != nil {
+		logger.Errorf("cannot stop UDP OpenTSDB server: %s", err)
+	}
+
+	// Wait until all the servers are stopped.
+	s.wg.Wait()
+	logger.Infof("TCP and UDP OpenTSDB servers at %q have been stopped", s.addr)
+}
+
+func serveTelnet(ln net.Listener) {
 	for {
 		c, err := ln.Accept()
 		if err != nil {
 			if ne, ok := err.(net.Error); ok {
 				if ne.Temporary() {
+					logger.Errorf("opentsdb: temporary error when listening for TCP addr %q: %s", ln.Addr(), err)
 					time.Sleep(time.Second)
 					continue
 				}
@@ -97,6 +145,7 @@ func serveUDP(ln net.PacketConn) {
 					writeErrorsUDP.Inc()
 					if ne, ok := err.(net.Error); ok {
 						if ne.Temporary() {
+							logger.Errorf("opentsdb: temporary error when listening for UDP addr %q: %s", ln.LocalAddr(), err)
 							time.Sleep(time.Second)
 							continue
 						}
@@ -119,20 +168,3 @@ func serveUDP(ln net.PacketConn) {
 	}
 	wg.Wait()
 }
-
-var (
-	listenerTCP net.Listener
-	listenerUDP net.PacketConn
-)
-
-// Stop stops the server.
-func Stop() {
-	logger.Infof("stopping TCP OpenTSDB server at %q...", listenerTCP.Addr())
-	if err := listenerTCP.Close(); err != nil {
-		logger.Errorf("cannot close TCP OpenTSDB server: %s", err)
-	}
-	logger.Infof("stopping UDP OpenTSDB server at %q...", listenerUDP.LocalAddr())
-	if err := listenerUDP.Close(); err != nil {
-		logger.Errorf("cannot close UDP OpenTSDB server: %s", err)
-	}
-}

app/vminsert/opentsdbhttp/parser.go

@@ -0,0 +1,198 @@
+package opentsdbhttp
+
+import (
+	"fmt"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fastjson"
+	"github.com/valyala/fastjson/fastfloat"
+)
+
+// Rows contains parsed OpenTSDB rows.
+type Rows struct {
+	Rows []Row
+
+	tagsPool []Tag
+}
+
+// Reset resets rs.
+func (rs *Rows) Reset() {
+	// Release references to objects, so they can be GC'ed.
+	for i := range rs.Rows {
+		rs.Rows[i].reset()
+	}
+	rs.Rows = rs.Rows[:0]
+
+	for i := range rs.tagsPool {
+		rs.tagsPool[i].reset()
+	}
+	rs.tagsPool = rs.tagsPool[:0]
+}
+
+// Unmarshal unmarshals OpenTSDB rows from av.
+//
+// See http://opentsdb.net/docs/build/html/api_http/put.html
+//
+// s must be unchanged until rs is in use.
+func (rs *Rows) Unmarshal(av *fastjson.Value) {
+	rs.Rows, rs.tagsPool = unmarshalRows(rs.Rows[:0], av, rs.tagsPool[:0])
+}
+
+// Row is a single OpenTSDB row.
+type Row struct {
+	Metric    string
+	Tags      []Tag
+	Value     float64
+	Timestamp int64
+}
+
+func (r *Row) reset() {
+	r.Metric = ""
+	r.Tags = nil
+	r.Value = 0
+	r.Timestamp = 0
+}
+
+func (r *Row) unmarshal(o *fastjson.Value, tagsPool []Tag) ([]Tag, error) {
+	r.reset()
+	m := o.GetStringBytes("metric")
+	if len(m) == 0 {
+		return tagsPool, fmt.Errorf("missing `metric` in %s", o)
+	}
+	r.Metric = bytesutil.ToUnsafeString(m)
+
+	rawTs := o.Get("timestamp")
+	if rawTs != nil {
+		ts, err := getFloat64(rawTs)
+		if err != nil {
+			return tagsPool, fmt.Errorf("invalid `timestamp` in %s: %s", o, err)
+		}
+		r.Timestamp = int64(ts)
+	} else {
+		// Allow missing timestamp. It is automatically populated
+		// with the current time in this case.
+		r.Timestamp = 0
+	}
+
+	rawV := o.Get("value")
+	if rawV == nil {
+		return tagsPool, fmt.Errorf("missing `value` in %s", o)
+	}
+	v, err := getFloat64(rawV)
+	if err != nil {
+		return tagsPool, fmt.Errorf("invalid `value` in %s: %s", o, err)
+	}
+	r.Value = v
+
+	vt := o.Get("tags")
+	if vt == nil {
+		// Allow empty tags.
+		return tagsPool, nil
+	}
+	rawTags, err := vt.Object()
+	if err != nil {
+		return tagsPool, fmt.Errorf("invalid `tags` in %s: %s", o, err)
+	}
+
+	tagsStart := len(tagsPool)
+	tagsPool, err = unmarshalTags(tagsPool, rawTags)
+	if err != nil {
+		return tagsPool, fmt.Errorf("cannot parse tags %s: %s", rawTags, err)
+	}
+	tags := tagsPool[tagsStart:]
+	r.Tags = tags[:len(tags):len(tags)]
+	return tagsPool, nil
+}
+
+func getFloat64(v *fastjson.Value) (float64, error) {
+	switch v.Type() {
+	case fastjson.TypeNumber:
+		return v.Float64()
+	case fastjson.TypeString:
+		vStr, _ := v.StringBytes()
+		vFloat := fastfloat.ParseBestEffort(bytesutil.ToUnsafeString(vStr))
+		if vFloat == 0 && string(vStr) != "0" && string(vStr) != "0.0" {
+			return 0, fmt.Errorf("invalid float64 value: %q", vStr)
+		}
+		return vFloat, nil
+	default:
+		return 0, fmt.Errorf("value doesn't contain float64; it contains %s", v.Type())
+	}
+}
+
+func unmarshalRows(dst []Row, av *fastjson.Value, tagsPool []Tag) ([]Row, []Tag) {
+	switch av.Type() {
+	case fastjson.TypeObject:
+		return unmarshalRow(dst, av, tagsPool)
+	case fastjson.TypeArray:
+		a, _ := av.Array()
+		for _, o := range a {
+			dst, tagsPool = unmarshalRow(dst, o, tagsPool)
+		}
+		return dst, tagsPool
+	default:
+		logger.Errorf("OpenTSDB JSON must be either object or array; got %s; body=%s", av.Type(), av)
+		invalidLines.Inc()
+		return dst, tagsPool
+	}
+}
+
+func unmarshalRow(dst []Row, o *fastjson.Value, tagsPool []Tag) ([]Row, []Tag) {
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	var err error
+	tagsPool, err = r.unmarshal(o, tagsPool)
+	if err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal OpenTSDB object %s: %s", o, err)
+		invalidLines.Inc()
+	}
+	return dst, tagsPool
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="opentsdb-http"}`)
+
+func unmarshalTags(dst []Tag, o *fastjson.Object) ([]Tag, error) {
+	var err error
+	o.Visit(func(k []byte, v *fastjson.Value) {
+		if v.Type() != fastjson.TypeString {
+			err = fmt.Errorf("tag value must be string; got %s; value=%s", v.Type(), v)
+			return
+		}
+		if len(k) == 0 {
+			// Skip empty tags
+			return
+		}
+		vStr, _ := v.StringBytes()
+		if len(vStr) == 0 {
+			// Skip empty tags
+			return
+		}
+		if cap(dst) > len(dst) {
+			dst = dst[:len(dst)+1]
+		} else {
+			dst = append(dst, Tag{})
+		}
+		tag := &dst[len(dst)-1]
+		tag.Key = bytesutil.ToUnsafeString(k)
+		tag.Value = bytesutil.ToUnsafeString(vStr)
+	})
+	return dst, err
+}
+
+// Tag is an OpenTSDB tag.
+type Tag struct {
+	Key   string
+	Value string
+}
+
+func (t *Tag) reset() {
+	t.Key = ""
+	t.Value = ""
+}

app/vminsert/opentsdbhttp/parser_test.go

@@ -0,0 +1,246 @@
+package opentsdbhttp
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestRowsUnmarshalFailure(t *testing.T) {
+	f := func(s string) {
+		t.Helper()
+		var rows Rows
+		p := parserPool.Get()
+		defer parserPool.Put(p)
+		v, err := p.Parse(s)
+		if err != nil {
+			// Expected JSON parser error
+			return
+		}
+		// Verify OpenTSDB body parsing error
+		rows.Unmarshal(v)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
+		}
+		// Try again
+		rows.Unmarshal(v)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("unexpected number of rows parsed; got %d; want 0", len(rows.Rows))
+		}
+	}
+
+	// invalid json
+	f("{g")
+
+	// Invalid json type
+	f(`1`)
+	f(`"foo"`)
+	f(`[1,2]`)
+	f(`null`)
+
+	// Incomplete object
+	f(`{}`)
+	f(`{"metric": "aaa"}`)
+	f(`{"metric": "aaa", "timestamp": 1122}`)
+	f(`{"metric": "aaa", "timestamp": "tststs"}`)
+	f(`{"timestamp": 1122, "value": 33}`)
+	f(`{"value": 33}`)
+	f(`{"value": 33, "tags": {"fooo":"bar"}}`)
+
+	// Invalid value
+	f(`{"metric": "aaa", "timestamp": 1122, "value": "0.0.0"}`)
+
+	// Invalid metric type
+	f(`{"metric": "", "timestamp": 1122, "value": 0.45, "tags": {"foo": "bar"}}`)
+	f(`{"metric": ["aaa"], "timestamp": 1122, "value": 0.45, "tags": {"foo": "bar"}}`)
+	f(`{"metric": {"aaa":1}, "timestamp": 1122, "value": 0.45, "tags": {"foo": "bar"}}`)
+	f(`{"metric": 1, "timestamp": 1122, "value": 0.45, "tags": {"foo": "bar"}}`)
+
+	// Invalid timestamp type
+	f(`{"metric": "aaa", "timestamp": "foobar", "value": 0.45, "tags": {"foo": "bar"}}`)
+	f(`{"metric": "aaa", "timestamp": [1,2], "value": 0.45, "tags": {"foo": "bar"}}`)
+	f(`{"metric": "aaa", "timestamp": {"a":1}, "value": 0.45, "tags": {"foo": "bar"}}`)
+
+	// Invalid value type
+	f(`{"metric": "aaa", "timestamp": 1122, "value": [0,1], "tags": {"foo":"bar"}}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": {"a":1}, "tags": {"foo":"bar"}}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": "foobar", "tags": {"foo":"bar"}}`)
+
+	// Invalid tags type
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": 1}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": [1,2]}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": "foo"}`)
+
+	// Invalid tag value type
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": {"foo": ["bar"]}}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": {"foo": {"bar":"baz"}}}`)
+	f(`{"metric": "aaa", "timestamp": 1122, "value": 0.45, "tags": {"foo": 1}}`)
+
+	// Invalid multiline
+	f(`[{"metric": "aaa", "timestamp": 1122, "value": "trt", "tags":{"foo":"bar"}}, {"metric": "aaa", "timestamp": [1122], "value": 111}]`)
+}
+
+func TestRowsUnmarshalSuccess(t *testing.T) {
+	f := func(s string, rowsExpected *Rows) {
+		t.Helper()
+		var rows Rows
+
+		p := parserPool.Get()
+		defer parserPool.Put(p)
+		v, err := p.Parse(s)
+		if err != nil {
+			t.Fatalf("cannot parse json %s: %s", s, err)
+		}
+		rows.Unmarshal(v)
+		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
+			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
+		}
+
+		// Try unmarshaling again
+		rows.Unmarshal(v)
+		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
+			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
+		}
+
+		rows.Reset()
+		if len(rows.Rows) != 0 {
+			t.Fatalf("non-empty rows after reset: %+v", rows.Rows)
+		}
+	}
+
+	// Normal line
+	f(`{"metric": "foobar", "timestamp": 789, "value": -123.456, "tags": {"a":"b"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 789,
+			Tags: []Tag{{
+				Key:   "a",
+				Value: "b",
+			}},
+		}},
+	})
+	// Timestamp as string
+	f(`{"metric": "foobar", "timestamp": "1789", "value": -123.456, "tags": {"a":"b"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 1789,
+			Tags: []Tag{{
+				Key:   "a",
+				Value: "b",
+			}},
+		}},
+	})
+	// Timestamp as float64 (it is truncated to integer)
+	f(`{"metric": "foobar", "timestamp": 17.89, "value": -123.456, "tags": {"a":"b"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 17,
+			Tags: []Tag{{
+				Key:   "a",
+				Value: "b",
+			}},
+		}},
+	})
+	// Empty tags
+	f(`{"metric": "foobar", "timestamp": 789, "value": -123.456, "tags": {}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 789,
+			Tags:      nil,
+		}},
+	})
+	// Missing tags
+	f(`{"metric": "foobar", "timestamp": 789, "value": -123.456}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 789,
+			Tags:      nil,
+		}},
+	})
+	// Empty tag value
+	f(`{"metric": "foobar", "timestamp": 123, "value": -123.456, "tags": {"a":"", "b":"c", "": "d"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -123.456,
+			Timestamp: 123,
+			Tags: []Tag{
+				{
+					Key:   "b",
+					Value: "c",
+				},
+			},
+		}},
+	})
+	// Value as string
+	f(`{"metric": "foobar", "timestamp": 789, "value": "-12.456", "tags": {"a":"b"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -12.456,
+			Timestamp: 789,
+			Tags: []Tag{{
+				Key:   "a",
+				Value: "b",
+			}},
+		}},
+	})
+	// Missing timestamp
+	f(`{"metric": "foobar", "value": "-12.456", "tags": {"a":"b"}}`, &Rows{
+		Rows: []Row{{
+			Metric:    "foobar",
+			Value:     -12.456,
+			Timestamp: 0,
+			Tags: []Tag{{
+				Key:   "a",
+				Value: "b",
+			}},
+		}},
+	})
+
+	// Multiple tags
+	f(`{"metric": "foo", "value": 1, "timestamp": 2, "tags": {"bar":"baz", "x": "y"}}`, &Rows{
+		Rows: []Row{{
+			Metric: "foo",
+			Tags: []Tag{
+				{
+					Key:   "bar",
+					Value: "baz",
+				},
+				{
+					Key:   "x",
+					Value: "y",
+				},
+			},
+			Value:     1,
+			Timestamp: 2,
+		}},
+	})
+
+	// Multi lines
+	f(`[{"metric": "foo", "value": "0.3", "timestamp": 2, "tags": {"a":"b"}},
+{"metric": "bar.baz", "value": 0.34, "timestamp": 43, "tags": {"a":"b"}}]`, &Rows{
+		Rows: []Row{
+			{
+				Metric:    "foo",
+				Value:     0.3,
+				Timestamp: 2,
+				Tags: []Tag{{
+					Key:   "a",
+					Value: "b",
+				}},
+			},
+			{
+				Metric:    "bar.baz",
+				Value:     0.34,
+				Timestamp: 43,
+				Tags: []Tag{{
+					Key:   "a",
+					Value: "b",
+				}},
+			},
+		},
+	})
+}

app/vminsert/opentsdbhttp/parser_timing_test.go

@@ -0,0 +1,33 @@
+package opentsdbhttp
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/valyala/fastjson"
+)
+
+func BenchmarkRowsUnmarshal(b *testing.B) {
+	s := `[{"metric": "cpu.usage_user", "timestamp": 1234556768, "value": 1.23, "tags": {"a":"b", "x": "y"}},
+{"metric": "cpu.usage_system", "timestamp": 1234556768, "value": 23.344, "tags": {"a":"b"}},
+{"metric": "cpu.usage_iowait", "timestamp": 1234556769, "value":3.3443, "tags": {"a":"b"}},
+{"metric": "cpu.usage_irq", "timestamp": 1234556768, "value": 0.34432, "tags": {"a":"b"}}
+]
+`
+	b.SetBytes(int64(len(s)))
+	b.ReportAllocs()
+	b.RunParallel(func(pb *testing.PB) {
+		var rows Rows
+		var p fastjson.Parser
+		for pb.Next() {
+			v, err := p.Parse(s)
+			if err != nil {
+				panic(fmt.Errorf("cannot parse %q: %s", s, err))
+			}
+			rows.Unmarshal(v)
+			if len(rows.Rows) != 4 {
+				panic(fmt.Errorf("unexpected number of rows unmarshaled; got %d; want 4", len(rows.Rows)))
+			}
+		}
+	})
+}

app/vminsert/opentsdbhttp/request_handler.go

@@ -0,0 +1,150 @@
+package opentsdbhttp
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"runtime"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fastjson"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb-http"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="opentsdb-http"}`)
+
+	readCalls       = metrics.NewCounter(`vm_read_calls_total{name="opentsdb-http"}`)
+	readErrors      = metrics.NewCounter(`vm_read_errors_total{name="opentsdb-http"}`)
+	unmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="opentsdb-http"}`)
+)
+
+// insertHandler processes HTTP OpenTSDB put requests.
+// See http://opentsdb.net/docs/build/html/api_http/put.html
+func insertHandler(req *http.Request, maxSize int64) error {
+	return concurrencylimiter.Do(func() error {
+		return insertHandlerInternal(req, maxSize)
+	})
+}
+
+func insertHandlerInternal(req *http.Request, maxSize int64) error {
+	readCalls.Inc()
+
+	r := req.Body
+	if req.Header.Get("Content-Encoding") == "gzip" {
+		zr, err := common.GetGzipReader(r)
+		if err != nil {
+			readErrors.Inc()
+			return fmt.Errorf("cannot read gzipped http protocol data: %s", err)
+		}
+		defer common.PutGzipReader(zr)
+		r = zr
+	}
+
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+
+	// Read the request in ctx.reqBuf
+	lr := io.LimitReader(r, maxSize+1)
+	reqLen, err := ctx.reqBuf.ReadFrom(lr)
+	if err != nil {
+		readErrors.Inc()
+		return fmt.Errorf("cannot read HTTP OpenTSDB request: %s", err)
+	}
+	if reqLen > maxSize {
+		readErrors.Inc()
+		return fmt.Errorf("too big HTTP OpenTSDB request; mustn't exceed %d bytes", maxSize)
+	}
+
+	// Unmarshal the request to ctx.Rows
+	p := parserPool.Get()
+	defer parserPool.Put(p)
+	v, err := p.ParseBytes(ctx.reqBuf.B)
+	if err != nil {
+		unmarshalErrors.Inc()
+		return fmt.Errorf("cannot parse HTTP OpenTSDB json: %s", err)
+	}
+	ctx.Rows.Unmarshal(v)
+
+	// Fill in missing timestamps
+	currentTimestamp := time.Now().Unix()
+	rows := ctx.Rows.Rows
+	for i := range rows {
+		r := &rows[i]
+		if r.Timestamp == 0 {
+			r.Timestamp = currentTimestamp
+		}
+	}
+
+	// Convert timestamps in seconds to milliseconds if needed.
+	// See http://opentsdb.net/docs/javadoc/net/opentsdb/core/Const.html#SECOND_MASK
+	for i := range rows {
+		r := &rows[i]
+		if r.Timestamp&secondMask == 0 {
+			r.Timestamp *= 1e3
+		}
+	}
+
+	// Insert ctx.Rows to db.
+	ic := &ctx.Common
+	ic.Reset(len(rows))
+	for i := range rows {
+		r := &rows[i]
+		ic.Labels = ic.Labels[:0]
+		ic.AddLabel("", r.Metric)
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			ic.AddLabel(tag.Key, tag.Value)
+		}
+		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
+	}
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return ic.FlushBufs()
+}
+
+const secondMask int64 = 0x7FFFFFFF00000000
+
+var parserPool fastjson.ParserPool
+
+type pushCtx struct {
+	Rows   Rows
+	Common common.InsertCtx
+
+	reqBuf bytesutil.ByteBuffer
+}
+
+func (ctx *pushCtx) reset() {
+	ctx.Rows.Reset()
+	ctx.Common.Reset(0)
+	ctx.reqBuf.Reset()
+}
+
+func getPushCtx() *pushCtx {
+	select {
+	case ctx := <-pushCtxPoolCh:
+		return ctx
+	default:
+		if v := pushCtxPool.Get(); v != nil {
+			return v.(*pushCtx)
+		}
+		return &pushCtx{}
+	}
+}
+
+func putPushCtx(ctx *pushCtx) {
+	ctx.reset()
+	select {
+	case pushCtxPoolCh <- ctx:
+	default:
+		pushCtxPool.Put(ctx)
+	}
+}
+
+var pushCtxPool sync.Pool
+var pushCtxPoolCh = make(chan *pushCtx, runtime.GOMAXPROCS(-1))

app/vminsert/opentsdbhttp/server.go

@@ -0,0 +1,101 @@
+package opentsdbhttp
+
+import (
+	"context"
+	"net"
+	"net/http"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	writeRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/put", protocol="opentsdb-http"}`)
+	writeErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/put", protocol="opentsdb-http"}`)
+)
+
+// Server represents HTTP OpenTSDB server.
+type Server struct {
+	s  *http.Server
+	ln net.Listener
+	wg sync.WaitGroup
+}
+
+// MustStart starts HTTP OpenTSDB server on the given addr.
+//
+// MustStop must be called on the returned server when it is no longer needed.
+func MustStart(addr string, maxRequestSize int64) *Server {
+	logger.Infof("starting HTTP OpenTSDB server at %q", addr)
+	lnTCP, err := netutil.NewTCPListener("opentsdbhttp", addr)
+	if err != nil {
+		logger.Fatalf("cannot start HTTP OpenTSDB collector at %q: %s", addr, err)
+	}
+	return MustServe(lnTCP, maxRequestSize)
+}
+
+// MustServe serves OpenTSDB HTTP put requests from ln with up to maxRequestSize size.
+//
+// MustStop must be called on the returned server when it is no longer needed.
+func MustServe(ln net.Listener, maxRequestSize int64) *Server {
+	h := newRequestHandler(maxRequestSize)
+	hs := &http.Server{
+		Handler:      h,
+		ReadTimeout:  30 * time.Second,
+		WriteTimeout: 10 * time.Second,
+	}
+	s := &Server{
+		s:  hs,
+		ln: ln,
+	}
+	s.wg.Add(1)
+	go func() {
+		defer s.wg.Done()
+		err := s.s.Serve(s.ln)
+		if err == http.ErrServerClosed {
+			return
+		}
+		if err != nil {
+			logger.Fatalf("error serving HTTP OpenTSDB at %q: %s", s.ln.Addr(), err)
+		}
+	}()
+	return s
+}
+
+// Wait waits until the server is stopped with MustStop.
+func (s *Server) Wait() {
+	s.wg.Wait()
+}
+
+// MustStop stops HTTP OpenTSDB server.
+func (s *Server) MustStop() {
+	logger.Infof("stopping HTTP OpenTSDB server at %q...", s.ln.Addr())
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := s.s.Shutdown(ctx); err != nil {
+		logger.Fatalf("cannot close HTTP OpenTSDB server at %q: %s", s.ln.Addr(), err)
+	}
+	s.wg.Wait()
+	logger.Infof("OpenTSDB HTTP server at %q has been stopped", s.ln.Addr())
+}
+
+func newRequestHandler(maxRequestSize int64) http.Handler {
+	rh := func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/put":
+			writeRequests.Inc()
+			if err := insertHandler(r, maxRequestSize); err != nil {
+				writeErrors.Inc()
+				httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+				return
+			}
+			w.WriteHeader(http.StatusNoContent)
+		default:
+			httpserver.Errorf(w, "unexpected path requested on HTTP OpenTSDB server: %q", r.URL.Path)
+		}
+	}
+	return http.HandlerFunc(rh)
+}

app/vminsert/vmimport/parser.go

@@ -0,0 +1,202 @@
+package vmimport
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fastjson"
+)
+
+// Rows contains parsed rows from `/api/v1/import` request.
+type Rows struct {
+	Rows []Row
+
+	tu tagsUnmarshaler
+}
+
+// Reset resets rs.
+func (rs *Rows) Reset() {
+	for i := range rs.Rows {
+		rs.Rows[i].reset()
+	}
+	rs.Rows = rs.Rows[:0]
+
+	rs.tu.reset()
+}
+
+// Unmarshal unmarshals influx line protocol rows from s.
+//
+// See https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/
+//
+// s must be unchanged until rs is in use.
+func (rs *Rows) Unmarshal(s string) {
+	rs.tu.reset()
+	rs.Rows = unmarshalRows(rs.Rows[:0], s, &rs.tu)
+}
+
+// Row is a single row from `/api/v1/import` request.
+type Row struct {
+	Tags       []Tag
+	Values     []float64
+	Timestamps []int64
+}
+
+func (r *Row) reset() {
+	r.Tags = nil
+	r.Values = r.Values[:0]
+	r.Timestamps = r.Timestamps[:0]
+}
+
+func (r *Row) unmarshal(s string, tu *tagsUnmarshaler) error {
+	r.reset()
+	v, err := tu.p.Parse(s)
+	if err != nil {
+		return fmt.Errorf("cannot parse json line: %s", err)
+	}
+
+	// Unmarshal tags
+	metric := v.GetObject("metric")
+	if metric == nil {
+		return fmt.Errorf("missing `metric` object")
+	}
+	tagsStart := len(tu.tagsPool)
+	if err := tu.unmarshalTags(metric); err != nil {
+		return fmt.Errorf("cannot unmarshal `metric`: %s", err)
+	}
+	tags := tu.tagsPool[tagsStart:]
+	r.Tags = tags[:len(tags):len(tags)]
+	if len(r.Tags) == 0 {
+		return fmt.Errorf("missing tags")
+	}
+
+	// Unmarshal values
+	values := v.GetArray("values")
+	if len(values) == 0 {
+		return fmt.Errorf("missing `values` array")
+	}
+	for i, v := range values {
+		f, err := v.Float64()
+		if err != nil {
+			return fmt.Errorf("cannot unmarshal value at position %d: %s", i, err)
+		}
+		r.Values = append(r.Values, f)
+	}
+
+	// Unmarshal timestamps
+	timestamps := v.GetArray("timestamps")
+	if len(timestamps) == 0 {
+		return fmt.Errorf("missing `timestamps` array")
+	}
+	for i, v := range timestamps {
+		ts, err := v.Int64()
+		if err != nil {
+			return fmt.Errorf("cannot unmarshal timestamp at position %d: %s", i, err)
+		}
+		r.Timestamps = append(r.Timestamps, ts)
+	}
+
+	if len(r.Timestamps) != len(r.Values) {
+		return fmt.Errorf("`timestamps` array size must match `values` array size; got %d; want %d", len(r.Timestamps), len(r.Values))
+	}
+	return nil
+}
+
+// Tag represents `/api/v1/import` tag.
+type Tag struct {
+	Key   []byte
+	Value []byte
+}
+
+func (tag *Tag) reset() {
+	// tag.Key and tag.Value point to tu.bytesPool, so there is no need in keeping these byte slices here.
+	tag.Key = nil
+	tag.Value = nil
+}
+
+type tagsUnmarshaler struct {
+	p         fastjson.Parser
+	tagsPool  []Tag
+	bytesPool []byte
+	err       error
+}
+
+func (tu *tagsUnmarshaler) reset() {
+	for i := range tu.tagsPool {
+		tu.tagsPool[i].reset()
+	}
+	tu.tagsPool = tu.tagsPool[:0]
+
+	tu.bytesPool = tu.bytesPool[:0]
+	tu.err = nil
+}
+
+func (tu *tagsUnmarshaler) addTag() *Tag {
+	dst := tu.tagsPool
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Tag{})
+	}
+	tag := &dst[len(dst)-1]
+	tu.tagsPool = dst
+	return tag
+}
+
+func (tu *tagsUnmarshaler) addBytes(b []byte) []byte {
+	bytesPoolLen := len(tu.bytesPool)
+	tu.bytesPool = append(tu.bytesPool, b...)
+	bCopy := tu.bytesPool[bytesPoolLen:]
+	return bCopy[:len(bCopy):len(bCopy)]
+}
+
+func (tu *tagsUnmarshaler) unmarshalTags(o *fastjson.Object) error {
+	tu.err = nil
+	o.Visit(func(key []byte, v *fastjson.Value) {
+		tag := tu.addTag()
+		tag.Key = tu.addBytes(key)
+		sb, err := v.StringBytes()
+		if err != nil && tu.err != nil {
+			tu.err = fmt.Errorf("cannot parse value for tag %q: %s", tag.Key, err)
+		}
+		tag.Value = tu.addBytes(sb)
+	})
+	return tu.err
+}
+
+func unmarshalRows(dst []Row, s string, tu *tagsUnmarshaler) []Row {
+	for len(s) > 0 {
+		n := strings.IndexByte(s, '\n')
+		if n < 0 {
+			// The last line.
+			return unmarshalRow(dst, s, tu)
+		}
+		dst = unmarshalRow(dst, s[:n], tu)
+		s = s[n+1:]
+	}
+	return dst
+}
+
+func unmarshalRow(dst []Row, s string, tu *tagsUnmarshaler) []Row {
+	if len(s) > 0 && s[len(s)-1] == '\r' {
+		s = s[:len(s)-1]
+	}
+	if len(s) == 0 {
+		return dst
+	}
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	if err := r.unmarshal(s, tu); err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal json line %q: %s; skipping it", s, err)
+		invalidLines.Inc()
+	}
+	return dst
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="vmimport"}`)

app/vminsert/vmimport/parser_test.go

@@ -0,0 +1,229 @@
+package vmimport
+
+import (
+	"math"
+	"reflect"
+	"testing"
+)
+
+func TestRowsUnmarshalFailure(t *testing.T) {
+	f := func(s string) {
+		t.Helper()
+		var rows Rows
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
+		}
+
+		// Try again
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
+		}
+	}
+
+	// Invalid json line
+	f("")
+	f("\n")
+	f("foo\n")
+	f("123")
+	f("[1,3]")
+	f("{}")
+	f("[]")
+	f(`{"foo":"bar"}`)
+
+	// Invalid metric
+	f(`{"metric":123,"values":[1,2],"timestamps":[3,4]}`)
+	f(`{"metric":[123],"values":[1,2],"timestamps":[3,4]}`)
+	f(`{"metric":[],"values":[1,2],"timestamps":[3,4]}`)
+	f(`{"metric":{},"values":[1,2],"timestamps":[3,4]}`)
+	f(`{"metric":null,"values":[1,2],"timestamps":[3,4]}`)
+	f(`{"values":[1,2],"timestamps":[3,4]}`)
+
+	// Invalid values
+	f(`{"metric":{"foo":"bar"},"values":1,"timestamps":[3,4]}`)
+	f(`{"metric":{"foo":"bar"},"values":{"x":1},"timestamps":[3,4]}`)
+	f(`{"metric":{"foo":"bar"},"values":{"x":1},"timestamps":[3,4]}`)
+	f(`{"metric":{"foo":"bar"},"values":null,"timestamps":[3,4]}`)
+	f(`{"metric":{"foo":"bar"},"timestamps":[3,4]}`)
+
+	// Invalid timestamps
+	f(`{"metric":{"foo":"bar"},"values":[1,2],"timestamps":3}`)
+	f(`{"metric":{"foo":"bar"},"values":[1,2],"timestamps":false}`)
+	f(`{"metric":{"foo":"bar"},"values":[1,2],"timestamps":{}}`)
+	f(`{"metric":{"foo":"bar"},"values":[1,2]}`)
+
+	// values and timestamps count mismatch
+	f(`{"metric":{"foo":"bar"},"values":[],"timestamps":[]}`)
+	f(`{"metric":{"foo":"bar"},"values":[],"timestamps":[1]}`)
+	f(`{"metric":{"foo":"bar"},"values":[2],"timestamps":[]}`)
+	f(`{"metric":{"foo":"bar"},"values":[2],"timestamps":[3,4]}`)
+	f(`{"metric":{"foo":"bar"},"values":[2,3],"timestamps":[4]}`)
+
+	// Garbage after the line
+	f(`{"metric":{"foo":"bar"},"values":[2],"timestamps":[4]}{}`)
+}
+
+func TestRowsUnmarshalSuccess(t *testing.T) {
+	f := func(s string, rowsExpected *Rows) {
+		t.Helper()
+		var rows Rows
+		rows.Unmarshal(s)
+		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
+			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
+		}
+
+		// Try unmarshaling again
+		rows.Unmarshal(s)
+		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
+			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
+		}
+
+		rows.Reset()
+		if len(rows.Rows) != 0 {
+			t.Fatalf("non-empty rows after reset: %+v", rows.Rows)
+		}
+	}
+
+	// Empty line
+	f("", &Rows{})
+	f("\n\n", &Rows{})
+	f("\n\r\n", &Rows{})
+
+	// Single line with a single tag
+	f(`{"metric":{"foo":"bar"},"values":[1.23],"timestamps":[456]}`, &Rows{
+		Rows: []Row{{
+			Tags: []Tag{{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			}},
+			Values:     []float64{1.23},
+			Timestamps: []int64{456},
+		}},
+	})
+
+	// Inf and nan values
+	f(`{"metric":{"foo":"bar"},"values":[Inf, -Inf],"timestamps":[456, 789]}`, &Rows{
+		Rows: []Row{{
+			Tags: []Tag{{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			}},
+			Values:     []float64{math.Inf(1), math.Inf(-1)},
+			Timestamps: []int64{456, 789},
+		}},
+	})
+
+	// Line with multiple tags
+	f(`{"metric":{"foo":"bar","baz":"xx"},"values":[1.23, -3.21],"timestamps" : [456,789]}`, &Rows{
+		Rows: []Row{{
+			Tags: []Tag{
+				{
+					Key:   []byte("foo"),
+					Value: []byte("bar"),
+				},
+				{
+					Key:   []byte("baz"),
+					Value: []byte("xx"),
+				},
+			},
+			Values:     []float64{1.23, -3.21},
+			Timestamps: []int64{456, 789},
+		}},
+	})
+
+	// Multiple lines
+	f(`{"metric":{"foo":"bar","baz":"xx"},"values":[1.23, -3.21],"timestamps" : [456,789]}
+{"metric":{"__name__":"xx"},"values":[34],"timestamps" : [11]}
+`, &Rows{
+		Rows: []Row{
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("foo"),
+						Value: []byte("bar"),
+					},
+					{
+						Key:   []byte("baz"),
+						Value: []byte("xx"),
+					},
+				},
+				Values:     []float64{1.23, -3.21},
+				Timestamps: []int64{456, 789},
+			},
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("__name__"),
+						Value: []byte("xx"),
+					},
+				},
+				Values:     []float64{34},
+				Timestamps: []int64{11},
+			},
+		},
+	})
+
+	// Multiple lines with invalid line in the middle.
+	f(`{"metric":{"xfoo":"bar","baz":"xx"},"values":[1.232, -3.21],"timestamps" : [456,7890]}
+garbage here
+{"metric":{"__name__":"xxy"},"values":[34],"timestamps" : [111]}`, &Rows{
+		Rows: []Row{
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("xfoo"),
+						Value: []byte("bar"),
+					},
+					{
+						Key:   []byte("baz"),
+						Value: []byte("xx"),
+					},
+				},
+				Values:     []float64{1.232, -3.21},
+				Timestamps: []int64{456, 7890},
+			},
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("__name__"),
+						Value: []byte("xxy"),
+					},
+				},
+				Values:     []float64{34},
+				Timestamps: []int64{111},
+			},
+		},
+	})
+
+	// No newline after the second line.
+	f(`{"metric":{"foo":"bar","baz":"xx"},"values":[1.23, -3.21],"timestamps" : [456,789]}
+{"metric":{"__name__":"xx"},"values":[34],"timestamps" : [11]}`, &Rows{
+		Rows: []Row{
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("foo"),
+						Value: []byte("bar"),
+					},
+					{
+						Key:   []byte("baz"),
+						Value: []byte("xx"),
+					},
+				},
+				Values:     []float64{1.23, -3.21},
+				Timestamps: []int64{456, 789},
+			},
+			{
+				Tags: []Tag{
+					{
+						Key:   []byte("__name__"),
+						Value: []byte("xx"),
+					},
+				},
+				Values:     []float64{34},
+				Timestamps: []int64{11},
+			},
+		},
+	})
+}

app/vminsert/vmimport/parser_timing_test.go

@@ -0,0 +1,25 @@
+package vmimport
+
+import (
+	"fmt"
+	"testing"
+)
+
+func BenchmarkRowsUnmarshal(b *testing.B) {
+	s := `{"metric":{"__name__":"up","job":"node_exporter","instance":"localhost:9100"},"values":[0,0,0],"timestamps":[1549891472010,1549891487724,1549891503438]}
+{"metric":{"__name__":"up","job":"prometheus","instance":"localhost:9090"},"values":[1,1,1],"timestamps":[1549891461511,1549891476511,1549891491511]}
+{"metric":{"__name__":"up","job":"node_exporter","instance":"foobar.com:9100"},"values":[0,0,0],"timestamps":[1549891472010,1549891487724,1549891503438]}
+{"metric":{"__name__":"up","job":"prometheus","instance":"xxx.yyy.zzz:9090"},"values":[1,1,1],"timestamps":[1549891461511,1549891476511,1549891491511]}
+`
+	b.SetBytes(int64(len(s)))
+	b.ReportAllocs()
+	b.RunParallel(func(pb *testing.PB) {
+		var rows Rows
+		for pb.Next() {
+			rows.Unmarshal(s)
+			if len(rows.Rows) != 4 {
+				panic(fmt.Errorf("unexpected number of rows parsed; got %d; want 4", len(rows.Rows)))
+			}
+		}
+	})
+}

app/vminsert/vmimport/request_handler.go

@@ -0,0 +1,160 @@
+package vmimport
+
+import (
+	"flag"
+	"fmt"
+	"io"
+	"net/http"
+	"runtime"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var maxLineLen = flag.Int("import.maxLineLen", 100*1024*1024, "The maximum length in bytes of a single line accepted by `/api/v1/import`")
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="vmimport"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="vmimport"}`)
+)
+
+// InsertHandler processes `/api/v1/import` request.
+//
+// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/6
+func InsertHandler(req *http.Request) error {
+	return concurrencylimiter.Do(func() error {
+		return insertHandlerInternal(req)
+	})
+}
+
+func insertHandlerInternal(req *http.Request) error {
+	readCalls.Inc()
+
+	r := req.Body
+	if req.Header.Get("Content-Encoding") == "gzip" {
+		zr, err := common.GetGzipReader(r)
+		if err != nil {
+			return fmt.Errorf("cannot read gzipped vmimport data: %s", err)
+		}
+		defer common.PutGzipReader(zr)
+		r = zr
+	}
+
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+	for ctx.Read(r) {
+		if err := ctx.InsertRows(); err != nil {
+			return err
+		}
+	}
+	return ctx.Error()
+}
+
+func (ctx *pushCtx) InsertRows() error {
+	rows := ctx.Rows.Rows
+	rowsLen := 0
+	for i := range rows {
+		rowsLen += len(rows[i].Values)
+	}
+	ic := &ctx.Common
+	ic.Reset(rowsLen)
+	rowsTotal := 0
+	for i := range rows {
+		r := &rows[i]
+		ic.Labels = ic.Labels[:0]
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			ic.AddLabelBytes(tag.Key, tag.Value)
+		}
+		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
+		values := r.Values
+		timestamps := r.Timestamps
+		_ = timestamps[len(values)-1]
+		for j, value := range values {
+			timestamp := timestamps[j]
+			ic.WriteDataPoint(ctx.metricNameBuf, nil, timestamp, value)
+		}
+		rowsTotal += len(values)
+	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	return ic.FlushBufs()
+}
+
+func (ctx *pushCtx) Read(r io.Reader) bool {
+	if ctx.err != nil {
+		return false
+	}
+	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlockExt(r, ctx.reqBuf, ctx.tailBuf, *maxLineLen)
+	if ctx.err != nil {
+		if ctx.err != io.EOF {
+			readErrors.Inc()
+			ctx.err = fmt.Errorf("cannot read vmimport data: %s", ctx.err)
+		}
+		return false
+	}
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
+	return true
+}
+
+var (
+	readCalls  = metrics.NewCounter(`vm_read_calls_total{name="vmimport"}`)
+	readErrors = metrics.NewCounter(`vm_read_errors_total{name="vmimport"}`)
+)
+
+type pushCtx struct {
+	Rows   Rows
+	Common common.InsertCtx
+
+	reqBuf        []byte
+	tailBuf       []byte
+	metricNameBuf []byte
+
+	err error
+}
+
+func (ctx *pushCtx) Error() error {
+	if ctx.err == io.EOF {
+		return nil
+	}
+	return ctx.err
+}
+
+func (ctx *pushCtx) reset() {
+	ctx.Rows.Reset()
+	ctx.Common.Reset(0)
+
+	ctx.reqBuf = ctx.reqBuf[:0]
+	ctx.tailBuf = ctx.tailBuf[:0]
+	ctx.metricNameBuf = ctx.metricNameBuf[:0]
+
+	ctx.err = nil
+}
+
+func getPushCtx() *pushCtx {
+	select {
+	case ctx := <-pushCtxPoolCh:
+		return ctx
+	default:
+		if v := pushCtxPool.Get(); v != nil {
+			return v.(*pushCtx)
+		}
+		return &pushCtx{}
+	}
+}
+
+func putPushCtx(ctx *pushCtx) {
+	ctx.reset()
+	select {
+	case pushCtxPoolCh <- ctx:
+	default:
+		pushCtxPool.Put(ctx)
+	}
+}
+
+var pushCtxPool sync.Pool
+var pushCtxPoolCh = make(chan *pushCtx, runtime.GOMAXPROCS(-1))

app/vmrestore/Makefile

@@ -0,0 +1,67 @@
+# All these commands must run from repository root.
+
+vmrestore:
+	APP_NAME=vmrestore $(MAKE) app-local
+
+vmrestore-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker
+
+vmrestore-pure-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-pure
+
+vmrestore-amd64-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-amd64
+
+vmrestore-arm-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-arm
+
+vmrestore-arm64-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-arm64
+
+vmrestore-ppc64le-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-ppc64le
+
+vmrestore-386-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-386
+
+package-vmrestore:
+	APP_NAME=vmrestore $(MAKE) package-via-docker
+
+package-vmrestore-pure:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-pure
+
+package-vmrestore-amd64:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-amd64
+
+package-vmrestore-arm:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-arm
+
+package-vmrestore-arm64:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-arm64
+
+package-vmrestore-ppc64le:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-ppc64le
+
+package-vmrestore-386:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-386
+
+publish-vmrestore:
+	APP_NAME=vmrestore $(MAKE) publish-via-docker
+
+vmrestore-pure:
+	APP_NAME=vmrestore $(MAKE) app-local-pure
+
+vmrestore-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-amd64 ./app/vmrestore
+
+vmrestore-arm:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-arm ./app/vmrestore
+
+vmrestore-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-arm64 ./app/vmrestore
+
+vmrestore-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-ppc64le ./app/vmrestore
+
+vmrestore-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-386 ./app/vmrestore

app/vmrestore/README.md

@@ -0,0 +1,86 @@
+## vmrestore
+
+`vmrestore` restores data from backups created by [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md).
+VictoriaMetrics `v1.29.0` and newer versions must be used for working with the restored data.
+
+Restore process can be interrupted at any time. It is automatically resumed from the inerruption point
+when restarting `vmrestore` with the same args.
+
+
+### Usage
+
+VictoriaMetrics must be stopped during the restore process.
+
+```
+vmrestore -src=gcs://<bucket>/<path/to/backup> -storageDataPath=<local/path/to/restore>
+
+```
+
+* `<bucket>` is [GCS bucket](https://cloud.google.com/storage/docs/creating-buckets) name.
+* `<path/to/backup>` is the path to backup made with [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md) on GCS bucket.
+* `<local/path/to/restore>` is the path to folder where data will be restored. This folder must be passed
+  to VictoriaMetrics in `-storageDataPath` command-line flag after the restore process is complete.
+
+The original `-storageDataPath` directory may contain old files. They will be susbstituted by the files from backup.
+
+
+### Troubleshooting
+
+* If `vmrestore` eats all the network bandwidth, then set `-maxBytesPerSecond` to the desired value.
+* If `vmrestore` has been interrupted due to temporary error, then just restart it with the same args. It will resume the restore process.
+
+
+### Advanced usage
+
+Run `vmrestore -help` in order to see all the available options:
+
+```
+  -concurrency int
+    	The number of concurrent workers. Higher concurrency may reduce restore duration (default 10)
+  -configFilePath string
+    	Path to file with S3 configs. Configs are loaded from default location if not set.
+    	See https://docs.aws.amazon.com/general/latest/gr/aws-security-credentials.html
+  -configProfile string
+    	Profile name for S3 configs (default "default")
+  -credsFilePath string
+    	Path to file with GCS or S3 credentials. Credentials are loaded from default locations if not set.
+    	See https://cloud.google.com/iam/docs/creating-managing-service-account-keys and https://docs.aws.amazon.com/general/latest/gr/aws-security-credentials.html
+  -customS3Endpoint string
+    	Custom S3 endpoint for use with S3-compatible storages (e.g. MinIO). S3 is used if not set
+  -loggerLevel string
+    	Minimum level of errors to log. Possible values: INFO, ERROR, FATAL, PANIC (default "INFO")
+  -maxBytesPerSecond int
+    	The maximum download speed. There is no limit if it is set to 0
+  -memory.allowedPercent float
+    	Allowed percent of system memory VictoriaMetrics caches may occupy (default 60)
+  -src string
+    	Source path with backup on the remote storage. Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir
+  -storageDataPath string
+    	Destination path where backup must be restored. VictoriaMetrics must be stopped when restoring from backup. -storageDataPath dir can be non-empty. In this case only missing data is downloaded from backup (default "victoria-metrics-data")
+  -version
+    	Show VictoriaMetrics version
+```
+
+
+### How to build from sources
+
+It is recommended using [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) - see `vmutils-*` archives there.
+
+
+#### Development build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make vmrestore` from the root folder of the repository.
+   It builds `vmrestore` binary and puts it into the `bin` folder.
+
+#### Production build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make vmrestore-prod` from the root folder of the repository.
+   It builds `vmrestore-prod` binary and puts it into the `bin` folder.
+
+#### Building docker images
+
+Run `make package-vmrestore`. It builds `victoriametrics/vmrestore:<PKG_TAG>` docker image locally.
+`<PKG_TAG>` is auto-generated image tag, which depends on source code in the repository.
+The `<PKG_TAG>` may be manually set via `PKG_TAG=foobar make package-vmrestore`.

app/vmrestore/deployment/Dockerfile

@@ -0,0 +1,7 @@
+ARG certs_image
+FROM $certs_image AS certs
+FROM scratch
+COPY --from=certs /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
+ARG src_binary
+COPY $src_binary ./vmrestore-prod
+ENTRYPOINT ["/vmrestore-prod"]

app/vmrestore/main.go

@@ -0,0 +1,80 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/actions"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/fslocal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+var (
+	src = flag.String("src", "", "Source path with backup on the remote storage. "+
+		"Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir")
+	storageDataPath = flag.String("storageDataPath", "victoria-metrics-data", "Destination path where backup must be restored. "+
+		"VictoriaMetrics must be stopped when restoring from backup. -storageDataPath dir can be non-empty. In this case only missing data is downloaded from backup")
+	concurrency             = flag.Int("concurrency", 10, "The number of concurrent workers. Higher concurrency may reduce restore duration")
+	maxBytesPerSecond       = flag.Int("maxBytesPerSecond", 0, "The maximum download speed. There is no limit if it is set to 0")
+	skipBackupCompleteCheck = flag.Bool("skipBackupCompleteCheck", false, "Whether to skip checking for `backup complete` file in `-src`. This may be useful for restoring from old backups, which were created without `backup complete` file")
+)
+
+func main() {
+	flag.Usage = usage
+	flag.Parse()
+	buildinfo.Init()
+
+	srcFS, err := newSrcFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	dstFS, err := newDstFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	a := &actions.Restore{
+		Concurrency:             *concurrency,
+		Src:                     srcFS,
+		Dst:                     dstFS,
+		SkipBackupCompleteCheck: *skipBackupCompleteCheck,
+	}
+	if err := a.Run(); err != nil {
+		logger.Fatalf("cannot restore from backup: %s", err)
+	}
+}
+
+func usage() {
+	const s = `
+vmrestore restores VictoriaMetrics data from backups made by vmbackup.
+
+See the docs at https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md .
+`
+
+	f := flag.CommandLine.Output()
+	fmt.Fprintf(f, "%s\n", s)
+	flag.PrintDefaults()
+}
+
+func newDstFS() (*fslocal.FS, error) {
+	if len(*storageDataPath) == 0 {
+		return nil, fmt.Errorf("`-storageDataPath` cannot be empty")
+	}
+	fs := &fslocal.FS{
+		Dir:               *storageDataPath,
+		MaxBytesPerSecond: *maxBytesPerSecond,
+	}
+	if err := fs.Init(); err != nil {
+		return nil, fmt.Errorf("cannot initialize local fs: %s", err)
+	}
+	return fs, nil
+}
+
+func newSrcFS() (common.RemoteFS, error) {
+	fs, err := actions.NewRemoteFS(*src)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-src`=%q: %s", *src, err)
+	}
+	return fs, nil
+}

app/vmselect/main.go

@@ -2,6 +2,7 @@ package vmselect
 
 import (
 	"flag"
+	"fmt"
 	"net/http"
 	"runtime"
 	"strings"
@@ -70,7 +71,11 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		case <-t.C:
 			timerpool.Put(t)
 			concurrencyLimitTimeout.Inc()
-			httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
+			err := &httpserver.ErrorWithStatusCode{
+				Err:        fmt.Errorf("cannot handle more than %d concurrent requests", cap(concurrencyCh)),
+				StatusCode: http.StatusServiceUnavailable,
+			}
+			httpserver.Errorf(w, "%s", err)
 			return true
 		}
 	}
@@ -162,6 +167,18 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 			return true
 		}
 		return true
+	case "/api/v1/rules":
+		// Return dumb placeholder
+		rulesRequests.Inc()
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprintf(w, "%s", `{"status":"success","data":{"groups":[]}}`)
+		return true
+	case "/api/v1/alerts":
+		// Return dumb placehloder
+		alertsRequests.Inc()
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprintf(w, "%s", `{"status":"success","data":{"alerts":[]}}`)
+		return true
 	case "/api/v1/admin/tsdb/delete_series":
 		deleteRequests.Inc()
 		authKey := r.FormValue("authKey")
@@ -185,7 +202,10 @@ func sendPrometheusError(w http.ResponseWriter, r *http.Request, err error) {
 	logger.Errorf("error in %q: %s", r.URL.Path, err)
 
 	w.Header().Set("Content-Type", "application/json")
-	statusCode := 422
+	statusCode := http.StatusUnprocessableEntity
+	if esc, ok := err.(*httpserver.ErrorWithStatusCode); ok {
+		statusCode = esc.StatusCode
+	}
 	w.WriteHeader(statusCode)
 	prometheus.WriteErrorResponse(w, statusCode, err)
 }
@@ -220,4 +240,7 @@ var (
 
 	federateRequests = metrics.NewCounter(`vm_http_requests_total{path="/federate"}`)
 	federateErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/federate"}`)
+
+	rulesRequests  = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/rules"}`)
+	alertsRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/alerts"}`)
 )

app/vmselect/netstorage/fadvise_darwin.go

@@ -1,9 +0,0 @@
-package netstorage
-
-import (
-	"os"
-)
-
-func mustFadviseRandomRead(f *os.File) {
-	// Do nothing :)
-}

app/vmselect/netstorage/fadvise_linux.go

@@ -1,15 +0,0 @@
-package netstorage
-
-import (
-	"os"
-
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
-	"golang.org/x/sys/unix"
-)
-
-func mustFadviseRandomRead(f *os.File) {
-	fd := int(f.Fd())
-	if err := unix.Fadvise(int(fd), 0, 0, unix.FADV_RANDOM|unix.FADV_WILLNEED); err != nil {
-		logger.Panicf("FATAL: error returned from unix.Fadvise(RANDOM|WILLNEED): %s", err)
-	}
-}

app/vmselect/netstorage/netstorage.go

@@ -19,9 +19,9 @@ import (
 )
 
 var (
-	maxTagKeysPerSearch   = flag.Int("search.maxTagKeys", 10e3, "The maximum number of tag keys returned per search")
-	maxTagValuesPerSearch = flag.Int("search.maxTagValues", 10e3, "The maximum number of tag values returned per search")
-	maxMetricsPerSearch   = flag.Int("search.maxUniqueTimeseries", 100e3, "The maximum number of unique time series each search can scan")
+	maxTagKeysPerSearch   = flag.Int("search.maxTagKeys", 100e3, "The maximum number of tag keys returned per search")
+	maxTagValuesPerSearch = flag.Int("search.maxTagValues", 100e3, "The maximum number of tag values returned per search")
+	maxMetricsPerSearch   = flag.Int("search.maxUniqueTimeseries", 300e3, "The maximum number of unique time series each search can scan")
 )
 
 // Result is a single timeseries result.
@@ -92,6 +92,7 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	doneCh := make(chan error)
 
 	// Start workers.
+	rowsProcessedTotal := uint64(0)
 	for i := 0; i < workersCount; i++ {
 		go func(workerID uint) {
 			rs := getResult()
@@ -99,6 +100,7 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 			maxWorkersCount := gomaxprocs / workersCount
 
 			var err error
+			rowsProcessed := 0
 			for pts := range workCh {
 				if time.Until(rss.deadline.Deadline) < 0 {
 					err = fmt.Errorf("timeout exceeded during query execution: %s", rss.deadline.Timeout)
@@ -111,8 +113,10 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 					// Skip empty blocks.
 					continue
 				}
+				rowsProcessed += len(rs.Values)
 				f(rs, workerID)
 			}
+			atomic.AddUint64(&rowsProcessedTotal, uint64(rowsProcessed))
 			// Drain the remaining work
 			for range workCh {
 			}
@@ -124,6 +128,7 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	for i := range rss.packedTimeseries {
 		workCh <- &rss.packedTimeseries[i]
 	}
+	seriesProcessedTotal := len(rss.packedTimeseries)
 	rss.packedTimeseries = rss.packedTimeseries[:0]
 	close(workCh)
 
@@ -134,6 +139,8 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 			errors = append(errors, err)
 		}
 	}
+	perQueryRowsProcessed.Update(float64(rowsProcessedTotal))
+	perQuerySeriesProcessed.Update(float64(seriesProcessedTotal))
 	if len(errors) > 0 {
 		// Return just the first error, since other errors
 		// is likely duplicate the first error.
@@ -142,6 +149,9 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	return nil
 }
 
+var perQueryRowsProcessed = metrics.NewHistogram(`vm_per_query_rows_processed_count`)
+var perQuerySeriesProcessed = metrics.NewHistogram(`vm_per_query_series_processed_count`)
+
 var gomaxprocs = runtime.GOMAXPROCS(-1)
 
 type packedTimeseries struct {
@@ -422,13 +432,10 @@ func GetLabelEntries(deadline Deadline) ([]storage.TagEntry, error) {
 	// Sort labelEntries by the number of label values in each entry.
 	sort.Slice(labelEntries, func(i, j int) bool {
 		a, b := labelEntries[i].Values, labelEntries[j].Values
-		if len(a) < len(b) {
-			return true
+		if len(a) != len(b) {
+			return len(a) > len(b)
 		}
-		if len(a) > len(b) {
-			return false
-		}
-		return labelEntries[i].Key < labelEntries[j].Key
+		return labelEntries[i].Key > labelEntries[j].Key
 	})
 
 	return labelEntries, nil
@@ -452,16 +459,12 @@ func getStorageSearch() *storage.Search {
 }
 
 func putStorageSearch(sr *storage.Search) {
-	n := atomic.LoadUint64(&sr.MissingMetricNamesForMetricID)
-	missingMetricNamesForMetricID.Add(int(n))
 	sr.MustClose()
 	ssPool.Put(sr)
 }
 
 var ssPool sync.Pool
 
-var missingMetricNamesForMetricID = metrics.NewCounter(`vm_missing_metric_names_for_metric_id_total`)
-
 // ProcessSearchQuery performs sq on storage nodes until the given deadline.
 func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadline) (*Results, error) {
 	// Setup search.
@@ -484,9 +487,12 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadli
 	tbf := getTmpBlocksFile()
 	m := make(map[string][]tmpBlockAddr)
 	blocksRead := 0
+	bb := tmpBufPool.Get()
+	defer tmpBufPool.Put(bb)
 	for sr.NextMetricBlock() {
 		blocksRead++
-		addr, err := tbf.WriteBlock(sr.MetricBlock.Block)
+		bb.B = storage.MarshalBlock(bb.B[:0], sr.MetricBlock.Block)
+		addr, err := tbf.WriteBlockData(bb.B)
 		if err != nil {
 			putTmpBlocksFile(tbf)
 			return nil, fmt.Errorf("cannot write data block #%d to temporary blocks file: %s", blocksRead, err)
@@ -520,6 +526,15 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadli
 		pts.metricName = metricName
 		pts.addrs = addrs
 	}
+
+	// Sort rss.packedTimeseries by the first addr offset in order
+	// to reduce the number of disk seeks during unpacking in RunParallel.
+	// In this case tmpBlocksFile must be read almost sequentially.
+	sort.Slice(rss.packedTimeseries, func(i, j int) bool {
+		pts := rss.packedTimeseries
+		return pts[i].addrs[0].offset < pts[j].addrs[0].offset
+	})
+
 	return &rss, nil
 }
 

app/vmselect/netstorage/tmp_blocks_file.go

@@ -1,7 +1,6 @@
 package netstorage
 
 import (
-	"bufio"
 	"fmt"
 	"io/ioutil"
 	"os"
@@ -10,6 +9,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -30,13 +30,23 @@ func InitTmpBlocksDir(tmpDirPath string) {
 
 var tmpBlocksDir string
 
-const maxInmemoryTmpBlocksFile = 512 * 1024
+func maxInmemoryTmpBlocksFile() int {
+	mem := memory.Allowed()
+	maxLen := mem / 1024
+	if maxLen < 64*1024 {
+		return 64 * 1024
+	}
+	return maxLen
+}
+
+var _ = metrics.NewGauge(`vm_tmp_blocks_max_inmemory_file_size_bytes`, func() float64 {
+	return float64(maxInmemoryTmpBlocksFile())
+})
 
 type tmpBlocksFile struct {
 	buf []byte
 
-	f  *os.File
-	bw *bufio.Writer
+	f *os.File
 
 	offset uint64
 }
@@ -44,7 +54,9 @@ type tmpBlocksFile struct {
 func getTmpBlocksFile() *tmpBlocksFile {
 	v := tmpBlocksFilePool.Get()
 	if v == nil {
-		return &tmpBlocksFile{}
+		return &tmpBlocksFile{
+			buf: make([]byte, 0, maxInmemoryTmpBlocksFile()),
+		}
 	}
 	return v.(*tmpBlocksFile)
 }
@@ -53,7 +65,6 @@ func putTmpBlocksFile(tbf *tmpBlocksFile) {
 	tbf.MustClose()
 	tbf.buf = tbf.buf[:0]
 	tbf.f = nil
-	tbf.bw = nil
 	tbf.offset = 0
 	tmpBlocksFilePool.Put(tbf)
 }
@@ -69,51 +80,34 @@ func (addr tmpBlockAddr) String() string {
 	return fmt.Sprintf("offset %d, size %d", addr.offset, addr.size)
 }
 
-func getBufioWriter(f *os.File) *bufio.Writer {
-	v := bufioWriterPool.Get()
-	if v == nil {
-		return bufio.NewWriterSize(f, maxInmemoryTmpBlocksFile*2)
-	}
-	bw := v.(*bufio.Writer)
-	bw.Reset(f)
-	return bw
-}
-
-func putBufioWriter(bw *bufio.Writer) {
-	bufioWriterPool.Put(bw)
-}
-
-var bufioWriterPool sync.Pool
-
 var tmpBlocksFilesCreated = metrics.NewCounter(`vm_tmp_blocks_files_created_total`)
 
-// WriteBlock writes b to tbf.
+// WriteBlockData writes b to tbf.
 //
 // It returns errors since the operation may fail on space shortage
 // and this must be handled.
-func (tbf *tmpBlocksFile) WriteBlock(b *storage.Block) (tmpBlockAddr, error) {
+func (tbf *tmpBlocksFile) WriteBlockData(b []byte) (tmpBlockAddr, error) {
 	var addr tmpBlockAddr
 	addr.offset = tbf.offset
-
-	tbfBufLen := len(tbf.buf)
-	tbf.buf = storage.MarshalBlock(tbf.buf, b)
-	addr.size = len(tbf.buf) - tbfBufLen
+	addr.size = len(b)
 	tbf.offset += uint64(addr.size)
-	if tbf.offset <= maxInmemoryTmpBlocksFile {
+	if len(tbf.buf)+len(b) <= cap(tbf.buf) {
+		// Fast path - the data fits tbf.buf
+		tbf.buf = append(tbf.buf, b...)
 		return addr, nil
 	}
 
+	// Slow path: flush the data from tbf.buf to file.
 	if tbf.f == nil {
 		f, err := ioutil.TempFile(tmpBlocksDir, "")
 		if err != nil {
 			return addr, err
 		}
 		tbf.f = f
-		tbf.bw = getBufioWriter(f)
 		tmpBlocksFilesCreated.Inc()
 	}
-	_, err := tbf.bw.Write(tbf.buf)
-	tbf.buf = tbf.buf[:0]
+	_, err := tbf.f.Write(tbf.buf)
+	tbf.buf = append(tbf.buf[:0], b...)
 	if err != nil {
 		return addr, fmt.Errorf("cannot write block to %q: %s", tbf.f.Name(), err)
 	}
@@ -124,15 +118,18 @@ func (tbf *tmpBlocksFile) Finalize() error {
 	if tbf.f == nil {
 		return nil
 	}
-
-	err := tbf.bw.Flush()
-	putBufioWriter(tbf.bw)
-	tbf.bw = nil
+	if _, err := tbf.f.Write(tbf.buf); err != nil {
+		return fmt.Errorf("cannot flush the remaining %d bytes to tmpBlocksFile: %s", len(tbf.buf), err)
+	}
+	tbf.buf = tbf.buf[:0]
 	if _, err := tbf.f.Seek(0, 0); err != nil {
 		logger.Panicf("FATAL: cannot seek to the start of file: %s", err)
 	}
-	mustFadviseRandomRead(tbf.f)
-	return err
+	// Hint the OS that the file is read almost sequentiallly.
+	// This should reduce the number of disk seeks, which is important
+	// for HDDs.
+	fs.MustFadviseSequentialRead(tbf.f, true)
+	return nil
 }
 
 func (tbf *tmpBlocksFile) MustReadBlockAt(dst *storage.Block, addr tmpBlockAddr) {
@@ -167,10 +164,6 @@ func (tbf *tmpBlocksFile) MustClose() {
 	if tbf.f == nil {
 		return
 	}
-	if tbf.bw != nil {
-		putBufioWriter(tbf.bw)
-		tbf.bw = nil
-	}
 	fname := tbf.f.Name()
 
 	// Remove the file at first, then close it.

app/vmselect/netstorage/tmp_blocks_file_test.go

@@ -30,7 +30,7 @@ func TestTmpBlocksFileSerial(t *testing.T) {
 }
 
 func TestTmpBlocksFileConcurrent(t *testing.T) {
-	concurrency := 4
+	concurrency := 3
 	ch := make(chan error, concurrency)
 	for i := 0; i < concurrency; i++ {
 		go func() {
@@ -69,7 +69,7 @@ func testTmpBlocksFile() error {
 		_, _, _ = b.MarshalData(0, 0)
 		return &b
 	}
-	for _, size := range []int{1024, 16 * 1024, maxInmemoryTmpBlocksFile / 2, 2 * maxInmemoryTmpBlocksFile} {
+	for _, size := range []int{1024, 16 * 1024, maxInmemoryTmpBlocksFile() / 2, 2 * maxInmemoryTmpBlocksFile()} {
 		err := func() error {
 			tbf := getTmpBlocksFile()
 			defer putTmpBlocksFile(tbf)
@@ -77,9 +77,12 @@ func testTmpBlocksFile() error {
 			// Write blocks until their summary size exceeds `size`.
 			var addrs []tmpBlockAddr
 			var blocks []*storage.Block
+			bb := tmpBufPool.Get()
+			defer tmpBufPool.Put(bb)
 			for tbf.offset < uint64(size) {
 				b := createBlock()
-				addr, err := tbf.WriteBlock(b)
+				bb.B = storage.MarshalBlock(bb.B[:0], b)
+				addr, err := tbf.WriteBlockData(bb.B)
 				if err != nil {
 					return fmt.Errorf("cannot write block at offset %d: %s", tbf.offset, err)
 				}
@@ -94,7 +97,7 @@ func testTmpBlocksFile() error {
 			}
 
 			// Read blocks in parallel and verify them
-			concurrency := 3
+			concurrency := 2
 			workCh := make(chan int)
 			doneCh := make(chan error)
 			for i := 0; i < concurrency; i++ {

app/vmselect/prometheus/export.qtpl

@@ -13,7 +13,7 @@
 	{% for i, ts := range rs.Timestamps %}
 		{%z= bb.B %}{% space %}
 		{%f= rs.Values[i] %}{% space %}
-		{%d= int(ts) %}{% newline %}
+		{%dl= ts %}{% newline %}
 	{% endfor %}
 	{% code quicktemplate.ReleaseByteBuffer(bb) %}
 {% endfunc %}
@@ -35,10 +35,10 @@
 		"timestamps":[
 			{% if len(rs.Timestamps) > 0 %}
 				{% code timestamps := rs.Timestamps %}
-				{%d= int(timestamps[0]) %}
+				{%dl= timestamps[0] %}
 				{% code timestamps = timestamps[1:] %}
 				{% for _, ts := range timestamps %}
-					,{%d= int(ts) %}
+					,{%dl= ts %}
 				{% endfor %}
 			{% endif %}
 		]

app/vmselect/prometheus/export.qtpl.go

@@ -49,7 +49,7 @@ func StreamExportPrometheusLine(qw422016 *qt422016.Writer, rs *netstorage.Result
 //line app/vmselect/prometheus/export.qtpl:15
 		qw422016.N().S(` `)
 //line app/vmselect/prometheus/export.qtpl:16
-		qw422016.N().D(int(ts))
+		qw422016.N().DL(ts)
 //line app/vmselect/prometheus/export.qtpl:16
 		qw422016.N().S(`
 `)
@@ -129,7 +129,7 @@ func StreamExportJSONLine(qw422016 *qt422016.Writer, rs *netstorage.Result) {
 		timestamps := rs.Timestamps
 
 //line app/vmselect/prometheus/export.qtpl:38
-		qw422016.N().D(int(timestamps[0]))
+		qw422016.N().DL(timestamps[0])
 //line app/vmselect/prometheus/export.qtpl:39
 		timestamps = timestamps[1:]
 
@@ -138,7 +138,7 @@ func StreamExportJSONLine(qw422016 *qt422016.Writer, rs *netstorage.Result) {
 //line app/vmselect/prometheus/export.qtpl:40
 			qw422016.N().S(`,`)
 //line app/vmselect/prometheus/export.qtpl:41
-			qw422016.N().D(int(ts))
+			qw422016.N().DL(ts)
 //line app/vmselect/prometheus/export.qtpl:42
 		}
 //line app/vmselect/prometheus/export.qtpl:43

app/vmselect/prometheus/federate.qtpl

@@ -10,7 +10,7 @@
 	{% if len(rs.Timestamps) == 0 || len(rs.Values) == 0 %}{% return %}{% endif %}
 	{%= prometheusMetricName(&rs.MetricName) %}{% space %}
 	{%f= rs.Values[len(rs.Values)-1] %}{% space %}
-	{%d= int(rs.Timestamps[len(rs.Timestamps)-1]) %}{% newline %}
+	{%dl= rs.Timestamps[len(rs.Timestamps)-1] %}{% newline %}
 {% endfunc %}
 
 {% endstripspace %}

app/vmselect/prometheus/federate.qtpl.go

@@ -41,7 +41,7 @@ func StreamFederate(qw422016 *qt422016.Writer, rs *netstorage.Result) {
 //line app/vmselect/prometheus/federate.qtpl:12
 	qw422016.N().S(` `)
 //line app/vmselect/prometheus/federate.qtpl:13
-	qw422016.N().D(int(rs.Timestamps[len(rs.Timestamps)-1]))
+	qw422016.N().DL(rs.Timestamps[len(rs.Timestamps)-1])
 //line app/vmselect/prometheus/federate.qtpl:13
 	qw422016.N().S(`
 `)

app/vmselect/prometheus/prometheus.go

@@ -15,23 +15,25 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/promql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/quicktemplate"
 )
 
 var (
-	maxQueryDuration = flag.Duration("search.maxQueryDuration", time.Second*30, "The maximum time for search query execution")
-	maxQueryLen      = flag.Int("search.maxQueryLen", 16*1024, "The maximum search query length in bytes")
+	latencyOffset = flag.Duration("search.latencyOffset", time.Second*30, "The time when data points become visible in query results after the colection. "+
+		"Too small value can result in incomplete last points for query results")
+	maxExportDuration = flag.Duration("search.maxExportDuration", 10*time.Minute, "The maximum duration for `/api/v1/export` call")
+	maxQueryDuration  = flag.Duration("search.maxQueryDuration", time.Second*30, "The maximum duration for search query execution")
+	maxQueryLen       = flag.Int("search.maxQueryLen", 16*1024, "The maximum search query length in bytes")
+	maxLookback       = flag.Duration("search.maxLookback", 0, "Synonim to `-search.lookback-delta` from Prometheus. "+
+		"The value is dynamically detected from interval between time series datapoints if not set. It can be overridden on per-query basis via `max_lookback` arg")
 )
 
 // Default step used if not set.
 const defaultStep = 5 * 60 * 1000
 
-// Latency for data processing pipeline, i.e. the time between data is ignested
-// into the system and the time it becomes visible to search.
-const latencyOffset = 60 * 1000
-
 // FederateHandler implements /federate . See https://prometheus.io/docs/prometheus/latest/federation/
 func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
@@ -43,11 +45,14 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		return fmt.Errorf("missing `match[]` arg")
 	}
-	maxLookback, err := getDuration(r, "max_lookback", defaultStep)
+	lookbackDelta, err := getMaxLookback(r)
 	if err != nil {
 		return err
 	}
-	start, err := getTime(r, "start", ct-maxLookback)
+	if lookbackDelta <= 0 {
+		lookbackDelta = defaultStep
+	}
+	start, err := getTime(r, "start", ct-lookbackDelta)
 	if err != nil {
 		return err
 	}
@@ -55,7 +60,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 	if start >= end {
 		start = end - defaultStep
 	}
@@ -126,9 +131,9 @@ func ExportHandler(w http.ResponseWriter, r *http.Request) error {
 		return err
 	}
 	format := r.FormValue("format")
-	deadline := getDeadline(r)
+	deadline := getDeadlineForExport(r)
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
 	if err := exportHandler(w, matches, start, end, format, deadline); err != nil {
 		return err
@@ -142,7 +147,7 @@ var exportDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 func exportHandler(w http.ResponseWriter, matches []string, start, end int64, format string, deadline netstorage.Deadline) error {
 	writeResponseFunc := WriteExportStdResponse
 	writeLineFunc := WriteExportJSONLine
-	contentType := "application/json"
+	contentType := "application/stream+json"
 	if format == "prometheus" {
 		contentType = "text/plain"
 		writeLineFunc = WriteExportPrometheusLine
@@ -232,7 +237,7 @@ var deleteDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#querying-label-values
 func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 
 	if err := r.ParseForm(); err != nil {
 		return fmt.Errorf("cannot parse form values: %s", err)
@@ -282,8 +287,15 @@ func labelValuesWithMatches(labelName string, matches []string, start, end int64
 	if err != nil {
 		return nil, err
 	}
+	for i, tfs := range tagFilterss {
+		// Add `labelName!=''` tag filter in order to filter out series without the labelName.
+		tagFilterss[i] = append(tfs, storage.TagFilter{
+			Key:        []byte(labelName),
+			IsNegative: true,
+		})
+	}
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
 	sq := &storage.SearchQuery{
 		MinTimestamp: start,
@@ -323,7 +335,7 @@ var labelValuesDuration = metrics.NewSummary(`vm_request_duration_seconds{path="
 // LabelsCountHandler processes /api/v1/labels/count request.
 func LabelsCountHandler(w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 	labelEntries, err := netstorage.GetLabelEntries(deadline)
 	if err != nil {
 		return fmt.Errorf(`cannot obtain label entries: %s`, err)
@@ -342,10 +354,38 @@ var labelsCountDuration = metrics.NewSummary(`vm_request_duration_seconds{path="
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#getting-label-names
 func LabelsHandler(w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
-	deadline := getDeadline(r)
-	labels, err := netstorage.GetLabels(deadline)
-	if err != nil {
-		return fmt.Errorf("cannot obtain labels: %s", err)
+	deadline := getDeadlineForQuery(r)
+
+	if err := r.ParseForm(); err != nil {
+		return fmt.Errorf("cannot parse form values: %s", err)
+	}
+	var labels []string
+	if len(r.Form["match[]"]) == 0 && len(r.Form["start"]) == 0 && len(r.Form["end"]) == 0 {
+		var err error
+		labels, err = netstorage.GetLabels(deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain labels: %s", err)
+		}
+	} else {
+		// Extended functionality that allows filtering by label filters and time range
+		// i.e. /api/v1/labels?match[]=foobar{baz="abc"}&start=...&end=...
+		matches := r.Form["match[]"]
+		if len(matches) == 0 {
+			matches = []string{"{__name__!=''}"}
+		}
+		ct := currentTime()
+		end, err := getTime(r, "end", ct)
+		if err != nil {
+			return err
+		}
+		start, err := getTime(r, "start", end-defaultStep)
+		if err != nil {
+			return err
+		}
+		labels, err = labelsWithMatches(matches, start, end, deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain labels for match[]=%q, start=%d, end=%d: %s", matches, start, end, err)
+		}
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -354,12 +394,57 @@ func LabelsHandler(w http.ResponseWriter, r *http.Request) error {
 	return nil
 }
 
+func labelsWithMatches(matches []string, start, end int64, deadline netstorage.Deadline) ([]string, error) {
+	if len(matches) == 0 {
+		logger.Panicf("BUG: matches must be non-empty")
+	}
+	tagFilterss, err := getTagFilterssFromMatches(matches)
+	if err != nil {
+		return nil, err
+	}
+	if start >= end {
+		end = start + defaultStep
+	}
+	sq := &storage.SearchQuery{
+		MinTimestamp: start,
+		MaxTimestamp: end,
+		TagFilterss:  tagFilterss,
+	}
+	rss, err := netstorage.ProcessSearchQuery(sq, false, deadline)
+	if err != nil {
+		return nil, fmt.Errorf("cannot fetch data for %q: %s", sq, err)
+	}
+
+	m := make(map[string]struct{})
+	var mLock sync.Mutex
+	err = rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
+		mLock.Lock()
+		tags := rs.MetricName.Tags
+		for i := range tags {
+			t := &tags[i]
+			m[string(t.Key)] = struct{}{}
+		}
+		m["__name__"] = struct{}{}
+		mLock.Unlock()
+	})
+	if err != nil {
+		return nil, fmt.Errorf("error when data fetching: %s", err)
+	}
+
+	labels := make([]string, 0, len(m))
+	for label := range m {
+		labels = append(labels, label)
+	}
+	sort.Strings(labels)
+	return labels, nil
+}
+
 var labelsDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/labels"}`)
 
 // SeriesCountHandler processes /api/v1/series/count request.
 func SeriesCountHandler(w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 	n, err := netstorage.GetSeriesCount(deadline)
 	if err != nil {
 		return fmt.Errorf("cannot obtain series count: %s", err)
@@ -399,14 +484,14 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 
 	tagFilterss, err := getTagFilterssFromMatches(matches)
 	if err != nil {
 		return err
 	}
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
 	sq := &storage.SearchQuery{
 		MinTimestamp: start,
@@ -463,34 +548,33 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	step, err := getDuration(r, "step", latencyOffset)
+	queryOffset := getLatencyOffsetMilliseconds()
+	step, err := getDuration(r, "step", queryOffset)
+	if err != nil {
+		return err
+	}
+	deadline := getDeadlineForQuery(r)
+	lookbackDelta, err := getMaxLookback(r)
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
 
 	if len(query) > *maxQueryLen {
 		return fmt.Errorf(`too long query; got %d bytes; mustn't exceed %d bytes`, len(query), *maxQueryLen)
 	}
-	if ct-start < latencyOffset {
-		start -= latencyOffset
+	if !getBool(r, "nocache") && ct-start < queryOffset {
+		// Adjust start time only if `nocache` arg isn't set.
+		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/241
+		start = ct - queryOffset
 	}
 	if childQuery, windowStr, offsetStr := promql.IsMetricSelectorWithRollup(query); childQuery != "" {
-		var window int64
-		if len(windowStr) > 0 {
-			var err error
-			window, err = promql.DurationValue(windowStr, step)
-			if err != nil {
-				return err
-			}
+		window, err := parsePositiveDuration(windowStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse window: %s", err)
 		}
-		var offset int64
-		if len(offsetStr) > 0 {
-			var err error
-			offset, err = promql.DurationValue(offsetStr, step)
-			if err != nil {
-				return err
-			}
+		offset, err := parseDuration(offsetStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse offset: %s", err)
 		}
 		start -= offset
 		end := start
@@ -501,12 +585,38 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 		queryDuration.UpdateDuration(startTime)
 		return nil
 	}
+	if childQuery, windowStr, stepStr, offsetStr := promql.IsRollup(query); childQuery != "" {
+		newStep, err := parsePositiveDuration(stepStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse step: %s", err)
+		}
+		if newStep > 0 {
+			step = newStep
+		}
+		window, err := parsePositiveDuration(windowStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse window: %s", err)
+		}
+		offset, err := parseDuration(offsetStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse offset: %s", err)
+		}
+		start -= offset
+		end := start
+		start = end - window
+		if err := queryRangeHandler(w, childQuery, start, end, step, r, ct); err != nil {
+			return err
+		}
+		queryDuration.UpdateDuration(startTime)
+		return nil
+	}
 
 	ec := promql.EvalConfig{
-		Start:    start,
-		End:      start,
-		Step:     step,
-		Deadline: deadline,
+		Start:         start,
+		End:           start,
+		Step:          step,
+		Deadline:      deadline,
+		LookbackDelta: lookbackDelta,
 	}
 	result, err := promql.Exec(&ec, query, true)
 	if err != nil {
@@ -521,6 +631,20 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 
 var queryDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/query"}`)
 
+func parseDuration(s string, step int64) (int64, error) {
+	if len(s) == 0 {
+		return 0, nil
+	}
+	return metricsql.DurationValue(s, step)
+}
+
+func parsePositiveDuration(s string, step int64) (int64, error) {
+	if len(s) == 0 {
+		return 0, nil
+	}
+	return metricsql.PositiveDurationValue(s, step)
+}
+
 // QueryRangeHandler processes /api/v1/query_range request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#range-queries
@@ -544,42 +668,92 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	if err := queryRangeHandler(w, query, start, end, step, r, ct); err != nil {
+		return err
+	}
+	queryRangeDuration.UpdateDuration(startTime)
+	return nil
+}
+
+func queryRangeHandler(w http.ResponseWriter, query string, start, end, step int64, r *http.Request, ct int64) error {
+	deadline := getDeadlineForQuery(r)
 	mayCache := !getBool(r, "nocache")
+	lookbackDelta, err := getMaxLookback(r)
+	if err != nil {
+		return err
+	}
 
 	// Validate input args.
 	if len(query) > *maxQueryLen {
 		return fmt.Errorf(`too long query; got %d bytes; mustn't exceed %d bytes`, len(query), *maxQueryLen)
 	}
 	if start > end {
-		start = end
+		end = start + defaultStep
 	}
 	if err := promql.ValidateMaxPointsPerTimeseries(start, end, step); err != nil {
 		return err
 	}
-	start, end = promql.AdjustStartEnd(start, end, step)
+	if mayCache {
+		start, end = promql.AdjustStartEnd(start, end, step)
+	}
 
 	ec := promql.EvalConfig{
-		Start:    start,
-		End:      end,
-		Step:     step,
-		Deadline: deadline,
-		MayCache: mayCache,
+		Start:         start,
+		End:           end,
+		Step:          step,
+		Deadline:      deadline,
+		MayCache:      mayCache,
+		LookbackDelta: lookbackDelta,
 	}
 	result, err := promql.Exec(&ec, query, false)
 	if err != nil {
 		return fmt.Errorf("cannot execute %q: %s", query, err)
 	}
-	if ct-end < latencyOffset {
+	queryOffset := getLatencyOffsetMilliseconds()
+	if ct-end < queryOffset {
 		result = adjustLastPoints(result)
 	}
 
+	// Remove NaN values as Prometheus does.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/153
+	removeNaNValuesInplace(result)
+
 	w.Header().Set("Content-Type", "application/json")
 	WriteQueryRangeResponse(w, result)
-	queryRangeDuration.UpdateDuration(startTime)
 	return nil
 }
 
+func removeNaNValuesInplace(tss []netstorage.Result) {
+	for i := range tss {
+		ts := &tss[i]
+		hasNaNs := false
+		for _, v := range ts.Values {
+			if math.IsNaN(v) {
+				hasNaNs = true
+				break
+			}
+		}
+		if !hasNaNs {
+			// Fast path: nothing to remove.
+			continue
+		}
+
+		// Slow path: remove NaNs.
+		srcTimestamps := ts.Timestamps
+		dstValues := ts.Values[:0]
+		dstTimestamps := ts.Timestamps[:0]
+		for j, v := range ts.Values {
+			if math.IsNaN(v) {
+				continue
+			}
+			dstValues = append(dstValues, v)
+			dstTimestamps = append(dstTimestamps, srcTimestamps[j])
+		}
+		ts.Values = dstValues
+		ts.Timestamps = dstTimestamps
+	}
+}
+
 var queryRangeDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/query_range"}`)
 
 // adjustLastPoints substitutes the last point values with the previous
@@ -689,12 +863,26 @@ func getDuration(r *http.Request, argKey string, defaultValue int64) (int64, err
 
 const maxDurationMsecs = 100 * 365 * 24 * 3600 * 1000
 
-func getDeadline(r *http.Request) netstorage.Deadline {
+func getMaxLookback(r *http.Request) (int64, error) {
+	d := int64(*maxLookback / time.Millisecond)
+	return getDuration(r, "max_lookback", d)
+}
+
+func getDeadlineForQuery(r *http.Request) netstorage.Deadline {
+	dMax := int64(maxQueryDuration.Seconds() * 1e3)
+	return getDeadlineWithMaxDuration(r, dMax)
+}
+
+func getDeadlineForExport(r *http.Request) netstorage.Deadline {
+	dMax := int64(maxExportDuration.Seconds() * 1e3)
+	return getDeadlineWithMaxDuration(r, dMax)
+}
+
+func getDeadlineWithMaxDuration(r *http.Request, dMax int64) netstorage.Deadline {
 	d, err := getDuration(r, "timeout", 0)
 	if err != nil {
 		d = 0
 	}
-	dMax := int64(maxQueryDuration.Seconds() * 1e3)
 	if d <= 0 || d > dMax {
 		d = dMax
 	}
@@ -727,3 +915,11 @@ func getTagFilterssFromMatches(matches []string) ([][]storage.TagFilter, error)
 	}
 	return tagFilterss, nil
 }
+
+func getLatencyOffsetMilliseconds() int64 {
+	d := int64(*latencyOffset / time.Millisecond)
+	if d <= 1000 {
+		d = 1000
+	}
+	return d
+}

app/vmselect/prometheus/prometheus_test.go

@@ -2,11 +2,48 @@ package prometheus
 
 import (
 	"fmt"
+	"math"
 	"net/http"
 	"net/url"
+	"reflect"
 	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 )
 
+func TestRemoveNaNValuesInplace(t *testing.T) {
+	f := func(tss []netstorage.Result, tssExpected []netstorage.Result) {
+		t.Helper()
+		removeNaNValuesInplace(tss)
+		if !reflect.DeepEqual(tss, tssExpected) {
+			t.Fatalf("unexpected result; got %v; want %v", tss, tssExpected)
+		}
+	}
+
+	nan := math.NaN()
+
+	f(nil, nil)
+	f([]netstorage.Result{
+		{
+			Timestamps: []int64{100, 200, 300},
+			Values:     []float64{1, 2, 3},
+		},
+		{
+			Timestamps: []int64{100, 200, 300, 400},
+			Values:     []float64{nan, nan, 3, nan},
+		},
+	}, []netstorage.Result{
+		{
+			Timestamps: []int64{100, 200, 300},
+			Values:     []float64{1, 2, 3},
+		},
+		{
+			Timestamps: []int64{300},
+			Values:     []float64{3},
+		},
+	})
+}
+
 func TestGetTimeSuccess(t *testing.T) {
 	f := func(s string, timestampExpected int64) {
 		t.Helper()

app/vmselect/prometheus/series_count_response.qtpl

@@ -3,7 +3,7 @@ SeriesCountResponse generates response for /api/v1/series/count .
 {% func SeriesCountResponse(n uint64) %}
 {
 	"status":"success",
-	"data":[{%d int(n) %}]
+	"data":[{%dl int64(n) %}]
 }
 {% endfunc %}
 {% endstripspace %}

app/vmselect/prometheus/series_count_response.qtpl.go

@@ -24,7 +24,7 @@ func StreamSeriesCountResponse(qw422016 *qt422016.Writer, n uint64) {
 //line app/vmselect/prometheus/series_count_response.qtpl:3
 	qw422016.N().S(`{"status":"success","data":[`)
 //line app/vmselect/prometheus/series_count_response.qtpl:6
-	qw422016.N().D(int(n))
+	qw422016.N().DL(int64(n))
 //line app/vmselect/prometheus/series_count_response.qtpl:6
 	qw422016.N().S(`]}`)
 //line app/vmselect/prometheus/series_count_response.qtpl:8

app/vmselect/promql/aggr.go

@@ -8,7 +8,10 @@ import (
 	"strings"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/histogram"
 )
 
 var aggrFuncs = map[string]aggrFunc{
@@ -25,19 +28,28 @@ var aggrFuncs = map[string]aggrFunc{
 	"topk":         newAggrFuncTopK(false),
 	"quantile":     aggrFuncQuantile,
 
-	// Extended PromQL funcs
-	"median":   aggrFuncMedian,
-	"limitk":   aggrFuncLimitK,
-	"distinct": newAggrFunc(aggrFuncDistinct),
-	"sum2":     newAggrFunc(aggrFuncSum2),
-	"geomean":  newAggrFunc(aggrFuncGeomean),
+	// PromQL extension funcs
+	"median":         aggrFuncMedian,
+	"limitk":         aggrFuncLimitK,
+	"distinct":       newAggrFunc(aggrFuncDistinct),
+	"sum2":           newAggrFunc(aggrFuncSum2),
+	"geomean":        newAggrFunc(aggrFuncGeomean),
+	"histogram":      newAggrFunc(aggrFuncHistogram),
+	"topk_min":       newAggrFuncRangeTopK(minValue, false),
+	"topk_max":       newAggrFuncRangeTopK(maxValue, false),
+	"topk_avg":       newAggrFuncRangeTopK(avgValue, false),
+	"topk_median":    newAggrFuncRangeTopK(medianValue, false),
+	"bottomk_min":    newAggrFuncRangeTopK(minValue, true),
+	"bottomk_max":    newAggrFuncRangeTopK(maxValue, true),
+	"bottomk_avg":    newAggrFuncRangeTopK(avgValue, true),
+	"bottomk_median": newAggrFuncRangeTopK(medianValue, true),
 }
 
 type aggrFunc func(afa *aggrFuncArg) ([]*timeseries, error)
 
 type aggrFuncArg struct {
 	args [][]*timeseries
-	ae   *aggrFuncExpr
+	ae   *metricsql.AggrFuncExpr
 	ec   *EvalConfig
 }
 
@@ -46,20 +58,6 @@ func getAggrFunc(s string) aggrFunc {
 	return aggrFuncs[s]
 }
 
-func isAggrFunc(s string) bool {
-	return getAggrFunc(s) != nil
-}
-
-func isAggrFuncModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	case "by", "without":
-		return true
-	default:
-		return false
-	}
-}
-
 func newAggrFunc(afe func(tss []*timeseries) []*timeseries) aggrFunc {
 	return func(afa *aggrFuncArg) ([]*timeseries, error) {
 		args := afa.args
@@ -70,7 +68,7 @@ func newAggrFunc(afe func(tss []*timeseries) []*timeseries) aggrFunc {
 	}
 }
 
-func removeGroupTags(metricName *storage.MetricName, modifier *modifierExpr) {
+func removeGroupTags(metricName *storage.MetricName, modifier *metricsql.ModifierExpr) {
 	groupOp := strings.ToLower(modifier.Op)
 	switch groupOp {
 	case "", "by":
@@ -82,7 +80,7 @@ func removeGroupTags(metricName *storage.MetricName, modifier *modifierExpr) {
 	}
 }
 
-func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeseries, modifier *modifierExpr, keepOriginal bool) ([]*timeseries, error) {
+func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeseries, modifier *metricsql.ModifierExpr, keepOriginal bool) ([]*timeseries, error) {
 	arg := copyTimeseriesMetricNames(argOrig)
 
 	// Perform grouping.
@@ -184,6 +182,38 @@ func aggrFuncGeomean(tss []*timeseries) []*timeseries {
 	return tss[:1]
 }
 
+func aggrFuncHistogram(tss []*timeseries) []*timeseries {
+	var h metrics.Histogram
+	m := make(map[string]*timeseries)
+	for i := range tss[0].Values {
+		h.Reset()
+		for _, ts := range tss {
+			v := ts.Values[i]
+			h.Update(v)
+		}
+		h.VisitNonZeroBuckets(func(vmrange string, count uint64) {
+			ts := m[vmrange]
+			if ts == nil {
+				ts = &timeseries{}
+				ts.CopyFromShallowTimestamps(tss[0])
+				ts.MetricName.RemoveTag("vmrange")
+				ts.MetricName.AddTag("vmrange", vmrange)
+				values := ts.Values
+				for k := range values {
+					values[k] = 0
+				}
+				m[vmrange] = ts
+			}
+			ts.Values[i] = float64(count)
+		})
+	}
+	rvs := make([]*timeseries, 0, len(m))
+	for _, ts := range m {
+		rvs = append(rvs, ts)
+	}
+	return vmrangeBucketsToLE(rvs)
+}
+
 func aggrFuncMin(tss []*timeseries) []*timeseries {
 	if len(tss) == 1 {
 		// Fast path - nothing to min.
@@ -353,6 +383,25 @@ func aggrFuncCountValues(afa *aggrFuncArg) ([]*timeseries, error) {
 	if err != nil {
 		return nil, err
 	}
+
+	// Remove dstLabel from grouping like Prometheus does.
+	modifier := &afa.ae.Modifier
+	switch strings.ToLower(modifier.Op) {
+	case "without":
+		modifier.Args = append(modifier.Args, dstLabel)
+	case "by":
+		dstArgs := modifier.Args[:0]
+		for _, arg := range modifier.Args {
+			if arg == dstLabel {
+				continue
+			}
+			dstArgs = append(dstArgs, arg)
+		}
+		modifier.Args = dstArgs
+	default:
+		// Do nothing
+	}
+
 	afe := func(tss []*timeseries) []*timeseries {
 		m := make(map[float64]bool)
 		for _, ts := range tss {
@@ -406,37 +455,138 @@ func newAggrFuncTopK(isReverse bool) aggrFunc {
 			return nil, err
 		}
 		afe := func(tss []*timeseries) []*timeseries {
-			rvs := tss
-			for n := range rvs[0].Values {
-				sort.Slice(rvs, func(i, j int) bool {
-					a := rvs[i].Values[n]
-					b := rvs[j].Values[n]
-					cmp := lessWithNaNs(a, b)
+			for n := range tss[0].Values {
+				sort.Slice(tss, func(i, j int) bool {
+					a := tss[i].Values[n]
+					b := tss[j].Values[n]
 					if isReverse {
-						cmp = !cmp
+						a, b = b, a
 					}
-					return cmp
+					return lessWithNaNs(a, b)
 				})
-				if math.IsNaN(ks[n]) {
-					ks[n] = 0
-				}
-				k := int(ks[n])
-				if k < 0 {
-					k = 0
-				}
-				if k > len(rvs) {
-					k = len(rvs)
-				}
-				for _, ts := range rvs[:len(rvs)-k] {
-					ts.Values[n] = nan
-				}
+				fillNaNsAtIdx(n, ks[n], tss)
 			}
-			return removeNaNs(rvs)
+			return removeNaNs(tss)
 		}
 		return aggrFuncExt(afe, args[1], &afa.ae.Modifier, true)
 	}
 }
 
+type tsWithValue struct {
+	ts    *timeseries
+	value float64
+}
+
+func newAggrFuncRangeTopK(f func(values []float64) float64, isReverse bool) aggrFunc {
+	return func(afa *aggrFuncArg) ([]*timeseries, error) {
+		args := afa.args
+		if err := expectTransformArgsNum(args, 2); err != nil {
+			return nil, err
+		}
+		ks, err := getScalar(args[0], 0)
+		if err != nil {
+			return nil, err
+		}
+		afe := func(tss []*timeseries) []*timeseries {
+			maxs := make([]tsWithValue, len(tss))
+			for i, ts := range tss {
+				value := f(ts.Values)
+				maxs[i] = tsWithValue{
+					ts:    ts,
+					value: value,
+				}
+			}
+			sort.Slice(maxs, func(i, j int) bool {
+				a := maxs[i].value
+				b := maxs[j].value
+				if isReverse {
+					a, b = b, a
+				}
+				return lessWithNaNs(a, b)
+			})
+			for i := range maxs {
+				tss[i] = maxs[i].ts
+			}
+			for i, k := range ks {
+				fillNaNsAtIdx(i, k, tss)
+			}
+			return removeNaNs(tss)
+		}
+		return aggrFuncExt(afe, args[1], &afa.ae.Modifier, true)
+	}
+}
+
+func fillNaNsAtIdx(idx int, k float64, tss []*timeseries) {
+	if math.IsNaN(k) {
+		k = 0
+	}
+	kn := int(k)
+	if kn < 0 {
+		kn = 0
+	}
+	if kn > len(tss) {
+		kn = len(tss)
+	}
+	for _, ts := range tss[:len(tss)-kn] {
+		ts.Values[idx] = nan
+	}
+}
+
+func minValue(values []float64) float64 {
+	if len(values) == 0 {
+		return nan
+	}
+	min := values[0]
+	for _, v := range values[1:] {
+		if v < min {
+			min = v
+		}
+	}
+	return min
+}
+
+func maxValue(values []float64) float64 {
+	if len(values) == 0 {
+		return nan
+	}
+	max := values[0]
+	for _, v := range values[1:] {
+		if v > max {
+			max = v
+		}
+	}
+	return max
+}
+
+func avgValue(values []float64) float64 {
+	sum := float64(0)
+	count := 0
+	for _, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		count++
+		sum += v
+	}
+	if count == 0 {
+		return nan
+	}
+	return sum / float64(count)
+}
+
+func medianValue(values []float64) float64 {
+	h := histogram.GetFast()
+	for _, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		h.Update(v)
+	}
+	value := h.Quantile(0.5)
+	histogram.PutFast(h)
+	return value
+}
+
 func aggrFuncLimitK(afa *aggrFuncArg) ([]*timeseries, error) {
 	args := afa.args
 	if err := expectTransformArgsNum(args, 2); err != nil {

app/vmselect/promql/aggr_incremental.go

@@ -4,10 +4,12 @@ import (
 	"math"
 	"strings"
 	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 )
 
 // callbacks for optimized incremental calculations for aggregate functions
-// over rollups over metricExpr.
+// over rollups over metricsql.MetricExpr.
 //
 // These calculations save RAM for aggregates over big number of time series.
 var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
@@ -49,7 +51,7 @@ var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
 }
 
 type incrementalAggrFuncContext struct {
-	ae *aggrFuncExpr
+	ae *metricsql.AggrFuncExpr
 
 	mLock sync.Mutex
 	m     map[uint]map[string]*incrementalAggrContext
@@ -57,7 +59,7 @@ type incrementalAggrFuncContext struct {
 	callbacks *incrementalAggrFuncCallbacks
 }
 
-func newIncrementalAggrFuncContext(ae *aggrFuncExpr, callbacks *incrementalAggrFuncCallbacks) *incrementalAggrFuncContext {
+func newIncrementalAggrFuncContext(ae *metricsql.AggrFuncExpr, callbacks *incrementalAggrFuncCallbacks) *incrementalAggrFuncContext {
 	return &incrementalAggrFuncContext{
 		ae:        ae,
 		m:         make(map[uint]map[string]*incrementalAggrContext),

app/vmselect/promql/aggr_incremental_test.go

@@ -7,6 +7,8 @@ import (
 	"runtime"
 	"sync"
 	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 )
 
 func TestIncrementalAggr(t *testing.T) {
@@ -42,7 +44,7 @@ func TestIncrementalAggr(t *testing.T) {
 	f := func(name string, valuesExpected []float64) {
 		t.Helper()
 		callbacks := getIncrementalAggrFuncCallbacks(name)
-		ae := &aggrFuncExpr{
+		ae := &metricsql.AggrFuncExpr{
 			Name: name,
 		}
 		tssExpected := []*timeseries{{
@@ -179,7 +181,8 @@ func compareValues(vs1, vs2 []float64) error {
 			}
 			continue
 		}
-		if v1 != v2 {
+		eps := math.Abs(v1 - v2)
+		if eps > 1e-14 {
 			return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
 		}
 	}

app/vmselect/promql/arch.go

@@ -0,0 +1,5 @@
+package promql
+
+import "unsafe"
+
+const maxByteSliceLen = 1<<(31+9*(unsafe.Sizeof(int(0))/8)) - 1

app/vmselect/promql/arch_amd64.go

@@ -1,3 +0,0 @@
-package promql
-
-const maxByteSliceLen = 1 << 40

app/vmselect/promql/arch_arm.go

@@ -1,3 +0,0 @@
-package promql
-
-const maxByteSliceLen = 1<<31 - 1

app/vmselect/promql/arch_arm64.go

@@ -1,3 +0,0 @@
-package promql
-
-const maxByteSliceLen = 1 << 40

app/vmselect/promql/binary_op.go

@@ -6,24 +6,26 @@ import (
 	"strings"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql/binaryop"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
 
 var binaryOpFuncs = map[string]binaryOpFunc{
-	"+": newBinaryOpArithFunc(binaryOpPlus),
-	"-": newBinaryOpArithFunc(binaryOpMinus),
-	"*": newBinaryOpArithFunc(binaryOpMul),
-	"/": newBinaryOpArithFunc(binaryOpDiv),
-	"%": newBinaryOpArithFunc(binaryOpMod),
-	"^": newBinaryOpArithFunc(binaryOpPow),
+	"+": newBinaryOpArithFunc(binaryop.Plus),
+	"-": newBinaryOpArithFunc(binaryop.Minus),
+	"*": newBinaryOpArithFunc(binaryop.Mul),
+	"/": newBinaryOpArithFunc(binaryop.Div),
+	"%": newBinaryOpArithFunc(binaryop.Mod),
+	"^": newBinaryOpArithFunc(binaryop.Pow),
 
 	// cmp ops
-	"==": newBinaryOpCmpFunc(binaryOpEq),
-	"!=": newBinaryOpCmpFunc(binaryOpNeq),
-	">":  newBinaryOpCmpFunc(binaryOpGt),
-	"<":  newBinaryOpCmpFunc(binaryOpLt),
-	">=": newBinaryOpCmpFunc(binaryOpGte),
-	"<=": newBinaryOpCmpFunc(binaryOpLte),
+	"==": newBinaryOpCmpFunc(binaryop.Eq),
+	"!=": newBinaryOpCmpFunc(binaryop.Neq),
+	">":  newBinaryOpCmpFunc(binaryop.Gt),
+	"<":  newBinaryOpCmpFunc(binaryop.Lt),
+	">=": newBinaryOpCmpFunc(binaryop.Gte),
+	"<=": newBinaryOpCmpFunc(binaryop.Lte),
 
 	// logical set ops
 	"and":    binaryOpAnd,
@@ -31,38 +33,9 @@ var binaryOpFuncs = map[string]binaryOpFunc{
 	"unless": binaryOpUnless,
 
 	// New op
-	"if":      newBinaryOpArithFunc(binaryOpIf),
-	"ifnot":   newBinaryOpArithFunc(binaryOpIfnot),
-	"default": newBinaryOpArithFunc(binaryOpDefault),
-}
-
-var binaryOpPriorities = map[string]int{
-	"default": -1,
-
-	"if":    0,
-	"ifnot": 0,
-
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#binary-operator-precedence
-	"or": 1,
-
-	"and":    2,
-	"unless": 2,
-
-	"==": 3,
-	"!=": 3,
-	"<":  3,
-	">":  3,
-	"<=": 3,
-	">=": 3,
-
-	"+": 4,
-	"-": 4,
-
-	"*": 5,
-	"/": 5,
-	"%": 5,
-
-	"^": 6,
+	"if":      newBinaryOpArithFunc(binaryop.If),
+	"ifnot":   newBinaryOpArithFunc(binaryop.Ifnot),
+	"default": newBinaryOpArithFunc(binaryop.Default),
 }
 
 func getBinaryOpFunc(op string) binaryOpFunc {
@@ -70,144 +43,8 @@ func getBinaryOpFunc(op string) binaryOpFunc {
 	return binaryOpFuncs[op]
 }
 
-func isBinaryOp(op string) bool {
-	return getBinaryOpFunc(op) != nil
-}
-
-func binaryOpPriority(op string) int {
-	op = strings.ToLower(op)
-	return binaryOpPriorities[op]
-}
-
-func scanBinaryOpPrefix(s string) int {
-	n := 0
-	for op := range binaryOpFuncs {
-		if len(s) < len(op) {
-			continue
-		}
-		ss := strings.ToLower(s[:len(op)])
-		if ss == op && len(op) > n {
-			n = len(op)
-		}
-	}
-	return n
-}
-
-func isRightAssociativeBinaryOp(op string) bool {
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#binary-operator-precedence
-	return op == "^"
-}
-
-func isBinaryOpGroupModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#vector-matching
-	case "on", "ignoring":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpJoinModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	case "group_left", "group_right":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpBoolModifier(s string) bool {
-	s = strings.ToLower(s)
-	return s == "bool"
-}
-
-func isBinaryOpCmp(op string) bool {
-	switch op {
-	case "==", "!=", ">", "<", ">=", "<=":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpLogicalSet(op string) bool {
-	op = strings.ToLower(op)
-	switch op {
-	case "and", "or", "unless":
-		return true
-	default:
-		return false
-	}
-}
-
-func binaryOpConstants(op string, left, right float64, isBool bool) float64 {
-	if isBinaryOpCmp(op) {
-		evalCmp := func(cf func(left, right float64) bool) float64 {
-			if isBool {
-				if cf(left, right) {
-					return 1
-				}
-				return 0
-			}
-			if cf(left, right) {
-				return left
-			}
-			return nan
-		}
-		switch op {
-		case "==":
-			left = evalCmp(binaryOpEq)
-		case "!=":
-			left = evalCmp(binaryOpNeq)
-		case ">":
-			left = evalCmp(binaryOpGt)
-		case "<":
-			left = evalCmp(binaryOpLt)
-		case ">=":
-			left = evalCmp(binaryOpGte)
-		case "<=":
-			left = evalCmp(binaryOpLte)
-		default:
-			logger.Panicf("BUG: unexpected comparison binaryOp: %q", op)
-		}
-	} else {
-		switch op {
-		case "+":
-			left = binaryOpPlus(left, right)
-		case "-":
-			left = binaryOpMinus(left, right)
-		case "*":
-			left = binaryOpMul(left, right)
-		case "/":
-			left = binaryOpDiv(left, right)
-		case "%":
-			left = binaryOpMod(left, right)
-		case "^":
-			left = binaryOpPow(left, right)
-		case "and":
-			// Nothing to do
-		case "or":
-			// Nothing to do
-		case "unless":
-			left = nan
-		case "default":
-			left = binaryOpDefault(left, right)
-		case "if":
-			left = binaryOpIf(left, right)
-		case "ifnot":
-			left = binaryOpIfnot(left, right)
-		default:
-			logger.Panicf("BUG: unexpected non-comparison binaryOp: %q", op)
-		}
-	}
-	return left
-}
-
 type binaryOpFuncArg struct {
-	be    *binaryOpExpr
+	be    *metricsql.BinaryOpExpr
 	left  []*timeseries
 	right []*timeseries
 }
@@ -267,7 +104,7 @@ func newBinaryOpFunc(bf func(left, right float64, isBool bool) float64) binaryOp
 	}
 }
 
-func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeseries, []*timeseries, []*timeseries, error) {
+func adjustBinaryOpTags(be *metricsql.BinaryOpExpr, left, right []*timeseries) ([]*timeseries, []*timeseries, []*timeseries, error) {
 	if len(be.GroupModifier.Op) == 0 && len(be.JoinModifier.Op) == 0 {
 		if isScalar(left) {
 			// Fast path: `scalar op vector`
@@ -292,24 +129,14 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	}
 
 	// Slow path: `vector op vector` or `a op {on|ignoring} {group_left|group_right} b`
-	ensureOneX := func(side string, tss []*timeseries) error {
-		if len(tss) == 0 {
-			logger.Panicf("BUG: tss must contain at least one value")
-		}
-		if len(tss) == 1 {
-			return nil
-		}
-		if mergeNonOverlappingTimeseries(tss) {
-			return nil
-		}
-		return fmt.Errorf(`duplicate timeseries on the %s side of %s %s: %s and %s`, side, be.Op, be.GroupModifier.AppendString(nil),
-			stringMetricTags(&tss[0].MetricName), stringMetricTags(&tss[1].MetricName))
-	}
-
 	var rvsLeft, rvsRight []*timeseries
 	mLeft, mRight := createTimeseriesMapByTagSet(be, left, right)
 	joinOp := strings.ToLower(be.JoinModifier.Op)
-	joinTags := be.JoinModifier.Args
+	groupOp := strings.ToLower(be.GroupModifier.Op)
+	if len(groupOp) == 0 {
+		groupOp = "ignoring"
+	}
+	groupTags := be.GroupModifier.Args
 	for k, tssLeft := range mLeft {
 		tssRight := mRight[k]
 		if len(tssRight) == 0 {
@@ -317,37 +144,38 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 		}
 		switch joinOp {
 		case "group_left":
-			if err := ensureOneX("right", tssRight); err != nil {
+			var err error
+			rvsLeft, rvsRight, err = groupJoin("right", be, rvsLeft, rvsRight, tssLeft, tssRight)
+			if err != nil {
 				return nil, nil, nil, err
 			}
-			src := tssRight[0]
-			for _, ts := range tssLeft {
-				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
-				rvsLeft = append(rvsLeft, ts)
-				rvsRight = append(rvsRight, src)
-			}
 		case "group_right":
-			if err := ensureOneX("left", tssLeft); err != nil {
+			var err error
+			rvsRight, rvsLeft, err = groupJoin("left", be, rvsRight, rvsLeft, tssRight, tssLeft)
+			if err != nil {
 				return nil, nil, nil, err
 			}
-			src := tssLeft[0]
-			for _, ts := range tssRight {
-				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
-				rvsLeft = append(rvsLeft, src)
-				rvsRight = append(rvsRight, ts)
-			}
 		case "":
-			if err := ensureOneX("left", tssLeft); err != nil {
+			if err := ensureSingleTimeseries("left", be, tssLeft); err != nil {
 				return nil, nil, nil, err
 			}
-			if err := ensureOneX("right", tssRight); err != nil {
+			if err := ensureSingleTimeseries("right", be, tssRight); err != nil {
 				return nil, nil, nil, err
 			}
-			resetMetricGroupIfRequired(be, tssLeft[0])
-			rvsLeft = append(rvsLeft, tssLeft[0])
+			tsLeft := tssLeft[0]
+			resetMetricGroupIfRequired(be, tsLeft)
+			switch groupOp {
+			case "on":
+				tsLeft.MetricName.RemoveTagsOn(groupTags)
+			case "ignoring":
+				tsLeft.MetricName.RemoveTagsIgnoring(groupTags)
+			default:
+				logger.Panicf("BUG: unexpected binary op modifier %q", groupOp)
+			}
+			rvsLeft = append(rvsLeft, tsLeft)
 			rvsRight = append(rvsRight, tssRight[0])
 		default:
-			return nil, nil, nil, fmt.Errorf(`unexpected join modifier %q`, joinOp)
+			logger.Panicf("BUG: unexpected join modifier %q", joinOp)
 		}
 	}
 	dst := rvsLeft
@@ -357,8 +185,92 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	return rvsLeft, rvsRight, dst, nil
 }
 
-func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
-	if isBinaryOpCmp(be.Op) && !be.Bool {
+func ensureSingleTimeseries(side string, be *metricsql.BinaryOpExpr, tss []*timeseries) error {
+	if len(tss) == 0 {
+		logger.Panicf("BUG: tss must contain at least one value")
+	}
+	for len(tss) > 1 {
+		if !mergeNonOverlappingTimeseries(tss[0], tss[len(tss)-1]) {
+			return fmt.Errorf(`duplicate time series on the %s side of %s %s: %s and %s`, side, be.Op, be.GroupModifier.AppendString(nil),
+				stringMetricTags(&tss[0].MetricName), stringMetricTags(&tss[len(tss)-1].MetricName))
+		}
+		tss = tss[:len(tss)-1]
+	}
+	return nil
+}
+
+func groupJoin(singleTimeseriesSide string, be *metricsql.BinaryOpExpr, rvsLeft, rvsRight, tssLeft, tssRight []*timeseries) ([]*timeseries, []*timeseries, error) {
+	joinTags := be.JoinModifier.Args
+	var m map[string]*timeseries
+	for _, tsLeft := range tssLeft {
+		resetMetricGroupIfRequired(be, tsLeft)
+		if len(tssRight) == 1 {
+			// Easy case - right part contains only a single matching time series.
+			tsLeft.MetricName.AddMissingTags(joinTags, &tssRight[0].MetricName)
+			rvsLeft = append(rvsLeft, tsLeft)
+			rvsRight = append(rvsRight, tssRight[0])
+			continue
+		}
+
+		// Hard case - right part contains multiple matching time series.
+		// Verify it doesn't result in duplicate MetricName values after adding missing tags.
+		if m == nil {
+			m = make(map[string]*timeseries, len(tssRight))
+		} else {
+			for k := range m {
+				delete(m, k)
+			}
+		}
+		bb := bbPool.Get()
+		for _, tsRight := range tssRight {
+			var tsCopy timeseries
+			tsCopy.CopyFromShallowTimestamps(tsLeft)
+			tsCopy.MetricName.AddMissingTags(joinTags, &tsRight.MetricName)
+			bb.B = marshalMetricTagsSorted(bb.B[:0], &tsCopy.MetricName)
+			if tsExisting := m[string(bb.B)]; tsExisting != nil {
+				// Try merging tsExisting with tsRight if they don't overlap.
+				if mergeNonOverlappingTimeseries(tsExisting, tsRight) {
+					continue
+				}
+				return nil, nil, fmt.Errorf("duplicate time series on the %s side of `%s %s %s`: %s and %s",
+					singleTimeseriesSide, be.Op, be.GroupModifier.AppendString(nil), be.JoinModifier.AppendString(nil),
+					stringMetricTags(&tsExisting.MetricName), stringMetricTags(&tsRight.MetricName))
+			}
+			m[string(bb.B)] = tsRight
+			rvsLeft = append(rvsLeft, &tsCopy)
+			rvsRight = append(rvsRight, tsRight)
+		}
+		bbPool.Put(bb)
+	}
+	return rvsLeft, rvsRight, nil
+}
+
+func mergeNonOverlappingTimeseries(dst, src *timeseries) bool {
+	// Verify whether the time series can be merged.
+	srcValues := src.Values
+	dstValues := dst.Values
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if math.IsNaN(v) {
+			continue
+		}
+		if !math.IsNaN(dstValues[i]) {
+			return false
+		}
+	}
+
+	// Time series can be merged. Merge them.
+	for i, v := range srcValues {
+		if math.IsNaN(v) {
+			continue
+		}
+		dstValues[i] = v
+	}
+	return true
+}
+
+func resetMetricGroupIfRequired(be *metricsql.BinaryOpExpr, ts *timeseries) {
+	if metricsql.IsBinaryOpCmp(be.Op) && !be.Bool {
 		// Do not reset MetricGroup for non-boolean `compare` binary ops like Prometheus does.
 		return
 	}
@@ -370,75 +282,6 @@ func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
 	ts.MetricName.ResetMetricGroup()
 }
 
-func binaryOpPlus(left, right float64) float64 {
-	return left + right
-}
-
-func binaryOpMinus(left, right float64) float64 {
-	return left - right
-}
-
-func binaryOpMul(left, right float64) float64 {
-	return left * right
-}
-
-func binaryOpDiv(left, right float64) float64 {
-	return left / right
-}
-
-func binaryOpMod(left, right float64) float64 {
-	return math.Mod(left, right)
-}
-
-func binaryOpPow(left, right float64) float64 {
-	return math.Pow(left, right)
-}
-
-func binaryOpDefault(left, right float64) float64 {
-	if math.IsNaN(left) {
-		return right
-	}
-	return left
-}
-
-func binaryOpIf(left, right float64) float64 {
-	if math.IsNaN(right) {
-		return nan
-	}
-	return left
-}
-
-func binaryOpIfnot(left, right float64) float64 {
-	if math.IsNaN(right) {
-		return left
-	}
-	return nan
-}
-
-func binaryOpEq(left, right float64) bool {
-	return left == right
-}
-
-func binaryOpNeq(left, right float64) bool {
-	return left != right
-}
-
-func binaryOpGt(left, right float64) bool {
-	return left > right
-}
-
-func binaryOpLt(left, right float64) bool {
-	return left < right
-}
-
-func binaryOpGte(left, right float64) bool {
-	return left >= right
-}
-
-func binaryOpLte(left, right float64) bool {
-	return left <= right
-}
-
 func binaryOpAnd(bfa *binaryOpFuncArg) ([]*timeseries, error) {
 	mLeft, mRight := createTimeseriesMapByTagSet(bfa.be, bfa.left, bfa.right)
 	var rvs []*timeseries
@@ -475,7 +318,7 @@ func binaryOpUnless(bfa *binaryOpFuncArg) ([]*timeseries, error) {
 	return rvs, nil
 }
 
-func createTimeseriesMapByTagSet(be *binaryOpExpr, left, right []*timeseries) (map[string][]*timeseries, map[string][]*timeseries) {
+func createTimeseriesMapByTagSet(be *metricsql.BinaryOpExpr, left, right []*timeseries) (map[string][]*timeseries, map[string][]*timeseries) {
 	groupTags := be.GroupModifier.Args
 	groupOp := strings.ToLower(be.GroupModifier.Op)
 	if len(groupOp) == 0 {
@@ -518,26 +361,3 @@ func isScalar(arg []*timeseries) bool {
 	}
 	return len(mn.Tags) == 0
 }
-
-func mergeNonOverlappingTimeseries(tss []*timeseries) bool {
-	if len(tss) < 2 {
-		logger.Panicf("BUG: expecting at least two timeseries. Got %d", len(tss))
-	}
-
-	// Check whether time series in tss overlap.
-	var dst timeseries
-	dst.CopyFromShallowTimestamps(tss[0])
-	dstValues := dst.Values
-	for _, ts := range tss[1:] {
-		for i, value := range ts.Values {
-			if math.IsNaN(dstValues[i]) {
-				dstValues[i] = value
-			} else if !math.IsNaN(value) {
-				// Time series overlap.
-				return false
-			}
-		}
-	}
-	tss[0].CopyFromShallowTimestamps(&dst)
-	return true
-}

app/vmselect/promql/eval.go

@@ -11,6 +11,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -57,6 +58,14 @@ func AdjustStartEnd(start, end, step int64) (int64, int64) {
 	if adjust > 0 {
 		end += step - adjust
 	}
+
+	// Make sure that the new number of points is the same as the initial number of points.
+	newPoints := (end-start)/step + 1
+	for newPoints > points {
+		end -= step
+		newPoints--
+	}
+
 	return start, end
 }
 
@@ -70,6 +79,9 @@ type EvalConfig struct {
 
 	MayCache bool
 
+	// LookbackDelta is analog to `-query.lookback-delta` from Prometheus.
+	LookbackDelta int64
+
 	timestamps     []int64
 	timestampsOnce sync.Once
 }
@@ -82,6 +94,7 @@ func newEvalConfig(src *EvalConfig) *EvalConfig {
 	ec.Step = src.Step
 	ec.Deadline = src.Deadline
 	ec.MayCache = src.MayCache
+	ec.LookbackDelta = src.LookbackDelta
 
 	// do not copy src.timestamps - they must be generated again.
 	return &ec
@@ -140,25 +153,25 @@ func getTimestamps(start, end, step int64) []int64 {
 	return timestamps
 }
 
-func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
-	if me, ok := e.(*metricExpr); ok {
-		re := &rollupExpr{
+func evalExpr(ec *EvalConfig, e metricsql.Expr) ([]*timeseries, error) {
+	if me, ok := e.(*metricsql.MetricExpr); ok {
+		re := &metricsql.RollupExpr{
 			Expr: me,
 		}
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, me.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if re, ok := e.(*rollupExpr); ok {
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
+	if re, ok := e.(*metricsql.RollupExpr); ok {
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, re.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if fe, ok := e.(*funcExpr); ok {
+	if fe, ok := e.(*metricsql.FuncExpr); ok {
 		nrf := getRollupFunc(fe.Name)
 		if nrf == nil {
 			args, err := evalExprs(ec, fe.Args)
@@ -188,17 +201,17 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		if err != nil {
 			return nil, err
 		}
-		rv, err := evalRollupFunc(ec, fe.Name, rf, re, nil)
+		rv, err := evalRollupFunc(ec, fe.Name, rf, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, fe.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if ae, ok := e.(*aggrFuncExpr); ok {
+	if ae, ok := e.(*metricsql.AggrFuncExpr); ok {
 		if callbacks := getIncrementalAggrFuncCallbacks(ae.Name); callbacks != nil {
 			fe, nrf := tryGetArgRollupFuncWithMetricExpr(ae)
 			if fe != nil {
-				// There is an optimized path for calculating aggrFuncExpr over rollupFunc over metricExpr.
+				// There is an optimized path for calculating metricsql.AggrFuncExpr over rollupFunc over metricsql.MetricExpr.
 				// The optimized path saves RAM for aggregates over big number of time series.
 				args, re, err := evalRollupFuncArgs(ec, fe)
 				if err != nil {
@@ -209,7 +222,7 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 					return nil, err
 				}
 				iafc := newIncrementalAggrFuncContext(ae, callbacks)
-				return evalRollupFunc(ec, fe.Name, rf, re, iafc)
+				return evalRollupFunc(ec, fe.Name, rf, e, re, iafc)
 			}
 		}
 		args, err := evalExprs(ec, ae.Args)
@@ -231,7 +244,7 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		}
 		return rv, nil
 	}
-	if be, ok := e.(*binaryOpExpr); ok {
+	if be, ok := e.(*metricsql.BinaryOpExpr); ok {
 		left, err := evalExpr(ec, be.Left)
 		if err != nil {
 			return nil, err
@@ -255,18 +268,18 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		}
 		return rv, nil
 	}
-	if ne, ok := e.(*numberExpr); ok {
+	if ne, ok := e.(*metricsql.NumberExpr); ok {
 		rv := evalNumber(ec, ne.N)
 		return rv, nil
 	}
-	if se, ok := e.(*stringExpr); ok {
+	if se, ok := e.(*metricsql.StringExpr); ok {
 		rv := evalString(ec, se.S)
 		return rv, nil
 	}
 	return nil, fmt.Errorf("unexpected expression %q", e.AppendString(nil))
 }
 
-func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFunc) {
+func tryGetArgRollupFuncWithMetricExpr(ae *metricsql.AggrFuncExpr) (*metricsql.FuncExpr, newRollupFunc) {
 	if len(ae.Args) != 1 {
 		return nil, nil
 	}
@@ -277,31 +290,31 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 	// - rollupFunc(metricExpr)
 	// - rollupFunc(metricExpr[d])
 
-	if me, ok := e.(*metricExpr); ok {
+	if me, ok := e.(*metricsql.MetricExpr); ok {
 		// e = metricExpr
 		if me.IsEmpty() {
 			return nil, nil
 		}
-		fe := &funcExpr{
+		fe := &metricsql.FuncExpr{
 			Name: "default_rollup",
-			Args: []expr{me},
+			Args: []metricsql.Expr{me},
 		}
 		nrf := getRollupFunc(fe.Name)
 		return fe, nrf
 	}
-	if re, ok := e.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+	if re, ok := e.(*metricsql.RollupExpr); ok {
+		if me, ok := re.Expr.(*metricsql.MetricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
-		// e = rollupExpr(metricExpr)
-		fe := &funcExpr{
+		// e = metricExpr[d]
+		fe := &metricsql.FuncExpr{
 			Name: "default_rollup",
-			Args: []expr{re},
+			Args: []metricsql.Expr{re},
 		}
 		nrf := getRollupFunc(fe.Name)
 		return fe, nrf
 	}
-	fe, ok := e.(*funcExpr)
+	fe, ok := e.(*metricsql.FuncExpr)
 	if !ok {
 		return nil, nil
 	}
@@ -311,25 +324,27 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 	}
 	rollupArgIdx := getRollupArgIdx(fe.Name)
 	arg := fe.Args[rollupArgIdx]
-	if me, ok := arg.(*metricExpr); ok {
+	if me, ok := arg.(*metricsql.MetricExpr); ok {
 		if me.IsEmpty() {
 			return nil, nil
 		}
-		return &funcExpr{
+		// e = rollupFunc(metricExpr)
+		return &metricsql.FuncExpr{
 			Name: fe.Name,
-			Args: []expr{me},
+			Args: []metricsql.Expr{me},
 		}, nrf
 	}
-	if re, ok := arg.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+	if re, ok := arg.(*metricsql.RollupExpr); ok {
+		if me, ok := re.Expr.(*metricsql.MetricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
+		// e = rollupFunc(metricExpr[d])
 		return fe, nrf
 	}
 	return nil, nil
 }
 
-func evalExprs(ec *EvalConfig, es []expr) ([][]*timeseries, error) {
+func evalExprs(ec *EvalConfig, es []metricsql.Expr) ([][]*timeseries, error) {
 	var rvs [][]*timeseries
 	for _, e := range es {
 		rv, err := evalExpr(ec, e)
@@ -341,9 +356,12 @@ func evalExprs(ec *EvalConfig, es []expr) ([][]*timeseries, error) {
 	return rvs, nil
 }
 
-func evalRollupFuncArgs(ec *EvalConfig, fe *funcExpr) ([]interface{}, *rollupExpr, error) {
-	var re *rollupExpr
+func evalRollupFuncArgs(ec *EvalConfig, fe *metricsql.FuncExpr) ([]interface{}, *metricsql.RollupExpr, error) {
+	var re *metricsql.RollupExpr
 	rollupArgIdx := getRollupArgIdx(fe.Name)
+	if len(fe.Args) <= rollupArgIdx {
+		return nil, nil, fmt.Errorf("expecting at least %d args to %q; got %d args; expr: %q", rollupArgIdx, fe.Name, len(fe.Args), fe.AppendString(nil))
+	}
 	args := make([]interface{}, len(fe.Args))
 	for i, arg := range fe.Args {
 		if i == rollupArgIdx {
@@ -360,57 +378,62 @@ func evalRollupFuncArgs(ec *EvalConfig, fe *funcExpr) ([]interface{}, *rollupExp
 	return args, re, nil
 }
 
-func getRollupExprArg(arg expr) *rollupExpr {
-	re, ok := arg.(*rollupExpr)
+func getRollupExprArg(arg metricsql.Expr) *metricsql.RollupExpr {
+	re, ok := arg.(*metricsql.RollupExpr)
 	if !ok {
-		// Wrap non-rollup arg into rollupExpr.
-		return &rollupExpr{
+		// Wrap non-rollup arg into metricsql.RollupExpr.
+		return &metricsql.RollupExpr{
 			Expr: arg,
 		}
 	}
-	if len(re.Step) == 0 && !re.InheritStep {
-		// Return standard rollup if it doesn't set step.
+	if !re.ForSubquery() {
+		// Return standard rollup if it doesn't contain subquery.
 		return re
 	}
-	me, ok := re.Expr.(*metricExpr)
+	me, ok := re.Expr.(*metricsql.MetricExpr)
 	if !ok {
 		// arg contains subquery.
 		return re
 	}
 	// Convert me[w:step] -> default_rollup(me)[w:step]
 	reNew := *re
-	reNew.Expr = &funcExpr{
+	reNew.Expr = &metricsql.FuncExpr{
 		Name: "default_rollup",
-		Args: []expr{
-			&rollupExpr{Expr: me},
+		Args: []metricsql.Expr{
+			&metricsql.RollupExpr{Expr: me},
 		},
 	}
 	return &reNew
 }
 
-func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr, iafc *incrementalAggrFuncContext) ([]*timeseries, error) {
+func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, expr metricsql.Expr, re *metricsql.RollupExpr, iafc *incrementalAggrFuncContext) ([]*timeseries, error) {
 	ecNew := ec
 	var offset int64
 	if len(re.Offset) > 0 {
 		var err error
-		offset, err = DurationValue(re.Offset, ec.Step)
+		offset, err = metricsql.DurationValue(re.Offset, ec.Step)
 		if err != nil {
 			return nil, err
 		}
 		ecNew = newEvalConfig(ec)
 		ecNew.Start -= offset
 		ecNew.End -= offset
-		ecNew.Start, ecNew.End = AdjustStartEnd(ecNew.Start, ecNew.End, ecNew.Step)
+		if ecNew.MayCache {
+			start, end := AdjustStartEnd(ecNew.Start, ecNew.End, ecNew.Step)
+			offset += ecNew.Start - start
+			ecNew.Start = start
+			ecNew.End = end
+		}
 	}
 	var rvs []*timeseries
 	var err error
-	if me, ok := re.Expr.(*metricExpr); ok {
-		rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, me, iafc, re.Window)
+	if me, ok := re.Expr.(*metricsql.MetricExpr); ok {
+		rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, expr, me, iafc, re.Window)
 	} else {
 		if iafc != nil {
 			logger.Panicf("BUG: iafc must be nil for rollup %q over subquery %q", name, re.AppendString(nil))
 		}
-		rvs, err = evalRollupFuncWithSubquery(ecNew, name, rf, re)
+		rvs, err = evalRollupFuncWithSubquery(ecNew, name, rf, expr, re)
 	}
 	if err != nil {
 		return nil, err
@@ -429,12 +452,12 @@ func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr,
 	return rvs, nil
 }
 
-func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr) ([]*timeseries, error) {
-	// Do not use rollupResultCacheV here, since it works only with metricExpr.
+func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, expr metricsql.Expr, re *metricsql.RollupExpr) ([]*timeseries, error) {
+	// TODO: determine whether to use rollupResultCacheV here.
 	var step int64
 	if len(re.Step) > 0 {
 		var err error
-		step, err = DurationValue(re.Step, ec.Step)
+		step, err = metricsql.PositiveDurationValue(re.Step, ec.Step)
 		if err != nil {
 			return nil, err
 		}
@@ -444,7 +467,7 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	var window int64
 	if len(re.Window) > 0 {
 		var err error
-		window, err = DurationValue(re.Window, ec.Step)
+		window, err = metricsql.PositiveDurationValue(re.Window, ec.Step)
 		if err != nil {
 			return nil, err
 		}
@@ -461,9 +484,19 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	if err != nil {
 		return nil, err
 	}
+	if len(tssSQ) == 0 {
+		if name == "absent_over_time" {
+			tss := evalNumber(ec, 1)
+			return tss, nil
+		}
+		return nil, nil
+	}
 
 	sharedTimestamps := getTimestamps(ec.Start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, ec.Start, ec.End, ec.Step, window, sharedTimestamps)
+	preFunc, rcs, err := getRollupConfigs(name, rf, expr, ec.Start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
+	if err != nil {
+		return nil, err
+	}
 	tss := make([]*timeseries, 0, len(tssSQ)*len(rcs))
 	var tssLock sync.Mutex
 	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
@@ -471,6 +504,13 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 		values, timestamps = removeNanValues(values[:0], timestamps[:0], tsSQ.Values, tsSQ.Timestamps)
 		preFunc(values, timestamps)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &tsSQ.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, values, timestamps)
+				tssLock.Lock()
+				tss = tsm.AppendTimeseriesTo(tss)
+				tssLock.Unlock()
+				continue
+			}
 			var ts timeseries
 			doRollupForTimeseries(rc, &ts, &tsSQ.MetricName, values, timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
@@ -538,21 +578,22 @@ var (
 	rollupResultCacheMiss        = metrics.NewCounter(`vm_rollup_result_cache_miss_total`)
 )
 
-func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me *metricExpr, iafc *incrementalAggrFuncContext, windowStr string) ([]*timeseries, error) {
+func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc,
+	expr metricsql.Expr, me *metricsql.MetricExpr, iafc *incrementalAggrFuncContext, windowStr string) ([]*timeseries, error) {
 	if me.IsEmpty() {
 		return evalNumber(ec, nan), nil
 	}
 	var window int64
 	if len(windowStr) > 0 {
 		var err error
-		window, err = DurationValue(windowStr, ec.Step)
+		window, err = metricsql.PositiveDurationValue(windowStr, ec.Step)
 		if err != nil {
 			return nil, err
 		}
 	}
 
 	// Search for partial results in cache.
-	tssCached, start := rollupResultCacheV.Get(name, ec, me, iafc, window)
+	tssCached, start := rollupResultCacheV.Get(ec, expr, window)
 	if start > ec.End {
 		// The result is fully cached.
 		rollupResultCacheFullHits.Inc()
@@ -565,10 +606,11 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	}
 
 	// Fetch the remaining part of the result.
+	tfs := toTagFilters(me.LabelFilters)
 	sq := &storage.SearchQuery{
 		MinTimestamp: start - window - maxSilenceInterval,
 		MaxTimestamp: ec.End + ec.Step,
-		TagFilterss:  [][]storage.TagFilter{me.TagFilters},
+		TagFilterss:  [][]storage.TagFilter{tfs},
 	}
 	rss, err := netstorage.ProcessSearchQuery(sq, true, ec.Deadline)
 	if err != nil {
@@ -577,19 +619,38 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	rssLen := rss.Len()
 	if rssLen == 0 {
 		rss.Cancel()
+		var tss []*timeseries
+		if name == "absent_over_time" {
+			tss = getAbsentTimeseries(ec, me)
+		}
 		// Add missing points until ec.End.
 		// Do not cache the result, since missing points
 		// may be backfilled in the future.
-		tss := mergeTimeseries(tssCached, nil, start, ec)
+		tss = mergeTimeseries(tssCached, tss, start, ec)
 		return tss, nil
 	}
 	sharedTimestamps := getTimestamps(start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, start, ec.End, ec.Step, window, sharedTimestamps)
+	preFunc, rcs, err := getRollupConfigs(name, rf, expr, start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
+	if err != nil {
+		rss.Cancel()
+		return nil, err
+	}
 
 	// Verify timeseries fit available memory after the rollup.
 	// Take into account points from tssCached.
 	pointsPerTimeseries := 1 + (ec.End-ec.Start)/ec.Step
-	rollupPoints := mulNoOverflow(pointsPerTimeseries, int64(rssLen*len(rcs)))
+	timeseriesLen := rssLen
+	if iafc != nil {
+		// Incremental aggregates require hold only GOMAXPROCS timeseries in memory.
+		timeseriesLen = runtime.GOMAXPROCS(-1)
+		if iafc.ae.Modifier.Op != "" {
+			// Increase the number of timeseries for non-empty group list: `aggr() by (something)`,
+			// since each group can have own set of time series in memory.
+			// Estimate the number of such groups is lower than 1000 :)
+			timeseriesLen *= 1000
+		}
+	}
+	rollupPoints := mulNoOverflow(pointsPerTimeseries, int64(timeseriesLen*len(rcs)))
 	rollupMemorySize := mulNoOverflow(rollupPoints, 16)
 	rml := getRollupMemoryLimiter()
 	if !rml.Get(uint64(rollupMemorySize)) {
@@ -605,16 +666,15 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
 	var tss []*timeseries
 	if iafc != nil {
-		tss, err = evalRollupWithIncrementalAggregate(iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+		tss, err = evalRollupWithIncrementalAggregate(name, iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	} else {
-		tss, err = evalRollupNoIncrementalAggregate(rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+		tss, err = evalRollupNoIncrementalAggregate(name, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	}
 	if err != nil {
 		return nil, err
 	}
 	tss = mergeTimeseries(tssCached, tss, start, ec)
-	rollupResultCacheV.Put(name, ec, me, iafc, window, tss)
-
+	rollupResultCacheV.Put(ec, expr, window, tss)
 	return tss, nil
 }
 
@@ -630,13 +690,20 @@ func getRollupMemoryLimiter() *memoryLimiter {
 	return &rollupMemoryLimiter
 }
 
-func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
+func evalRollupWithIncrementalAggregate(name string, iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
 	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		ts := getTimeseries()
 		defer putTimeseries(ts)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &rs.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, rs.Values, rs.Timestamps)
+				for _, ts := range tsm.m {
+					iafc.updateTimeseries(ts, workerID)
+				}
+				continue
+			}
 			ts.Reset()
 			doRollupForTimeseries(rc, ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
 			iafc.updateTimeseries(ts, workerID)
@@ -653,13 +720,20 @@ func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *n
 	return tss, nil
 }
 
-func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConfig,
+func evalRollupNoIncrementalAggregate(name string, rss *netstorage.Results, rcs []*rollupConfig,
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
 	tss := make([]*timeseries, 0, rss.Len()*len(rcs))
 	var tssLock sync.Mutex
 	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &rs.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, rs.Values, rs.Timestamps)
+				tssLock.Lock()
+				tss = tsm.AppendTimeseriesTo(tss)
+				tssLock.Unlock()
+				continue
+			}
 			var ts timeseries
 			doRollupForTimeseries(rc, &ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
@@ -687,60 +761,6 @@ func doRollupForTimeseries(rc *rollupConfig, tsDst *timeseries, mnSrc *storage.M
 	tsDst.denyReuse = true
 }
 
-func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64, sharedTimestamps []int64) (func(values []float64, timestamps []int64), []*rollupConfig) {
-	preFunc := func(values []float64, timestamps []int64) {}
-	if rollupFuncsRemoveCounterResets[name] {
-		preFunc = func(values []float64, timestamps []int64) {
-			removeCounterResets(values)
-		}
-	}
-	newRollupConfig := func(rf rollupFunc, tagValue string) *rollupConfig {
-		return &rollupConfig{
-			TagValue:        tagValue,
-			Func:            rf,
-			Start:           start,
-			End:             end,
-			Step:            step,
-			Window:          window,
-			MayAdjustWindow: rollupFuncsMayAdjustWindow[name],
-			Timestamps:      sharedTimestamps,
-		}
-	}
-	appendRollupConfigs := func(dst []*rollupConfig) []*rollupConfig {
-		dst = append(dst, newRollupConfig(rollupMin, "min"))
-		dst = append(dst, newRollupConfig(rollupMax, "max"))
-		dst = append(dst, newRollupConfig(rollupAvg, "avg"))
-		return dst
-	}
-	var rcs []*rollupConfig
-	switch name {
-	case "rollup":
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_rate", "rollup_deriv":
-		preFuncPrev := preFunc
-		preFunc = func(values []float64, timestamps []int64) {
-			preFuncPrev(values, timestamps)
-			derivValues(values, timestamps)
-		}
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_increase", "rollup_delta":
-		preFuncPrev := preFunc
-		preFunc = func(values []float64, timestamps []int64) {
-			preFuncPrev(values, timestamps)
-			deltaValues(values)
-		}
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_candlestick":
-		rcs = append(rcs, newRollupConfig(rollupFirst, "open"))
-		rcs = append(rcs, newRollupConfig(rollupLast, "close"))
-		rcs = append(rcs, newRollupConfig(rollupMin, "low"))
-		rcs = append(rcs, newRollupConfig(rollupMax, "high"))
-	default:
-		rcs = append(rcs, newRollupConfig(rf, ""))
-	}
-	return preFunc, rcs
-}
-
 var bbPool bytesutil.ByteBufferPool
 
 func evalNumber(ec *EvalConfig, n float64) []*timeseries {
@@ -779,3 +799,23 @@ func mulNoOverflow(a, b int64) int64 {
 	}
 	return a * b
 }
+
+func toTagFilters(lfs []metricsql.LabelFilter) []storage.TagFilter {
+	tfs := make([]storage.TagFilter, len(lfs))
+	for i := range lfs {
+		toTagFilter(&tfs[i], &lfs[i])
+	}
+	return tfs
+}
+
+func toTagFilter(dst *storage.TagFilter, src *metricsql.LabelFilter) {
+	if src.Label != "__name__" {
+		dst.Key = []byte(src.Label)
+	} else {
+		// This is required for storage.Search.
+		dst.Key = nil
+	}
+	dst.Value = []byte(src.Value)
+	dst.IsRegexp = src.IsRegexp
+	dst.IsNegative = src.IsNegative
+}

app/vmselect/promql/exec.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/metrics"
 )
 
@@ -18,17 +19,6 @@ var logSlowQueryDuration = flag.Duration("search.logSlowQueryDuration", 5*time.S
 
 var slowQueries = metrics.NewCounter(`vm_slow_queries_total`)
 
-// ExpandWithExprs expands WITH expressions inside q and returns the resulting
-// PromQL without WITH expressions.
-func ExpandWithExprs(q string) (string, error) {
-	e, err := parsePromQLWithCache(q)
-	if err != nil {
-		return "", err
-	}
-	buf := e.AppendString(nil)
-	return string(buf), nil
-}
-
 // Exec executes q for the given ec.
 func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result, error) {
 	if *logSlowQueryDuration > 0 {
@@ -85,12 +75,12 @@ func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result,
 	return result, err
 }
 
-func maySortResults(e expr, tss []*timeseries) bool {
+func maySortResults(e metricsql.Expr, tss []*timeseries) bool {
 	if len(tss) > 100 {
 		// There is no sense in sorting a lot of results
 		return false
 	}
-	fe, ok := e.(*funcExpr)
+	fe, ok := e.(*metricsql.FuncExpr)
 	if !ok {
 		return true
 	}
@@ -105,14 +95,14 @@ func maySortResults(e expr, tss []*timeseries) bool {
 func timeseriesToResult(tss []*timeseries, maySort bool) ([]netstorage.Result, error) {
 	tss = removeNaNs(tss)
 	result := make([]netstorage.Result, len(tss))
-	m := make(map[string]bool)
+	m := make(map[string]struct{}, len(tss))
 	bb := bbPool.Get()
 	for i, ts := range tss {
 		bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
-		if m[string(bb.B)] {
-			return nil, fmt.Errorf(`duplicate output timeseries: %s%s`, ts.MetricName.MetricGroup, stringMetricName(&ts.MetricName))
+		if _, ok := m[string(bb.B)]; ok {
+			return nil, fmt.Errorf(`duplicate output timeseries: %s`, stringMetricName(&ts.MetricName))
 		}
-		m[string(bb.B)] = true
+		m[string(bb.B)] = struct{}{}
 
 		rs := &result[i]
 		rs.MetricNameMarshaled = append(rs.MetricNameMarshaled[:0], bb.B...)
@@ -154,10 +144,10 @@ func removeNaNs(tss []*timeseries) []*timeseries {
 	return rvs
 }
 
-func parsePromQLWithCache(q string) (expr, error) {
+func parsePromQLWithCache(q string) (metricsql.Expr, error) {
 	pcv := parseCacheV.Get(q)
 	if pcv == nil {
-		e, err := parsePromQL(q)
+		e, err := metricsql.Parse(q)
 		pcv = &parseCacheValue{
 			e:   e,
 			err: err,
@@ -189,16 +179,19 @@ var parseCacheV = func() *parseCache {
 const parseCacheMaxLen = 10e3
 
 type parseCacheValue struct {
-	e   expr
+	e   metricsql.Expr
 	err error
 }
 
 type parseCache struct {
-	m  map[string]*parseCacheValue
-	mu sync.RWMutex
+	// Move atomic counters to the top of struct for 8-byte alignment on 32-bit arch.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
 
 	requests uint64
 	misses   uint64
+
+	m  map[string]*parseCacheValue
+	mu sync.RWMutex
 }
 
 func (pc *parseCache) Requests() uint64 {

app/vmselect/promql/parser_test.go

@@ -47,650 +47,3 @@ func TestParseMetricSelectorError(t *testing.T) {
 	f(`foo[5m]`)
 	f(`foo offset 5m`)
 }
-
-func TestParsePromQLSuccess(t *testing.T) {
-	another := func(s string, sExpected string) {
-		t.Helper()
-
-		e, err := parsePromQL(s)
-		if err != nil {
-			t.Fatalf("unexpected error when parsing %q: %s", s, err)
-		}
-		res := e.AppendString(nil)
-		if string(res) != sExpected {
-			t.Fatalf("unexpected string constructed;\ngot\n%q\nwant\n%q", res, sExpected)
-		}
-	}
-	same := func(s string) {
-		t.Helper()
-		another(s, s)
-	}
-
-	// metricExpr
-	same(`{}`)
-	same(`{}[5m]`)
-	same(`{}[5m:]`)
-	same(`{}[:]`)
-	another(`{}[: ]`, `{}[:]`)
-	same(`{}[:3s]`)
-	another(`{}[: 3s ]`, `{}[:3s]`)
-	same(`{}[5m:3s]`)
-	another(`{}[ 5m : 3s ]`, `{}[5m:3s]`)
-	same(`{} offset 5m`)
-	same(`{}[5m] offset 10y`)
-	same(`{}[5.3m:3.4s] offset 10y`)
-	same(`{}[:3.4s] offset 10y`)
-	same(`{Foo="bAR"}`)
-	same(`{foo="bar"}`)
-	same(`{foo="bar"}[5m]`)
-	same(`{foo="bar"}[5m:]`)
-	same(`{foo="bar"}[5m:3s]`)
-	same(`{foo="bar"} offset 10y`)
-	same(`{foo="bar"}[5m] offset 10y`)
-	same(`{foo="bar"}[5m:3s] offset 10y`)
-	another(`{foo="bar"}[5m] oFFSEt 10y`, `{foo="bar"}[5m] offset 10y`)
-	same("METRIC")
-	same("metric")
-	same("m_e:tri44:_c123")
-	another("-metric", "0 - metric")
-	same(`metric offset 10h`)
-	same("metric[5m]")
-	same("metric[5m:3s]")
-	same("metric[5m] offset 10h")
-	same("metric[5m:3s] offset 10h")
-	same("metric[5i:3i] offset 10i")
-	same(`metric{foo="bar"}`)
-	same(`metric{foo="bar"} offset 10h`)
-	same(`metric{foo!="bar"}[2d]`)
-	same(`metric{foo="bar"}[2d] offset 10h`)
-	same(`metric{foo="bar", b="sdfsdf"}[2d:3h] offset 10h`)
-	another(`  metric  {  foo  = "bar"  }  [  2d ]   offset   10h  `, `metric{foo="bar"}[2d] offset 10h`)
-	// metric name matching keywords
-	same("rate")
-	same("RATE")
-	same("by")
-	same("BY")
-	same("bool")
-	same("BOOL")
-	same("unless")
-	same("UNLESS")
-	same("Ignoring")
-	same("with")
-	same("WITH")
-	same("With")
-	same("alias")
-	same(`alias{foo="bar"}`)
-	same(`aLIas{alias="aa"}`)
-	another(`al\ias`, `alias`)
-	// identifiers with with escape chars
-	same(`foo\ bar`)
-	same(`foo\-bar\{{baz\+bar="aa"}`)
-	another(`\x2E\x2ef\oo{b\xEF\ar="aa"}`, `\x2e.foo{b\xefar="aa"}`)
-	// Duplicate filters
-	same(`foo{__name__="bar"}`)
-	same(`foo{a="b", a="c", __name__="aaa", b="d"}`)
-	// Metric filters ending with comma
-	another(`m{foo="bar",}`, `m{foo="bar"}`)
-	// String concat in tag value
-	another(`m{foo="bar" + "baz"}`, `m{foo="barbaz"}`)
-
-	// Valid regexp
-	same(`foo{bar=~"x"}`)
-	same(`foo{bar=~"^x"}`)
-	same(`foo{bar=~"^x$"}`)
-	same(`foo{bar=~"^(a[bc]|d)$"}`)
-	same(`foo{bar!~"x"}`)
-	same(`foo{bar!~"^x"}`)
-	same(`foo{bar!~"^x$"}`)
-	same(`foo{bar!~"^(a[bc]|d)$"}`)
-
-	// stringExpr
-	same(`""`)
-	same(`"\n\t\r 12:{}[]()44"`)
-	another(`''`, `""`)
-	another("``", `""`)
-	another("   `foo\"b'ar`  ", "\"foo\\\"b'ar\"")
-	another(`  'foo\'bar"BAZ'  `, `"foo'bar\"BAZ"`)
-	// string concat
-	another(`"foo"+'bar'`, `"foobar"`)
-
-	// numberExpr
-	same(`1`)
-	same(`1.23`)
-	same(`0.23`)
-	same(`1.2e+45`)
-	same(`1.2e-45`)
-	same(`-1`)
-	same(`-1.23`)
-	same(`-0.23`)
-	same(`-1.2e+45`)
-	same(`-1.2e-45`)
-	same(`-1.2e-45`)
-	another(`12.5E34`, `1.25e+35`)
-	another(`-.2`, `-0.2`)
-	another(`-.2E-2`, `-0.002`)
-	same(`NaN`)
-	another(`Inf`, `+Inf`)
-	another(`+Inf`, `+Inf`)
-	another(`-Inf`, `-Inf`)
-
-	// binaryOpExpr
-	another(`NaN + 2 *3 * Inf`, `NaN`)
-	another(`Inf - Inf`, `NaN`)
-	another(`Inf + Inf`, `+Inf`)
-	another(`-m`, `0 - m`)
-	same(`m + ignoring () n[5m]`)
-	another(`M + IGNORING () N[5m]`, `M + ignoring () N[5m]`)
-	same(`m + on (foo) n[5m]`)
-	another(`m + ON (Foo) n[5m]`, `m + on (Foo) n[5m]`)
-	same(`m + ignoring (a, b) n[5m]`)
-	another(`1 or 2`, `1`)
-	another(`1 and 2`, `1`)
-	another(`1 unless 2`, `NaN`)
-	another(`1 default 2`, `1`)
-	another(`1 default NaN`, `1`)
-	another(`NaN default 2`, `2`)
-	another(`1 > 2`, `NaN`)
-	another(`1 > bool 2`, `0`)
-	another(`3 >= 2`, `3`)
-	another(`3 <= bool 2`, `0`)
-	another(`1 + -2 - 3`, `-4`)
-	another(`1 / 0 + 2`, `+Inf`)
-	another(`2 + -1 / 0`, `-Inf`)
-	another(`-1 ^ 0.5`, `NaN`)
-	another(`512.5 - (1 + 3) * (2 ^ 2) ^ 3`, `256.5`)
-	another(`1 == bool 1 != bool 24 < bool 4 > bool -1`, `1`)
-	another(`1 == bOOl 1 != BOOL 24 < Bool 4 > booL -1`, `1`)
-	another(`m1+on(foo)group_left m2`, `m1 + on (foo) group_left () m2`)
-	another(`M1+ON(FOO)GROUP_left M2`, `M1 + on (FOO) group_left () M2`)
-	same(`m1 + on (foo) group_right () m2`)
-	same(`m1 + on (foo, bar) group_right (x, y) m2`)
-	another(`m1 + on (foo, bar,) group_right (x, y,) m2`, `m1 + on (foo, bar) group_right (x, y) m2`)
-	same(`m1 == bool on (foo, bar) group_right (x, y) m2`)
-	another(`5 - 1 + 3 * 2 ^ 2 ^ 3 - 2  OR Metric {Bar= "Baz", aaa!="bb",cc=~"dd" ,zz !~"ff" } `,
-		`770 or Metric{Bar="Baz", aaa!="bb", cc=~"dd", zz!~"ff"}`)
-	same(`"foo" + bar()`)
-	same(`"foo" + bar{x="y"}`)
-	same(`("foo"[3s] + bar{x="y"})[5m:3s] offset 10s`)
-	same(`("foo"[3s] + bar{x="y"})[5i:3i] offset 10i`)
-	same(`bar + "foo" offset 3s`)
-	same(`bar + "foo" offset 3i`)
-	another(`1+2 if 2>3`, `NaN`)
-	another(`1+4 if 2<3`, `5`)
-	another(`2+6 default 3 if 2>3`, `8`)
-	another(`2+6 if 2>3 default NaN`, `NaN`)
-	another(`42 if 3>2 if 2+2<5`, `42`)
-	another(`42 if 3>2 if 2+2>=5`, `NaN`)
-	another(`1+2 ifnot 2>3`, `3`)
-	another(`1+4 ifnot 2<3`, `NaN`)
-	another(`2+6 default 3 ifnot 2>3`, `8`)
-	another(`2+6 ifnot 2>3 default NaN`, `8`)
-	another(`42 if 3>2 ifnot 2+2<5`, `NaN`)
-	another(`42 if 3>2 ifnot 2+2>=5`, `42`)
-
-	// parensExpr
-	another(`(-foo + ((bar) / (baz))) + ((23))`, `((0 - foo) + (bar / baz)) + 23`)
-	another(`(FOO + ((Bar) / (baZ))) + ((23))`, `(FOO + (Bar / baZ)) + 23`)
-	same(`(foo, bar)`)
-	another(`1+(foo, bar,)`, `1 + (foo, bar)`)
-	another(`((foo(bar,baz)), (1+(2)+(3,4)+()))`, `(foo(bar, baz), (3 + (3, 4)) + ())`)
-	same(`()`)
-
-	// funcExpr
-	same(`f()`)
-	another(`f(x,)`, `f(x)`)
-	another(`-f()-Ff()`, `(0 - f()) - Ff()`)
-	same(`F()`)
-	another(`+F()`, `F()`)
-	another(`++F()`, `F()`)
-	another(`--F()`, `0 - (0 - F())`)
-	same(`f(http_server_request)`)
-	same(`f(http_server_request)[4s:5m] offset 10m`)
-	same(`f(http_server_request)[4i:5i] offset 10i`)
-	same(`F(HttpServerRequest)`)
-	same(`f(job, foo)`)
-	same(`F(Job, Foo)`)
-	another(` FOO (bar) + f  (  m  (  ),ff(1 + (  2.5)) ,M[5m ]  , "ff"  )`, `FOO(bar) + f(m(), ff(3.5), M[5m], "ff")`)
-	// funcName matching keywords
-	same(`by(2)`)
-	same(`BY(2)`)
-	same(`or(2)`)
-	same(`OR(2)`)
-	same(`bool(2)`)
-	same(`BOOL(2)`)
-	same(`rate(rate(m))`)
-	same(`rate(rate(m[5m]))`)
-	same(`rate(rate(m[5m])[1h:])`)
-	same(`rate(rate(m[5m])[1h:3s])`)
-	// funcName with escape chars
-	same(`foo\(ba\-r()`)
-
-	// aggrFuncExpr
-	same(`sum(http_server_request) by ()`)
-	same(`sum(http_server_request) by (job)`)
-	same(`sum(http_server_request) without (job, foo)`)
-	another(`sum(x,y,) without (a,b,)`, `sum(x, y) without (a, b)`)
-	another(`sum by () (xx)`, `sum(xx) by ()`)
-	another(`sum by (s) (xx)[5s]`, `(sum(xx) by (s))[5s]`)
-	another(`SUM BY (ZZ, aa) (XX)`, `sum(XX) by (ZZ, aa)`)
-	another(`sum without (a, b) (xx,2+2)`, `sum(xx, 4) without (a, b)`)
-	another(`Sum WIthout (a, B) (XX,2+2)`, `sum(XX, 4) without (a, B)`)
-	same(`sum(a) or sum(b)`)
-	same(`sum(a) by () or sum(b) without (x, y)`)
-	same(`sum(a) + sum(b)`)
-	same(`sum(x) * (1 + sum(a))`)
-
-	// All the above
-	another(`Sum(Ff(M) * M{X=""}[5m] Offset 7m - 123, 35) BY (X, y) * F2("Test")`,
-		`sum((Ff(M) * M{X=""}[5m] offset 7m) - 123, 35) by (X, y) * F2("Test")`)
-	another(`# comment
-		Sum(Ff(M) * M{X=""}[5m] Offset 7m - 123, 35) BY (X, y) # yet another comment
-		* F2("Test")`,
-		`sum((Ff(M) * M{X=""}[5m] offset 7m) - 123, 35) by (X, y) * F2("Test")`)
-
-	// withExpr
-	another(`with () x`, `x`)
-	another(`with (x=1,) x`, `1`)
-	another(`with (x = m offset 5h) x + x`, `m offset 5h + m offset 5h`)
-	another(`with (x = m offset 5i) x + x`, `m offset 5i + m offset 5i`)
-	another(`with (foo = bar{x="x"}) 1`, `1`)
-	another(`with (foo = bar{x="x"}) "x"`, `"x"`)
-	another(`with (f="x") f`, `"x"`)
-	another(`with (foo = bar{x="x"}) x{x="y"}`, `x{x="y"}`)
-	another(`with (foo = bar{x="x"}) 1+1`, `2`)
-	another(`with (foo = bar{x="x"}) f()`, `f()`)
-	another(`with (foo = bar{x="x"}) sum(x)`, `sum(x)`)
-	another(`with (foo = bar{x="x"}) baz{foo="bar"}`, `baz{foo="bar"}`)
-	another(`with (foo = bar) baz`, `baz`)
-	another(`with (foo = bar) foo + foo{a="b"}`, `bar + bar{a="b"}`)
-	another(`with (foo = bar, bar=baz + f()) test`, `test`)
-	another(`with (ct={job="test"}) a{ct} + ct() + f({ct="x"})`, `(a{job="test"} + {job="test"}) + f({ct="x"})`)
-	another(`with (ct={job="test", i="bar"}) ct + {ct, x="d"} + foo{ct, ct} + ctx(1)`,
-		`(({job="test", i="bar"} + {job="test", i="bar", x="d"}) + foo{job="test", i="bar"}) + ctx(1)`)
-	another(`with (foo = bar) {__name__=~"foo"}`, `{__name__=~"foo"}`)
-	another(`with (foo = bar) foo{__name__="foo"}`, `bar`)
-	another(`with (foo = bar) {__name__="foo", x="y"}`, `bar{x="y"}`)
-	another(`with (foo(bar) = {__name__!="bar"}) foo(x)`, `{__name__!="bar"}`)
-	another(`with (foo(bar) = bar{__name__="bar"}) foo(x)`, `x`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar((x,y))`, `(x, y) + (x, y)`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x*y)`, `(x * y) + (x * y)`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x\*y)`, `x\*y + x\*y`)
-	another(`with (foo\-bar(b\ az) = b\ az + b\ az) foo\-bar(x\*y)`, `x\*y + x\*y`)
-	// override ttf to something new.
-	another(`with (ttf = a) ttf + b`, `a + b`)
-	// override ttf to ru
-	another(`with (ttf = ru(m, n)) ttf`, `(clamp_min(n - clamp_min(m, 0), 0) / clamp_min(n, 0)) * 100`)
-
-	// Verify withExpr recursion and forward reference
-	another(`with (x = x+y, y = x+x) y ^ 2`, `((x + y) + (x + y)) ^ 2`)
-	another(`with (f1(x)=f2(x), f2(x)=f1(x)^2) f1(foobar)`, `f2(foobar)`)
-	another(`with (f1(x)=f2(x), f2(x)=f1(x)^2) f2(foobar)`, `f2(foobar) ^ 2`)
-
-	// Verify withExpr funcs
-	another(`with (x() = y+1) x`, `y + 1`)
-	another(`with (x(foo) = foo+1) x(a)`, `a + 1`)
-	another(`with (x(a, b) = a + b) x(foo, bar)`, `foo + bar`)
-	another(`with (x(a, b) = a + b) x(foo, x(1, 2))`, `foo + 3`)
-	another(`with (x(a) = sum(a) by (b)) x(xx) / x(y)`, `sum(xx) by (b) / sum(y) by (b)`)
-	another(`with (f(a,f,x)=ff(x,f,a)) f(f(x,y,z),1,2)`, `ff(2, 1, ff(z, y, x))`)
-	another(`with (f(x)=1+f(x)) f(foo{bar="baz"})`, `1 + f(foo{bar="baz"})`)
-	another(`with (a=foo, y=bar, f(a)= a+a+y) f(x)`, `(x + x) + bar`)
-	another(`with (f(a, b) = m{a, b}) f({a="x", b="y"}, {c="d"})`, `m{a="x", b="y", c="d"}`)
-	another(`with (xx={a="x"}, f(a, b) = m{a, b}) f({xx, b="y"}, {c="d"})`, `m{a="x", b="y", c="d"}`)
-	another(`with (x() = {b="c"}) foo{x}`, `foo{b="c"}`)
-	another(`with (f(x)=x{foo="bar"} offset 5m) f(m offset 10m)`, `(m{foo="bar"} offset 10m) offset 5m`)
-	another(`with (f(x)=x{foo="bar",bas="a"}[5m]) f(m[10m] offset 3s)`, `(m{foo="bar", bas="a"}[10m] offset 3s)[5m]`)
-	another(`with (f(x)=x{foo="bar"}[5m] offset 10m) f(m{x="y"})`, `m{x="y", foo="bar"}[5m] offset 10m`)
-	another(`with (f(x)=x{foo="bar"}[5m] offset 10m) f({x="y", foo="bar", foo="bar"})`, `{x="y", foo="bar"}[5m] offset 10m`)
-	another(`with (f(m, x)=m{x}[5m] offset 10m) f(foo, {})`, `foo[5m] offset 10m`)
-	another(`with (f(m, x)=m{x, bar="baz"}[5m] offset 10m) f(foo, {})`, `foo{bar="baz"}[5m] offset 10m`)
-	another(`with (f(x)=x[5m] offset 3s) f(foo[3m]+bar)`, `(foo[3m] + bar)[5m] offset 3s`)
-	another(`with (f(x)=x[5m:3s] oFFsEt 1.5m) f(sum(s) by (a,b))`, `(sum(s) by (a, b))[5m:3s] offset 1.5m`)
-	another(`with (x="a", y=x) y+"bc"`, `"abc"`)
-	another(`with (x="a", y="b"+x) "we"+y+"z"+f()`, `"webaz" + f()`)
-	another(`with (f(x) = m{foo=x+"y", bar="y"+x, baz=x} + x) f("qwe")`, `m{foo="qwey", bar="yqwe", baz="qwe"} + "qwe"`)
-	another(`with (f(a)=a) f`, `f`)
-	another(`with (f\q(a)=a) f\q`, `fq`)
-
-	// Verify withExpr for aggr func modifiers
-	another(`with (f(x) = x, y = sum(m) by (f)) y`, `sum(m) by (f)`)
-	another(`with (f(x) = sum(m) by (x)) f(foo)`, `sum(m) by (foo)`)
-	another(`with (f(x) = sum(m) by (x)) f((foo, bar, foo))`, `sum(m) by (foo, bar)`)
-	another(`with (f(x) = sum(m) without (x,y)) f((a, b))`, `sum(m) without (a, b, y)`)
-	another(`with (f(x) = sum(m) without (y,x)) f((a, y))`, `sum(m) without (y, a)`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f(foo,())`, `a + on (foo) group_left (bar) b`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f((foo),())`, `a + on (foo) group_left (bar) b`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f((foo,xx),())`, `a + on (foo, xx) group_left (bar) b`)
-
-	// Verify nested with exprs
-	another(`with (f(x) = (with(x=y) x) + x) f(z)`, `y + z`)
-	another(`with (x=foo) f(a, with (y=x) y)`, `f(a, foo)`)
-	another(`with (x=foo) a * x + (with (y=x) y) / y`, `(a * foo) + (foo / y)`)
-	another(`with (x = with (y = foo) y + x) x/x`, `(foo + x) / (foo + x)`)
-	another(`with (
-		x = {foo="bar"},
-		q = m{x, y="1"},
-		f(x) =
-			with (
-				z(y) = x + y * q
-			)
-			z(foo) / f(x)
-	)
-	f(a)`, `(a + (foo * m{foo="bar", y="1"})) / f(a)`)
-
-	// complex withExpr
-	another(`WITH (
-		treshold = (0.9),
-		commonFilters = {job="cacher", instance=~"1.2.3.4"},
-		hits = rate(cache{type="hit", commonFilters}[5m]),
-		miss = rate(cache{type="miss", commonFilters}[5m]),
-		sumByInstance(arg) = sum(arg) by (instance),
-		hitRatio = sumByInstance(hits) / sumByInstance(hits + miss)
-	)
-	hitRatio < treshold`,
-		`(sum(rate(cache{type="hit", job="cacher", instance=~"1.2.3.4"}[5m])) by (instance) / sum(rate(cache{type="hit", job="cacher", instance=~"1.2.3.4"}[5m]) + rate(cache{type="miss", job="cacher", instance=~"1.2.3.4"}[5m])) by (instance)) < 0.9`)
-	another(`WITH (
-		x2(x) = x^2,
-		f(x, y) = x2(x) + x*y + x2(y)
-	)
-	f(a, 3)
-	`, `((a ^ 2) + (a * 3)) + 9`)
-	another(`WITH (
-		x2(x) = x^2,
-		f(x, y) = x2(x) + x*y + x2(y)
-	)
-	f(2, 3)
-	`, `19`)
-	another(`WITH (
-		commonFilters = {instance="foo"},
-		timeToFuckup(currv, maxv) = (maxv - currv) / rate(currv)
-	)
-	timeToFuckup(diskUsage{commonFilters}, maxDiskSize{commonFilters})`,
-		`(maxDiskSize{instance="foo"} - diskUsage{instance="foo"}) / rate(diskUsage{instance="foo"})`)
-	another(`WITH (
-	       commonFilters = {job="foo", instance="bar"},
-	       sumRate(m, cf) = sum(rate(m{cf})) by (job, instance),
-	       hitRate(hits, misses) = sumRate(hits, commonFilters) / (sumRate(hits, commonFilters) + sumRate(misses, commonFilters))
-	   )
-	   hitRate(cacheHits, cacheMisses)`,
-		`sum(rate(cacheHits{job="foo", instance="bar"})) by (job, instance) / (sum(rate(cacheHits{job="foo", instance="bar"})) by (job, instance) + sum(rate(cacheMisses{job="foo", instance="bar"})) by (job, instance))`)
-	another(`with(y=123,z=5) union(with(y=3,f(x)=x*y) f(2) + f(3), with(x=5,y=2) x*y*z)`, `union(15, 50)`)
-}
-
-func TestParsePromQLError(t *testing.T) {
-	f := func(s string) {
-		t.Helper()
-
-		e, err := parsePromQL(s)
-		if err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
-		}
-		if e != nil {
-			t.Fatalf("expecting nil expr when parsing %q", s)
-		}
-	}
-
-	// an empty string
-	f("")
-	f("  \t\b\r\n  ")
-
-	// invalid metricExpr
-	f(`{__name__="ff"} offset 55`)
-	f(`{__name__="ff"} offset -5m`)
-	f(`foo[55]`)
-	f(`m[-5m]`)
-	f(`{`)
-	f(`foo{`)
-	f(`foo{bar`)
-	f(`foo{bar=`)
-	f(`foo{bar="baz"`)
-	f(`foo{bar="baz",  `)
-	f(`foo{123="23"}`)
-	f(`foo{foo}`)
-	f(`foo{,}`)
-	f(`foo{,foo="bar"}`)
-	f(`foo{foo=}`)
-	f(`foo{foo="ba}`)
-	f(`foo{"foo"="bar"}`)
-	f(`foo{$`)
-	f(`foo{a $`)
-	f(`foo{a="b",$`)
-	f(`foo{a="b"}$`)
-	f(`[`)
-	f(`[]`)
-	f(`f[5m]$`)
-	f(`[5m]`)
-	f(`[5m] offset 4h`)
-	f(`m[5m] offset $`)
-	f(`m[5m] offset 5h $`)
-	f(`m[]`)
-	f(`m[-5m]`)
-	f(`m[5m:`)
-	f(`m[5m:-`)
-	f(`m[5m:-1`)
-	f(`m[5m:-1]`)
-	f(`m[:`)
-	f(`m[:-`)
-	f(`m[:1]`)
-	f(`m[:-1m]`)
-	f(`m[5]`)
-	f(`m[[5m]]`)
-	f(`m[foo]`)
-	f(`m["ff"]`)
-	f(`m[10m`)
-	f(`m[123`)
-	f(`m["ff`)
-	f(`m[(f`)
-	f(`fd}`)
-	f(`]`)
-	f(`m $`)
-	f(`m{,}`)
-	f(`m{x=y}`)
-	f(`m{x=y/5}`)
-	f(`m{x=y+5}`)
-
-	// Invalid regexp
-	f(`foo{bar=~"x["}`)
-	f(`foo{bar=~"x("}`)
-	f(`foo{bar=~"x)"}`)
-	f(`foo{bar!~"x["}`)
-	f(`foo{bar!~"x("}`)
-	f(`foo{bar!~"x)"}`)
-
-	// invalid stringExpr
-	f(`'`)
-	f(`"`)
-	f("`")
-	f(`"foo`)
-	f(`'foo`)
-	f("`foo")
-	f(`"foo\"bar`)
-	f(`'foo\'bar`)
-	f("`foo\\`bar")
-	f(`"" $`)
-	f(`"foo" +`)
-	f(`n{"foo" + m`)
-
-	// invalid numberExpr
-	f(`12.`)
-	f(`1.2e`)
-	f(`23e-`)
-	f(`23E+`)
-	f(`.`)
-	f(`-12.`)
-	f(`-1.2e`)
-	f(`-23e-`)
-	f(`-23E+`)
-	f(`-.`)
-	f(`-1$$`)
-	f(`-$$`)
-	f(`+$$`)
-	f(`23 $$`)
-
-	// invalid binaryOpExpr
-	f(`+`)
-	f(`1 +`)
-	f(`1 + 2.`)
-	f(`3 unless`)
-	f(`23 + on (foo)`)
-	f(`m + on (,) m`)
-	f(`3 * ignoring`)
-	f(`m * on (`)
-	f(`m * on (foo`)
-	f(`m * on (foo,`)
-	f(`m * on (foo,)`)
-	f(`m * on (,foo)`)
-	f(`m * on (,)`)
-	f(`m == bool (bar) baz`)
-	f(`m == bool () baz`)
-	f(`m * by (baz) n`)
-	f(`m + bool group_left m2`)
-	f(`m + on () group_left (`)
-	f(`m + on () group_left (,`)
-	f(`m + on () group_left (,foo`)
-	f(`m + on () group_left (foo,)`)
-	f(`m + on () group_left (,foo)`)
-	f(`m + on () group_left (foo)`)
-	f(`m + on () group_right (foo) (m`)
-	f(`m or ignoring () group_left () n`)
-	f(`1 + bool 2`)
-	f(`m % bool n`)
-	f(`m * bool baz`)
-	f(`M * BOoL BaZ`)
-	f(`foo unless ignoring (bar) group_left xxx`)
-	f(`foo or bool bar`)
-	f(`foo == bool $$`)
-	f(`"foo" + bar`)
-
-	// invalid parensExpr
-	f(`(`)
-	f(`($`)
-	f(`(+`)
-	f(`(1`)
-	f(`(m+`)
-	f(`1)`)
-	f(`(,)`)
-	f(`(1)$`)
-
-	// invalid funcExpr
-	f(`f $`)
-	f(`f($)`)
-	f(`f[`)
-	f(`f()$`)
-	f(`f(`)
-	f(`f(foo`)
-	f(`f(f,`)
-	f(`f(,`)
-	f(`f(,)`)
-	f(`f(,foo)`)
-	f(`f(,foo`)
-	f(`f(foo,$`)
-	f(`f() by (a)`)
-	f(`f without (x) (y)`)
-	f(`f() foo (a)`)
-	f(`f bar (x) (b)`)
-	f(`f bar (x)`)
-
-	// invalid aggrFuncExpr
-	f(`sum(`)
-	f(`sum $`)
-	f(`sum [`)
-	f(`sum($)`)
-	f(`sum()$`)
-	f(`sum(foo) ba`)
-	f(`sum(foo) ba()`)
-	f(`sum(foo) by`)
-	f(`sum(foo) without x`)
-	f(`sum(foo) aaa`)
-	f(`sum(foo) aaa x`)
-	f(`sum() by $`)
-	f(`sum() by (`)
-	f(`sum() by ($`)
-	f(`sum() by (a`)
-	f(`sum() by (a $`)
-	f(`sum() by (a ]`)
-	f(`sum() by (a)$`)
-	f(`sum() by (,`)
-	f(`sum() by (a,$`)
-	f(`sum() by (,)`)
-	f(`sum() by (,a`)
-	f(`sum() by (,a)`)
-	f(`sum() on (b)`)
-	f(`sum() bool`)
-	f(`sum() group_left`)
-	f(`sum() group_right(x)`)
-	f(`sum ba`)
-	f(`sum ba ()`)
-	f(`sum by (`)
-	f(`sum by (a`)
-	f(`sum by (,`)
-	f(`sum by (,)`)
-	f(`sum by (,a`)
-	f(`sum by (,a)`)
-	f(`sum by (a)`)
-	f(`sum by (a) (`)
-	f(`sum by (a) [`)
-	f(`sum by (a) {`)
-	f(`sum by (a) (b`)
-	f(`sum by (a) (b,`)
-	f(`sum by (a) (,)`)
-	f(`avg by (a) (,b)`)
-	f(`sum by (x) (y) by (z)`)
-	f(`sum(m) by (1)`)
-
-	// invalid withExpr
-	f(`with $`)
-	f(`with a`)
-	f(`with a=b c`)
-	f(`with (`)
-	f(`with (x=b)$`)
-	f(`with ($`)
-	f(`with (foo`)
-	f(`with (foo $`)
-	f(`with (x y`)
-	f(`with (x =`)
-	f(`with (x = $`)
-	f(`with (x= y`)
-	f(`with (x= y $`)
-	f(`with (x= y)`)
-	f(`with (x=(`)
-	f(`with (x=[)`)
-	f(`with (x=() x)`)
-	f(`with ($$)`)
-	f(`with (x $$`)
-	f(`with (x = $$)`)
-	f(`with (x = foo) bar{x}`)
-	f(`with (x = {foo="bar"}[5m]) bar{x}`)
-	f(`with (x = {foo="bar"} offset 5m) bar{x}`)
-	f(`with (x = a, x = b) c`)
-	f(`with (x(a, a) = b) c`)
-	f(`with (x=m{f="x"}) foo{x}`)
-	f(`with (sum = x) y`)
-	f(`with (rate(a) = b) c`)
-	f(`with (clamp_min=x) y`)
-	f(`with (f()`)
-	f(`with (a=b c=d) e`)
-	f(`with (f(x)=x^2) m{x}`)
-	f(`with (f(x)=ff()) m{x}`)
-	f(`with (f(x`)
-	f(`with (x=m) a{x} + b`)
-	f(`with (x=m) b + a{x}`)
-	f(`with (x=m) f(b, a{x})`)
-	f(`with (x=m) sum(a{x})`)
-	f(`with (x=m) (a{x})`)
-	f(`with (f(a)=a) f(1, 2)`)
-	f(`with (f(x)=x{foo="bar"}) f(1)`)
-	f(`with (f(x)=x{foo="bar"}) f(m + n)`)
-	f(`with (f = with`)
-	f(`with (,)`)
-	f(`with (1) 2`)
-	f(`with (f(1)=2) 3`)
-	f(`with (f(,)=x) x`)
-	f(`with (x(a) = {b="c"}) foo{x}`)
-	f(`with (f(x) = m{foo=xx}) f("qwe")`)
-	f(`a + with(f(x)=x) f(1,2)`)
-	f(`with (f(x) = sum(m) by (x)) f({foo="bar"})`)
-	f(`with (f(x) = sum(m) by (x)) f((xx(), {foo="bar"}))`)
-	f(`with (f(x) = m + on (x) n) f(xx())`)
-	f(`with (f(x) = m + on (a) group_right (x) n) f(xx())`)
-}

app/vmselect/promql/rollup.go

@@ -3,12 +3,14 @@ package promql
 import (
 	"fmt"
 	"math"
-	"sort"
 	"strings"
 	"sync"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/histogram"
 )
 
@@ -23,8 +25,8 @@ var rollupFuncs = map[string]newRollupFunc{
 	"deriv_fast":         newRollupFuncOneArg(rollupDerivFast),
 	"holt_winters":       newRollupHoltWinters,
 	"idelta":             newRollupFuncOneArg(rollupIdelta),
-	"increase":           newRollupFuncOneArg(rollupDelta),  // + rollupFuncsRemoveCounterResets
-	"irate":              newRollupFuncOneArg(rollupIderiv), // + rollupFuncsRemoveCounterResets
+	"increase":           newRollupFuncOneArg(rollupIncrease), // + rollupFuncsRemoveCounterResets
+	"irate":              newRollupFuncOneArg(rollupIderiv),   // + rollupFuncsRemoveCounterResets
 	"predict_linear":     newRollupPredictLinear,
 	"rate":               newRollupFuncOneArg(rollupDerivFast), // + rollupFuncsRemoveCounterResets
 	"resets":             newRollupFuncOneArg(rollupResets),
@@ -36,21 +38,71 @@ var rollupFuncs = map[string]newRollupFunc{
 	"quantile_over_time": newRollupQuantile,
 	"stddev_over_time":   newRollupFuncOneArg(rollupStddev),
 	"stdvar_over_time":   newRollupFuncOneArg(rollupStdvar),
+	"absent_over_time":   newRollupFuncOneArg(rollupAbsent),
 
 	// Additional rollup funcs.
-	"sum2_over_time":     newRollupFuncOneArg(rollupSum2),
-	"geomean_over_time":  newRollupFuncOneArg(rollupGeomean),
-	"first_over_time":    newRollupFuncOneArg(rollupFirst),
-	"last_over_time":     newRollupFuncOneArg(rollupLast),
-	"distinct_over_time": newRollupFuncOneArg(rollupDistinct),
-	"integrate":          newRollupFuncOneArg(rollupIntegrate),
-	"ideriv":             newRollupFuncOneArg(rollupIderiv),
-	"rollup":             newRollupFuncOneArg(rollupFake),
-	"rollup_rate":        newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
-	"rollup_deriv":       newRollupFuncOneArg(rollupFake),
-	"rollup_delta":       newRollupFuncOneArg(rollupFake),
-	"rollup_increase":    newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
-	"rollup_candlestick": newRollupFuncOneArg(rollupFake),
+	"sum2_over_time":      newRollupFuncOneArg(rollupSum2),
+	"geomean_over_time":   newRollupFuncOneArg(rollupGeomean),
+	"first_over_time":     newRollupFuncOneArg(rollupFirst),
+	"last_over_time":      newRollupFuncOneArg(rollupLast),
+	"distinct_over_time":  newRollupFuncOneArg(rollupDistinct),
+	"increases_over_time": newRollupFuncOneArg(rollupIncreases),
+	"decreases_over_time": newRollupFuncOneArg(rollupDecreases),
+	"integrate":           newRollupFuncOneArg(rollupIntegrate),
+	"ideriv":              newRollupFuncOneArg(rollupIderiv),
+	"lifetime":            newRollupFuncOneArg(rollupLifetime),
+	"lag":                 newRollupFuncOneArg(rollupLag),
+	"scrape_interval":     newRollupFuncOneArg(rollupScrapeInterval),
+	"tmin_over_time":      newRollupFuncOneArg(rollupTmin),
+	"tmax_over_time":      newRollupFuncOneArg(rollupTmax),
+	"share_le_over_time":  newRollupShareLE,
+	"share_gt_over_time":  newRollupShareGT,
+	"histogram_over_time": newRollupFuncOneArg(rollupHistogram),
+	"rollup":              newRollupFuncOneArg(rollupFake),
+	"rollup_rate":         newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
+	"rollup_deriv":        newRollupFuncOneArg(rollupFake),
+	"rollup_delta":        newRollupFuncOneArg(rollupFake),
+	"rollup_increase":     newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
+	"rollup_candlestick":  newRollupFuncOneArg(rollupFake),
+	"aggr_over_time":      newRollupFuncTwoArgs(rollupFake),
+}
+
+// rollupAggrFuncs are functions that can be passed to `aggr_over_time()`
+var rollupAggrFuncs = map[string]rollupFunc{
+	// Standard rollup funcs from PromQL.
+	"changes":          rollupChanges,
+	"delta":            rollupDelta,
+	"deriv":            rollupDerivSlow,
+	"deriv_fast":       rollupDerivFast,
+	"idelta":           rollupIdelta,
+	"increase":         rollupIncrease,  // + rollupFuncsRemoveCounterResets
+	"irate":            rollupIderiv,    // + rollupFuncsRemoveCounterResets
+	"rate":             rollupDerivFast, // + rollupFuncsRemoveCounterResets
+	"resets":           rollupResets,
+	"avg_over_time":    rollupAvg,
+	"min_over_time":    rollupMin,
+	"max_over_time":    rollupMax,
+	"sum_over_time":    rollupSum,
+	"count_over_time":  rollupCount,
+	"stddev_over_time": rollupStddev,
+	"stdvar_over_time": rollupStdvar,
+	"absent_over_time": rollupAbsent,
+
+	// Additional rollup funcs.
+	"sum2_over_time":      rollupSum2,
+	"geomean_over_time":   rollupGeomean,
+	"first_over_time":     rollupFirst,
+	"last_over_time":      rollupLast,
+	"distinct_over_time":  rollupDistinct,
+	"increases_over_time": rollupIncreases,
+	"decreases_over_time": rollupDecreases,
+	"integrate":           rollupIntegrate,
+	"ideriv":              rollupIderiv,
+	"lifetime":            rollupLifetime,
+	"lag":                 rollupLag,
+	"scrape_interval":     rollupScrapeInterval,
+	"tmin_over_time":      rollupTmin,
+	"tmax_over_time":      rollupTmax,
 }
 
 var rollupFuncsMayAdjustWindow = map[string]bool{
@@ -61,6 +113,9 @@ var rollupFuncsMayAdjustWindow = map[string]bool{
 	"deriv_fast":      true,
 	"irate":           true,
 	"rate":            true,
+	"lifetime":        true,
+	"lag":             true,
+	"scrape_interval": true,
 }
 
 var rollupFuncsRemoveCounterResets = map[string]bool{
@@ -81,15 +136,137 @@ var rollupFuncsKeepMetricGroup = map[string]bool{
 	"geomean_over_time":  true,
 }
 
+func getRollupAggrFuncNames(expr metricsql.Expr) ([]string, error) {
+	afe, ok := expr.(*metricsql.AggrFuncExpr)
+	if ok {
+		// This is for incremental aggregate function case:
+		//
+		//     sum(aggr_over_time(...))
+		//
+		// See aggr_incremental.go for details.
+		expr = afe.Args[0]
+	}
+	fe, ok := expr.(*metricsql.FuncExpr)
+	if !ok {
+		logger.Panicf("BUG: unexpected expression; want metricsql.FuncExpr; got %T; value: %s", expr, expr.AppendString(nil))
+	}
+	if fe.Name != "aggr_over_time" {
+		logger.Panicf("BUG: unexpected function name: %q; want `aggr_over_time`", fe.Name)
+	}
+	if len(fe.Args) != 2 {
+		return nil, fmt.Errorf("unexpected number of args to aggr_over_time(); got %d; want %d", len(fe.Args), 2)
+	}
+	arg := fe.Args[0]
+	var aggrFuncNames []string
+	if se, ok := arg.(*metricsql.StringExpr); ok {
+		aggrFuncNames = append(aggrFuncNames, se.S)
+	} else {
+		fe, ok := arg.(*metricsql.FuncExpr)
+		if !ok || fe.Name != "" {
+			return nil, fmt.Errorf("%s cannot be passed to aggr_over_time(); expecting quoted aggregate function name or a list of quoted aggregate function names",
+				arg.AppendString(nil))
+		}
+		for _, e := range fe.Args {
+			se, ok := e.(*metricsql.StringExpr)
+			if !ok {
+				return nil, fmt.Errorf("%s cannot be passed here; expecting quoted aggregate function name", e.AppendString(nil))
+			}
+			aggrFuncNames = append(aggrFuncNames, se.S)
+		}
+	}
+	if len(aggrFuncNames) == 0 {
+		return nil, fmt.Errorf("aggr_over_time() must contain at least a single aggregate function name")
+	}
+	for _, s := range aggrFuncNames {
+		if rollupAggrFuncs[s] == nil {
+			return nil, fmt.Errorf("%q cannot be used in `aggr_over_time` function; expecting quoted aggregate function name", s)
+		}
+	}
+	return aggrFuncNames, nil
+}
+
 func getRollupArgIdx(funcName string) int {
 	funcName = strings.ToLower(funcName)
 	if rollupFuncs[funcName] == nil {
 		logger.Panicf("BUG: getRollupArgIdx is called for non-rollup func %q", funcName)
 	}
-	if funcName == "quantile_over_time" {
+	switch funcName {
+	case "quantile_over_time", "aggr_over_time":
 		return 1
+	default:
+		return 0
 	}
-	return 0
+}
+
+func getRollupConfigs(name string, rf rollupFunc, expr metricsql.Expr, start, end, step, window int64, lookbackDelta int64, sharedTimestamps []int64) (
+	func(values []float64, timestamps []int64), []*rollupConfig, error) {
+	preFunc := func(values []float64, timestamps []int64) {}
+	if rollupFuncsRemoveCounterResets[name] {
+		preFunc = func(values []float64, timestamps []int64) {
+			removeCounterResets(values)
+		}
+	}
+	newRollupConfig := func(rf rollupFunc, tagValue string) *rollupConfig {
+		return &rollupConfig{
+			TagValue:        tagValue,
+			Func:            rf,
+			Start:           start,
+			End:             end,
+			Step:            step,
+			Window:          window,
+			MayAdjustWindow: rollupFuncsMayAdjustWindow[name],
+			LookbackDelta:   lookbackDelta,
+			Timestamps:      sharedTimestamps,
+		}
+	}
+	appendRollupConfigs := func(dst []*rollupConfig) []*rollupConfig {
+		dst = append(dst, newRollupConfig(rollupMin, "min"))
+		dst = append(dst, newRollupConfig(rollupMax, "max"))
+		dst = append(dst, newRollupConfig(rollupAvg, "avg"))
+		return dst
+	}
+	var rcs []*rollupConfig
+	switch name {
+	case "rollup":
+		rcs = appendRollupConfigs(rcs)
+	case "rollup_rate", "rollup_deriv":
+		preFuncPrev := preFunc
+		preFunc = func(values []float64, timestamps []int64) {
+			preFuncPrev(values, timestamps)
+			derivValues(values, timestamps)
+		}
+		rcs = appendRollupConfigs(rcs)
+	case "rollup_increase", "rollup_delta":
+		preFuncPrev := preFunc
+		preFunc = func(values []float64, timestamps []int64) {
+			preFuncPrev(values, timestamps)
+			deltaValues(values)
+		}
+		rcs = appendRollupConfigs(rcs)
+	case "rollup_candlestick":
+		rcs = append(rcs, newRollupConfig(rollupFirst, "open"))
+		rcs = append(rcs, newRollupConfig(rollupLast, "close"))
+		rcs = append(rcs, newRollupConfig(rollupMin, "low"))
+		rcs = append(rcs, newRollupConfig(rollupMax, "high"))
+	case "aggr_over_time":
+		aggrFuncNames, err := getRollupAggrFuncNames(expr)
+		if err != nil {
+			return nil, nil, fmt.Errorf("invalid args to %s: %s", expr.AppendString(nil), err)
+		}
+		for _, aggrFuncName := range aggrFuncNames {
+			if rollupFuncsRemoveCounterResets[aggrFuncName] {
+				// There is no need to save the previous preFunc, since it is either empty or the same.
+				preFunc = func(values []float64, timestamps []int64) {
+					removeCounterResets(values)
+				}
+			}
+			rf := rollupAggrFuncs[aggrFuncName]
+			rcs = append(rcs, newRollupConfig(rf, aggrFuncName))
+		}
+	default:
+		rcs = append(rcs, newRollupConfig(rf, ""))
+	}
+	return preFunc, rcs, nil
 }
 
 func getRollupFunc(funcName string) newRollupFunc {
@@ -97,18 +274,21 @@ func getRollupFunc(funcName string) newRollupFunc {
 	return rollupFuncs[funcName]
 }
 
-func isRollupFunc(funcName string) bool {
-	return getRollupFunc(funcName) != nil
-}
-
 type rollupFuncArg struct {
 	prevValue     float64
 	prevTimestamp int64
 	values        []float64
 	timestamps    []int64
 
-	idx  int
-	step int64
+	currTimestamp int64
+	idx           int
+	step          int64
+
+	// Real previous value even if it is located too far from the current window.
+	// It matches prevValue if prevValue is not nan.
+	realPrevValue float64
+
+	tsm *timeseriesMap
 }
 
 func (rfa *rollupFuncArg) reset() {
@@ -116,8 +296,11 @@ func (rfa *rollupFuncArg) reset() {
 	rfa.prevTimestamp = 0
 	rfa.values = nil
 	rfa.timestamps = nil
+	rfa.currTimestamp = 0
 	rfa.idx = 0
 	rfa.step = 0
+	rfa.realPrevValue = nan
+	rfa.tsm = nil
 }
 
 // rollupFunc must return rollup value for the given rfa.
@@ -143,6 +326,9 @@ type rollupConfig struct {
 	MayAdjustWindow bool
 
 	Timestamps []int64
+
+	// LoookbackDelta is the analog to `-query.lookback-delta` from Prometheus world.
+	LookbackDelta int64
 }
 
 var (
@@ -153,6 +339,54 @@ var (
 // The maximum interval without previous rows.
 const maxSilenceInterval = 5 * 60 * 1000
 
+type timeseriesMap struct {
+	origin    *timeseries
+	labelName string
+	h         metrics.Histogram
+	m         map[string]*timeseries
+}
+
+func newTimeseriesMap(funcName string, sharedTimestamps []int64, mnSrc *storage.MetricName) *timeseriesMap {
+	if funcName != "histogram_over_time" {
+		return nil
+	}
+
+	values := make([]float64, len(sharedTimestamps))
+	for i := range values {
+		values[i] = nan
+	}
+	var origin timeseries
+	origin.MetricName.CopyFrom(mnSrc)
+	origin.MetricName.ResetMetricGroup()
+	origin.Timestamps = sharedTimestamps
+	origin.Values = values
+	return &timeseriesMap{
+		origin:    &origin,
+		labelName: "vmrange",
+		m:         make(map[string]*timeseries),
+	}
+}
+
+func (tsm *timeseriesMap) AppendTimeseriesTo(dst []*timeseries) []*timeseries {
+	for _, ts := range tsm.m {
+		dst = append(dst, ts)
+	}
+	return dst
+}
+
+func (tsm *timeseriesMap) GetOrCreateTimeseries(labelValue string) *timeseries {
+	ts := tsm.m[labelValue]
+	if ts != nil {
+		return ts
+	}
+	ts = &timeseries{}
+	ts.CopyFromShallowTimestamps(tsm.origin)
+	ts.MetricName.RemoveTag(tsm.labelName)
+	ts.MetricName.AddTag(tsm.labelName, labelValue)
+	tsm.m[labelValue] = ts
+	return ts
+}
+
 // Do calculates rollups for the given timestamps and values, appends
 // them to dstValues and returns results.
 //
@@ -160,8 +394,19 @@ const maxSilenceInterval = 5 * 60 * 1000
 //
 // timestamps must cover time range [rc.Start - rc.Window - maxSilenceInterval ... rc.End + rc.Step].
 //
-// Cannot be called from concurrent goroutines.
+// Do cannot be called from concurrent goroutines.
 func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []int64) []float64 {
+	return rc.doInternal(dstValues, nil, values, timestamps)
+}
+
+// DoTimeseriesMap calculates rollups for the given timestamps and values and puts them to tsm.
+func (rc *rollupConfig) DoTimeseriesMap(tsm *timeseriesMap, values []float64, timestamps []int64) {
+	ts := getTimeseries()
+	ts.Values = rc.doInternal(ts.Values[:0], tsm, values, timestamps)
+	putTimeseries(ts)
+}
+
+func (rc *rollupConfig) doInternal(dstValues []float64, tsm *timeseriesMap, values []float64, timestamps []int64) []float64 {
 	// Sanity checks.
 	if rc.Step <= 0 {
 		logger.Panicf("BUG: Step must be bigger than 0; got %d", rc.Step)
@@ -179,7 +424,11 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	// Extend dstValues in order to remove mallocs below.
 	dstValues = decimal.ExtendFloat64sCapacity(dstValues, len(rc.Timestamps))
 
-	maxPrevInterval := getMaxPrevInterval(timestamps)
+	scrapeInterval := getScrapeInterval(timestamps)
+	maxPrevInterval := getMaxPrevInterval(scrapeInterval)
+	if rc.LookbackDelta > 0 && maxPrevInterval > rc.LookbackDelta {
+		maxPrevInterval = rc.LookbackDelta
+	}
 	window := rc.Window
 	if window <= 0 {
 		window = rc.Step
@@ -190,32 +439,36 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	rfa := getRollupFuncArg()
 	rfa.idx = 0
 	rfa.step = rc.Step
+	rfa.realPrevValue = nan
+	rfa.tsm = tsm
 
 	i := 0
 	j := 0
+	ni := 0
+	nj := 0
 	for _, tEnd := range rc.Timestamps {
 		tStart := tEnd - window
-		n := sort.Search(len(timestamps)-i, func(n int) bool {
-			return timestamps[i+n] > tStart
-		})
-		i += n
+		ni = seekFirstTimestampIdxAfter(timestamps[i:], tStart, ni)
+		i += ni
 		if j < i {
 			j = i
 		}
-		n = sort.Search(len(timestamps)-j, func(n int) bool {
-			return timestamps[j+n] > tEnd
-		})
-		j += n
+		nj = seekFirstTimestampIdxAfter(timestamps[j:], tEnd, nj)
+		j += nj
 
 		rfa.prevValue = nan
 		rfa.prevTimestamp = tStart - maxPrevInterval
-		if i > 0 && timestamps[i-1] > rfa.prevTimestamp {
+		if i < len(timestamps) && i > 0 && timestamps[i-1] > rfa.prevTimestamp {
 			rfa.prevValue = values[i-1]
 			rfa.prevTimestamp = timestamps[i-1]
 		}
 
 		rfa.values = values[i:j]
 		rfa.timestamps = timestamps[i:j]
+		rfa.currTimestamp = tEnd
+		if i > 0 {
+			rfa.realPrevValue = values[i-1]
+		}
 		value := rc.Func(rfa)
 		rfa.idx++
 		dstValues = append(dstValues, value)
@@ -225,33 +478,102 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	return dstValues
 }
 
-func getMaxPrevInterval(timestamps []int64) int64 {
+func seekFirstTimestampIdxAfter(timestamps []int64, seekTimestamp int64, nHint int) int {
+	if len(timestamps) == 0 || timestamps[0] > seekTimestamp {
+		return 0
+	}
+	startIdx := nHint - 2
+	if startIdx < 0 {
+		startIdx = 0
+	}
+	if startIdx >= len(timestamps) {
+		startIdx = len(timestamps) - 1
+	}
+	endIdx := nHint + 2
+	if endIdx > len(timestamps) {
+		endIdx = len(timestamps)
+	}
+	if startIdx > 0 && timestamps[startIdx] <= seekTimestamp {
+		timestamps = timestamps[startIdx:]
+		endIdx -= startIdx
+	} else {
+		startIdx = 0
+	}
+	if endIdx < len(timestamps) && timestamps[endIdx] > seekTimestamp {
+		timestamps = timestamps[:endIdx]
+	}
+	if len(timestamps) < 16 {
+		// Fast path: the number of timestamps to search is small, so scan them all.
+		for i, timestamp := range timestamps {
+			if timestamp > seekTimestamp {
+				return startIdx + i
+			}
+		}
+		return startIdx + len(timestamps)
+	}
+	// Slow path: too big len(timestamps), so use binary search.
+	i := binarySearchInt64(timestamps, seekTimestamp+1)
+	return startIdx + int(i)
+}
+
+func binarySearchInt64(a []int64, v int64) uint {
+	// Copy-pasted sort.Search from https://golang.org/src/sort/search.go?s=2246:2286#L49
+	i, j := uint(0), uint(len(a))
+	for i < j {
+		h := (i + j) >> 1
+		if h < uint(len(a)) && a[h] < v {
+			i = h + 1
+		} else {
+			j = h
+		}
+	}
+	return i
+}
+
+func getScrapeInterval(timestamps []int64) int64 {
 	if len(timestamps) < 2 {
 		return int64(maxSilenceInterval)
 	}
-	d := (timestamps[len(timestamps)-1] - timestamps[0]) / int64(len(timestamps)-1)
-	if d <= 0 {
+
+	// Estimate scrape interval as 0.6 quantile for the first 100 intervals.
+	h := histogram.GetFast()
+	tsPrev := timestamps[0]
+	timestamps = timestamps[1:]
+	if len(timestamps) > 100 {
+		timestamps = timestamps[:100]
+	}
+	for _, ts := range timestamps {
+		h.Update(float64(ts - tsPrev))
+		tsPrev = ts
+	}
+	scrapeInterval := int64(h.Quantile(0.6))
+	histogram.PutFast(h)
+	if scrapeInterval <= 0 {
 		return int64(maxSilenceInterval)
 	}
-	// Increase d more for smaller scrape intervals in order to hide possible gaps
+	return scrapeInterval
+}
+
+func getMaxPrevInterval(scrapeInterval int64) int64 {
+	// Increase scrapeInterval more for smaller scrape intervals in order to hide possible gaps
 	// when high jitter is present.
 	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/139 .
-	if d <= 2*1000 {
-		return d + 4*d
+	if scrapeInterval <= 2*1000 {
+		return scrapeInterval + 4*scrapeInterval
 	}
-	if d <= 4*1000 {
-		return d + 2*d
+	if scrapeInterval <= 4*1000 {
+		return scrapeInterval + 2*scrapeInterval
 	}
-	if d <= 8*1000 {
-		return d + d
+	if scrapeInterval <= 8*1000 {
+		return scrapeInterval + scrapeInterval
 	}
-	if d <= 16*1000 {
-		return d + d/2
+	if scrapeInterval <= 16*1000 {
+		return scrapeInterval + scrapeInterval/2
 	}
-	if d <= 32*1000 {
-		return d + d/4
+	if scrapeInterval <= 32*1000 {
+		return scrapeInterval + scrapeInterval/4
 	}
-	return d + d/8
+	return scrapeInterval + scrapeInterval/8
 }
 
 func removeCounterResets(values []float64) {
@@ -330,6 +652,15 @@ func newRollupFuncOneArg(rf rollupFunc) newRollupFunc {
 	}
 }
 
+func newRollupFuncTwoArgs(rf rollupFunc) newRollupFunc {
+	return func(args []interface{}) (rollupFunc, error) {
+		if err := expectRollupArgsNum(args, 2); err != nil {
+			return nil, err
+		}
+		return rf, nil
+	}
+}
+
 func newRollupHoltWinters(args []interface{}) (rollupFunc, error) {
 	if err := expectRollupArgsNum(args, 3); err != nil {
 		return nil, err
@@ -438,6 +769,56 @@ func linearRegression(rfa *rollupFuncArg) (float64, float64) {
 	return v, k
 }
 
+func newRollupShareLE(args []interface{}) (rollupFunc, error) {
+	return newRollupShareFilter(args, countFilterLE)
+}
+
+func countFilterLE(values []float64, le float64) int {
+	n := 0
+	for _, v := range values {
+		if v <= le {
+			n++
+		}
+	}
+	return n
+}
+
+func newRollupShareGT(args []interface{}) (rollupFunc, error) {
+	return newRollupShareFilter(args, countFilterGT)
+}
+
+func countFilterGT(values []float64, gt float64) int {
+	n := 0
+	for _, v := range values {
+		if v > gt {
+			n++
+		}
+	}
+	return n
+}
+
+func newRollupShareFilter(args []interface{}, countFilter func(values []float64, limit float64) int) (rollupFunc, error) {
+	if err := expectRollupArgsNum(args, 2); err != nil {
+		return nil, err
+	}
+	limits, err := getScalar(args[1], 1)
+	if err != nil {
+		return nil, err
+	}
+	rf := func(rfa *rollupFuncArg) float64 {
+		// There is no need in handling NaNs here, since they must be cleaned up
+		// before calling rollup funcs.
+		values := rfa.values
+		if len(values) == 0 {
+			return nan
+		}
+		limit := limits[rfa.idx]
+		n := countFilter(values, limit)
+		return float64(n) / float64(len(values))
+	}
+	return rf, nil
+}
+
 func newRollupQuantile(args []interface{}) (rollupFunc, error) {
 	if err := expectRollupArgsNum(args, 2); err != nil {
 		return nil, err
@@ -469,6 +850,21 @@ func newRollupQuantile(args []interface{}) (rollupFunc, error) {
 	return rf, nil
 }
 
+func rollupHistogram(rfa *rollupFuncArg) float64 {
+	values := rfa.values
+	tsm := rfa.tsm
+	tsm.h.Reset()
+	for _, v := range values {
+		tsm.h.Update(v)
+	}
+	idx := rfa.idx
+	tsm.h.VisitNonZeroBuckets(func(vmrange string, count uint64) {
+		ts := tsm.GetOrCreateTimeseries(vmrange)
+		ts.Values[idx] = float64(count)
+	})
+	return nan
+}
+
 func rollupAvg(rfa *rollupFuncArg) float64 {
 	// Do not use `Rapid calculation methods` at https://en.wikipedia.org/wiki/Standard_deviation,
 	// since it is slower and has no significant benefits in precision.
@@ -491,7 +887,10 @@ func rollupMin(rfa *rollupFuncArg) float64 {
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return rfa.prevValue
+		// Do not take into account rfa.prevValue, since it may lead
+		// to inconsistent results comparing to Prometheus on broken time series
+		// with irregular data points.
+		return nan
 	}
 	minValue := values[0]
 	for _, v := range values {
@@ -507,7 +906,10 @@ func rollupMax(rfa *rollupFuncArg) float64 {
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return rfa.prevValue
+		// Do not take into account rfa.prevValue, since it may lead
+		// to inconsistent results comparing to Prometheus on broken time series
+		// with irregular data points.
+		return nan
 	}
 	maxValue := values[0]
 	for _, v := range values {
@@ -518,12 +920,53 @@ func rollupMax(rfa *rollupFuncArg) float64 {
 	return maxValue
 }
 
+func rollupTmin(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	timestamps := rfa.timestamps
+	if len(values) == 0 {
+		return nan
+	}
+	minValue := values[0]
+	minTimestamp := timestamps[0]
+	for i, v := range values {
+		if v < minValue {
+			minValue = v
+			minTimestamp = timestamps[i]
+		}
+	}
+	return float64(minTimestamp) * 1e-3
+}
+
+func rollupTmax(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	timestamps := rfa.timestamps
+	if len(values) == 0 {
+		return nan
+	}
+	maxValue := values[0]
+	maxTimestamp := timestamps[0]
+	for i, v := range values {
+		if v > maxValue {
+			maxValue = v
+			maxTimestamp = timestamps[i]
+		}
+	}
+	return float64(maxTimestamp) * 1e-3
+}
+
 func rollupSum(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return rfa.prevValue
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	var sum float64
 	for _, v := range values {
@@ -560,6 +1003,13 @@ func rollupGeomean(rfa *rollupFuncArg) float64 {
 	return math.Pow(p, 1/float64(len(values)))
 }
 
+func rollupAbsent(rfa *rollupFuncArg) float64 {
+	if len(rfa.values) == 0 {
+		return 1
+	}
+	return nan
+}
+
 func rollupCount(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
@@ -607,6 +1057,14 @@ func rollupStdvar(rfa *rollupFuncArg) float64 {
 }
 
 func rollupDelta(rfa *rollupFuncArg) float64 {
+	return rollupDeltaInternal(rfa, false)
+}
+
+func rollupIncrease(rfa *rollupFuncArg) float64 {
+	return rollupDeltaInternal(rfa, true)
+}
+
+func rollupDeltaInternal(rfa *rollupFuncArg, canUseRealPrevValue bool) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
@@ -615,10 +1073,14 @@ func rollupDelta(rfa *rollupFuncArg) float64 {
 		if len(values) == 0 {
 			return nan
 		}
-		prevValue = values[0]
-		values = values[1:]
+		// Assume that the previous non-existing value was 0.
+		prevValue = 0
+		if canUseRealPrevValue && !math.IsNaN(rfa.prevValue) {
+			prevValue = rfa.prevValue
+		}
 	}
 	if len(values) == 0 {
+		// Assume that the value didn't change on the given interval.
 		return 0
 	}
 	return values[len(values)-1] - prevValue
@@ -632,6 +1094,7 @@ func rollupIdelta(rfa *rollupFuncArg) float64 {
 		if math.IsNaN(rfa.prevValue) {
 			return nan
 		}
+		// Assume that the value didn't change on the given interval.
 		return 0
 	}
 	lastValue := values[len(values)-1]
@@ -639,7 +1102,8 @@ func rollupIdelta(rfa *rollupFuncArg) float64 {
 	if len(values) == 0 {
 		prevValue := rfa.prevValue
 		if math.IsNaN(prevValue) {
-			return 0
+			// Assume that the previous non-existing value was 0.
+			return lastValue
 		}
 		return lastValue - prevValue
 	}
@@ -664,12 +1128,23 @@ func rollupDerivFast(rfa *rollupFuncArg) float64 {
 		if len(values) == 0 {
 			return nan
 		}
+		if len(values) == 1 {
+			// It is impossible to determine the duration during which the value changed
+			// from 0 to the current value.
+			// The following attempts didn't work well:
+			// - using scrape interval as the duration. It fails on Prometheus restarts when it
+			//   skips scraping for the counter. This results in too high rate() value for the first point
+			//   after Prometheus restarts.
+			// - using window or step as the duration. It results in too small rate() values for the first
+			//   points of time series.
+			//
+			// So just return nan
+			return nan
+		}
 		prevValue = values[0]
 		prevTimestamp = timestamps[0]
-		values = values[1:]
-		timestamps = timestamps[1:]
-	}
-	if len(values) == 0 {
+	} else if len(values) == 0 {
+		// Assume that the value didn't change on the given interval.
 		return 0
 	}
 	vEnd := values[len(values)-1]
@@ -684,11 +1159,24 @@ func rollupIderiv(rfa *rollupFuncArg) float64 {
 	// before calling rollup funcs.
 	values := rfa.values
 	timestamps := rfa.timestamps
-	if len(values) == 0 {
-		if math.IsNaN(rfa.prevValue) {
+	if len(values) < 2 {
+		if len(values) == 0 {
 			return nan
 		}
-		return 0
+		if math.IsNaN(rfa.prevValue) {
+			// It is impossible to determine the duration during which the value changed
+			// from 0 to the current value.
+			// The following attempts didn't work well:
+			// - using scrape interval as the duration. It fails on Prometheus restarts when it
+			//   skips scraping for the counter. This results in too high rate() value for the first point
+			//   after Prometheus restarts.
+			// - using window or step as the duration. It results in too small rate() values for the first
+			//   points of time series.
+			//
+			// So just return nan
+			return nan
+		}
+		return (values[0] - rfa.prevValue) / (float64(timestamps[0]-rfa.prevTimestamp) * 1e-3)
 	}
 	vEnd := values[len(values)-1]
 	tEnd := timestamps[len(timestamps)-1]
@@ -712,7 +1200,49 @@ func rollupIderiv(rfa *rollupFuncArg) float64 {
 	}
 	dv := vEnd - vStart
 	dt := tEnd - tStart
-	return dv / (float64(dt) / 1000)
+	return dv / (float64(dt) * 1e-3)
+}
+
+func rollupLifetime(rfa *rollupFuncArg) float64 {
+	// Calculate the duration between the first and the last data points.
+	timestamps := rfa.timestamps
+	if math.IsNaN(rfa.prevValue) {
+		if len(timestamps) < 2 {
+			return nan
+		}
+		return float64(timestamps[len(timestamps)-1]-timestamps[0]) * 1e-3
+	}
+	if len(timestamps) == 0 {
+		return nan
+	}
+	return float64(timestamps[len(timestamps)-1]-rfa.prevTimestamp) * 1e-3
+}
+
+func rollupLag(rfa *rollupFuncArg) float64 {
+	// Calculate the duration between the current timestamp and the last data point.
+	timestamps := rfa.timestamps
+	if len(timestamps) == 0 {
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return float64(rfa.currTimestamp-rfa.prevTimestamp) * 1e-3
+	}
+	return float64(rfa.currTimestamp-timestamps[len(timestamps)-1]) * 1e-3
+}
+
+func rollupScrapeInterval(rfa *rollupFuncArg) float64 {
+	// Calculate the average interval between data points.
+	timestamps := rfa.timestamps
+	if math.IsNaN(rfa.prevValue) {
+		if len(timestamps) < 2 {
+			return nan
+		}
+		return float64(timestamps[len(timestamps)-1]-timestamps[0]) * 1e-3 / float64(len(timestamps)-1)
+	}
+	if len(timestamps) == 0 {
+		return nan
+	}
+	return (float64(timestamps[len(timestamps)-1]-rfa.prevTimestamp) * 1e-3) / float64(len(timestamps))
 }
 
 func rollupChanges(rfa *rollupFuncArg) float64 {
@@ -738,6 +1268,37 @@ func rollupChanges(rfa *rollupFuncArg) float64 {
 	return float64(n)
 }
 
+func rollupIncreases(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	if len(values) == 0 {
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
+	}
+	prevValue := rfa.prevValue
+	if math.IsNaN(prevValue) {
+		prevValue = values[0]
+		values = values[1:]
+	}
+	if len(values) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range values {
+		if v > prevValue {
+			n++
+		}
+		prevValue = v
+	}
+	return float64(n)
+}
+
+// `decreases_over_time` logic is the same as `resets` logic.
+var rollupDecreases = rollupResets
+
 func rollupResets(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
@@ -767,16 +1328,13 @@ func rollupResets(rfa *rollupFuncArg) float64 {
 }
 
 func rollupFirst(rfa *rollupFuncArg) float64 {
-	// See https://prometheus.io/docs/prometheus/latest/querying/basics/#staleness
-	v := rfa.prevValue
-	if !math.IsNaN(v) {
-		return v
-	}
-
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
+		// Do not take into account rfa.prevValue, since it may lead
+		// to inconsistent results comparing to Prometheus on broken time series
+		// with irregular data points.
 		return nan
 	}
 	return values[0]
@@ -789,7 +1347,10 @@ func rollupLast(rfa *rollupFuncArg) float64 {
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return rfa.prevValue
+		// Do not take into account rfa.prevValue, since it may lead
+		// to inconsistent results comparing to Prometheus on broken time series
+		// with irregular data points.
+		return nan
 	}
 	return values[len(values)-1]
 }
@@ -840,6 +1401,8 @@ func rollupIntegrate(rfa *rollupFuncArg) float64 {
 		timestamp := timestamps[i]
 		dt := float64(timestamp-prevTimestamp) * 1e-3
 		sum += 0.5 * (v + prevValue) * dt
+		prevTimestamp = timestamp
+		prevValue = v
 	}
 	return sum
 }

app/vmselect/promql/rollup_result_cache.go

@@ -4,14 +4,16 @@ import (
 	"crypto/rand"
 	"flag"
 	"fmt"
-	"runtime"
 	"sync"
 	"sync/atomic"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -19,7 +21,7 @@ import (
 var disableCache = flag.Bool("search.disableCache", false, "Whether to disable response caching. This may be useful during data backfilling")
 
 var rollupResultCacheV = &rollupResultCache{
-	fastcache.New(1024 * 1024), // This is a cache for testing.
+	c: workingsetcache.New(1024*1024, time.Hour), // This is a cache for testing.
 }
 var rollupResultCachePath string
 
@@ -43,12 +45,13 @@ var (
 func InitRollupResultCache(cachePath string) {
 	rollupResultCachePath = cachePath
 	startTime := time.Now()
-	var c *fastcache.Cache
+	cacheSize := getRollupResultCacheSize()
+	var c *workingsetcache.Cache
 	if len(rollupResultCachePath) > 0 {
 		logger.Infof("loading rollupResult cache from %q...", rollupResultCachePath)
-		c = fastcache.LoadFromFileOrNew(rollupResultCachePath, getRollupResultCacheSize())
+		c = workingsetcache.Load(rollupResultCachePath, cacheSize, time.Hour)
 	} else {
-		c = fastcache.New(getRollupResultCacheSize())
+		c = workingsetcache.New(cacheSize, time.Hour)
 	}
 	if *disableCache {
 		c.Reset()
@@ -96,25 +99,26 @@ func InitRollupResultCache(cachePath string) {
 // StopRollupResultCache closes the rollupResult cache.
 func StopRollupResultCache() {
 	if len(rollupResultCachePath) == 0 {
-		rollupResultCacheV.c.Reset()
+		rollupResultCacheV.c.Stop()
+		rollupResultCacheV.c = nil
 		return
 	}
-	gomaxprocs := runtime.GOMAXPROCS(-1)
 	logger.Infof("saving rollupResult cache to %q...", rollupResultCachePath)
 	startTime := time.Now()
-	if err := rollupResultCacheV.c.SaveToFileConcurrent(rollupResultCachePath, gomaxprocs); err != nil {
+	if err := rollupResultCacheV.c.Save(rollupResultCachePath); err != nil {
 		logger.Errorf("cannot close rollupResult cache at %q: %s", rollupResultCachePath, err)
-	} else {
-		var fcs fastcache.Stats
-		rollupResultCacheV.c.UpdateStats(&fcs)
-		rollupResultCacheV.c.Reset()
-		logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, sizeBytes: %d",
-			rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
+		return
 	}
+	var fcs fastcache.Stats
+	rollupResultCacheV.c.UpdateStats(&fcs)
+	rollupResultCacheV.c.Stop()
+	rollupResultCacheV.c = nil
+	logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, sizeBytes: %d",
+		rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
 }
 
 type rollupResultCache struct {
-	c *fastcache.Cache
+	c *workingsetcache.Cache
 }
 
 var rollupResultCacheResets = metrics.NewCounter(`vm_cache_resets_total{type="promql/rollupResult"}`)
@@ -125,7 +129,7 @@ func ResetRollupResultCache() {
 	rollupResultCacheV.c.Reset()
 }
 
-func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64) (tss []*timeseries, newStart int64) {
+func (rrc *rollupResultCache) Get(ec *EvalConfig, expr metricsql.Expr, window int64) (tss []*timeseries, newStart int64) {
 	if *disableCache || !ec.mayCache() {
 		return nil, ec.Start
 	}
@@ -134,7 +138,7 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	bb := bbPool.Get()
 	defer bbPool.Put(bb)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	if len(metainfoBuf) == 0 {
 		return nil, ec.Start
@@ -148,15 +152,23 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 		return nil, ec.Start
 	}
 	bb.B = key.Marshal(bb.B[:0])
-	resultBuf := rrc.c.GetBig(nil, bb.B)
-	if len(resultBuf) == 0 {
+	compressedResultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(compressedResultBuf)
+	compressedResultBuf.B = rrc.c.GetBig(compressedResultBuf.B[:0], bb.B)
+	if len(compressedResultBuf.B) == 0 {
 		mi.RemoveKey(key)
 		metainfoBuf = mi.Marshal(metainfoBuf[:0])
-		bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+		bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 		rrc.c.Set(bb.B, metainfoBuf)
 		return nil, ec.Start
 	}
-	tss, err := unmarshalTimeseriesFast(resultBuf)
+	// Decompress into newly allocated byte slice, since tss returned from unmarshalTimeseriesFast
+	// refers to the byte slice, so it cannot be returned to the resultBufPool.
+	resultBuf, err := encoding.DecompressZSTD(nil, compressedResultBuf.B)
+	if err != nil {
+		logger.Panicf("BUG: cannot decompress resultBuf from rollupResultCache: %s; it looks like it was improperly saved", err)
+	}
+	tss, err = unmarshalTimeseriesFast(resultBuf)
 	if err != nil {
 		logger.Panicf("BUG: cannot unmarshal timeseries from rollupResultCache: %s; it looks like it was improperly saved", err)
 	}
@@ -196,7 +208,9 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	return tss, newStart
 }
 
-func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64, tss []*timeseries) {
+var resultBufPool bytesutil.ByteBufferPool
+
+func (rrc *rollupResultCache) Put(ec *EvalConfig, expr metricsql.Expr, window int64, tss []*timeseries) {
 	if *disableCache || len(tss) == 0 || !ec.mayCache() {
 		return
 	}
@@ -227,11 +241,16 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 
 	// Store tss in the cache.
 	maxMarshaledSize := getRollupResultCacheSize() / 4
-	tssMarshaled := marshalTimeseriesFast(tss, maxMarshaledSize, ec.Step)
-	if tssMarshaled == nil {
+	resultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(resultBuf)
+	resultBuf.B = marshalTimeseriesFast(resultBuf.B[:0], tss, maxMarshaledSize, ec.Step)
+	if len(resultBuf.B) == 0 {
 		tooBigRollupResults.Inc()
 		return
 	}
+	compressedResultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(compressedResultBuf)
+	compressedResultBuf.B = encoding.CompressZSTDLevel(compressedResultBuf.B[:0], resultBuf.B, 1)
 
 	bb := bbPool.Get()
 	defer bbPool.Put(bb)
@@ -240,9 +259,9 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 	key.prefix = rollupResultCacheKeyPrefix
 	key.suffix = atomic.AddUint64(&rollupResultCacheKeySuffix, 1)
 	bb.B = key.Marshal(bb.B[:0])
-	rrc.c.SetBig(bb.B, tssMarshaled)
+	rrc.c.SetBig(bb.B, compressedResultBuf.B)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	var mi rollupResultCacheMetainfo
 	if len(metainfoBuf) > 0 {
@@ -270,23 +289,13 @@ var (
 var tooBigRollupResults = metrics.NewCounter("vm_too_big_rollup_results_total")
 
 // Increment this value every time the format of the cache changes.
-const rollupResultCacheVersion = 5
+const rollupResultCacheVersion = 7
 
-func marshalRollupResultCacheKey(dst []byte, funcName string, me *metricExpr, iafc *incrementalAggrFuncContext, window, step int64) []byte {
+func marshalRollupResultCacheKey(dst []byte, expr metricsql.Expr, window, step int64) []byte {
 	dst = append(dst, rollupResultCacheVersion)
-	if iafc == nil {
-		dst = append(dst, 0)
-	} else {
-		dst = append(dst, 1)
-		dst = iafc.ae.AppendString(dst)
-	}
-	dst = encoding.MarshalUint64(dst, uint64(len(funcName)))
-	dst = append(dst, funcName...)
 	dst = encoding.MarshalInt64(dst, window)
 	dst = encoding.MarshalInt64(dst, step)
-	for i := range me.TagFilters {
-		dst = me.TagFilters[i].Marshal(dst)
-	}
+	dst = expr.AppendString(dst)
 	return dst
 }
 

app/vmselect/promql/rollup_result_cache_test.go

@@ -3,12 +3,12 @@ package promql
 import (
 	"testing"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
 
 func TestRollupResultCache(t *testing.T) {
 	ResetRollupResultCache()
-	funcName := "foo"
 	window := int64(456)
 	ec := &EvalConfig{
 		Start: 1000,
@@ -17,21 +17,24 @@ func TestRollupResultCache(t *testing.T) {
 
 		MayCache: true,
 	}
-	me := &metricExpr{
-		TagFilters: []storage.TagFilter{{
-			Key:   []byte("aaa"),
-			Value: []byte("xxx"),
+	me := &metricsql.MetricExpr{
+		LabelFilters: []metricsql.LabelFilter{{
+			Label: "aaa",
+			Value: "xxx",
 		}},
 	}
-	iafc := &incrementalAggrFuncContext{
-		ae: &aggrFuncExpr{
-			Name: "foobar",
-		},
+	fe := &metricsql.FuncExpr{
+		Name: "foo",
+		Args: []metricsql.Expr{me},
+	}
+	ae := &metricsql.AggrFuncExpr{
+		Name: "foobar",
+		Args: []metricsql.Expr{fe},
 	}
 
 	// Try obtaining an empty value.
 	t.Run("empty", func(t *testing.T) {
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != ec.Start {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, ec.Start)
 		}
@@ -41,7 +44,7 @@ func TestRollupResultCache(t *testing.T) {
 	})
 
 	// Store timeseries overlapping with start
-	t.Run("start-overlap-no-iafc", func(t *testing.T) {
+	t.Run("start-overlap-no-ae", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
 			{
@@ -49,8 +52,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1400 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
 		}
@@ -62,7 +65,7 @@ func TestRollupResultCache(t *testing.T) {
 		}
 		testTimeseriesEqual(t, tss, tssExpected)
 	})
-	t.Run("start-overlap-with-iafc", func(t *testing.T) {
+	t.Run("start-overlap-with-ae", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
 			{
@@ -70,8 +73,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, iafc, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, iafc, window)
+		rollupResultCacheV.Put(ec, ae, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, ae, window)
 		if newStart != 1400 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
 		}
@@ -93,8 +96,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{333, 0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -112,8 +115,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -131,8 +134,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -150,8 +153,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -169,8 +172,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2, 3, 4, 5, 6, 7},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
@@ -192,8 +195,8 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
@@ -217,8 +220,8 @@ func TestRollupResultCache(t *testing.T) {
 			}
 			tss = append(tss, ts)
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
-		tssResult, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss)
+		tssResult, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
@@ -246,10 +249,10 @@ func TestRollupResultCache(t *testing.T) {
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss1)
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss2)
-		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss3)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
+		rollupResultCacheV.Put(ec, fe, window, tss1)
+		rollupResultCacheV.Put(ec, fe, window, tss2)
+		rollupResultCacheV.Put(ec, fe, window, tss3)
+		tss, newStart := rollupResultCacheV.Get(ec, fe, window)
 		if newStart != 1400 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
 		}
@@ -388,7 +391,7 @@ func testTimeseriesEqual(t *testing.T, tss, tssExpected []*timeseries) {
 	}
 	for i, ts := range tss {
 		tsExpected := tssExpected[i]
-		testMetricNamesEqual(t, &ts.MetricName, &tsExpected.MetricName)
+		testMetricNamesEqual(t, &ts.MetricName, &tsExpected.MetricName, i)
 		testRowsEqual(t, ts.Values, ts.Timestamps, tsExpected.Values, tsExpected.Timestamps)
 	}
 }

app/vmselect/promql/rollup_test.go

@@ -3,6 +3,8 @@ package promql
 import (
 	"math"
 	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 )
 
 var (
@@ -45,8 +47,19 @@ func TestRollupIderivDuplicateTimestamps(t *testing.T) {
 		timestamps: []int64{100},
 	}
 	n = rollupIderiv(rfa)
-	if n != 0 {
-		t.Fatalf("unexpected value; got %v; want %v", n, 0)
+	if !math.IsNaN(n) {
+		t.Fatalf("unexpected value; got %v; want %v", n, nan)
+	}
+
+	rfa = &rollupFuncArg{
+		prevTimestamp: 90,
+		prevValue:     10,
+		values:        []float64{15},
+		timestamps:    []int64{100},
+	}
+	n = rollupIderiv(rfa)
+	if n != 500 {
+		t.Fatalf("unexpected value; got %v; want %v", n, 500)
 	}
 
 	rfa = &rollupFuncArg{
@@ -146,7 +159,7 @@ func TestDerivValues(t *testing.T) {
 	testRowsEqual(t, values, timestamps, valuesExpected, timestamps)
 }
 
-func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpected *metricExpr, vExpected float64) {
+func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpected *metricsql.MetricExpr, vExpected float64) {
 	t.Helper()
 	nrf := getRollupFunc(funcName)
 	if nrf == nil {
@@ -171,13 +184,60 @@ func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpecte
 				t.Fatalf("unexpected value; got %v; want %v", v, vExpected)
 			}
 		} else {
-			if v != vExpected {
+			eps := math.Abs(v - vExpected)
+			if eps > 1e-14 {
 				t.Fatalf("unexpected value; got %v; want %v", v, vExpected)
 			}
 		}
 	}
 }
 
+func TestRollupShareLEOverTime(t *testing.T) {
+	f := func(le, vExpected float64) {
+		t.Helper()
+		les := []*timeseries{{
+			Values:     []float64{le},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, les}
+		testRollupFunc(t, "share_le_over_time", args, &me, vExpected)
+	}
+
+	f(-123, 0)
+	f(0, 0)
+	f(10, 0)
+	f(12, 0.08333333333333333)
+	f(30, 0.16666666666666666)
+	f(50, 0.75)
+	f(100, 0.9166666666666666)
+	f(123, 1)
+	f(1000, 1)
+}
+
+func TestRollupShareGTOverTime(t *testing.T) {
+	f := func(gt, vExpected float64) {
+		t.Helper()
+		gts := []*timeseries{{
+			Values:     []float64{gt},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, gts}
+		testRollupFunc(t, "share_gt_over_time", args, &me, vExpected)
+	}
+
+	f(-123, 1)
+	f(0, 1)
+	f(10, 1)
+	f(12, 0.9166666666666666)
+	f(30, 0.8333333333333334)
+	f(50, 0.25)
+	f(100, 0.08333333333333333)
+	f(123, 0)
+	f(1000, 0)
+}
+
 func TestRollupQuantileOverTime(t *testing.T) {
 	f := func(phi, vExpected float64) {
 		t.Helper()
@@ -185,8 +245,8 @@ func TestRollupQuantileOverTime(t *testing.T) {
 			Values:     []float64{phi},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{phis, &rollupExpr{Expr: &me}}
+		var me metricsql.MetricExpr
+		args := []interface{}{phis, &metricsql.RollupExpr{Expr: &me}}
 		testRollupFunc(t, "quantile_over_time", args, &me, vExpected)
 	}
 
@@ -207,8 +267,8 @@ func TestRollupPredictLinear(t *testing.T) {
 			Values:     []float64{sec},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}, secs}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, secs}
 		testRollupFunc(t, "predict_linear", args, &me, vExpected)
 	}
 
@@ -229,8 +289,8 @@ func TestRollupHoltWinters(t *testing.T) {
 			Values:     []float64{tf},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}, sfs, tfs}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, sfs, tfs}
 		testRollupFunc(t, "holt_winters", args, &me, vExpected)
 	}
 
@@ -253,24 +313,26 @@ func TestRollupHoltWinters(t *testing.T) {
 func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f := func(funcName string, vExpected float64) {
 		t.Helper()
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}}
 		testRollupFunc(t, funcName, args, &me, vExpected)
 	}
 
 	f("default_rollup", 34)
 	f("changes", 11)
-	f("delta", -89)
+	f("delta", 34)
 	f("deriv", -266.85860231406065)
 	f("deriv_fast", -712)
 	f("idelta", 0)
-	f("increase", 275)
+	f("increase", 398)
 	f("irate", 0)
 	f("rate", 2200)
 	f("resets", 5)
 	f("avg_over_time", 47.083333333333336)
 	f("min_over_time", 12)
 	f("max_over_time", 123)
+	f("tmin_over_time", 0.08)
+	f("tmax_over_time", 0.005)
 	f("sum_over_time", 565)
 	f("sum2_over_time", 37951)
 	f("geomean_over_time", 39.33466603189148)
@@ -279,9 +341,11 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f("stdvar_over_time", 945.7430555555555)
 	f("first_over_time", 123)
 	f("last_over_time", 34)
-	f("integrate", 61.0275)
+	f("integrate", 5.4705)
 	f("distinct_over_time", 8)
 	f("ideriv", 0)
+	f("decreases_over_time", 5)
+	f("increases_over_time", 5)
 }
 
 func TestRollupNewRollupFuncError(t *testing.T) {
@@ -313,7 +377,7 @@ func TestRollupNewRollupFuncError(t *testing.T) {
 		Values:     []float64{321},
 		Timestamps: []int64{123},
 	}}
-	me := &metricExpr{}
+	me := &metricsql.MetricExpr{}
 	f("holt_winters", []interface{}{123, 123, 321})
 	f("holt_winters", []interface{}{me, 123, 321})
 	f("holt_winters", []interface{}{me, scalarTs, 321})
@@ -347,7 +411,7 @@ func TestRollupNoWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{2, 0, 0, 0, 0, 0, 0, 0}
+		valuesExpected := []float64{2, 0, 0, 0, nan, nan, nan, nan}
 		timestampsExpected := []int64{120, 124, 128, 132, 136, 140, 144, 148}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -378,7 +442,7 @@ func TestRollupWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{34, 34, 34, nan}
+		valuesExpected := []float64{nan, nan, nan, nan}
 		timestampsExpected := []int64{161, 171, 181, 191}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -395,7 +459,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 123, 123, 123, 34, 34}
+		valuesExpected := []float64{nan, 123, nan, 34, nan, 44}
 		timestampsExpected := []int64{0, 5, 10, 15, 20, 25}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -409,7 +473,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{12, 44, 34, 34}
+		valuesExpected := []float64{44, 32, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -423,7 +487,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, nan, 123, 54, 44}
+		valuesExpected := []float64{nan, nan, 123, 34, 32}
 		timestampsExpected := []int64{-50, 0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -454,7 +518,7 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{44, 34, 34, 34}
+		valuesExpected := []float64{44, 34, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -468,12 +532,57 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 54, 44, 34}
+		valuesExpected := []float64{nan, 54, 44, nan}
 		timestampsExpected := []int64{0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 }
 
+func TestRollupFuncsLookbackDelta(t *testing.T) {
+	t.Run("1", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 1,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("7", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 7,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("0", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+}
+
 func TestRollupFuncsNoWindow(t *testing.T) {
 	t.Run("first", func(t *testing.T) {
 		rc := rollupConfig{
@@ -485,7 +594,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 123, 21, 12, 34}
+		valuesExpected := []float64{nan, 123, 54, 44, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -555,7 +664,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, -102, -9, 22, 0}
+		valuesExpected := []float64{nan, nan, -9, 22, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -569,10 +678,80 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{0, 33, -87, 0}
+		valuesExpected := []float64{123, 33, -87, 0}
 		timestampsExpected := []int64{10, 50, 90, 130}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("lag", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLag,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.004, 0, 0, 0.03}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("lifetime_1", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLifetime,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.031, 0.044, 0.04, 0.01}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("lifetime_2", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLifetime,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 200,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.031, 0.075, 0.115, 0.125}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("scrape_interval_1", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupScrapeInterval,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.010333333333333333, 0.011, 0.013333333333333334, 0.01}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("scrape_interval_2", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupScrapeInterval,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 80,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.010333333333333333, 0.010714285714285714, 0.012, 0.0125}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("changes", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupChanges,
@@ -643,6 +822,20 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("deriv_fast", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupDerivFast,
+			Start:  0,
+			End:    20,
+			Step:   4,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, nan, nan, 0, -8900, 0}
+		timestampsExpected := []int64{0, 4, 8, 12, 16, 20}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("ideriv", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupIderiv,
@@ -681,11 +874,11 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 4.6035, 4.3934999999999995, 2.166, 0.34}
+		valuesExpected := []float64{nan, 1.526, 2.2795, 1.325, 0.34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("distinct", func(t *testing.T) {
+	t.Run("distinct_over_time_1", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupDistinct,
 			Start:  0,
@@ -699,6 +892,41 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("distinct_over_time_2", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupDistinct,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 80,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 4, 7, 6, 3}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+}
+
+func TestRollupBigNumberOfValues(t *testing.T) {
+	const srcValuesCount = 1e4
+	rc := rollupConfig{
+		Func:   rollupDefault,
+		End:    srcValuesCount,
+		Step:   srcValuesCount / 5,
+		Window: srcValuesCount / 4,
+	}
+	rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+	srcValues := make([]float64, srcValuesCount)
+	srcTimestamps := make([]int64, srcValuesCount)
+	for i := 0; i < srcValuesCount; i++ {
+		srcValues[i] = float64(i)
+		srcTimestamps[i] = int64(i / 2)
+	}
+	values := rc.Do(nil, srcValues, srcTimestamps)
+	valuesExpected := []float64{1, 4001, 8001, 9999, nan, nan}
+	timestampsExpected := []int64{0, 2000, 4000, 6000, 8000, 10000}
+	testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 }
 
 func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExpected []float64, timestampsExpected []int64) {
@@ -729,7 +957,7 @@ func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExp
 			}
 			continue
 		}
-		if v != vExpected {
+		if math.Abs(v-vExpected) > 1e-15 {
 			t.Fatalf("unexpected value at values[%d]; got %f; want %f\nvalues=\n%v\nvaluesExpected=\n%v",
 				i, v, vExpected, values, valuesExpected)
 		}

app/vmselect/promql/timeseries.go

@@ -76,7 +76,7 @@ func putTimeseries(ts *timeseries) {
 
 var timeseriesPool sync.Pool
 
-func marshalTimeseriesFast(tss []*timeseries, maxSize int, step int64) []byte {
+func marshalTimeseriesFast(dst []byte, tss []*timeseries, maxSize int, step int64) []byte {
 	if len(tss) == 0 {
 		logger.Panicf("BUG: tss cannot be empty")
 	}
@@ -92,13 +92,13 @@ func marshalTimeseriesFast(tss []*timeseries, maxSize int, step int64) []byte {
 
 	if size > maxSize {
 		// Do not marshal tss, since it would occupy too much space
-		return nil
+		return dst
 	}
 
 	// Allocate the buffer for the marshaled tss before its' marshaling.
 	// This should reduce memory fragmentation and memory usage.
-	dst := make([]byte, 0, size)
-	dst = marshalFastTimestamps(dst, tss[0].Timestamps)
+	dst = bytesutil.Resize(dst, size)
+	dst = marshalFastTimestamps(dst[:0], tss[0].Timestamps)
 	for _, ts := range tss {
 		dst = ts.marshalFastNoTimestamps(dst)
 	}
@@ -288,7 +288,6 @@ func marshalMetricTagsFast(dst []byte, tags []storage.Tag) []byte {
 }
 
 func marshalMetricNameSorted(dst []byte, mn *storage.MetricName) []byte {
-	// Do not marshal AccountID and ProjectID, since they are unused.
 	dst = marshalBytesFast(dst, mn.MetricGroup)
 	sortMetricTags(mn.Tags)
 	dst = marshalMetricTagsFast(dst, mn.Tags)

app/vmselect/promql/timeseries_test.go

@@ -74,7 +74,7 @@ func TestTimeseriesMarshalUnmarshalFast(t *testing.T) {
 
 			tssOrig = append(tssOrig, &ts)
 		}
-		buf := marshalTimeseriesFast(tssOrig, 1e6, 123)
+		buf := marshalTimeseriesFast(nil, tssOrig, 1e6, 123)
 		tssGot, err := unmarshalTimeseriesFast(buf)
 		if err != nil {
 			t.Fatalf("error in unmarshalTimeseriesFast: %s", err)