lib/fs: add MustStopDirRemover for waiting until pending directories are removed on graceful shutdown

This patch is mainly required for laggy NFS. See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/162

.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,30 @@
+---
+name: Bug report
+about: Create a report to help us improve
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+**Describe the bug**
+A clear and concise description of what the bug is.
+
+**To Reproduce**
+Steps to reproduce the behavior
+
+**Expected behavior**
+A clear and concise description of what you expected to happen.
+
+**Screenshots**
+If applicable, add screenshots to help explain your problem.
+
+**Version**
+The line returned when passing `--version` command line flag to binary. For example:
+```
+$ ./victoria-metrics-prod --version
+victoria-metrics-20190730-121249-heads-single-node-0-g671d9e55
+```
+
+**Additional context**
+Add any other context about the problem here such as error logs, `/metrics` output, screenshots from [the official Grafana dashboard for VictoriaMetrics](https://grafana.com/dashboards/10229).

.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,20 @@
+---
+name: Feature request
+about: Suggest an idea for this project
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+**Is your feature request related to a problem? Please describe.**
+A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]
+
+**Describe the solution you'd like**
+A clear and concise description of what you want to happen.
+
+**Describe alternatives you've considered**
+A clear and concise description of any alternative solutions or features you've considered.
+
+**Additional context**
+Add any other context or screenshots about the feature request here.

.github/workflows/main.yml

@@ -0,0 +1,42 @@
+name: main
+on:
+  - push
+jobs:
+  build:
+    name: Build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Setup Go
+        uses: actions/setup-go@v1
+        with:
+          go-version: 1.12
+        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 victoria-metrics
+            make victoria-metrics-pure
+            make victoria-metrics-arm
+            make victoria-metrics-arm64
+            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@v1.0.0
+        with:
+          token: ${{secrets.CODECOV_TOKEN}}
+          file: ./coverage.txt
+

.gitignore

@@ -9,3 +9,7 @@
 /victoria-metrics-data
 /vmstorage-data
 /vmselect-cache
+/package/temp-deb-*
+/package/temp-rpm-*
+/package/*.deb
+/package/*.rpm

.travis.yml

@@ -1,22 +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 victoria-metrics
-
-after_success:
-  - bash <(curl -s https://codecov.io/bash)

CODE_OF_CONDUCT.md

@@ -0,0 +1,76 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to making participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+ advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+ address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+ professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies both within project spaces and in public spaces
+when an individual is representing the project or its community. Examples of
+representing a project or community include using an official project e-mail
+address, posting via an official social media account, or acting as an appointed
+representative at an online or offline event. Representation of a project may be
+further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at info@victoriametrics.com. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

CONTRIBUTING.md

@@ -0,0 +1,16 @@
+If you like VictoriaMetrics and want to contribute, then we need the following:
+
+- Filing issues and feature requests [here](https://github.com/VictoriaMetrics/VictoriaMetrics/issues).
+- Spreading a word about VictoriaMetrics: conference talks, articles, comments, experience sharing with colleagues.
+- Updating documentation.
+
+We are open to third-party pull requests provided they follow [KISS design principle](https://en.wikipedia.org/wiki/KISS_principle):
+
+- Prefer simple code and architecture.
+- Avoid complex abstractions.
+- Avoid magic code and fancy algorithms.
+- Avoid [big external dependencies](https://medium.com/@valyala/stripping-dependency-bloat-in-victoriametrics-docker-image-983fb5912b0d).
+- Minimize the number of moving parts in the distributed system.
+- Avoid automated decisions, which may hurt cluster availability, consistency or performance.
+
+Adhering `KISS` principle simplifies the resulting code and architecture, so it can be reviewed, understood and verified by many people.

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),)
@@ -50,25 +50,37 @@ errcheck: install-errcheck
 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 -mod=vendor ./lib/...
-	GO111MODULE=on go test -mod=vendor ./app/...
+	GO111MODULE=on go test -tags=integration -mod=vendor ./lib/... ./app/...
 
-test_full:
+test-pure:
+	GO111MODULE=on CGO_ENABLED=0 go test -tags=integration -mod=vendor ./lib/... ./app/...
+
+test-full:
 	GO111MODULE=on 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/...
 
+benchmark-pure:
+	GO111MODULE=on CGO_ENABLED=0 go test -mod=vendor -bench=. ./lib/...
+	GO111MODULE=on CGO_ENABLED=0 go test -mod=vendor -bench=. ./app/...
+
 vendor-update:
 	GO111MODULE=on go get -u ./lib/...
 	GO111MODULE=on go get -u ./app/...
 	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
 

README.md

@@ -1,14 +1,15 @@
 [![Latest Release](https://img.shields.io/github/release/VictoriaMetrics/VictoriaMetrics.svg?style=flat-square)](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/latest)
+[![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 is fast, cost-effective and scalable time series database. It can be used as a long-term remote storage for Prometheus.
+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).
@@ -20,14 +21,14 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
 
 * 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.
+* 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).
 * 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 a 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).
+  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).
@@ -37,13 +38,14 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   * 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:
+* 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)
     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.
+  * [HTTP OpenTSDB /api/put requests](http://opentsdb.net/docs/build/html/api_http/put.html) if `-opentsdbHTTPListenAddr` is set.
+* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars, industrial telemetry and various Enterprise workloads.
 * Has open source [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
 
 
@@ -52,20 +54,24 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
 
 ### Table of contents
 
-  - [How to build from sources](#how-to-build-from-sources)
-    - [Development build](#development-build)
-    - [Production build](#production-build)
-    - [Building docker images](#building-docker-images)
   - [How to start VictoriaMetrics](#how-to-start-victoriametrics)
-  - [Setting up service](#setting-up-service)
-  - [Third-party contributions](#third-party-contributions)
   - [Prometheus setup](#prometheus-setup)
   - [Grafana setup](#grafana-setup)
   - [How to upgrade VictoriaMetrics?](#how-to-upgrade-victoriametrics)
   - [How to apply new config to VictoriaMetrics?](#how-to-apply-new-config-to-victoriametrics)
   - [How to send data from InfluxDB-compatible agents such as Telegraf?](#how-to-send-data-from-influxdb-compatible-agents-such-as-telegraf)
   - [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)
+  - [How to build from sources](#how-to-build-from-sources)
+    - [Development build](#development-build)
+    - [Production build](#production-build)
+    - [ARM build](#arm-build)
+    - [Pure Go build (CGO_ENABLED=0)](#pure-go-build-cgo_enabled0)
+    - [Building docker images](#building-docker-images)
+  - [Start with docker-compose](#start-with-docker-compose)
+  - [Setting up service](#setting-up-service)
+  - [Third-party contributions](#third-party-contributions)
   - [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)
@@ -81,6 +87,9 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [Tuning](#tuning)
   - [Monitoring](#monitoring)
   - [Troubleshooting](#troubleshooting)
+  - [Backfilling](#backfilling)
+  - [Profiling](#profiling)
+- [Roadmap](#roadmap)
 - [Contacts](#contacts)
 - [Community and contributions](#community-and-contributions)
 - [Reporting bugs](#reporting-bugs)
@@ -91,56 +100,23 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [We kindly ask:](#we-kindly-ask)
 
 
-### How to build from sources
-
-We recommend using either [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) or
-[docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) instead of building VictoriaMetrics
-from sources. Building from sources is reasonable when developing an additional features specific
-to your needs.
-
-
-#### Development build
-
-1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
-2. Run `make victoria-metrics` from the root folder of the repository.
-   It will build `victoria-metrics` binary and put it into the `bin` folder.
-
-#### Production build
-
-1. [Install docker](https://docs.docker.com/install/).
-2. Run `make victoria-metrics-prod` from the root folder of the repository.
-   It will build `victoria-metrics-prod` binary and put it into the `bin` folder.
-
-#### Building docker images
-
-Run `make package-victoria-metrics`. It will build `victoriametrics/victoria-metrics:<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-victoria-metrics`.
-
-
 ### How to start VictoriaMetrics
 
-Just start VictoriaMetrics executable or docker image with the desired command-line flags.
+Just start VictoriaMetrics [executable](https://github.com/VictoriaMetrics/VictoriaMetrics/releases)
+or [docker image](https://hub.docker.com/r/victoriametrics/victoria-metrics/) with the desired command-line flags.
 
-The following command line flags are used the most:
+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.
-* `-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.
+* `-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 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.
 
-
-### Setting up service
-
-Read [these instructions](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/43) on how to set up VictoriaMetrics as a service in your OS.
-
-
-### Third-party contributions
-
-* [Unofficial yum repository](https://copr.fedorainfracloud.org/coprs/antonpatsev/VictoriaMetrics/) ([source code](https://github.com/patsevanton/victoriametrics-rpm))
+It is recommended setting up [monitoring](#monitoring) for VictoriaMetrics.
 
 
 ### Prometheus setup
@@ -152,7 +128,7 @@ remote_write:
   - url: http://<victoriametrics-addr>:8428/api/v1/write
     queue_config:
       max_samples_per_send: 10000
-      max_shards: 100
+      max_shards: 30
 ```
 
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
@@ -162,11 +138,11 @@ Then apply the new config via the following command:
 kill -HUP `pidof prometheus`
 ```
 
-Prometheus writes incoming data to local storage and to remote storage in parallel.
+Prometheus writes incoming data to local storage and replicates it to remote storage in parallel.
 This means the data remains available in local storage for `--storage.tsdb.retention.time` duration
-if remote storage stops working.
+even if remote storage is unavailable.
 
-If you plan sending data to VictoriaMetrics from multiple Prometheus instances, then add the following lines into `global` section
+If you plan to send data to VictoriaMetrics from multiple Prometheus instances, then add the following lines into `global` section
 of [Prometheus config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#configuration-file):
 
 ```yml
@@ -177,7 +153,7 @@ global:
 
 This instructs Prometheus to add `datacenter=dc-123` label to each time series sent to remote storage.
 The label name may be arbitrary - `datacenter` is just an example. The label value must be unique
-across Prometheus instances, so time series may be filtered and grouped by this label.
+across Prometheus instances, so those time series may be filtered and grouped by this label.
 
 
 It is recommended upgrading Prometheus to [v2.10.0](https://github.com/prometheus/prometheus/releases) or newer,
@@ -210,6 +186,9 @@ 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?
 
@@ -217,7 +196,10 @@ VictoriaMetrics must be restarted for applying new config:
 
 1) Send `SIGINT` signal to VictoriaMetrics process in order to gracefully stop it.
 2) Wait until the process stops. This can take a few seconds.
-3) Start VictoriaMetrics with new config.
+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/)?
@@ -233,7 +215,8 @@ 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.
@@ -249,8 +232,8 @@ 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/)
@@ -260,11 +243,11 @@ to local VictoriaMetrics using `curl`:
 curl -d 'measurement,tag1=value1,tag2=value2 field1=123,field2=1.23' -X POST 'http://localhost:8428/write'
 ```
 
-Arbitrary number of lines delimited by '\n' may be sent in a single request.
+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__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -274,6 +257,9 @@ The `/api/v1/export` endpoint should return the following response:
 {"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)?
 
@@ -294,12 +280,12 @@ Example for writing data with Graphite plaintext protocol to local VictoriaMetri
 echo "foo.bar.baz;tag1=value1;tag2=value2 123 `date +%s`" | nc -N localhost 2003
 ```
 
-VictoriaMetrics sets the current time if timestamp is omitted.
-Arbitrary number of lines delimited by `\n` may be sent in one go.
+VictoriaMetrics sets the current time if the timestamp is omitted.
+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={__name__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -309,10 +295,23 @@ 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/)
+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
@@ -327,11 +326,11 @@ Example for writing data with OpenTSDB protocol to local VictoriaMetrics using `
 echo "put foo.bar.baz `date +%s` 123 tag1=value1 tag2=value2" | nc -N localhost 4242
 ```
 
-Arbitrary number of lines delimited by `\n` may be sent in one go.
+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={__name__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -341,9 +340,117 @@ 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]}
+```
+
+
+### How to build from sources
+
+We recommend using either [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) or
+[docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) instead of building VictoriaMetrics
+from sources. Building from sources is reasonable when developing additional features specific
+to your needs.
+
+
+#### Development build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics` from the root folder of the repository.
+   It builds `victoria-metrics` binary and puts it into the `bin` folder.
+
+#### Production build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make victoria-metrics-prod` from the root folder of the repository.
+   It builds `victoria-metrics-prod` binary and puts it into the `bin` folder.
+
+#### ARM build
+
+ARM build may run on Raspberry Pi or on [energy-efficient ARM servers](https://blog.cloudflare.com/arm-takes-wing/).
+
+#### Development ARM build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics-arm` or `make victoria-metrics-arm64` from the root folder of the repository.
+   It builds `victoria-metrics-arm` or `victoria-metrics-arm64` binary respectively and puts it into the `bin` folder.
+
+#### Production ARM build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make victoria-metrics-arm-prod` or `make victoria-metrics-arm64-prod` from the root folder of the repository.
+   It builds `victoria-metrics-arm-prod` or `victoria-metrics-arm64-prod` binary respectively and puts it into the `bin` folder.
+
+#### Pure Go build (CGO_ENABLED=0)
+
+`Pure Go` mode builds only Go code without [cgo](https://golang.org/cmd/cgo/) dependencies.
+This is an experimental mode, which may result in a lower compression ratio and slower decompression performance.
+Use it with caution!
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics-pure` from the root folder of the repository.
+   It builds `victoria-metrics-pure` binary and puts it into the `bin` folder.
+
+#### Building docker images
+
+Run `make package-victoria-metrics`. It builds `victoriametrics/victoria-metrics:<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-victoria-metrics`.
+
+
+### Start with docker-compose
+
+[Docker-compose](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/docker-compose.yml)
+helps to spin up VictoriaMetrics, Prometheus and Grafana with one command.
+More details may be found [here](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/master/deployment/docker#folder-contains-basic-images-and-tools-for-building-and-running-victoria-metrics-in-docker).
+
+
+### Setting up service
+
+Read [these instructions](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/43) on how to set up VictoriaMetrics as a service in your OS.
+
+
+### Third-party contributions
+
+* [Unofficial yum repository](https://copr.fedorainfracloud.org/coprs/antonpatsev/VictoriaMetrics/) ([source code](https://github.com/patsevanton/victoriametrics-rpm))
+
+
 ### How to work with snapshots?
 
-VictoriaMetrics is able to create [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
+VictoriaMetrics can create [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
 for all the data stored under `-storageDataPath` directory.
 Navigate to `http://<victoriametrics-addr>:8428/snapshot/create` in order to create an instant snapshot.
 The page will return the following JSON response:
@@ -377,6 +484,9 @@ 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?
 
@@ -400,15 +510,15 @@ VictoriaMetrics exports [Prometheus-compatible federation data](https://promethe
 at `http://<victoriametrics-addr>:8428/federate?match[]=<timeseries_selector_for_federation>`.
 
 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. Default value for `max_lookback` is `5m` (5 minutes), but can be overridden.
+`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.
 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`.
 
 
 ### Capacity planning
 
-Rough estimation of the required resources:
+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.
@@ -422,12 +532,29 @@ Rough estimation of the required resources:
   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.
 
-* Storage size: less than a byte per data point on average. So, ~260GB is required for storing a month-long insert stream
+* Storage space: less than a byte per data point on average. So, ~260GB is required for storing a month-long insert stream
   of 100K data points per second.
   The actual storage size heavily depends on data randomness (entropy). Higher randomness means higher storage size requirements.
   Read [this article](https://medium.com/faun/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932)
   for details.
 
+* Network usage: outbound traffic is negligible. Ingress traffic is ~100 bytes per ingested data point via
+  [Prometheus remote_write API](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write).
+  The actual ingress bandwidth usage depends on the average number of labels per ingested metric and the average size
+  of label values. The higher number of per-metric labels and longer label values mean the higher ingress bandwidth.
+
+
+The required resources for query path:
+
+* RAM size: depends on the number of time series to scan in each query and the `step`
+  argument passed to [/api/v1/query_range](https://prometheus.io/docs/prometheus/latest/querying/api/#range-queries).
+  The higher number of scanned time series and lower `step` argument results in the higher RAM usage.
+
+* CPU cores: a CPU core per 30 millions of scanned data points per second.
+
+* Network usage: depends on the frequency and the type of incoming requests. Typical Grafana dashboards usually
+  require negligible network bandwidth.
+
 
 ### High availability
 
@@ -457,7 +584,7 @@ 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
@@ -477,7 +604,7 @@ 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 implementing downsampling in the future.
+These properties reduce the need in downsampling. We plan to implement downsampling in the future.
 See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/36) for details.
 
 
@@ -489,7 +616,7 @@ Single-node VictoriaMetrics doesn't support multi-tenancy. Use [cluster version]
 ### Scalability and cluster version
 
 Though single-node VictoriaMetrics cannot scale to multiple nodes, it is optimized for resource usage - storage size / bandwidth / IOPS, RAM, CPU.
-This means that a single-node VictoriaMetrics may scale vertically and substitute moderately sized cluster built with competing solutions
+This means that a single-node VictoriaMetrics may scale vertically and substitute a moderately sized cluster built with competing solutions
 such as Thanos, Uber M3, InfluxDB or TimescaleDB. See [vertical scalability benchmarks](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
 
 So try single-node VictoriaMetrics at first and then [switch to cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster) if you still need
@@ -505,7 +632,7 @@ on [Prometheus side](https://prometheus.io/docs/alerting/overview/) or on [Grafa
 
 ### Security
 
-Do not forget protecting sensitive endpoints in VictoriaMetrics when exposing it to untrusted networks such as internet.
+Do not forget protecting sensitive endpoints in VictoriaMetrics when exposing it to untrusted networks such as the internet.
 Consider setting the following command-line flags:
 
 * `-tls`, `-tlsCertFile` and `-tlsKeyFile` for switching from HTTP to HTTPS.
@@ -520,10 +647,10 @@ 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 in 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.
-  The only option is increasing the limit on [the number open files in the OS](https://medium.com/@muhammadtriwibowo/set-permanently-ulimit-n-open-files-in-ubuntu-4d61064429a),
+* There is no need in 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.
 
 
@@ -558,12 +685,56 @@ The most interesting metrics are:
   Another option is to increase `-memory.allowedPercent` command-line flag value. Be careful with this
   option, since too big value for `-memory.allowedPercent` may result in high I/O usage.
 
+* VictoriaMetrics requires free disk space for [merging data files to bigger ones](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
+  It may slow down when there is no enough free space left. So make sure `-storageDataPath` directory
+  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
-  of data loss stored in the broken parts. In the future `vmrecover` tool will be created
+  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).
+
+
+## Roadmap
+
+- [ ] Replication [#118](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/118)
+- [ ] Support of Object Storages (GCS, S3, Azure Storage) [#38](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/38)
+- [ ] Data downsampling [#36](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/36)
+- [ ] Alert Manager Integration [#119](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/119) 
+- [ ] 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.
+
+
 ## Contacts
 
 Contact us with any questions regarding VictoriaMetrics at [info@victoriametrics.com](mailto:info@victoriametrics.com).
@@ -579,7 +750,7 @@ Feel free asking any questions regarding VictoriaMetrics:
 - [google groups](https://groups.google.com/forum/#!forum/victorametrics-users)
 
 
-If you like VictoriaMetrics and want contributing, then we need the following:
+If you like VictoriaMetrics and want to contribute, then we need the following:
 
 - Filing issues and feature requests [here](https://github.com/VictoriaMetrics/VictoriaMetrics/issues).
 - Spreading a word about VictoriaMetrics: conference talks, articles, comments, experience sharing with colleagues.

app/victoria-metrics/Makefile

@@ -1,7 +1,7 @@
 # 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
@@ -21,7 +21,52 @@ run-victoria-metrics:
 	$(MAKE) run-via-docker
 
 victoria-metrics-arm:
-	CC=arm-linux-gnueabi-gcc CGO_ENABLED=1 GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm ./app/victoria-metrics
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(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:
-	CC=aarch64-linux-gnu-gcc CGO_ENABLED=1 GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm64 ./app/victoria-metrics
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(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-pure:
+	APP_NAME=victoria-metrics $(MAKE) app-local-pure
+
+victoria-metrics-pure-prod:
+	APP_NAME=victoria-metrics APP_SUFFIX='-pure' DOCKER_OPTS='--env CGO_ENABLED=0' $(MAKE) app-via-docker
+
+### Packaging as DEB - amd64
+victoria-metrics-package-deb: victoria-metrics-prod
+	./package/package_deb.sh amd64
+
+### Packaging as DEB - arm64
+victoria-metrics-package-deb-arm64: victoria-metrics-arm64-prod
+	./package/package_deb.sh arm64
+
+### Packaging as DEB - all
+victoria-metrics-package-deb-all: \
+        victoria-metrics-package-deb \
+        victoria-metrics-package-deb-arm64
+
+### Packaging as RPM - amd64
+victoria-metrics-package-rpm: victoria-metrics-prod
+	./package/package_rpm.sh amd64
+
+### Packaging as RPM - arm64
+victoria-metrics-package-rpm-arm64: victoria-metrics-arm64-prod
+	./package/package_rpm.sh arm64
+
+### Packaging as RPM - all
+victoria-metrics-package-rpm-all: \
+        victoria-metrics-package-rpm \
+        victoria-metrics-package-rpm-arm64
+
+### Packaging as both DEB and RPM - all
+victoria-metrics-package-deb-rpm-all: \
+        victoria-metrics-package-deb \
+        victoria-metrics-package-deb-arm64 \
+        victoria-metrics-package-rpm \
+        victoria-metrics-package-rpm-arm64

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"
@@ -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

@@ -21,6 +21,7 @@ import (
 	"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"
 )
@@ -92,7 +93,7 @@ func setUp() {
 
 func processFlags() {
 	flag.Parse()
-	for _, fs := range []struct {
+	for _, fv := range []struct {
 		flag  string
 		value string
 	}{
@@ -103,8 +104,8 @@ func processFlags() {
 		{flag: "loggerLevel", value: testLogLevel},
 	} {
 		// 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)
 		}
 	}
 }
@@ -121,13 +122,14 @@ func waitFor(timeout time.Duration, f func() bool) error {
 }
 
 func tearDown() {
-	vminsert.Stop()
-	vmstorage.Stop()
-	vmselect.Stop()
 	if err := httpserver.Stop(*httpListenAddr); err != nil {
 		log.Fatalf("cannot stop the webservice: %s", err)
 	}
-	os.RemoveAll(storagePath)
+	vminsert.Stop()
+	vmstorage.Stop()
+	vmselect.Stop()
+	fs.MustRemoveAll(storagePath)
+	fs.MustStopDirRemover()
 }
 
 func TestWriteRead(t *testing.T) {

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"
 )
@@ -99,7 +101,10 @@ 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/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"
 )
@@ -32,6 +34,17 @@ func Init() {
 func Do(f func() error) error {
 	// Limit the number of conurrent f calls in order to prevent from excess
 	// memory usage and CPU trashing.
+	select {
+	case ch <- struct{}{}:
+		err := f()
+		<-ch
+		return err
+	default:
+	}
+
+	// All the workers are busy.
+	// Sleep for up to waitDuration.
+	concurrencyLimitReached.Inc()
 	t := timerpool.Get(waitDuration)
 	select {
 	case ch <- struct{}{}:
@@ -41,9 +54,22 @@ func Do(f func() error) error {
 		return err
 	case <-t.C:
 		timerpool.Put(t)
-		concurrencyLimitErrors.Inc()
-		return fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase -maxConcurrentInserts or reduce the load", cap(ch))
+		concurrencyLimitTimeout.Inc()
+		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,
+		}
 	}
 }
 
-var concurrencyLimitErrors = metrics.NewCounter(`vm_concurrency_limit_errors_total`)
+var (
+	concurrencyLimitReached = metrics.NewCounter(`vm_concurrent_insert_limit_reached_total`)
+	concurrencyLimitTimeout = metrics.NewCounter(`vm_concurrent_insert_limit_timeout_total`)
+
+	_ = metrics.NewGauge(`vm_concurrent_insert_capacity`, func() float64 {
+		return float64(cap(ch))
+	})
+	_ = metrics.NewGauge(`vm_concurrent_insert_current`, func() float64 {
+		return float64(len(ch))
+	})
+)

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 {
@@ -95,39 +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 {
-				return dst, tagsPool, err
-			}
-			return dst, tagsPool, nil
-		}
-		var err error
-		tagsPool, err = r.unmarshal(s[:n], tagsPool)
-		if err != nil {
-			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) {
@@ -143,12 +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]
+		}
 	}
 }
 
@@ -170,9 +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{
@@ -98,7 +97,8 @@ 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 +106,6 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 					Key:   "bar",
 					Value: "baz",
 				},
-				{
-					Key:   "aa",
-					Value: "",
-				},
 				{
 					Key:   "x",
 					Value: "y",
@@ -139,4 +135,20 @@ 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

@@ -14,7 +14,10 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="graphite"}`)
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="graphite"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="graphite"}`)
+)
 
 // insertHandler processes remote write for graphite plaintext protocol.
 //
@@ -51,6 +54,7 @@ func (ctx *pushCtx) InsertRows() error {
 		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
 	return ic.FlushBufs()
 }
 
@@ -81,11 +85,7 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			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()
@@ -132,9 +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"}`)
+	graphiteReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
+	graphiteReadErrors = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/graphite/server.go

@@ -9,6 +9,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
 	"github.com/VictoriaMetrics/metrics"
 )
 
@@ -23,7 +24,7 @@ var (
 // Serve starts graphite server on the given addr.
 func Serve(addr string) {
 	logger.Infof("starting TCP Graphite server at %q", addr)
-	lnTCP, err := net.Listen("tcp4", addr)
+	lnTCP, err := netutil.NewTCPListener("graphite", addr)
 	if err != nil {
 		logger.Fatalf("cannot start TCP Graphite server at %q: %s", addr, 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)
@@ -141,9 +137,6 @@ 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
 }
@@ -177,39 +170,51 @@ 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 {
-				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 {
-			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) {
@@ -221,14 +226,22 @@ 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

@@ -74,13 +74,15 @@ 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))
 		}
 	}
 
@@ -94,12 +96,8 @@ 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 +120,13 @@ 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 +140,12 @@ 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{})
 
 	// Minimal line without tags and timestamp
 	f("foo bar=123", &Rows{
@@ -157,6 +157,15 @@ 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 +225,7 @@ 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 +233,6 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 					Key:   "tag1",
 					Value: "xyz",
 				},
-				{
-					Key:   "tagN",
-					Value: "",
-				},
 				{
 					Key:   "tag2",
 					Value: "43as",
@@ -309,11 +314,11 @@ 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{{
@@ -348,6 +353,34 @@ 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 +407,24 @@ 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"
@@ -22,7 +21,10 @@ var (
 	skipSingleField           = flag.Bool("influxSkipSingleField", false, "Uses `{measurement}` instead of `{measurement}{separator}{field_name}` for metic name if Influx line contains only a single field")
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="influx"}`)
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="influx"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="influx"}`)
+)
 
 // InsertHandler processes remote write for influx line protocol.
 //
@@ -38,11 +40,11 @@ func insertHandlerInternal(req *http.Request) error {
 
 	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
 	}
 
@@ -84,14 +86,21 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	}
 	ic := &ctx.Common
 	ic.Reset(rowsLen)
+	rowsTotal := 0
 	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
@@ -109,30 +118,13 @@ func (ctx *pushCtx) InsertRows(db string) error {
 			ic.AddLabel("", metricGroup)
 			ic.WriteDataPoint(ctx.metricNameBuf, ic.Labels[:1], r.Timestamp, f.Value)
 		}
-		rowsInserted.Add(len(r.Fields))
+		rowsTotal += len(r.Fields)
 	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
 	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
@@ -145,11 +137,7 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 		}
 		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
@@ -164,6 +152,7 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 		}
 	} else if tsMultiplier < 0 {
 		tsMultiplier = -tsMultiplier
+		currentTs -= currentTs % tsMultiplier
 		for i := range ctx.Rows.Rows {
 			row := &ctx.Rows.Rows[i]
 			if row.Timestamp == 0 {
@@ -177,9 +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"}`)
+	influxReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
+	influxReadErrors = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
 )
 
 type pushCtx struct {

app/vminsert/main.go

@@ -10,19 +10,25 @@ 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/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 put messages. 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")
 )
 
 // Init initializes vminsert.
 func Init() {
+	storage.SetMaxLabelsPerTimeseries(*maxLabelsPerTimeseries)
+
 	concurrencylimiter.Init()
 	if len(*graphiteListenAddr) > 0 {
 		go graphite.Serve(*graphiteListenAddr)
@@ -30,6 +36,9 @@ func Init() {
 	if len(*opentsdbListenAddr) > 0 {
 		go opentsdb.Serve(*opentsdbListenAddr)
 	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		go opentsdbhttp.Serve(*opentsdbHTTPListenAddr, int64(*maxInsertRequestSize))
+	}
 }
 
 // Stop stops vminsert.
@@ -40,6 +49,9 @@ func Stop() {
 	if len(*opentsdbListenAddr) > 0 {
 		opentsdb.Stop()
 	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttp.Stop()
+	}
 }
 
 // RequestHandler is a handler for Prometheus remote storage write API

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 {
@@ -92,39 +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 {
-				return dst, tagsPool, err
-			}
-			return dst, tagsPool, nil
-		}
-		var err error
-		tagsPool, err = r.unmarshal(s[:n], tagsPool)
-		if err != nil {
-			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) {
@@ -140,12 +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]
+		}
 	}
 }
 
@@ -167,9 +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

@@ -14,7 +14,10 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb"}`)
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="opentsdb"}`)
+)
 
 // insertHandler processes remote write for OpenTSDB put protocol.
 //
@@ -51,6 +54,7 @@ func (ctx *pushCtx) InsertRows() error {
 		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
 	return ic.FlushBufs()
 }
 
@@ -81,15 +85,21 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			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
 }
@@ -121,9 +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"}`)
+	opentsdbReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
+	opentsdbReadErrors = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/opentsdb/server.go

@@ -9,6 +9,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
 	"github.com/VictoriaMetrics/metrics"
 )
 
@@ -23,7 +24,7 @@ var (
 // Serve starts OpenTSDB collector on the given addr.
 func Serve(addr string) {
 	logger.Infof("starting TCP OpenTSDB collector at %q", addr)
-	lnTCP, err := net.Listen("tcp4", addr)
+	lnTCP, err := netutil.NewTCPListener("opentsdb", addr)
 	if err != nil {
 		logger.Fatalf("cannot start TCP OpenTSDB collector at %q: %s", addr, 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"}`)
+
+	opentsdbReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="opentsdb-http"}`)
+	opentsdbReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="opentsdb-http"}`)
+	opentsdbUnmarshalErrors = 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 {
+	opentsdbReadCalls.Inc()
+
+	r := req.Body
+	if req.Header.Get("Content-Encoding") == "gzip" {
+		zr, err := common.GetGzipReader(r)
+		if err != nil {
+			opentsdbReadErrors.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 {
+		opentsdbReadErrors.Inc()
+		return fmt.Errorf("cannot read HTTP OpenTSDB request: %s", err)
+	}
+	if reqLen > maxSize {
+		opentsdbReadErrors.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 {
+		opentsdbUnmarshalErrors.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 +1,70 @@
+package opentsdbhttp
+
+import (
+	"context"
+	"net/http"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"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"}`)
+)
+
+var (
+	httpServer     *http.Server
+	httpAddr       string
+	maxRequestSize int64
+)
+
+// Serve starts HTTP OpenTSDB server on the given addr.
+func Serve(addr string, maxReqSize int64) {
+	logger.Infof("starting HTTP OpenTSDB server at %q", addr)
+	httpAddr = addr
+	maxRequestSize = maxReqSize
+	httpServer = &http.Server{
+		Addr:         addr,
+		Handler:      http.HandlerFunc(requestHandler),
+		ReadTimeout:  30 * time.Second,
+		WriteTimeout: 10 * time.Second,
+	}
+	go func() {
+		err := httpServer.ListenAndServe()
+		if err == http.ErrServerClosed {
+			return
+		}
+		if err != nil {
+			logger.Fatalf("FATAL: error serving HTTP OpenTSDB: %s", err)
+		}
+	}()
+}
+
+// requestHandler handles HTTP OpenTSDB insert request.
+func requestHandler(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)
+	}
+}
+
+// Stop stops HTTP OpenTSDB server.
+func Stop() {
+	logger.Infof("stopping HTTP OpenTSDB server at %q...", httpAddr)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := httpServer.Shutdown(ctx); err != nil {
+		logger.Fatalf("FATAL: cannot close HTTP OpenTSDB server: %s", err)
+	}
+}

app/vminsert/prometheus/request_handler.go

@@ -12,7 +12,10 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="prometheus"}`)
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="prometheus"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="prometheus"}`)
+)
 
 // InsertHandler processes remote write for prometheus.
 func InsertHandler(r *http.Request, maxSize int64) error {
@@ -34,6 +37,7 @@ func insertHandlerInternal(r *http.Request, maxSize int64) error {
 	}
 	ic := &ctx.Common
 	ic.Reset(rowsLen)
+	rowsTotal := 0
 	for i := range timeseries {
 		ts := &timeseries[i]
 		var metricNameRaw []byte
@@ -41,8 +45,10 @@ func insertHandlerInternal(r *http.Request, maxSize int64) error {
 			r := &ts.Samples[i]
 			metricNameRaw = ic.WriteDataPointExt(metricNameRaw, ts.Labels, r.Timestamp, r.Value)
 		}
-		rowsInserted.Add(len(ts.Samples))
+		rowsTotal += len(ts.Samples)
 	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
 	return ic.FlushBufs()
 }
 

app/vmselect/main.go

@@ -2,6 +2,7 @@ package vmselect
 
 import (
 	"flag"
+	"fmt"
 	"net/http"
 	"runtime"
 	"strings"
@@ -30,29 +31,53 @@ func Init() {
 	fs.RemoveDirContents(tmpDirPath)
 	netstorage.InitTmpBlocksDir(tmpDirPath)
 	promql.InitRollupResultCache(*vmstorage.DataPath + "/cache/rollupResult")
+
 	concurrencyCh = make(chan struct{}, *maxConcurrentRequests)
 }
 
-var concurrencyCh chan struct{}
-
 // Stop stops vmselect
 func Stop() {
 	promql.StopRollupResultCache()
 }
 
+var concurrencyCh chan struct{}
+
+var (
+	concurrencyLimitReached = metrics.NewCounter(`vm_concurrent_select_limit_reached_total`)
+	concurrencyLimitTimeout = metrics.NewCounter(`vm_concurrent_select_limit_timeout_total`)
+
+	_ = metrics.NewGauge(`vm_concurrent_select_capacity`, func() float64 {
+		return float64(cap(concurrencyCh))
+	})
+	_ = metrics.NewGauge(`vm_concurrent_select_current`, func() float64 {
+		return float64(len(concurrencyCh))
+	})
+)
+
 // RequestHandler handles remote read API requests for Prometheus
 func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	// Limit the number of concurrent queries.
-	// Sleep for a while until giving up. This should resolve short bursts in requests.
-	t := timerpool.Get(*maxQueueDuration)
 	select {
 	case concurrencyCh <- struct{}{}:
-		timerpool.Put(t)
 		defer func() { <-concurrencyCh }()
-	case <-t.C:
-		timerpool.Put(t)
-		httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
-		return true
+	default:
+		// Sleep for a while until giving up. This should resolve short bursts in requests.
+		concurrencyLimitReached.Inc()
+		t := timerpool.Get(*maxQueueDuration)
+		select {
+		case concurrencyCh <- struct{}{}:
+			timerpool.Put(t)
+			defer func() { <-concurrencyCh }()
+		case <-t.C:
+			timerpool.Put(t)
+			concurrencyLimitTimeout.Inc()
+			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
+		}
 	}
 
 	path := strings.Replace(r.URL.Path, "//", "/", -1)
@@ -165,7 +190,10 @@ 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)
 }

app/vmselect/netstorage/fadvise_freebsd.go

@@ -0,0 +1,15 @@
+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.
@@ -49,8 +49,9 @@ func (r *Result) reset() {
 
 // Results holds results returned from ProcessSearchQuery.
 type Results struct {
-	tr       storage.TimeRange
-	deadline Deadline
+	tr        storage.TimeRange
+	fetchData bool
+	deadline  Deadline
 
 	tbf *tmpBlocksFile
 
@@ -71,9 +72,10 @@ func (rss *Results) Cancel() {
 // RunParallel runs in parallel f for all the results from rss.
 //
 // f shouldn't hold references to rs after returning.
+// workerID is the id of the worker goroutine that calls f.
 //
 // rss becomes unusable after the call to RunParallel.
-func (rss *Results) RunParallel(f func(rs *Result)) error {
+func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	defer func() {
 		putTmpBlocksFile(rss.tbf)
 		rss.tbf = nil
@@ -91,7 +93,7 @@ func (rss *Results) RunParallel(f func(rs *Result)) error {
 
 	// Start workers.
 	for i := 0; i < workersCount; i++ {
-		go func() {
+		go func(workerID uint) {
 			rs := getResult()
 			defer putResult(rs)
 			maxWorkersCount := gomaxprocs / workersCount
@@ -102,20 +104,20 @@ func (rss *Results) RunParallel(f func(rs *Result)) error {
 					err = fmt.Errorf("timeout exceeded during query execution: %s", rss.deadline.Timeout)
 					break
 				}
-				if err = pts.Unpack(rss.tbf, rs, rss.tr, maxWorkersCount); err != nil {
+				if err = pts.Unpack(rss.tbf, rs, rss.tr, rss.fetchData, maxWorkersCount); err != nil {
 					break
 				}
-				if len(rs.Timestamps) == 0 {
+				if len(rs.Timestamps) == 0 && rss.fetchData {
 					// Skip empty blocks.
 					continue
 				}
-				f(rs)
+				f(rs, workerID)
 			}
 			// Drain the remaining work
 			for range workCh {
 			}
 			doneCh <- err
-		}()
+		}(uint(i))
 	}
 
 	// Feed workers with work.
@@ -148,7 +150,7 @@ type packedTimeseries struct {
 }
 
 // Unpack unpacks pts to dst.
-func (pts *packedTimeseries) Unpack(tbf *tmpBlocksFile, dst *Result, tr storage.TimeRange, maxWorkersCount int) error {
+func (pts *packedTimeseries) Unpack(tbf *tmpBlocksFile, dst *Result, tr storage.TimeRange, fetchData bool, maxWorkersCount int) error {
 	dst.reset()
 
 	if err := dst.MetricName.Unmarshal(bytesutil.ToUnsafeBytes(pts.metricName)); err != nil {
@@ -175,7 +177,7 @@ func (pts *packedTimeseries) Unpack(tbf *tmpBlocksFile, dst *Result, tr storage.
 			var err error
 			for addr := range workCh {
 				sb := getSortBlock()
-				if err = sb.unpackFrom(tbf, addr, tr); err != nil {
+				if err = sb.unpackFrom(tbf, addr, tr, fetchData); err != nil {
 					break
 				}
 
@@ -294,10 +296,12 @@ func (sb *sortBlock) reset() {
 	sb.NextIdx = 0
 }
 
-func (sb *sortBlock) unpackFrom(tbf *tmpBlocksFile, addr tmpBlockAddr, tr storage.TimeRange) error {
+func (sb *sortBlock) unpackFrom(tbf *tmpBlocksFile, addr tmpBlockAddr, tr storage.TimeRange, fetchData bool) error {
 	tbf.MustReadBlockAt(&sb.b, addr)
-	if err := sb.b.UnmarshalData(); err != nil {
-		return fmt.Errorf("cannot unmarshal block: %s", err)
+	if fetchData {
+		if err := sb.b.UnmarshalData(); err != nil {
+			return fmt.Errorf("cannot unmarshal block: %s", err)
+		}
 	}
 	timestamps := sb.b.Timestamps()
 
@@ -459,7 +463,7 @@ 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, deadline Deadline) (*Results, error) {
+func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadline) (*Results, error) {
 	// Setup search.
 	tfss, err := setupTfss(sq.TagFilterss)
 	if err != nil {
@@ -475,35 +479,38 @@ func ProcessSearchQuery(sq *storage.SearchQuery, deadline Deadline) (*Results, e
 
 	sr := getStorageSearch()
 	defer putStorageSearch(sr)
-	sr.Init(vmstorage.Storage, tfss, tr, *maxMetricsPerSearch)
+	sr.Init(vmstorage.Storage, tfss, tr, fetchData, *maxMetricsPerSearch)
 
 	tbf := getTmpBlocksFile()
 	m := make(map[string][]tmpBlockAddr)
+	blocksRead := 0
 	for sr.NextMetricBlock() {
+		blocksRead++
 		addr, err := tbf.WriteBlock(sr.MetricBlock.Block)
 		if err != nil {
 			putTmpBlocksFile(tbf)
-			return nil, fmt.Errorf("cannot write data to temporary blocks file: %s", err)
+			return nil, fmt.Errorf("cannot write data block #%d to temporary blocks file: %s", blocksRead, err)
 		}
 		if time.Until(deadline.Deadline) < 0 {
 			putTmpBlocksFile(tbf)
-			return nil, fmt.Errorf("timeout exceeded while fetching data from storage: %s", deadline.Timeout)
+			return nil, fmt.Errorf("timeout exceeded while fetching data block #%d from storage: %s", blocksRead, deadline.Timeout)
 		}
 		metricName := sr.MetricBlock.MetricName
 		m[string(metricName)] = append(m[string(metricName)], addr)
 	}
 	if err := sr.Error(); err != nil {
 		putTmpBlocksFile(tbf)
-		return nil, fmt.Errorf("search error: %s", err)
+		return nil, fmt.Errorf("search error after reading %d data blocks: %s", blocksRead, err)
 	}
 	if err := tbf.Finalize(); err != nil {
 		putTmpBlocksFile(tbf)
-		return nil, fmt.Errorf("cannot finalize temporary blocks file: %s", err)
+		return nil, fmt.Errorf("cannot finalize temporary blocks file with %d blocks: %s", blocksRead, err)
 	}
 
 	var rss Results
 	rss.packedTimeseries = make([]packedTimeseries, len(m))
 	rss.tr = tr
+	rss.fetchData = fetchData
 	rss.deadline = deadline
 	rss.tbf = tbf
 	i := 0

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)
 }
@@ -69,22 +80,6 @@ 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.
@@ -92,28 +87,31 @@ var tmpBlocksFilesCreated = metrics.NewCounter(`vm_tmp_blocks_files_created_tota
 // 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) {
+	bb := tmpBufPool.Get()
+	defer tmpBufPool.Put(bb)
+	bb.B = storage.MarshalBlock(bb.B[:0], b)
+
 	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(bb.B)
 	tbf.offset += uint64(addr.size)
-	if tbf.offset <= maxInmemoryTmpBlocksFile {
+	if len(tbf.buf)+len(bb.B) <= cap(tbf.buf) {
+		// Fast path - the data fits tbf.buf
+		tbf.buf = append(tbf.buf, bb.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], bb.B...)
 	if err != nil {
 		return addr, fmt.Errorf("cannot write block to %q: %s", tbf.f.Name(), err)
 	}
@@ -124,15 +122,15 @@ 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
+	return nil
 }
 
 func (tbf *tmpBlocksFile) MustReadBlockAt(dst *storage.Block, addr tmpBlockAddr) {
@@ -167,10 +165,6 @@ 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)
@@ -94,7 +94,7 @@ 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/prometheus.go

@@ -6,12 +6,15 @@ import (
 	"math"
 	"net/http"
 	"runtime"
+	"sort"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"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/storage"
 	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/quicktemplate"
@@ -65,7 +68,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 		MaxTimestamp: end,
 		TagFilterss:  tagFilterss,
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, true, deadline)
 	if err != nil {
 		return fmt.Errorf("cannot fetch data for %q: %s", sq, err)
 	}
@@ -73,7 +76,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	resultsCh := make(chan *quicktemplate.ByteBuffer)
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			WriteFederate(bb, rs)
 			resultsCh <- bb
@@ -157,7 +160,7 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 		MaxTimestamp: end,
 		TagFilterss:  tagFilterss,
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, true, deadline)
 	if err != nil {
 		return fmt.Errorf("cannot fetch data for %q: %s", sq, err)
 	}
@@ -165,7 +168,7 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 	resultsCh := make(chan *quicktemplate.ByteBuffer, runtime.GOMAXPROCS(-1))
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			writeLineFunc(bb, rs)
 			resultsCh <- bb
@@ -230,9 +233,39 @@ var deleteDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
 	deadline := getDeadline(r)
-	labelValues, err := netstorage.GetLabelValues(labelName, deadline)
-	if err != nil {
-		return fmt.Errorf(`cannot obtain label values for %q: %s`, labelName, err)
+
+	if err := r.ParseForm(); err != nil {
+		return fmt.Errorf("cannot parse form values: %s", err)
+	}
+	var labelValues []string
+	if len(r.Form["match[]"]) == 0 && len(r.Form["start"]) == 0 && len(r.Form["end"]) == 0 {
+		var err error
+		labelValues, err = netstorage.GetLabelValues(labelName, deadline)
+		if err != nil {
+			return fmt.Errorf(`cannot obtain label values for %q: %s`, labelName, err)
+		}
+	} else {
+		// Extended functionality that allows filtering by label filters and time range
+		// i.e. /api/v1/label/foo/values?match[]=foobar{baz="abc"}&start=...&end=...
+		// is equivalent to `label_values(foobar{baz="abc"}, foo)` call on the selected
+		// time range in Grafana templating.
+		matches := r.Form["match[]"]
+		if len(matches) == 0 {
+			matches = []string{fmt.Sprintf("{%s!=''}", labelName)}
+		}
+		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
+		}
+		labelValues, err = labelValuesWithMatches(labelName, matches, start, end, deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain label values for %q, match[]=%q, start=%d, end=%d: %s", labelName, matches, start, end, err)
+		}
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -241,6 +274,50 @@ func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request
 	return nil
 }
 
+func labelValuesWithMatches(labelName string, 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 {
+		start = end - 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) {
+		labelValue := rs.MetricName.GetTagValue(labelName)
+		if len(labelValue) == 0 {
+			return
+		}
+		mLock.Lock()
+		m[string(labelValue)] = struct{}{}
+		mLock.Unlock()
+	})
+	if err != nil {
+		return nil, fmt.Errorf("error when data fetching: %s", err)
+	}
+
+	labelValues := make([]string, 0, len(m))
+	for labelValue := range m {
+		labelValues = append(labelValues, labelValue)
+	}
+	sort.Strings(labelValues)
+	return labelValues, nil
+}
+
 var labelValuesDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/label/{}/values"}`)
 
 // LabelsCountHandler processes /api/v1/labels/count request.
@@ -309,13 +386,16 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		return fmt.Errorf("missing `match[]` arg")
 	}
-	// Set start to minTimeMsecs by default as Prometheus does.
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/91
-	start, err := getTime(r, "start", minTimeMsecs)
+	end, err := getTime(r, "end", ct)
 	if err != nil {
 		return err
 	}
-	end, err := getTime(r, "end", ct)
+	// Do not set start to minTimeMsecs by default as Prometheus does,
+	// since this leads to fetching and scanning all the data from the storage,
+	// which can take a lot of time for big storages.
+	// It is better setting start as end-defaultStep by default.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/91
+	start, err := getTime(r, "start", end-defaultStep)
 	if err != nil {
 		return err
 	}
@@ -333,7 +413,7 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 		MaxTimestamp: end,
 		TagFilterss:  tagFilterss,
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, false, deadline)
 	if err != nil {
 		return fmt.Errorf("cannot fetch data for %q: %s", sq, err)
 	}
@@ -341,7 +421,7 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	resultsCh := make(chan *quicktemplate.ByteBuffer)
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			writemetricNameObject(bb, &rs.MetricName)
 			resultsCh <- bb
@@ -477,7 +557,9 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	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,
@@ -494,12 +576,47 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 		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

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/promql/aggr.go

@@ -312,7 +312,11 @@ func aggrFuncCount(tss []*timeseries) []*timeseries {
 			}
 			count++
 		}
-		dst.Values[i] = float64(count)
+		v := float64(count)
+		if count == 0 {
+			v = nan
+		}
+		dst.Values[i] = v
 	}
 	return tss[:1]
 }
@@ -349,6 +353,25 @@ 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 {

app/vmselect/promql/aggr_incremental.go

@@ -13,30 +13,37 @@ import (
 var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
 	"sum": {
 		updateAggrFunc:   updateAggrSum,
+		mergeAggrFunc:    mergeAggrSum,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"min": {
 		updateAggrFunc:   updateAggrMin,
+		mergeAggrFunc:    mergeAggrMin,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"max": {
 		updateAggrFunc:   updateAggrMax,
+		mergeAggrFunc:    mergeAggrMax,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"avg": {
 		updateAggrFunc:   updateAggrAvg,
+		mergeAggrFunc:    mergeAggrAvg,
 		finalizeAggrFunc: finalizeAggrAvg,
 	},
 	"count": {
 		updateAggrFunc:   updateAggrCount,
+		mergeAggrFunc:    mergeAggrCount,
 		finalizeAggrFunc: finalizeAggrCount,
 	},
 	"sum2": {
 		updateAggrFunc:   updateAggrSum2,
+		mergeAggrFunc:    mergeAggrSum2,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"geomean": {
 		updateAggrFunc:   updateAggrGeomean,
+		mergeAggrFunc:    mergeAggrGeomean,
 		finalizeAggrFunc: finalizeAggrGeomean,
 	},
 }
@@ -44,8 +51,8 @@ var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
 type incrementalAggrFuncContext struct {
 	ae *aggrFuncExpr
 
-	mu sync.Mutex
-	m  map[string]*incrementalAggrContext
+	mLock sync.Mutex
+	m     map[uint]map[string]*incrementalAggrContext
 
 	callbacks *incrementalAggrFuncCallbacks
 }
@@ -53,17 +60,24 @@ type incrementalAggrFuncContext struct {
 func newIncrementalAggrFuncContext(ae *aggrFuncExpr, callbacks *incrementalAggrFuncCallbacks) *incrementalAggrFuncContext {
 	return &incrementalAggrFuncContext{
 		ae:        ae,
-		m:         make(map[string]*incrementalAggrContext, 1),
+		m:         make(map[uint]map[string]*incrementalAggrContext),
 		callbacks: callbacks,
 	}
 }
 
-func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries) {
+func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries, workerID uint) {
+	iafc.mLock.Lock()
+	m := iafc.m[workerID]
+	if m == nil {
+		m = make(map[string]*incrementalAggrContext, 1)
+		iafc.m[workerID] = m
+	}
+	iafc.mLock.Unlock()
+
 	removeGroupTags(&ts.MetricName, &iafc.ae.Modifier)
 	bb := bbPool.Get()
 	bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
-	iafc.mu.Lock()
-	iac := iafc.m[string(bb.B)]
+	iac := m[string(bb.B)]
 	if iac == nil {
 		tsAggr := &timeseries{
 			Values:     make([]float64, len(ts.Values)),
@@ -75,19 +89,30 @@ func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries) {
 			ts:     tsAggr,
 			values: make([]float64, len(ts.Values)),
 		}
-		iafc.m[string(bb.B)] = iac
+		m[string(bb.B)] = iac
 	}
-	iafc.callbacks.updateAggrFunc(iac, ts.Values)
-	iafc.mu.Unlock()
 	bbPool.Put(bb)
+	iafc.callbacks.updateAggrFunc(iac, ts.Values)
 }
 
 func (iafc *incrementalAggrFuncContext) finalizeTimeseries() []*timeseries {
-	// There is no need in iafc.mu.Lock here, since getTimeseries must be called
+	// There is no need in iafc.mLock.Lock here, since finalizeTimeseries must be called
 	// without concurrent goroutines touching iafc.
-	tss := make([]*timeseries, 0, len(iafc.m))
+	mGlobal := make(map[string]*incrementalAggrContext)
+	mergeAggrFunc := iafc.callbacks.mergeAggrFunc
+	for _, m := range iafc.m {
+		for k, iac := range m {
+			iacGlobal := mGlobal[k]
+			if iacGlobal == nil {
+				mGlobal[k] = iac
+				continue
+			}
+			mergeAggrFunc(iacGlobal, iac)
+		}
+	}
+	tss := make([]*timeseries, 0, len(mGlobal))
 	finalizeAggrFunc := iafc.callbacks.finalizeAggrFunc
-	for _, iac := range iafc.m {
+	for _, iac := range mGlobal {
 		finalizeAggrFunc(iac)
 		tss = append(tss, iac.ts)
 	}
@@ -96,6 +121,7 @@ func (iafc *incrementalAggrFuncContext) finalizeTimeseries() []*timeseries {
 
 type incrementalAggrFuncCallbacks struct {
 	updateAggrFunc   func(iac *incrementalAggrContext, values []float64)
+	mergeAggrFunc    func(dst, src *incrementalAggrContext)
 	finalizeAggrFunc func(iac *incrementalAggrContext)
 }
 
@@ -129,8 +155,33 @@ func updateAggrSum(iac *incrementalAggrContext, values []float64) {
 		if math.IsNaN(v) {
 			continue
 		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
+	}
+}
+
+func mergeAggrSum(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
 		dstValues[i] += v
-		dstCounts[i] = 1
 	}
 }
 
@@ -154,6 +205,29 @@ func updateAggrMin(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrMin(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v < dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
 func updateAggrMax(iac *incrementalAggrContext, values []float64) {
 	dstValues := iac.ts.Values
 	dstCounts := iac.values
@@ -174,6 +248,29 @@ func updateAggrMax(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrMax(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v > dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
 func updateAggrAvg(iac *incrementalAggrContext, values []float64) {
 	// Do not use `Rapid calculation methods` at https://en.wikipedia.org/wiki/Standard_deviation,
 	// since it is slower and has no obvious benefits in increased precision.
@@ -195,6 +292,28 @@ func updateAggrAvg(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrAvg(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = srcCounts[i]
+			continue
+		}
+		dstValues[i] += v
+		dstCounts[i] += srcCounts[i]
+	}
+}
+
 func finalizeAggrAvg(iac *incrementalAggrContext) {
 	dstValues := iac.ts.Values
 	counts := iac.values
@@ -219,8 +338,22 @@ func updateAggrCount(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrCount(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		dstValues[i] += v
+	}
+}
+
 func finalizeAggrCount(iac *incrementalAggrContext) {
-	// Nothing to do
+	dstValues := iac.ts.Values
+	for i, v := range dstValues {
+		if v == 0 {
+			dstValues[i] = nan
+		}
+	}
 }
 
 func updateAggrSum2(iac *incrementalAggrContext, values []float64) {
@@ -232,8 +365,33 @@ func updateAggrSum2(iac *incrementalAggrContext, values []float64) {
 		if math.IsNaN(v) {
 			continue
 		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v * v
+			dstCounts[i] = 1
+			continue
+		}
 		dstValues[i] += v * v
-		dstCounts[i] = 1
+	}
+}
+
+func mergeAggrSum2(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
 	}
 }
 
@@ -256,6 +414,28 @@ func updateAggrGeomean(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrGeomean(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = srcCounts[i]
+			continue
+		}
+		dstValues[i] *= v
+		dstCounts[i] += srcCounts[i]
+	}
+}
+
 func finalizeAggrGeomean(iac *incrementalAggrContext) {
 	dstValues := iac.ts.Values
 	counts := iac.values

app/vmselect/promql/aggr_incremental_test.go

@@ -0,0 +1,188 @@
+package promql
+
+import (
+	"fmt"
+	"math"
+	"reflect"
+	"runtime"
+	"sync"
+	"testing"
+)
+
+func TestIncrementalAggr(t *testing.T) {
+	defaultTimestamps := []int64{100e3, 200e3, 300e3, 400e3}
+	values := [][]float64{
+		{1, nan, 2, nan},
+		{3, nan, nan, 4},
+		{nan, nan, 5, 6},
+		{7, nan, 8, 9},
+		{4, nan, nan, nan},
+		{2, nan, 3, 2},
+		{0, nan, 1, 1},
+	}
+	tssSrc := make([]*timeseries, len(values))
+	for i, vs := range values {
+		ts := &timeseries{
+			Timestamps: defaultTimestamps,
+			Values:     vs,
+		}
+		tssSrc[i] = ts
+	}
+
+	copyTimeseries := func(tssSrc []*timeseries) []*timeseries {
+		tssDst := make([]*timeseries, len(tssSrc))
+		for i, tsSrc := range tssSrc {
+			var tsDst timeseries
+			tsDst.CopyFromShallowTimestamps(tsSrc)
+			tssDst[i] = &tsDst
+		}
+		return tssDst
+	}
+
+	f := func(name string, valuesExpected []float64) {
+		t.Helper()
+		callbacks := getIncrementalAggrFuncCallbacks(name)
+		ae := &aggrFuncExpr{
+			Name: name,
+		}
+		tssExpected := []*timeseries{{
+			Timestamps: defaultTimestamps,
+			Values:     valuesExpected,
+		}}
+		// run the test multiple times to make sure there are no side effects on concurrency
+		for i := 0; i < 10; i++ {
+			iafc := newIncrementalAggrFuncContext(ae, callbacks)
+			tssSrcCopy := copyTimeseries(tssSrc)
+			if err := testIncrementalParallelAggr(iafc, tssSrcCopy, tssExpected); err != nil {
+				t.Fatalf("unexpected error on iteration %d: %s", i, err)
+			}
+		}
+	}
+
+	t.Run("sum", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{17, nan, 19, 22}
+		f("sum", valuesExpected)
+	})
+	t.Run("min", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{0, nan, 1, 1}
+		f("min", valuesExpected)
+	})
+	t.Run("max", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{7, nan, 8, 9}
+		f("max", valuesExpected)
+	})
+	t.Run("avg", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{2.8333333333333335, nan, 3.8, 4.4}
+		f("avg", valuesExpected)
+	})
+	t.Run("count", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{6, nan, 5, 5}
+		f("count", valuesExpected)
+	})
+	t.Run("sum2", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{79, nan, 103, 138}
+		f("sum2", valuesExpected)
+	})
+	t.Run("geomean", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{0, nan, 2.9925557394776896, 3.365865436338599}
+		f("geomean", valuesExpected)
+	})
+}
+
+func testIncrementalParallelAggr(iafc *incrementalAggrFuncContext, tssSrc, tssExpected []*timeseries) error {
+	const workersCount = 3
+	tsCh := make(chan *timeseries)
+	var wg sync.WaitGroup
+	wg.Add(workersCount)
+	for i := 0; i < workersCount; i++ {
+		go func(workerID uint) {
+			defer wg.Done()
+			for ts := range tsCh {
+				runtime.Gosched() // allow other goroutines performing the work
+				iafc.updateTimeseries(ts, workerID)
+			}
+		}(uint(i))
+	}
+	for _, ts := range tssSrc {
+		tsCh <- ts
+	}
+	close(tsCh)
+	wg.Wait()
+	tssActual := iafc.finalizeTimeseries()
+	if err := expectTimeseriesEqual(tssActual, tssExpected); err != nil {
+		return fmt.Errorf("%s; tssActual=%v, tssExpected=%v", err, tssActual, tssExpected)
+	}
+	return nil
+}
+
+func expectTimeseriesEqual(actual, expected []*timeseries) error {
+	if len(actual) != len(expected) {
+		return fmt.Errorf("unexpected number of time series; got %d; want %d", len(actual), len(expected))
+	}
+	mActual := timeseriesToMap(actual)
+	mExpected := timeseriesToMap(expected)
+	if len(mActual) != len(mExpected) {
+		return fmt.Errorf("unexpected number of time series after converting to map; got %d; want %d", len(mActual), len(mExpected))
+	}
+	for k, tsExpected := range mExpected {
+		tsActual := mActual[k]
+		if tsActual == nil {
+			return fmt.Errorf("missing time series for key=%q", k)
+		}
+		if err := expectTsEqual(tsActual, tsExpected); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func timeseriesToMap(tss []*timeseries) map[string]*timeseries {
+	m := make(map[string]*timeseries, len(tss))
+	for _, ts := range tss {
+		k := ts.MetricName.Marshal(nil)
+		m[string(k)] = ts
+	}
+	return m
+}
+
+func expectTsEqual(actual, expected *timeseries) error {
+	mnActual := actual.MetricName.Marshal(nil)
+	mnExpected := expected.MetricName.Marshal(nil)
+	if string(mnActual) != string(mnExpected) {
+		return fmt.Errorf("unexpected metric name; got %q; want %q", mnActual, mnExpected)
+	}
+	if !reflect.DeepEqual(actual.Timestamps, expected.Timestamps) {
+		return fmt.Errorf("unexpected timestamps; got %v; want %v", actual.Timestamps, expected.Timestamps)
+	}
+	if err := compareValues(actual.Values, expected.Values); err != nil {
+		return fmt.Errorf("%s; actual %v; expected %v", err, actual.Values, expected.Values)
+	}
+	return nil
+}
+
+func compareValues(vs1, vs2 []float64) error {
+	if len(vs1) != len(vs2) {
+		return fmt.Errorf("unexpected number of values; got %d; want %d", len(vs1), len(vs2))
+	}
+	for i, v1 := range vs1 {
+		v2 := vs2[i]
+		if math.IsNaN(v1) {
+			if !math.IsNaN(v2) {
+				return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
+			}
+			continue
+		}
+		eps := math.Abs(v1 - v2)
+		if eps > 1e-14 {
+			return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
+		}
+	}
+	return nil
+}

app/vmselect/promql/binary_op.go

@@ -260,6 +260,9 @@ func newBinaryOpFunc(bf func(left, right float64, isBool bool) float64) binaryOp
 				dstValues[j] = bf(a, b, isBool)
 			}
 		}
+		// Optimization: remove time series containing only NaNs.
+		// This is quite common after applying filters like `q > 0`.
+		dst = removeNaNs(dst)
 		return dst, nil
 	}
 }
@@ -319,6 +322,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 			}
 			src := tssRight[0]
 			for _, ts := range tssLeft {
+				resetMetricGroupIfRequired(be, ts)
 				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
 				rvsLeft = append(rvsLeft, ts)
 				rvsRight = append(rvsRight, src)
@@ -329,6 +333,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 			}
 			src := tssLeft[0]
 			for _, ts := range tssRight {
+				resetMetricGroupIfRequired(be, ts)
 				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
 				rvsLeft = append(rvsLeft, src)
 				rvsRight = append(rvsRight, ts)
@@ -413,10 +418,25 @@ func binaryOpIfnot(left, right float64) float64 {
 }
 
 func binaryOpEq(left, right float64) bool {
+	// Special handling for nan == nan.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
+	if math.IsNaN(left) {
+		return math.IsNaN(right)
+	}
+
 	return left == right
 }
 
 func binaryOpNeq(left, right float64) bool {
+	// Special handling for comparison with nan.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
+	if math.IsNaN(left) {
+		return !math.IsNaN(right)
+	}
+	if math.IsNaN(right) {
+		return true
+	}
+
 	return left != right
 }
 

app/vmselect/promql/eval.go

@@ -466,31 +466,19 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	preFunc, rcs := getRollupConfigs(name, rf, ec.Start, ec.End, ec.Step, window, sharedTimestamps)
 	tss := make([]*timeseries, 0, len(tssSQ)*len(rcs))
 	var tssLock sync.Mutex
+	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
 	doParallel(tssSQ, func(tsSQ *timeseries, values []float64, timestamps []int64) ([]float64, []int64) {
 		values, timestamps = removeNanValues(values[:0], timestamps[:0], tsSQ.Values, tsSQ.Timestamps)
 		preFunc(values, timestamps)
 		for _, rc := range rcs {
 			var ts timeseries
-			ts.MetricName.CopyFrom(&tsSQ.MetricName)
-			if len(rc.TagValue) > 0 {
-				ts.MetricName.AddTag("rollup", rc.TagValue)
-			}
-			ts.Values = rc.Do(ts.Values[:0], values, timestamps)
-			ts.Timestamps = sharedTimestamps
-			ts.denyReuse = true
-
+			doRollupForTimeseries(rc, &ts, &tsSQ.MetricName, values, timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
 			tss = append(tss, &ts)
 			tssLock.Unlock()
 		}
 		return values, timestamps
 	})
-	if !rollupFuncsKeepMetricGroup[name] {
-		tss = copyTimeseriesMetricNames(tss)
-		for _, ts := range tss {
-			ts.MetricName.ResetMetricGroup()
-		}
-	}
 	return tss, nil
 }
 
@@ -582,7 +570,7 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 		MaxTimestamp: ec.End + ec.Step,
 		TagFilterss:  [][]storage.TagFilter{me.TagFilters},
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, ec.Deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, true, ec.Deadline)
 	if err != nil {
 		return nil, err
 	}
@@ -614,22 +602,16 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	defer rml.Put(uint64(rollupMemorySize))
 
 	// Evaluate rollup
+	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
 	var tss []*timeseries
 	if iafc != nil {
-		tss, err = evalRollupWithIncrementalAggregate(iafc, rss, rcs, preFunc, sharedTimestamps)
+		tss, err = evalRollupWithIncrementalAggregate(iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	} else {
-		tss, err = evalRollupNoIncrementalAggregate(rss, rcs, preFunc, sharedTimestamps)
+		tss, err = evalRollupNoIncrementalAggregate(rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	}
 	if err != nil {
 		return nil, err
 	}
-
-	if !rollupFuncsKeepMetricGroup[name] {
-		tss = copyTimeseriesMetricNames(tss)
-		for _, ts := range tss {
-			ts.MetricName.ResetMetricGroup()
-		}
-	}
 	tss = mergeTimeseries(tssCached, tss, start, ec)
 	rollupResultCacheV.Put(name, ec, me, iafc, window, tss)
 
@@ -649,21 +631,19 @@ func getRollupMemoryLimiter() *memoryLimiter {
 }
 
 func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
-	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64) ([]*timeseries, error) {
-	err := rss.RunParallel(func(rs *netstorage.Result) {
+	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 {
 			ts.Reset()
-			ts.MetricName.CopyFrom(&rs.MetricName)
-			if len(rc.TagValue) > 0 {
-				ts.MetricName.AddTag("rollup", rc.TagValue)
-			}
-			ts.Values = rc.Do(ts.Values[:0], rs.Values, rs.Timestamps)
-			ts.Timestamps = sharedTimestamps
-			iafc.updateTimeseries(ts)
+			doRollupForTimeseries(rc, ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
+			iafc.updateTimeseries(ts, workerID)
+
+			// ts.Timestamps points to sharedTimestamps. Zero it, so it can be re-used.
 			ts.Timestamps = nil
+			ts.denyReuse = false
 		}
 	})
 	if err != nil {
@@ -674,21 +654,14 @@ func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *n
 }
 
 func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConfig,
-	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64) ([]*timeseries, error) {
+	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) {
+	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		for _, rc := range rcs {
 			var ts timeseries
-			ts.MetricName.CopyFrom(&rs.MetricName)
-			if len(rc.TagValue) > 0 {
-				ts.MetricName.AddTag("rollup", rc.TagValue)
-			}
-			ts.Values = rc.Do(ts.Values[:0], rs.Values, rs.Timestamps)
-			ts.Timestamps = sharedTimestamps
-			ts.denyReuse = true
-
+			doRollupForTimeseries(rc, &ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
 			tss = append(tss, &ts)
 			tssLock.Unlock()
@@ -700,6 +673,20 @@ func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConf
 	return tss, nil
 }
 
+func doRollupForTimeseries(rc *rollupConfig, tsDst *timeseries, mnSrc *storage.MetricName, valuesSrc []float64, timestampsSrc []int64,
+	sharedTimestamps []int64, removeMetricGroup bool) {
+	tsDst.MetricName.CopyFrom(mnSrc)
+	if len(rc.TagValue) > 0 {
+		tsDst.MetricName.AddTag("rollup", rc.TagValue)
+	}
+	if removeMetricGroup {
+		tsDst.MetricName.ResetMetricGroup()
+	}
+	tsDst.Values = rc.Do(tsDst.Values[:0], valuesSrc, timestampsSrc)
+	tsDst.Timestamps = sharedTimestamps
+	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] {

app/vmselect/promql/exec.go

@@ -16,6 +16,8 @@ import (
 
 var logSlowQueryDuration = flag.Duration("search.logSlowQueryDuration", 5*time.Second, "Log queries with execution time exceeding this value. Zero disables slow query logging")
 
+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) {
@@ -36,6 +38,7 @@ func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result,
 			if d >= *logSlowQueryDuration {
 				logger.Infof("slow query according to -search.logSlowQueryDuration=%s: duration=%s, start=%d, end=%d, step=%d, query=%q",
 					*logSlowQueryDuration, d, ec.Start/1000, ec.End/1000, ec.Step/1000, q)
+				slowQueries.Inc()
 			}
 		}()
 	}
@@ -102,14 +105,14 @@ 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)] {
+		if _, ok := m[string(bb.B)]; ok {
 			return nil, fmt.Errorf(`duplicate output timeseries: %s%s`, ts.MetricName.MetricGroup, stringMetricName(&ts.MetricName))
 		}
-		m[string(bb.B)] = true
+		m[string(bb.B)] = struct{}{}
 
 		rs := &result[i]
 		rs.MetricNameMarshaled = append(rs.MetricNameMarshaled[:0], bb.B...)
@@ -131,18 +134,23 @@ func timeseriesToResult(tss []*timeseries, maySort bool) ([]netstorage.Result, e
 func removeNaNs(tss []*timeseries) []*timeseries {
 	rvs := tss[:0]
 	for _, ts := range tss {
-		nans := 0
+		allNans := true
 		for _, v := range ts.Values {
-			if math.IsNaN(v) {
-				nans++
+			if !math.IsNaN(v) {
+				allNans = false
+				break
 			}
 		}
-		if nans == len(ts.Values) {
+		if allNans {
 			// Skip timeseries with all NaNs.
 			continue
 		}
 		rvs = append(rvs, ts)
 	}
+	for i := len(rvs); i < len(tss); i++ {
+		// Zero unused time series, so GC could reclaim them.
+		tss[i] = nil
+	}
 	return rvs
 }
 

app/vmselect/promql/exec_test.go

@@ -1302,6 +1302,44 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`label_value()`, func(t *testing.T) {
+		t.Parallel()
+		q := `with (
+			x = (
+				label_set(time(), "foo", "123.456", "__name__", "aaa"),
+				label_set(-time(), "foo", "bar", "__name__", "bbb"),
+				label_set(-time(), "__name__", "bxs"),
+				label_set(-time(), "foo", "45", "bar", "xs"),
+			)
+		)
+		sort(x + label_value(x, "foo"))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{-955, -1155, -1355, -1555, -1755, -1955},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("bar"),
+				Value: []byte("xs"),
+			},
+			{
+				Key:   []byte("foo"),
+				Value: []byte("45"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1123.456, 1323.456, 1523.456, 1723.456, 1923.456, 2123.456},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("123.456"),
+		}}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
 	t.Run(`label_transform(mismatch)`, func(t *testing.T) {
 		t.Parallel()
 		q := `label_transform(time(), "__name__", "foobar", "xx")`
@@ -1821,9 +1859,9 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
-	t.Run(`vector * on(foo) group_left() duplicate_timeseries`, func(t *testing.T) {
+	t.Run(`vector * on(foo) group_left() duplicate_nonoverlapping_timeseries`, func(t *testing.T) {
 		t.Parallel()
-		q := `label_set(time()/10, "foo", "bar") + on(foo) group_left() (
+		q := `label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_left() (
 			label_set(time() < 1400, "foo", "bar", "op", "le"),
 			label_set(time() >= 1400, "foo", "bar", "op", "ge"),
 		)`
@@ -1832,13 +1870,85 @@ func TestExecSuccess(t *testing.T) {
 			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
 			Timestamps: timestampsExpected,
 		}
-		r1.MetricName.Tags = []storage.Tag{{
-			Key:   []byte("foo"),
-			Value: []byte("bar"),
-		}}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
 		resultExpected := []netstorage.Result{r1}
 		f(q, resultExpected)
 	})
+	t.Run(`vector * on(foo) group_left(__name__)`, func(t *testing.T) {
+		t.Parallel()
+		q := `label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_left(__name__)
+			label_set(time(), "foo", "bar", "__name__", "aaa")`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.MetricGroup = []byte("aaa")
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1}
+		f(q, resultExpected)
+	})
+	t.Run(`vector * on(foo) group_right()`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort(label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_right(xx) (
+			label_set(time(), "foo", "bar", "__name__", "aaa"),
+			label_set(time()+3, "foo", "bar", "__name__", "yyy","ppp", "123"),
+		))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1103, 1323, 1543, 1763, 1983, 2203},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("ppp"),
+				Value: []byte("123"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
 	t.Run(`vector * on() group_left scalar`, func(t *testing.T) {
 		t.Parallel()
 		q := `sort_desc((label_set(time(), "foo", "bar") or label_set(10, "foo", "qwert")) * on() group_left 2)`
@@ -2158,6 +2268,83 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r1, r2}
 		f(q, resultExpected)
 	})
+	t.Run(`histogram_quantile(negative-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(90, "foo", "bar", "le", "10")
+			or label_set(-100, "foo", "bar", "le", "30")
+			or label_set(300, "foo", "bar", "le", "+Inf")
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{30, 30, 30, 30, 30, 30},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(90, "foo", "bar", "le", "10")
+			or label_set(NaN, "foo", "bar", "le", "30")
+			or label_set(300, "foo", "bar", "le", "+Inf")
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{30, 30, 30, 30, 30, 30},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.2,
+			label_set(0, "foo", "bar", "le", "10")
+			or label_set(100, "foo", "bar", "le", "30")
+			or label_set(300, "foo", "bar", "le", "+Inf")
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{22, 22, 22, 22, 22, 22},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(zero-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(0, "foo", "bar", "le", "10")
+			or label_set(0, "foo", "bar", "le", "30")
+			or label_set(0, "foo", "bar", "le", "+Inf")
+		)`
+		resultExpected := []netstorage.Result{}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(nan, "foo", "bar", "le", "10")
+			or label_set(nan, "foo", "bar", "le", "30")
+			or label_set(nan, "foo", "bar", "le", "+Inf")
+		)`
+		resultExpected := []netstorage.Result{}
+		f(q, resultExpected)
+	})
 	t.Run(`median_over_time()`, func(t *testing.T) {
 		t.Parallel()
 		q := `median_over_time({})`
@@ -2343,10 +2530,10 @@ func TestExecSuccess(t *testing.T) {
 	})
 	t.Run(`count(multi-vector)`, func(t *testing.T) {
 		t.Parallel()
-		q := `count(label_set(10, "foo", "bar") or label_set((15-time()/100)^0.5, "baz", "sss"))`
+		q := `count(label_set(time()<1500, "foo", "bar") or label_set(time()<1800, "baz", "sss"))`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{2, 2, 2, 1, 1, 1},
+			Values:     []float64{2, 2, 2, 1, nan, nan},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -3703,6 +3890,107 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r1, r2, r3, r4, r5, r6}
 		f(q, resultExpected)
 	})
+	t.Run(`count_values by (xxx)`, func(t *testing.T) {
+		t.Parallel()
+		q := `count_values("xxx", label_set(10, "foo", "bar", "xxx", "aaa") or label_set(floor(time()/600), "foo", "bar", "baz", "xx")) by (xxx)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, nan, nan, nan, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("1"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, 1, 1, 1, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("2"),
+			},
+		}
+		r3 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, nan, nan, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r3.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("3"),
+			},
+		}
+		r4 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, 1, 1, 1, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r4.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("10"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2, r3, r4}
+		f(q, resultExpected)
+	})
+	t.Run(`count_values without (baz)`, func(t *testing.T) {
+		t.Parallel()
+		q := `count_values("xxx", label_set(floor(time()/600), "foo", "bar")) without (baz)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, nan, nan, nan, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("1"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, 1, 1, 1, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("2"),
+			},
+		}
+		r3 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, nan, nan, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r3.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("3"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2, r3}
+		f(q, resultExpected)
+	})
 }
 
 func TestExecError(t *testing.T) {

app/vmselect/promql/lexer.go

@@ -149,12 +149,6 @@ func scanString(s string) (string, error) {
 }
 
 func scanPositiveNumber(s string) (string, error) {
-	if strings.HasPrefix(s, "Inf") {
-		return "Inf", nil
-	}
-	if strings.HasPrefix(s, "NaN") {
-		return "NaN", nil
-	}
 	// Scan integer part. It may be empty if fractional part exists.
 	i := 0
 	for i < len(s) && isDecimalChar(s[i]) {
@@ -333,6 +327,14 @@ func scanTagFilterOpPrefix(s string) int {
 	return -1
 }
 
+func isInfOrNaN(s string) bool {
+	if len(s) != 3 {
+		return false
+	}
+	s = strings.ToLower(s)
+	return s == "inf" || s == "nan"
+}
+
 func isOffset(s string) bool {
 	s = strings.ToLower(s)
 	return s == "offset"
@@ -361,7 +363,7 @@ func isPositiveNumberPrefix(s string) bool {
 
 	// Check for .234 numbers
 	if s[0] != '.' || len(s) < 2 {
-		return strings.HasPrefix(s, "Inf") || strings.HasPrefix(s, "NaN")
+		return false
 	}
 	return isDecimalChar(s[1])
 }

app/vmselect/promql/parser.go

@@ -373,7 +373,7 @@ func (p *parser) parseSingleExpr() (expr, error) {
 }
 
 func (p *parser) parseSingleExprWithoutRollupSuffix() (expr, error) {
-	if isPositiveNumberPrefix(p.lex.Token) {
+	if isPositiveNumberPrefix(p.lex.Token) || isInfOrNaN(p.lex.Token) {
 		return p.parsePositiveNumberExpr()
 	}
 	if isStringPrefix(p.lex.Token) {
@@ -417,7 +417,7 @@ func (p *parser) parseSingleExprWithoutRollupSuffix() (expr, error) {
 }
 
 func (p *parser) parsePositiveNumberExpr() (*numberExpr, error) {
-	if !isPositiveNumberPrefix(p.lex.Token) {
+	if !isPositiveNumberPrefix(p.lex.Token) && !isInfOrNaN(p.lex.Token) {
 		return nil, fmt.Errorf(`positiveNumberExpr: unexpected token %q; want "number"`, p.lex.Token)
 	}
 

app/vmselect/promql/parser_test.go

@@ -170,14 +170,34 @@ func TestParsePromQLSuccess(t *testing.T) {
 	another(`-.2`, `-0.2`)
 	another(`-.2E-2`, `-0.002`)
 	same(`NaN`)
+	another(`nan`, `NaN`)
+	another(`NAN`, `NaN`)
+	another(`nAN`, `NaN`)
 	another(`Inf`, `+Inf`)
+	another(`INF`, `+Inf`)
+	another(`inf`, `+Inf`)
 	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(`nan == nan`, `NaN`)
+	another(`nan ==bool nan`, `1`)
+	another(`nan !=bool nan`, `0`)
+	another(`nan !=bool 2`, `1`)
+	another(`2 !=bool nan`, `1`)
+	another(`nan >bool nan`, `0`)
+	another(`nan <bool nan`, `0`)
+	another(`1 ==bool nan`, `0`)
+	another(`NaN !=bool 1`, `1`)
+	another(`inf >=bool 2`, `1`)
+	another(`-1 >bool -inf`, `1`)
+	another(`-1 <bool -inf`, `0`)
+	another(`nan + 2 *3 * inf`, `NaN`)
+	another(`INF - Inf`, `NaN`)
+	another(`Inf + inf`, `+Inf`)
+	another(`1/0`, `+Inf`)
+	another(`0/0`, `NaN`)
 	another(`-m`, `0 - m`)
 	same(`m + ignoring () n[5m]`)
 	another(`M + IGNORING () N[5m]`, `M + ignoring () N[5m]`)

app/vmselect/promql/rollup.go

@@ -45,6 +45,8 @@ var rollupFuncs = map[string]newRollupFunc{
 	"distinct_over_time": newRollupFuncOneArg(rollupDistinct),
 	"integrate":          newRollupFuncOneArg(rollupIntegrate),
 	"ideriv":             newRollupFuncOneArg(rollupIderiv),
+	"lifetime":           newRollupFuncOneArg(rollupLifetime),
+	"scrape_interval":    newRollupFuncOneArg(rollupScrapeInterval),
 	"rollup":             newRollupFuncOneArg(rollupFake),
 	"rollup_rate":        newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
 	"rollup_deriv":       newRollupFuncOneArg(rollupFake),
@@ -54,10 +56,15 @@ var rollupFuncs = map[string]newRollupFunc{
 }
 
 var rollupFuncsMayAdjustWindow = map[string]bool{
-	"deriv":      true,
-	"deriv_fast": true,
-	"irate":      true,
-	"rate":       true,
+	"default_rollup":  true,
+	"first_over_time": true,
+	"last_over_time":  true,
+	"deriv":           true,
+	"deriv_fast":      true,
+	"irate":           true,
+	"rate":            true,
+	"lifetime":        true,
+	"scrape_interval": true,
 }
 
 var rollupFuncsRemoveCounterResets = map[string]bool{
@@ -190,23 +197,21 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 
 	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]
 		}
@@ -222,16 +227,73 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	return dstValues
 }
 
+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 := sort.Search(len(timestamps), func(n int) bool {
+		return n >= 0 && n < len(timestamps) && timestamps[n] > seekTimestamp
+	})
+	return startIdx + i
+}
+
 func getMaxPrevInterval(timestamps []int64) int64 {
 	if len(timestamps) < 2 {
 		return int64(maxSilenceInterval)
 	}
 	d := (timestamps[len(timestamps)-1] - timestamps[0]) / int64(len(timestamps)-1)
 	if d <= 0 {
-		return 1
+		return int64(maxSilenceInterval)
 	}
-	// Slightly increase d in order to handle possible jitter in scrape interval.
-	return d + (d / 16)
+	// Increase d 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 d <= 4*1000 {
+		return d + 2*d
+	}
+	if d <= 8*1000 {
+		return d + d
+	}
+	if d <= 16*1000 {
+		return d + d/2
+	}
+	if d <= 32*1000 {
+		return d + d/4
+	}
+	return d + d/8
 }
 
 func removeCounterResets(values []float64) {
@@ -595,10 +657,15 @@ func rollupDelta(rfa *rollupFuncArg) float64 {
 		if len(values) == 0 {
 			return nan
 		}
+		if len(values) == 1 {
+			// Assume that the previous non-existing value was 0.
+			return values[0]
+		}
 		prevValue = values[0]
 		values = values[1:]
 	}
 	if len(values) == 0 {
+		// Assume that the value didn't change on the given interval.
 		return 0
 	}
 	return values[len(values)-1] - prevValue
@@ -612,6 +679,7 @@ 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]
@@ -619,7 +687,8 @@ 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
 	}
@@ -641,7 +710,8 @@ func rollupDerivFast(rfa *rollupFuncArg) float64 {
 	prevValue := rfa.prevValue
 	prevTimestamp := rfa.prevTimestamp
 	if math.IsNaN(prevValue) {
-		if len(values) == 0 {
+		if len(values) < 2 {
+			// It is impossible to calculate derivative on 0 or 1 values.
 			return nan
 		}
 		prevValue = values[0]
@@ -650,6 +720,7 @@ func rollupDerivFast(rfa *rollupFuncArg) float64 {
 		timestamps = timestamps[1:]
 	}
 	if len(values) == 0 {
+		// Assume that the value didn't change on the given interval.
 		return 0
 	}
 	vEnd := values[len(values)-1]
@@ -664,11 +735,12 @@ 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 || math.IsNaN(rfa.prevValue) {
+			// It is impossible to calculate derivative on 0 or 1 values.
 			return nan
 		}
-		return 0
+		return (values[0] - rfa.prevValue) / (float64(timestamps[0]-rfa.prevTimestamp) * 1e-3)
 	}
 	vEnd := values[len(values)-1]
 	tEnd := timestamps[len(timestamps)-1]
@@ -692,7 +764,37 @@ 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 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 {

app/vmselect/promql/rollup_result_cache.go

@@ -4,14 +4,15 @@ 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/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -19,7 +20,7 @@ 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 +44,13 @@ 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 +98,26 @@ 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"}`)
@@ -148,15 +151,23 @@ 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)
 		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)
 	}
@@ -196,6 +207,8 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	return tss, newStart
 }
 
+var resultBufPool bytesutil.ByteBufferPool
+
 func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64, tss []*timeseries) {
 	if *disableCache || len(tss) == 0 || !ec.mayCache() {
 		return
@@ -227,11 +240,16 @@ 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)
@@ -240,7 +258,7 @@ 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)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
@@ -270,7 +288,7 @@ 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 = 6
 
 func marshalRollupResultCacheKey(dst []byte, funcName string, me *metricExpr, iafc *incrementalAggrFuncContext, window, step int64) []byte {
 	dst = append(dst, rollupResultCacheVersion)

app/vmselect/promql/rollup_test.go

@@ -45,8 +45,19 @@ 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, 0.5)
 	}
 
 	rfa = &rollupFuncArg{
@@ -171,7 +182,8 @@ 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)
 			}
 		}
@@ -347,7 +359,7 @@ 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, nan}
+		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)
 	})
@@ -371,15 +383,15 @@ func TestRollupWindowNoPoints(t *testing.T) {
 	t.Run("afterEnd", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupFirst,
-			Start:  141,
-			End:    171,
+			Start:  161,
+			End:    191,
 			Step:   10,
 			Window: 3,
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{34, nan, nan, nan}
-		timestampsExpected := []int64{141, 151, 161, 171}
+		valuesExpected := []float64{nan, nan, nan, nan}
+		timestampsExpected := []int64{161, 171, 181, 191}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 }
@@ -409,7 +421,7 @@ 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{12, 44, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -468,7 +480,7 @@ 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)
 	})
@@ -569,10 +581,66 @@ 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("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,
@@ -685,7 +753,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		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 +767,20 @@ 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 testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExpected []float64, timestampsExpected []int64) {

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)
 	}

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)

app/vmselect/promql/transform.go

@@ -63,6 +63,7 @@ var transformFuncs = map[string]transformFunc{
 	"label_copy":         transformLabelCopy,
 	"label_move":         transformLabelMove,
 	"label_transform":    transformLabelTransform,
+	"label_value":        transformLabelValue,
 	"union":              transformUnion,
 	"":                   transformUnion, // empty func is a synonim to union
 	"keep_last_value":    transformKeepLastValue,
@@ -308,8 +309,16 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 	bbPool.Put(bb)
 
 	// Calculate quantile for each group in m
-	lastNonInf := func(xss []x) float64 {
-		for len(xss) > 0 && math.IsInf(xss[len(xss)-1].le, 0) {
+
+	lastNonInf := func(i int, xss []x) float64 {
+		for len(xss) > 0 {
+			xsLast := xss[len(xss)-1]
+			if xsLast.ts.Values[i] == 0 {
+				return nan
+			}
+			if !math.IsInf(xsLast.le, 0) {
+				break
+			}
 			xss = xss[:len(xss)-1]
 		}
 		if len(xss) == 0 {
@@ -318,42 +327,57 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 		return xss[len(xss)-1].le
 	}
 	quantile := func(i int, phis []float64, xss []x) float64 {
-		vPrev := float64(0)
-		lePrev := float64(0)
 		phi := phis[i]
 		if math.IsNaN(phi) {
 			return nan
 		}
+		// Fix broken buckets.
+		// They are already sorted by le, so their values must be in ascending order,
+		// since the next bucket value includes all the previous buckets.
+		vPrev := float64(0)
+		for _, xs := range xss {
+			v := xs.ts.Values[i]
+			if math.IsNaN(v) || v < vPrev {
+				xs.ts.Values[i] = vPrev
+			} else {
+				vPrev = v
+			}
+		}
+		if len(xss) == 0 {
+			return nan
+		}
 		if phi < 0 {
 			return -inf
 		}
 		if phi > 1 {
 			return inf
 		}
-		vReq := xss[len(xss)-1].ts.Values[i] * phi
+		vLast := xss[len(xss)-1].ts.Values[i]
+		if vLast == 0 {
+			return nan
+		}
+		vReq := vLast * phi
+		vPrev = 0
+		lePrev := float64(0)
 		for _, xs := range xss {
 			v := xs.ts.Values[i]
 			le := xs.le
-			if v <= vPrev {
-				v = vPrev
-				le = lePrev
-			}
 			if v < vReq {
 				vPrev = v
 				lePrev = le
 				continue
 			}
 			if math.IsInf(le, 0) {
-				return lastNonInf(xss)
+				return lastNonInf(i, xss)
 			}
 			if v == vPrev {
 				return lePrev
 			}
 			return lePrev + (le-lePrev)*(vReq-vPrev)/(v-vPrev)
 		}
-		return lastNonInf(xss)
+		return lastNonInf(i, xss)
 	}
-	var rvs []*timeseries
+	rvs := make([]*timeseries, 0, len(m))
 	for _, xss := range m {
 		sort.Slice(xss, func(i, j int) bool {
 			return xss[i].le < xss[j].le
@@ -877,6 +901,33 @@ func labelReplace(tss []*timeseries, srcLabel string, r *regexp.Regexp, dstLabel
 	return tss, nil
 }
 
+func transformLabelValue(tfa *transformFuncArg) ([]*timeseries, error) {
+	args := tfa.args
+	if err := expectTransformArgsNum(args, 2); err != nil {
+		return nil, err
+	}
+	labelName, err := getString(args[1], 1)
+	if err != nil {
+		return nil, fmt.Errorf("cannot get label name: %s", err)
+	}
+	rvs := args[0]
+	for _, ts := range rvs {
+		ts.MetricName.ResetMetricGroup()
+		labelValue := ts.MetricName.GetTagValue(labelName)
+		v, err := strconv.ParseFloat(string(labelValue), 64)
+		if err != nil {
+			v = nan
+		}
+		values := ts.Values
+		for i := range values {
+			values[i] = v
+		}
+	}
+	// Do not remove timeseries with only NaN values, so `default` could be applied to them:
+	// label_value(q, "label") default 123
+	return rvs, nil
+}
+
 func transformLn(v float64) float64 {
 	return math.Log(v)
 }

app/vmstorage/main.go

@@ -358,6 +358,29 @@ func registerStorageMetrics() {
 		return float64(idbm().SizeBytes)
 	})
 
+	metrics.NewGauge(`vm_rows_ignored_total{reason="big_timestamp"}`, func() float64 {
+		return float64(m().TooBigTimestampRows)
+	})
+	metrics.NewGauge(`vm_rows_ignored_total{reason="small_timestamp"}`, func() float64 {
+		return float64(m().TooSmallTimestampRows)
+	})
+
+	metrics.NewGauge(`vm_concurrent_addrows_limit_reached_total`, func() float64 {
+		return float64(m().AddRowsConcurrencyLimitReached)
+	})
+	metrics.NewGauge(`vm_concurrent_addrows_limit_timeout_total`, func() float64 {
+		return float64(m().AddRowsConcurrencyLimitTimeout)
+	})
+	metrics.NewGauge(`vm_concurrent_addrows_dropped_rows_total`, func() float64 {
+		return float64(m().AddRowsConcurrencyDroppedRows)
+	})
+	metrics.NewGauge(`vm_concurrent_addrows_capacity`, func() float64 {
+		return float64(m().AddRowsConcurrencyCapacity)
+	})
+	metrics.NewGauge(`vm_concurrent_addrows_current`, func() float64 {
+		return float64(m().AddRowsConcurrencyCurrent)
+	})
+
 	metrics.NewGauge(`vm_rows{type="storage/big"}`, func() float64 {
 		return float64(tm().BigRowsCount)
 	})

dashboards/victoriametrics.json

@@ -60,12 +60,12 @@
       }
     ]
   },
-  "description": "Overview for single node VictoriaMetrics",
+  "description": "Overview for single node VictoriaMetrics v1.22.2 or higher",
   "editable": true,
   "gnetId": 10229,
   "graphTooltip": 0,
   "id": null,
-  "iteration": 1562261153865,
+  "iteration": 1563651131627,
   "links": [
     {
       "icon": "doc",
@@ -1693,7 +1693,7 @@
       },
       "id": 46,
       "panels": [],
-      "title": "Go runtime",
+      "title": "Resource usage",
       "type": "row"
     },
     {
@@ -1805,7 +1805,6 @@
       "bars": false,
       "dashLength": 10,
       "dashes": false,
-      "datasource": "${DS_PROMETHEUS}",
       "fill": 1,
       "gridPos": {
         "h": 8,
@@ -1813,13 +1812,13 @@
         "x": 12,
         "y": 69
       },
-      "id": 42,
+      "id": 57,
       "legend": {
         "avg": false,
         "current": false,
         "max": false,
         "min": false,
-        "show": false,
+        "show": true,
         "total": false,
         "values": false
       },
@@ -1838,10 +1837,10 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(rate(go_gc_duration_seconds_sum{job=\"$job\"}[2m]))",
+          "expr": "rate(process_cpu_seconds_total{job=\"$job\"}[5m])",
           "format": "time_series",
-          "intervalFactor": 2,
-          "legendFormat": "gc duration",
+          "intervalFactor": 1,
+          "legendFormat": "Rate 5m",
           "refId": "A"
         }
       ],
@@ -1849,7 +1848,7 @@
       "timeFrom": null,
       "timeRegions": [],
       "timeShift": null,
-      "title": "GC duration",
+      "title": "CPU",
       "tooltip": {
         "shared": true,
         "sort": 0,
@@ -1865,7 +1864,7 @@
       },
       "yaxes": [
         {
-          "format": "s",
+          "format": "short",
           "label": null,
           "logBase": 1,
           "max": null,
@@ -1986,6 +1985,92 @@
         "x": 12,
         "y": 77
       },
+      "id": 42,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": false,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {},
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(rate(go_gc_duration_seconds_sum{job=\"$job\"}[2m]))",
+          "format": "time_series",
+          "intervalFactor": 2,
+          "legendFormat": "gc duration",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "GC duration",
+      "tooltip": {
+        "shared": true,
+        "sort": 0,
+        "value_type": "individual"
+      },
+      "type": "graph",
+      "xaxis": {
+        "buckets": null,
+        "mode": "time",
+        "name": null,
+        "show": true,
+        "values": []
+      },
+      "yaxes": [
+        {
+          "format": "s",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        },
+        {
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        }
+      ],
+      "yaxis": {
+        "align": false,
+        "alignLevel": null
+      }
+    },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "fill": 1,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 0,
+        "y": 85
+      },
       "id": 48,
       "legend": {
         "avg": false,
@@ -2011,7 +2096,7 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(go_threads{job=\"$job\"})",
+          "expr": "sum(process_num_threads{job=\"$job\"})",
           "format": "time_series",
           "intervalFactor": 2,
           "legendFormat": "threads",
@@ -2145,5 +2230,5 @@
   "timezone": "",
   "title": "VictoriaMetrics",
   "uid": "wNf0q_kZk",
-  "version": 1
+  "version": 2
 }

deployment/docker/Makefile

@@ -1,5 +1,5 @@
 DOCKER_NAMESPACE := victoriametrics
-BUILDER_IMAGE := local/builder:go1.12.6
+BUILDER_IMAGE := local/builder:go1.13.0
 CERTS_IMAGE := local/certs:1.0.2
 
 package-certs:
@@ -19,8 +19,9 @@ app-via-docker: package-certs package-builder
 		--mount type=bind,src="$(shell pwd)/gocache-for-docker",dst=/gocache \
 		--env GOCACHE=/gocache \
 		--env GO111MODULE=on \
+		$(DOCKER_OPTS) \
 		$(BUILDER_IMAGE) \
-		go build $(RACE) -mod=vendor -ldflags "-s -w -extldflags '-static' $(GO_BUILDINFO)" -tags 'netgo osusergo' -o bin/$(APP_NAME)-prod $(PKG_PREFIX)/app/$(APP_NAME)
+		go build $(RACE) -mod=vendor -ldflags "-s -w -extldflags '-static' $(GO_BUILDINFO)" -tags 'netgo osusergo' -o bin/$(APP_NAME)$(APP_SUFFIX)-prod $(PKG_PREFIX)/app/$(APP_NAME)
 
 package-via-docker:
 	(docker image ls --format '{{.Repository}}:{{.Tag}}' | grep -q '$(DOCKER_NAMESPACE)/$(APP_NAME):$(PKG_TAG)$(RACE)') || (\

deployment/docker/builder/Dockerfile

@@ -1,2 +1,2 @@
-FROM golang:1.12.6
+FROM golang:1.13.0
 STOPSIGNAL SIGINT

go.mod

@@ -2,15 +2,17 @@ module github.com/VictoriaMetrics/VictoriaMetrics
 
 require (
 	github.com/VictoriaMetrics/fastcache v1.5.1
-	github.com/VictoriaMetrics/metrics v1.6.2
+	github.com/VictoriaMetrics/metrics v1.7.1
 	github.com/cespare/xxhash/v2 v2.0.1-0.20190104013014-3767db7a7e18
 	github.com/golang/snappy v0.0.1
+	github.com/google/go-cmp v0.3.0 // indirect
+	github.com/klauspost/compress v1.7.6
 	github.com/spaolacci/murmur3 v1.1.0 // indirect
 	github.com/valyala/fastjson v1.4.1
-	github.com/valyala/gozstd v1.5.1
+	github.com/valyala/gozstd v1.6.1
 	github.com/valyala/histogram v1.0.1
-	github.com/valyala/quicktemplate v1.1.1
-	golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f
+	github.com/valyala/quicktemplate v1.2.0
+	golang.org/x/sys v0.0.0-20190813064441-fde4db37ae7a
 )
 
 go 1.12

go.sum

@@ -3,8 +3,8 @@ github.com/OneOfOne/xxhash v1.2.5 h1:zl/OfRA6nftbBK9qTohYBJ5xvw6C/oNKizR7cZGl3cI
 github.com/OneOfOne/xxhash v1.2.5/go.mod h1:eZbhyaAYD41SGSSsnmcpxVoRiQ/MPUTjUdIIOT9Um7Q=
 github.com/VictoriaMetrics/fastcache v1.5.1 h1:qHgHjyoNFV7jgucU8QZUuU4gcdhfs8QW1kw68OD2Lag=
 github.com/VictoriaMetrics/fastcache v1.5.1/go.mod h1:+jv9Ckb+za/P1ZRg/sulP5Ni1v49daAVERr0H3CuscE=
-github.com/VictoriaMetrics/metrics v1.6.2 h1:VMe8c8ZBPgNVZkPoT06LsoU2nb+8e7iPaOWbVRNhxjo=
-github.com/VictoriaMetrics/metrics v1.6.2/go.mod h1:LU2j9qq7xqZYXz8tF3/RQnB2z2MbZms5TDiIg9/NHiQ=
+github.com/VictoriaMetrics/metrics v1.7.1 h1:g2qrY6Upn8rvlvR40cGHFY0crwi4hpqF0n9vJMNsCSg=
+github.com/VictoriaMetrics/metrics v1.7.1/go.mod h1:LU2j9qq7xqZYXz8tF3/RQnB2z2MbZms5TDiIg9/NHiQ=
 github.com/allegro/bigcache v1.2.1-0.20190218064605-e24eb225f156 h1:eMwmnE/GDgah4HI848JfFxHt+iPb26b4zyfspmqY0/8=
 github.com/allegro/bigcache v1.2.1-0.20190218064605-e24eb225f156/go.mod h1:Cb/ax3seSYIx7SuZdm2G2xzfwmv3TPSk2ucNfQESPXM=
 github.com/cespare/xxhash v1.1.0 h1:a6HrQnmkObjyL+Gs60czilIUGqrzKutQD6XZog3p+ko=
@@ -16,9 +16,14 @@ github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/golang/snappy v0.0.1 h1:Qgr9rKW7uDUkrbSmQeiDsGa8SjGyCOGtuasMWwvp2P4=
 github.com/golang/snappy v0.0.1/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
+github.com/google/go-cmp v0.3.0 h1:crn/baboCvb5fXaQ0IJ1SGTsTVrWpDsCWC8EGETZijY=
+github.com/google/go-cmp v0.3.0/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU=
 github.com/klauspost/compress v1.4.0/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
 github.com/klauspost/compress v1.4.1/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
+github.com/klauspost/compress v1.7.6 h1:GH2karLOcuZtA5a3+KuzSU33A2cvcHGbtEWM6K4t7oU=
+github.com/klauspost/compress v1.7.6/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
 github.com/klauspost/cpuid v0.0.0-20180405133222-e7e905edc00e/go.mod h1:Pj4uuM528wm8OyEC2QMXAi2YiTZ96dNQPGgoMS4s3ek=
+github.com/klauspost/cpuid v1.2.0 h1:NMpwD2G9JSFOE1/TJjGSo5zG7Yb2bTe7eq1jH+irmeE=
 github.com/klauspost/cpuid v1.2.0/go.mod h1:Pj4uuM528wm8OyEC2QMXAi2YiTZ96dNQPGgoMS4s3ek=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
@@ -36,13 +41,13 @@ github.com/valyala/fastjson v1.4.1 h1:hrltpHpIpkaxll8QltMU8c3QZ5+qIiCL8yKqPFJI/y
 github.com/valyala/fastjson v1.4.1/go.mod h1:nV6MsjxL2IMJQUoHDIrjEI7oLyeqK6aBD7EFWPsvP8o=
 github.com/valyala/fastrand v1.0.0 h1:LUKT9aKer2dVQNUi3waewTbKV+7H17kvWFNKs2ObdkI=
 github.com/valyala/fastrand v1.0.0/go.mod h1:HWqCzkrkg6QXT8V2EXWvXCoow7vLwOFN002oeRzjapQ=
-github.com/valyala/gozstd v1.5.1 h1:ZLepItgu2g+B2CfVQy6KCV/as8lnJ7ef1KU6DPxQSS0=
-github.com/valyala/gozstd v1.5.1/go.mod h1:oYOS+oJovjw9ewtrwEYb9+ybolEXd6pHyLMuAWN5zts=
+github.com/valyala/gozstd v1.6.1 h1:oFN2mNW0kOr1fEKJuLpDwakNb6Y9fElVEBZmPEsFTUw=
+github.com/valyala/gozstd v1.6.1/go.mod h1:y5Ew47GLlP37EkTB+B4s7r6A5rdaeB7ftbl9zoYiIPQ=
 github.com/valyala/histogram v1.0.1 h1:FzA7n2Tz/wKRMejgu3PV1vw3htAklTjjuoI6z3d4KDg=
 github.com/valyala/histogram v1.0.1/go.mod h1:lQy0xA4wUz2+IUnf97SivorsJIp8FxsnRd6x25q7Mto=
-github.com/valyala/quicktemplate v1.1.1 h1:C58y/wN0FMTi2PR0n3onltemfFabany53j7M6SDDB8k=
-github.com/valyala/quicktemplate v1.1.1/go.mod h1:EH+4AkTd43SvgIbQHYu59/cJyxDoOVRUAfrukLPuGJ4=
+github.com/valyala/quicktemplate v1.2.0 h1:BaO1nHTkspYzmAjPXj0QiDJxai96tlcZyKcI9dyEGvM=
+github.com/valyala/quicktemplate v1.2.0/go.mod h1:EH+4AkTd43SvgIbQHYu59/cJyxDoOVRUAfrukLPuGJ4=
 github.com/valyala/tcplisten v0.0.0-20161114210144-ceec8f93295a/go.mod h1:v3UYOV9WzVtRmSR+PDvWpU/qWl4Wa5LApYYX4ZtKbio=
 golang.org/x/net v0.0.0-20180911220305-26e67e76b6c3/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
-golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f h1:25KHgbfyiSm6vwQLbM3zZIe1v9p/3ea4Rz+nnM5K/i4=
-golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20190813064441-fde4db37ae7a h1:aYOabOQFp6Vj6W1F80affTUvO9UxmJRx8K0gsfABByQ=
+golang.org/x/sys v0.0.0-20190813064441-fde4db37ae7a/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

lib/decimal/decimal_test.go

@@ -108,60 +108,60 @@ func testCalibrateScale(t *testing.T, a, b []int64, ae, be int16, aExpected, bEx
 }
 
 func TestMaxUpExponent(t *testing.T) {
-	testMaxUpExponent(t, 0, 1024)
-	testMaxUpExponent(t, -1<<63, 0)
-	testMaxUpExponent(t, (-1<<63)+1, 0)
-	testMaxUpExponent(t, (1<<63)-1, 0)
-	testMaxUpExponent(t, 1, 18)
-	testMaxUpExponent(t, 12, 17)
-	testMaxUpExponent(t, 123, 16)
-	testMaxUpExponent(t, 1234, 15)
-	testMaxUpExponent(t, 12345, 14)
-	testMaxUpExponent(t, 123456, 13)
-	testMaxUpExponent(t, 1234567, 12)
-	testMaxUpExponent(t, 12345678, 11)
-	testMaxUpExponent(t, 123456789, 10)
-	testMaxUpExponent(t, 1234567890, 9)
-	testMaxUpExponent(t, 12345678901, 8)
-	testMaxUpExponent(t, 123456789012, 7)
-	testMaxUpExponent(t, 1234567890123, 6)
-	testMaxUpExponent(t, 12345678901234, 5)
-	testMaxUpExponent(t, 123456789012345, 4)
-	testMaxUpExponent(t, 1234567890123456, 3)
-	testMaxUpExponent(t, 12345678901234567, 2)
-	testMaxUpExponent(t, 123456789012345678, 1)
-	testMaxUpExponent(t, 1234567890123456789, 0)
-	testMaxUpExponent(t, 923456789012345678, 0)
-	testMaxUpExponent(t, 92345678901234567, 1)
-	testMaxUpExponent(t, 9234567890123456, 2)
-	testMaxUpExponent(t, 923456789012345, 3)
-	testMaxUpExponent(t, 92345678901234, 4)
-	testMaxUpExponent(t, 9234567890123, 5)
-	testMaxUpExponent(t, 923456789012, 6)
-	testMaxUpExponent(t, 92345678901, 7)
-	testMaxUpExponent(t, 9234567890, 8)
-	testMaxUpExponent(t, 923456789, 9)
-	testMaxUpExponent(t, 92345678, 10)
-	testMaxUpExponent(t, 9234567, 11)
-	testMaxUpExponent(t, 923456, 12)
-	testMaxUpExponent(t, 92345, 13)
-	testMaxUpExponent(t, 9234, 14)
-	testMaxUpExponent(t, 923, 15)
-	testMaxUpExponent(t, 92, 17)
-	testMaxUpExponent(t, 9, 18)
-}
+	f := func(v int64, eExpected int16) {
+		t.Helper()
 
-func testMaxUpExponent(t *testing.T, v int64, eExpected int16) {
-	t.Helper()
+		e := maxUpExponent(v)
+		if e != eExpected {
+			t.Fatalf("unexpected e for v=%d; got %d; epxecting %d", v, e, eExpected)
+		}
+		e = maxUpExponent(-v)
+		if e != eExpected {
+			t.Fatalf("unexpected e for v=%d; got %d; expecting %d", -v, e, eExpected)
+		}
+	}
 
-	e := maxUpExponent(v)
-	if e != eExpected {
-		t.Fatalf("unexpected e for v=%d; got %d; epxecting %d", v, e, eExpected)
-	}
-	e = maxUpExponent(-v)
-	if e != eExpected {
-		t.Fatalf("unexpected e for v=%d; got %d; expecting %d", -v, e, eExpected)
-	}
+	f(0, 1024)
+	f(-1<<63, 0)
+	f((-1<<63)+1, 0)
+	f((1<<63)-1, 0)
+	f(1, 18)
+	f(12, 17)
+	f(123, 16)
+	f(1234, 15)
+	f(12345, 14)
+	f(123456, 13)
+	f(1234567, 12)
+	f(12345678, 11)
+	f(123456789, 10)
+	f(1234567890, 9)
+	f(12345678901, 8)
+	f(123456789012, 7)
+	f(1234567890123, 6)
+	f(12345678901234, 5)
+	f(123456789012345, 4)
+	f(1234567890123456, 3)
+	f(12345678901234567, 2)
+	f(123456789012345678, 1)
+	f(1234567890123456789, 0)
+	f(923456789012345678, 0)
+	f(92345678901234567, 1)
+	f(9234567890123456, 2)
+	f(923456789012345, 3)
+	f(92345678901234, 4)
+	f(9234567890123, 5)
+	f(923456789012, 6)
+	f(92345678901, 7)
+	f(9234567890, 8)
+	f(923456789, 9)
+	f(92345678, 10)
+	f(9234567, 11)
+	f(923456, 12)
+	f(92345, 13)
+	f(9234, 14)
+	f(923, 15)
+	f(92, 17)
+	f(9, 18)
 }
 
 func TestAppendFloatToDecimal(t *testing.T) {
@@ -207,83 +207,103 @@ func testAppendFloatToDecimal(t *testing.T, fa []float64, daExpected []int64, eE
 }
 
 func TestFloatToDecimal(t *testing.T) {
-	testFloatToDecimal(t, 0, 0, 0)
-	testFloatToDecimal(t, 1, 1, 0)
-	testFloatToDecimal(t, -1, -1, 0)
-	testFloatToDecimal(t, 0.9, 9, -1)
-	testFloatToDecimal(t, 0.99, 99, -2)
-	testFloatToDecimal(t, 9, 9, 0)
-	testFloatToDecimal(t, 99, 99, 0)
-	testFloatToDecimal(t, 20, 2, 1)
-	testFloatToDecimal(t, 100, 1, 2)
-	testFloatToDecimal(t, 3000, 3, 3)
-
-	testFloatToDecimal(t, 0.123, 123, -3)
-	testFloatToDecimal(t, -0.123, -123, -3)
-	testFloatToDecimal(t, 1.2345, 12345, -4)
-	testFloatToDecimal(t, -1.2345, -12345, -4)
-	testFloatToDecimal(t, 12000, 12, 3)
-	testFloatToDecimal(t, -12000, -12, 3)
-	testFloatToDecimal(t, 1e-30, 1, -30)
-	testFloatToDecimal(t, -1e-30, -1, -30)
-	testFloatToDecimal(t, 1e-260, 1, -260)
-	testFloatToDecimal(t, -1e-260, -1, -260)
-	testFloatToDecimal(t, 321e260, 321, 260)
-	testFloatToDecimal(t, -321e260, -321, 260)
-	testFloatToDecimal(t, 1234567890123, 1234567890123, 0)
-	testFloatToDecimal(t, -1234567890123, -1234567890123, 0)
-	testFloatToDecimal(t, 123e5, 123, 5)
-	testFloatToDecimal(t, 15e18, 15, 18)
-
-	testFloatToDecimal(t, math.Inf(1), vInfPos, 0)
-	testFloatToDecimal(t, math.Inf(-1), vInfNeg, 0)
-	testFloatToDecimal(t, 1<<63-1, 922337203685, 7)
-	testFloatToDecimal(t, -1<<63, -922337203685, 7)
-}
-
-func testFloatToDecimal(t *testing.T, f float64, vExpected int64, eExpected int16) {
-	t.Helper()
-
-	v, e := FromFloat(f)
-	if v != vExpected {
-		t.Fatalf("unexpected v for f=%e; got %d; expecting %d", f, v, vExpected)
-	}
-	if e != eExpected {
-		t.Fatalf("unexpected e for f=%e; got %d; expecting %d", f, e, eExpected)
+	f := func(f float64, vExpected int64, eExpected int16) {
+		t.Helper()
+		v, e := FromFloat(f)
+		if v != vExpected {
+			t.Fatalf("unexpected v for f=%e; got %d; expecting %d", f, v, vExpected)
+		}
+		if e != eExpected {
+			t.Fatalf("unexpected e for f=%e; got %d; expecting %d", f, e, eExpected)
+		}
 	}
+
+	f(0, 0, 0)
+	f(1, 1, 0)
+	f(-1, -1, 0)
+	f(0.9, 9, -1)
+	f(0.99, 99, -2)
+	f(9, 9, 0)
+	f(99, 99, 0)
+	f(20, 2, 1)
+	f(100, 1, 2)
+	f(3000, 3, 3)
+
+	f(0.123, 123, -3)
+	f(-0.123, -123, -3)
+	f(1.2345, 12345, -4)
+	f(-1.2345, -12345, -4)
+	f(12000, 12, 3)
+	f(-12000, -12, 3)
+	f(1e-30, 1, -30)
+	f(-1e-30, -1, -30)
+	f(1e-260, 1, -260)
+	f(-1e-260, -1, -260)
+	f(321e260, 321, 260)
+	f(-321e260, -321, 260)
+	f(1234567890123, 1234567890123, 0)
+	f(-1234567890123, -1234567890123, 0)
+	f(123e5, 123, 5)
+	f(15e18, 15, 18)
+
+	f(math.Inf(1), vInfPos, 0)
+	f(math.Inf(-1), vInfNeg, 0)
+	f(1<<63-1, 922337203685, 7)
+	f(-1<<63, -922337203685, 7)
+
+	// Test precision loss due to conversionPrecision.
+	f(0.1234567890123456, 12345678901234, -14)
+	f(-123456.7890123456, -12345678901234, -8)
 }
 
 func TestFloatToDecimalRoundtrip(t *testing.T) {
-	testFloatToDecimalRoundtrip(t, 0)
-	testFloatToDecimalRoundtrip(t, 1)
-	testFloatToDecimalRoundtrip(t, 0.123)
-	testFloatToDecimalRoundtrip(t, 1.2345)
-	testFloatToDecimalRoundtrip(t, 12000)
-	testFloatToDecimalRoundtrip(t, 1e-30)
-	testFloatToDecimalRoundtrip(t, 1e-260)
-	testFloatToDecimalRoundtrip(t, 321e260)
-	testFloatToDecimalRoundtrip(t, 1234567890123)
-	testFloatToDecimalRoundtrip(t, 12.34567890125)
-	testFloatToDecimalRoundtrip(t, 15e18)
+	f := func(f float64) {
+		t.Helper()
 
-	testFloatToDecimalRoundtrip(t, math.Inf(1))
-	testFloatToDecimalRoundtrip(t, math.Inf(-1))
-	testFloatToDecimalRoundtrip(t, 1<<63-1)
-	testFloatToDecimalRoundtrip(t, -1<<63)
+		v, e := FromFloat(f)
+		fNew := ToFloat(v, e)
+		if !equalFloat(fNew, f) {
+			t.Fatalf("unexpected fNew for v=%d, e=%d; got %g; expecting %g", v, e, fNew, f)
+		}
+
+		v, e = FromFloat(-f)
+		fNew = ToFloat(v, e)
+		if !equalFloat(fNew, -f) {
+			t.Fatalf("unexepcted fNew for v=%d, e=%d; got %g; expecting %g", v, e, fNew, -f)
+		}
+	}
+
+	f(0)
+	f(1)
+	f(0.123)
+	f(1.2345)
+	f(12000)
+	f(1e-30)
+	f(1e-260)
+	f(321e260)
+	f(1234567890123)
+	f(12.34567890125)
+	f(-1234567.8901256789)
+	f(15e18)
+
+	f(math.Inf(1))
+	f(math.Inf(-1))
+	f(1<<63 - 1)
+	f(-1 << 63)
 
 	for i := 0; i < 1e4; i++ {
-		f := rand.NormFloat64()
-		testFloatToDecimalRoundtrip(t, f)
-		testFloatToDecimalRoundtrip(t, f*1e-6)
-		testFloatToDecimalRoundtrip(t, f*1e6)
+		v := rand.NormFloat64()
+		f(v)
+		f(v * 1e-6)
+		f(v * 1e6)
 
-		testFloatToDecimalRoundtrip(t, roundFloat(f, 20))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, 10))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, 5))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, 0))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, -5))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, -10))
-		testFloatToDecimalRoundtrip(t, roundFloat(f, -20))
+		f(roundFloat(v, 20))
+		f(roundFloat(v, 10))
+		f(roundFloat(v, 5))
+		f(roundFloat(v, 0))
+		f(roundFloat(v, -5))
+		f(roundFloat(v, -10))
+		f(roundFloat(v, -20))
 	}
 }
 
@@ -292,22 +312,6 @@ func roundFloat(f float64, exp int) float64 {
 	return math.Trunc(f) * math.Pow10(exp)
 }
 
-func testFloatToDecimalRoundtrip(t *testing.T, f float64) {
-	t.Helper()
-
-	v, e := FromFloat(f)
-	fNew := ToFloat(v, e)
-	if !equalFloat(fNew, f) {
-		t.Fatalf("unexpected fNew for v=%d, e=%d; got %g; expecting %g", v, e, fNew, f)
-	}
-
-	v, e = FromFloat(-f)
-	fNew = ToFloat(v, e)
-	if !equalFloat(fNew, -f) {
-		t.Fatalf("unexepcted fNew for v=%d, e=%d; got %g; expecting %g", v, e, fNew, -f)
-	}
-}
-
 func equalFloat(f1, f2 float64) bool {
 	if math.IsInf(f1, 0) {
 		return math.IsInf(f1, 1) == math.IsInf(f2, 1) || math.IsInf(f1, -1) == math.IsInf(f2, -1)

lib/encoding/compress.go

@@ -1,8 +1,8 @@
 package encoding
 
 import (
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding/zstd"
 	"github.com/VictoriaMetrics/metrics"
-	"github.com/valyala/gozstd"
 )
 
 // CompressZSTDLevel appends compressed src to dst and returns
@@ -13,7 +13,7 @@ func CompressZSTDLevel(dst, src []byte, compressLevel int) []byte {
 	compressCalls.Inc()
 	originalBytes.Add(len(src))
 	dstLen := len(dst)
-	dst = gozstd.CompressLevel(dst, src, compressLevel)
+	dst = zstd.CompressLevel(dst, src, compressLevel)
 	compressedBytes.Add(len(dst) - dstLen)
 	return dst
 }
@@ -22,7 +22,7 @@ func CompressZSTDLevel(dst, src []byte, compressLevel int) []byte {
 // the appended dst.
 func DecompressZSTD(dst, src []byte) ([]byte, error) {
 	decompressCalls.Inc()
-	return gozstd.Decompress(dst, src)
+	return zstd.Decompress(dst, src)
 }
 
 var (

lib/encoding/encoding.go

@@ -296,7 +296,7 @@ func isDeltaConst(a []int64) bool {
 // i.e. arbitrary changing values.
 //
 // It is OK if a few gauges aren't detected (i.e. detected as counters),
-// since misdetected counters as gauges are much worse condition.
+// since misdetected counters as gauges leads to worser compression ratio.
 func isGauge(a []int64) bool {
 	// Check all the items in a, since a part of items may lead
 	// to incorrect gauge detection.
@@ -305,32 +305,36 @@ func isGauge(a []int64) bool {
 		return false
 	}
 
-	extremes := 0
-	plus := a[0] <= a[1]
-	v1 := a[1]
-	for _, v2 := range a[2:] {
-		if plus {
-			if v2 < v1 {
-				extremes++
-				plus = false
-			}
-		} else {
-			if v2 > v1 {
-				extremes++
-				plus = true
-			}
-		}
-		v1 = v2
+	resets := 0
+	vPrev := a[0]
+	if vPrev < 0 {
+		// Counter values cannot be negative.
+		return true
 	}
-	if extremes <= 2 {
-		// Probably counter reset.
+	for _, v := range a[1:] {
+		if v < vPrev {
+			if v < 0 {
+				// Counter values cannot be negative.
+				return true
+			}
+			if v > (vPrev >> 3) {
+				// Decreasing sequence detected.
+				// This is a gauge.
+				return true
+			}
+			// Possible counter reset.
+			resets++
+		}
+		vPrev = v
+	}
+	if resets <= 2 {
+		// Counter with a few resets.
 		return false
 	}
 
-	// A few extremes may indicate counter resets.
-	// Let it be a gauge if extremes exceed len(a)/32,
-	// otherwise assume counter reset.
-	return extremes > (len(a) >> 5)
+	// Let it be a gauge if resets exceeds len(a)/8,
+	// otherwise assume counter.
+	return resets > (len(a) >> 3)
 }
 
 func getCompressLevel(itemsCount int) int {

lib/encoding/encoding_cgo_test.go

@@ -0,0 +1,83 @@
+// +build cgo
+
+package encoding
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func TestMarshalUnmarshalInt64Array(t *testing.T) {
+	var va []int64
+	var v int64
+
+	// Verify nearest delta encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 8*1024; i++ {
+		v += int64(rand.NormFloat64() * 1e6)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 23; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta)
+	}
+	for precisionBits := uint8(23); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
+	}
+
+	// Verify nearest delta2 encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 8*1024; i++ {
+		v += 30e6 + int64(rand.NormFloat64()*1e6)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 24; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta2)
+	}
+	for precisionBits := uint8(24); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
+	}
+
+	// Verify nearest delta encoding.
+	va = va[:0]
+	v = 1000
+	for i := 0; i < 6; i++ {
+		v += int64(rand.NormFloat64() * 100)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
+	}
+
+	// Verify nearest delta2 encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 6; i++ {
+		v += 3000 + int64(rand.NormFloat64()*100)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(5); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
+	}
+}
+
+func TestMarshalInt64ArraySize(t *testing.T) {
+	var va []int64
+	v := int64(rand.Float64() * 1e9)
+	for i := 0; i < 8*1024; i++ {
+		va = append(va, v)
+		v += 30e3 + int64(rand.NormFloat64()*1e3)
+	}
+
+	testMarshalInt64ArraySize(t, va, 1, 500, 1300)
+	testMarshalInt64ArraySize(t, va, 2, 500, 1400)
+	testMarshalInt64ArraySize(t, va, 3, 800, 1800)
+	testMarshalInt64ArraySize(t, va, 4, 1300, 2100)
+	testMarshalInt64ArraySize(t, va, 5, 2000, 3200)
+	testMarshalInt64ArraySize(t, va, 6, 3000, 4800)
+	testMarshalInt64ArraySize(t, va, 7, 4000, 6400)
+	testMarshalInt64ArraySize(t, va, 8, 6000, 8000)
+	testMarshalInt64ArraySize(t, va, 9, 7000, 8800)
+	testMarshalInt64ArraySize(t, va, 10, 8000, 10000)
+}

lib/encoding/encoding_pure_test.go

@@ -0,0 +1,83 @@
+// +build !cgo
+
+package encoding
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func TestMarshalUnmarshalInt64Array(t *testing.T) {
+	var va []int64
+	var v int64
+
+	// Verify nearest delta encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 8*1024; i++ {
+		v += int64(rand.NormFloat64() * 1e6)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 17; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta)
+	}
+	for precisionBits := uint8(23); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
+	}
+
+	// Verify nearest delta2 encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 8*1024; i++ {
+		v += 30e6 + int64(rand.NormFloat64()*1e6)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 15; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta2)
+	}
+	for precisionBits := uint8(24); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
+	}
+
+	// Verify nearest delta encoding.
+	va = va[:0]
+	v = 1000
+	for i := 0; i < 6; i++ {
+		v += int64(rand.NormFloat64() * 100)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(1); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
+	}
+
+	// Verify nearest delta2 encoding.
+	va = va[:0]
+	v = 0
+	for i := 0; i < 6; i++ {
+		v += 3000 + int64(rand.NormFloat64()*100)
+		va = append(va, v)
+	}
+	for precisionBits := uint8(5); precisionBits < 65; precisionBits++ {
+		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
+	}
+}
+
+func TestMarshalInt64ArraySize(t *testing.T) {
+	var va []int64
+	v := int64(rand.Float64() * 1e9)
+	for i := 0; i < 8*1024; i++ {
+		va = append(va, v)
+		v += 30e3 + int64(rand.NormFloat64()*1e3)
+	}
+
+	testMarshalInt64ArraySize(t, va, 1, 500, 1700)
+	testMarshalInt64ArraySize(t, va, 2, 600, 1800)
+	testMarshalInt64ArraySize(t, va, 3, 900, 2100)
+	testMarshalInt64ArraySize(t, va, 4, 1300, 2200)
+	testMarshalInt64ArraySize(t, va, 5, 2000, 3300)
+	testMarshalInt64ArraySize(t, va, 6, 3000, 5000)
+	testMarshalInt64ArraySize(t, va, 7, 4000, 6500)
+	testMarshalInt64ArraySize(t, va, 8, 6000, 8000)
+	testMarshalInt64ArraySize(t, va, 9, 7000, 8800)
+	testMarshalInt64ArraySize(t, va, 10, 8000, 17000)
+}

lib/encoding/encoding_test.go

@@ -43,27 +43,29 @@ func TestIsDeltaConst(t *testing.T) {
 }
 
 func TestIsGauge(t *testing.T) {
-	testIsGauge(t, []int64{}, false)
-	testIsGauge(t, []int64{0}, false)
-	testIsGauge(t, []int64{1, 2}, false)
-	testIsGauge(t, []int64{0, 1, 2, 3, 4, 5}, false)
-	testIsGauge(t, []int64{0, -1, -2, -3, -4}, false)
-	testIsGauge(t, []int64{0, 0, 0, 0, 0, 0, 0}, false)
-	testIsGauge(t, []int64{1, 1, 1, 1, 1}, false)
-	testIsGauge(t, []int64{1, 1, 2, 2, 2, 2}, false)
-	testIsGauge(t, []int64{1, 5, 2, 3}, false) // a single counter reset
-	testIsGauge(t, []int64{1, 5, 2, 3, 2}, true)
-	testIsGauge(t, []int64{-1, -5, -2, -3}, false) // a single counter reset
-	testIsGauge(t, []int64{-1, -5, -2, -3, -2}, true)
-}
-
-func testIsGauge(t *testing.T, a []int64, okExpected bool) {
-	t.Helper()
-
-	ok := isGauge(a)
-	if ok != okExpected {
-		t.Fatalf("unexpected result for isGauge(%d); got %v; expecting %v", a, ok, okExpected)
+	f := func(a []int64, okExpected bool) {
+		t.Helper()
+		ok := isGauge(a)
+		if ok != okExpected {
+			t.Fatalf("unexpected result for isGauge(%d); got %v; expecting %v", a, ok, okExpected)
+		}
 	}
+	f([]int64{}, false)
+	f([]int64{0}, false)
+	f([]int64{1, 2}, false)
+	f([]int64{0, 1, 2, 3, 4, 5}, false)
+	f([]int64{0, -1, -2, -3, -4}, true)
+	f([]int64{0, 0, 0, 0, 0, 0, 0}, false)
+	f([]int64{1, 1, 1, 1, 1}, false)
+	f([]int64{1, 1, 2, 2, 2, 2}, false)
+	f([]int64{1, 17, 2, 3}, false) // a single counter reset
+	f([]int64{1, 5, 2, 3}, true)
+	f([]int64{1, 5, 2, 3, 2}, true)
+	f([]int64{-1, -5, -2, -3}, true)
+	f([]int64{-1, -5, -2, -3, -2}, true)
+	f([]int64{5, 6, 4, 3, 2}, true)
+	f([]int64{4, 5, 6, 5, 4, 3, 2}, true)
+	f([]int64{1064, 1132, 1083, 1062, 856, 747}, true)
 }
 
 func TestEnsureNonDecreasingSequence(t *testing.T) {
@@ -87,73 +89,6 @@ func testEnsureNonDecreasingSequence(t *testing.T, a []int64, vMin, vMax int64,
 	}
 }
 
-func TestMarshalUnmarshalInt64Array(t *testing.T) {
-	testMarshalUnmarshalInt64Array(t, []int64{1, 20, 234}, 4, MarshalTypeNearestDelta2)
-	testMarshalUnmarshalInt64Array(t, []int64{1, 20, -2345, 678934, 342}, 4, MarshalTypeNearestDelta)
-	testMarshalUnmarshalInt64Array(t, []int64{1, 20, 2345, 6789, 12342}, 4, MarshalTypeNearestDelta2)
-
-	// Constant encoding
-	testMarshalUnmarshalInt64Array(t, []int64{1}, 4, MarshalTypeConst)
-	testMarshalUnmarshalInt64Array(t, []int64{1, 2}, 4, MarshalTypeDeltaConst)
-	testMarshalUnmarshalInt64Array(t, []int64{-1, 0, 1, 2, 3, 4, 5}, 4, MarshalTypeDeltaConst)
-	testMarshalUnmarshalInt64Array(t, []int64{-10, -1, 8, 17, 26}, 4, MarshalTypeDeltaConst)
-	testMarshalUnmarshalInt64Array(t, []int64{0, 0, 0, 0, 0, 0}, 4, MarshalTypeConst)
-	testMarshalUnmarshalInt64Array(t, []int64{100, 100, 100, 100}, 4, MarshalTypeConst)
-
-	var va []int64
-	var v int64
-
-	// Verify nearest delta encoding.
-	va = va[:0]
-	v = 0
-	for i := 0; i < 8*1024; i++ {
-		v += int64(rand.NormFloat64() * 1e6)
-		va = append(va, v)
-	}
-	for precisionBits := uint8(1); precisionBits < 23; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta)
-	}
-	for precisionBits := uint8(23); precisionBits < 65; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
-	}
-
-	// Verify nearest delta2 encoding.
-	va = va[:0]
-	v = 0
-	for i := 0; i < 8*1024; i++ {
-		v += 30e6 + int64(rand.NormFloat64()*1e6)
-		va = append(va, v)
-	}
-	for precisionBits := uint8(1); precisionBits < 24; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeZSTDNearestDelta2)
-	}
-	for precisionBits := uint8(24); precisionBits < 65; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
-	}
-
-	// Verify nearest delta encoding.
-	va = va[:0]
-	v = 1000
-	for i := 0; i < 6; i++ {
-		v += int64(rand.NormFloat64() * 100)
-		va = append(va, v)
-	}
-	for precisionBits := uint8(1); precisionBits < 65; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta)
-	}
-
-	// Verify nearest delta2 encoding.
-	va = va[:0]
-	v = 0
-	for i := 0; i < 6; i++ {
-		v += 3000 + int64(rand.NormFloat64()*100)
-		va = append(va, v)
-	}
-	for precisionBits := uint8(5); precisionBits < 65; precisionBits++ {
-		testMarshalUnmarshalInt64Array(t, va, precisionBits, MarshalTypeNearestDelta2)
-	}
-}
-
 func testMarshalUnmarshalInt64Array(t *testing.T, va []int64, precisionBits uint8, mtExpected MarshalType) {
 	t.Helper()
 
@@ -257,24 +192,18 @@ func TestMarshalUnmarshalValues(t *testing.T) {
 	}
 }
 
-func TestMarshalInt64ArraySize(t *testing.T) {
-	var va []int64
-	v := int64(rand.Float64() * 1e9)
-	for i := 0; i < 8*1024; i++ {
-		va = append(va, v)
-		v += 30e3 + int64(rand.NormFloat64()*1e3)
-	}
+func TestMarshalUnmarshalInt64ArrayGeneric(t *testing.T) {
+	testMarshalUnmarshalInt64Array(t, []int64{1, 20, 234}, 4, MarshalTypeNearestDelta2)
+	testMarshalUnmarshalInt64Array(t, []int64{1, 20, -2345, 678934, 342}, 4, MarshalTypeNearestDelta)
+	testMarshalUnmarshalInt64Array(t, []int64{1, 20, 2345, 6789, 12342}, 4, MarshalTypeNearestDelta2)
 
-	testMarshalInt64ArraySize(t, va, 1, 500, 1300)
-	testMarshalInt64ArraySize(t, va, 2, 600, 1400)
-	testMarshalInt64ArraySize(t, va, 3, 900, 1800)
-	testMarshalInt64ArraySize(t, va, 4, 1300, 2100)
-	testMarshalInt64ArraySize(t, va, 5, 2000, 3200)
-	testMarshalInt64ArraySize(t, va, 6, 3000, 4800)
-	testMarshalInt64ArraySize(t, va, 7, 4000, 6400)
-	testMarshalInt64ArraySize(t, va, 8, 6000, 8000)
-	testMarshalInt64ArraySize(t, va, 9, 7000, 8800)
-	testMarshalInt64ArraySize(t, va, 10, 8000, 10000)
+	// Constant encoding
+	testMarshalUnmarshalInt64Array(t, []int64{1}, 4, MarshalTypeConst)
+	testMarshalUnmarshalInt64Array(t, []int64{1, 2}, 4, MarshalTypeDeltaConst)
+	testMarshalUnmarshalInt64Array(t, []int64{-1, 0, 1, 2, 3, 4, 5}, 4, MarshalTypeDeltaConst)
+	testMarshalUnmarshalInt64Array(t, []int64{-10, -1, 8, 17, 26}, 4, MarshalTypeDeltaConst)
+	testMarshalUnmarshalInt64Array(t, []int64{0, 0, 0, 0, 0, 0}, 4, MarshalTypeConst)
+	testMarshalUnmarshalInt64Array(t, []int64{100, 100, 100, 100}, 4, MarshalTypeConst)
 }
 
 func testMarshalInt64ArraySize(t *testing.T, va []int64, precisionBits uint8, minSizeExpected, maxSizeExpected int) {

lib/encoding/zstd/zstd_cgo.go

@@ -0,0 +1,19 @@
+// +build cgo
+
+package zstd
+
+import (
+	"github.com/valyala/gozstd"
+)
+
+// Decompress appends decompressed src to dst and returns the result.
+func Decompress(dst, src []byte) ([]byte, error) {
+	return gozstd.Decompress(dst, src)
+}
+
+// CompressLevel appends compressed src to dst and returns the result.
+//
+// The given compressionLevel is used for the compression.
+func CompressLevel(dst, src []byte, compressionLevel int) []byte {
+	return gozstd.CompressLevel(dst, src, compressionLevel)
+}

lib/encoding/zstd/zstd_pure.go

@@ -0,0 +1,78 @@
+// +build !cgo
+
+package zstd
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/klauspost/compress/zstd"
+)
+
+var (
+	decoder *zstd.Decoder
+
+	mu sync.Mutex
+	av atomic.Value
+)
+
+type registry map[int]*zstd.Encoder
+
+func init() {
+	r := make(registry)
+	av.Store(r)
+
+	var err error
+	decoder, err = zstd.NewReader(nil)
+	if err != nil {
+		logger.Panicf("BUG: failed to create ZSTD reader: %s", err)
+	}
+}
+
+// Decompress appends decompressed src to dst and returns the result.
+func Decompress(dst, src []byte) ([]byte, error) {
+	return decoder.DecodeAll(src, dst)
+}
+
+// CompressLevel appends compressed src to dst and returns the result.
+//
+// The given compressionLevel is used for the compression.
+func CompressLevel(dst, src []byte, compressionLevel int) []byte {
+	e := getEncoder(compressionLevel)
+	return e.EncodeAll(src, dst)
+}
+
+func getEncoder(compressionLevel int) *zstd.Encoder {
+	r := av.Load().(registry)
+	e := r[compressionLevel]
+	if e != nil {
+		return e
+	}
+
+	mu.Lock()
+	// Create the encoder under lock in order to prevent from wasted work
+	// when concurrent goroutines create encoder for the same compressionLevel.
+	e = newEncoder(compressionLevel)
+	r1 := av.Load().(registry)
+	r2 := make(registry)
+	for k, v := range r1 {
+		r2[k] = v
+	}
+	r2[compressionLevel] = e
+	av.Store(r2)
+	mu.Unlock()
+
+	return e
+}
+
+func newEncoder(compressionLevel int) *zstd.Encoder {
+	level := zstd.EncoderLevelFromZstd(compressionLevel)
+	e, err := zstd.NewWriter(nil,
+		zstd.WithEncoderCRC(false), // Disable CRC for performance reasons.
+		zstd.WithEncoderLevel(level))
+	if err != nil {
+		logger.Panicf("BUG: failed to create ZSTD writer: %s", err)
+	}
+	return e
+}

lib/encoding/zstd/zstd_test.go

@@ -0,0 +1,96 @@
+// +build cgo
+
+package zstd
+
+import (
+	"math/rand"
+	"testing"
+
+	pure "github.com/klauspost/compress/zstd"
+	cgo "github.com/valyala/gozstd"
+)
+
+func TestCompressDecompress(t *testing.T) {
+	testCrossCompressDecompress(t, []byte("a"))
+	testCrossCompressDecompress(t, []byte("foobarbaz"))
+
+	var b []byte
+	for i := 0; i < 64*1024; i++ {
+		b = append(b, byte(rand.Int31n(256)))
+	}
+	testCrossCompressDecompress(t, b)
+}
+
+func testCrossCompressDecompress(t *testing.T, b []byte) {
+	testCompressDecompress(t, pureCompress, pureDecompress, b)
+	testCompressDecompress(t, cgoCompress, cgoDecompress, b)
+	testCompressDecompress(t, pureCompress, cgoDecompress, b)
+	testCompressDecompress(t, cgoCompress, pureDecompress, b)
+}
+
+func testCompressDecompress(t *testing.T, compress compressFn, decompress decompressFn, b []byte) {
+	bc, err := compress(nil, b, 5)
+	if err != nil {
+		t.Fatalf("unexpected error when compressing b=%x: %s", b, err)
+	}
+	bNew, err := decompress(nil, bc)
+	if err != nil {
+		t.Fatalf("unexpected error when decompressing b=%x from bc=%x: %s", b, bc, err)
+	}
+	if string(bNew) != string(b) {
+		t.Fatalf("invalid bNew; got\n%x; expecting\n%x", bNew, b)
+	}
+
+	prefix := []byte{1, 2, 33}
+	bcNew, err := compress(prefix, b, 5)
+	if err != nil {
+		t.Fatalf("unexpected error when compressing b=%x: %s", bcNew, err)
+	}
+	if string(bcNew[:len(prefix)]) != string(prefix) {
+		t.Fatalf("invalid prefix for b=%x; got\n%x; expecting\n%x", b, bcNew[:len(prefix)], prefix)
+	}
+	if string(bcNew[len(prefix):]) != string(bc) {
+		t.Fatalf("invalid prefixed bcNew for b=%x; got\n%x; expecting\n%x", b, bcNew[len(prefix):], bc)
+	}
+
+	bNew, err = decompress(prefix, bc)
+	if err != nil {
+		t.Fatalf("unexpected error when decompressing b=%x from bc=%x with prefix: %s", b, bc, err)
+	}
+	if string(bNew[:len(prefix)]) != string(prefix) {
+		t.Fatalf("invalid bNew prefix when decompressing bc=%x; got\n%x; expecting\n%x", bc, bNew[:len(prefix)], prefix)
+	}
+	if string(bNew[len(prefix):]) != string(b) {
+		t.Fatalf("invalid prefixed bNew; got\n%x; expecting\n%x", bNew[len(prefix):], b)
+	}
+}
+
+type compressFn func(dst, src []byte, compressionLevel int) ([]byte, error)
+
+func pureCompress(dst, src []byte, _ int) ([]byte, error) {
+	w, err := pure.NewWriter(nil,
+		pure.WithEncoderCRC(false), // Disable CRC for performance reasons.
+		pure.WithEncoderLevel(pure.SpeedBestCompression))
+	if err != nil {
+		return nil, err
+	}
+	return w.EncodeAll(src, dst), nil
+}
+
+func cgoCompress(dst, src []byte, compressionLevel int) ([]byte, error) {
+	return cgo.CompressLevel(dst, src, compressionLevel), nil
+}
+
+type decompressFn func(dst, src []byte) ([]byte, error)
+
+func pureDecompress(dst, src []byte) ([]byte, error) {
+	decoder, err := pure.NewReader(nil)
+	if err != nil {
+		return nil, err
+	}
+	return decoder.DecodeAll(src, dst)
+}
+
+func cgoDecompress(dst, src []byte) ([]byte, error) {
+	return cgo.Decompress(dst, src)
+}

lib/filestream/filestream_freebsd.go

@@ -0,0 +1,64 @@
+package filestream
+
+import (
+	"fmt"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+)
+
+func (st *streamTracker) adviseDontNeed(n int, fdatasync bool) error {
+	st.length += uint64(n)
+	if st.fd == 0 {
+		return nil
+	}
+	if st.length < dontNeedBlockSize {
+		return nil
+	}
+	blockSize := st.length - (st.length % dontNeedBlockSize)
+	if fdatasync {
+		if err := unixFdatasync(int(st.fd)); err != nil {
+			return fmt.Errorf("unix.Fdatasync error: %s", err)
+		}
+	}
+	if err := unix.Fadvise(int(st.fd), int64(st.offset), int64(blockSize), unix.FADV_DONTNEED); err != nil {
+		return fmt.Errorf("unix.Fadvise(FADV_DONTNEEDED, %d, %d) error: %s", st.offset, blockSize, err)
+	}
+	st.offset += blockSize
+	st.length -= blockSize
+	return nil
+}
+
+func (st *streamTracker) close() error {
+	if st.fd == 0 {
+		return nil
+	}
+	// Advise the whole file as it shouldn't be cached.
+	if err := unix.Fadvise(int(st.fd), 0, 0, unix.FADV_DONTNEED); err != nil {
+		return fmt.Errorf("unix.Fadvise(FADV_DONTNEEDED, 0, 0) error: %s", err)
+	}
+	return nil
+}
+
+// unix.Fdatasync is missing, so put it here
+func unixFdatasync(fd int) (err error) {
+	_, _, e1 := unix.Syscall(unix.SYS_FDATASYNC, uintptr(fd), 0, 0)
+	if e1 != 0 {
+		err = errnoErr(e1)
+	}
+	return
+}
+
+func errnoErr(e syscall.Errno) error {
+	switch e {
+	case 0:
+		return nil
+	case unix.EAGAIN:
+		return syscall.EAGAIN
+	case unix.EINVAL:
+		return syscall.EINVAL
+	case unix.ENOENT:
+		return syscall.ENOENT
+	}
+	return e
+}

lib/fs/dir_remover.go

@@ -0,0 +1,111 @@
+package fs
+
+import (
+	"os"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+func mustRemoveAll(path string) bool {
+	err := os.RemoveAll(path)
+	if err == nil {
+		// Make sure the parent directory doesn't contain references
+		// to the current directory.
+		mustSyncParentDirIfExists(path)
+		return true
+	}
+	if !isTemporaryNFSError(err) {
+		logger.Panicf("FATAL: cannot remove %q: %s", path, err)
+	}
+	// NFS prevents from removing directories with open files.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
+	// Schedule for later directory removal.
+	nfsDirRemoveFailedAttempts.Inc()
+	select {
+	case removeDirCh <- path:
+	default:
+		logger.Panicf("FATAL: cannot schedule %s for removal, since the removal queue is full (%d entries)", path, cap(removeDirCh))
+	}
+	return false
+}
+
+var nfsDirRemoveFailedAttempts = metrics.NewCounter(`vm_nfs_dir_remove_failed_attempts_total`)
+
+var removeDirCh = make(chan string, 1024)
+
+func dirRemover() {
+	const minSleepTime = 100 * time.Millisecond
+	const maxSleepTime = time.Second
+	sleepTime := minSleepTime
+	for {
+		var path string
+		select {
+		case path = <-removeDirCh:
+		default:
+			if atomic.LoadUint64(&stopDirRemover) != 0 {
+				return
+			}
+			time.Sleep(minSleepTime)
+			continue
+		}
+		if mustRemoveAll(path) {
+			sleepTime = minSleepTime
+			continue
+		}
+
+		// Couldn't remove the directory at the path because of NFS lock.
+		// Sleep for a while and try again.
+		// Do not limit the amount of time required for deleting the directory,
+		// since this may break on laggy NFS.
+		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/162 .
+		time.Sleep(sleepTime)
+		if sleepTime < maxSleepTime {
+			sleepTime *= 2
+		} else {
+			logger.Errorf("failed to remove directory %q due to NFS lock; retrying later", path)
+		}
+	}
+}
+
+func isTemporaryNFSError(err error) bool {
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 for details.
+	errStr := err.Error()
+	return strings.Contains(errStr, "directory not empty") || strings.Contains(errStr, "device or resource busy")
+}
+
+var dirRemoverWG sync.WaitGroup
+var stopDirRemover uint64
+
+func init() {
+	dirRemoverWG.Add(1)
+	go func() {
+		defer dirRemoverWG.Done()
+		dirRemover()
+	}()
+}
+
+// MustStopDirRemover must be called in the end of graceful shutdown
+// in order to wait for removing the remaining directories from removeDirCh.
+//
+// It is expected that nobody calls MustRemoveAll when MustStopDirRemover
+// is called.
+func MustStopDirRemover() {
+	atomic.StoreUint64(&stopDirRemover, 1)
+	doneCh := make(chan struct{})
+	go func() {
+		dirRemoverWG.Wait()
+		close(doneCh)
+	}()
+	const maxWaitTime = 5 * time.Second
+	select {
+	case <-doneCh:
+		return
+	case <-time.After(maxWaitTime):
+		logger.Panicf("FATAL: cannot stop dirRemover in %s", maxWaitTime)
+	}
+}

lib/fs/fs.go

@@ -5,12 +5,13 @@ import (
 	"io"
 	"os"
 	"path/filepath"
-	"strings"
-	"time"
+	"regexp"
+	"sync/atomic"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/metrics"
+	"golang.org/x/sys/unix"
 )
 
 // ReadAtCloser is rand-access read interface.
@@ -87,26 +88,42 @@ func MustSyncPath(path string) {
 	}
 }
 
-// WriteFile writes data to the given file path.
+var tmpFileNum uint64
+
+// WriteFileAtomically atomically writes data to the given file path.
 //
-// WriteFile returns only after the file is fully written
+// WriteFile returns only after the file is fully written and synced
 // to the underlying storage.
-func WriteFile(path string, data []byte) error {
+func WriteFileAtomically(path string, data []byte) error {
+	// Check for the existing file. It is expected that
+	// the WriteFileAtomically function cannot be called concurrently
+	// with the same `path`.
 	if IsPathExist(path) {
 		return fmt.Errorf("cannot create file %q, since it already exists", path)
 	}
-	f, err := filestream.Create(path, false)
+
+	n := atomic.AddUint64(&tmpFileNum, 1)
+	tmpPath := fmt.Sprintf("%s.tmp.%d", path, n)
+	f, err := filestream.Create(tmpPath, false)
 	if err != nil {
-		return fmt.Errorf("cannot create file %q: %s", path, err)
+		return fmt.Errorf("cannot create file %q: %s", tmpPath, err)
 	}
 	if _, err := f.Write(data); err != nil {
 		f.MustClose()
-		return fmt.Errorf("cannot write %d bytes to file %q: %s", len(data), path, err)
+		MustRemoveAll(tmpPath)
+		return fmt.Errorf("cannot write %d bytes to file %q: %s", len(data), tmpPath, err)
 	}
 
 	// Sync and close the file.
 	f.MustClose()
 
+	// Atomically move the file from tmpPath to path.
+	if err := os.Rename(tmpPath, path); err != nil {
+		// do not call MustRemoveAll(tmpPath) here, so the user could inspect
+		// the file contents during investigating the issue.
+		return fmt.Errorf("cannot move %q to %q: %s", tmpPath, path, err)
+	}
+
 	// Sync the containing directory, so the file is guaranteed to appear in the directory.
 	// See https://www.quora.com/When-should-you-fsync-the-containing-directory-in-addition-to-the-file-itself
 	absPath, err := filepath.Abs(path)
@@ -119,6 +136,15 @@ func WriteFile(path string, data []byte) error {
 	return nil
 }
 
+// IsTemporaryFileName returns true if fn matches temporary file name pattern
+// from WriteFileAtomically.
+func IsTemporaryFileName(fn string) bool {
+	return tmpFileNameRe.MatchString(fn)
+}
+
+// tmpFileNameRe is regexp for temporary file name - see WriteFileAtomically for details.
+var tmpFileNameRe = regexp.MustCompile(`\.tmp\.\d+$`)
+
 // MkdirAllIfNotExist creates the given path dir if it isn't exist.
 func MkdirAllIfNotExist(path string) error {
 	if IsPathExist(path) {
@@ -220,62 +246,7 @@ func mustSyncParentDirIfExists(path string) {
 //
 // It properly handles NFS issue https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
 func MustRemoveAll(path string) {
-	err := os.RemoveAll(path)
-	if err == nil {
-		// Make sure the parent directory doesn't contain references
-		// to the current directory.
-		mustSyncParentDirIfExists(path)
-		return
-	}
-	if !isTemporaryNFSError(err) {
-		logger.Panicf("FATAL: cannot remove %q: %s", path, err)
-	}
-	// NFS prevents from removing directories with open files.
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
-	// Schedule for later directory removal.
-	select {
-	case removeDirCh <- path:
-	default:
-		logger.Panicf("FATAL: cannot schedule %s for removal, since the removal queue is full (%d entries)", path, cap(removeDirCh))
-	}
-}
-
-var removeDirCh = make(chan string, 1024)
-
-func dirRemover() {
-	for path := range removeDirCh {
-		attempts := 0
-		for {
-			err := os.RemoveAll(path)
-			if err == nil {
-				break
-			}
-			if !isTemporaryNFSError(err) {
-				logger.Panicf("FATAL: cannot remove %q: %s", path, err)
-			}
-			// NFS prevents from removing directories with open files.
-			// Sleep for a while and try again in the hope open files will be closed.
-			// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
-			attempts++
-			if attempts > 10 {
-				logger.Panicf("FATAL: cannot remove %q in %d attempts: %s", path, attempts, err)
-			}
-			time.Sleep(100 * time.Millisecond)
-		}
-		// Make sure the parent directory doesn't contain references
-		// to the current directory.
-		mustSyncParentDirIfExists(path)
-	}
-}
-
-func isTemporaryNFSError(err error) bool {
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 for details.
-	errStr := err.Error()
-	return strings.Contains(errStr, "directory not empty") || strings.Contains(errStr, "device or resource busy")
-}
-
-func init() {
-	go dirRemover()
+	_ = mustRemoveAll(path)
 }
 
 // HardLinkFiles makes hard links for all the files from srcDir in dstDir.
@@ -358,3 +329,34 @@ func MustWriteData(w io.Writer, data []byte) {
 		logger.Panicf("BUG: writer wrote %d bytes instead of %d bytes", n, len(data))
 	}
 }
+
+// CreateFlockFile creates flock.lock file in the directory dir
+// and returns the handler to the file.
+func CreateFlockFile(dir string) (*os.File, error) {
+	flockFile := dir + "/flock.lock"
+	flockF, err := os.Create(flockFile)
+	if err != nil {
+		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
+	}
+	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
+		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+	}
+	return flockF, nil
+}
+
+// MustGetFreeSpace returns free space for the given directory path.
+func MustGetFreeSpace(path string) uint64 {
+	d, err := os.Open(path)
+	if err != nil {
+		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", path, err)
+	}
+	defer MustClose(d)
+
+	fd := d.Fd()
+	var stat unix.Statfs_t
+	if err := unix.Fstatfs(int(fd), &stat); err != nil {
+		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", path, err)
+	}
+	freeSpace := uint64(stat.Bavail) * uint64(stat.Bsize)
+	return freeSpace
+}

lib/fs/fs_test.go

@@ -0,0 +1,24 @@
+package fs
+
+import (
+	"testing"
+)
+
+func TestIsTemporaryFileName(t *testing.T) {
+	f := func(s string, resultExpected bool) {
+		t.Helper()
+		result := IsTemporaryFileName(s)
+		if result != resultExpected {
+			t.Fatalf("unexpected IsTemporaryFileName(%q); got %v; want %v", s, result, resultExpected)
+		}
+	}
+	f("", false)
+	f(".", false)
+	f(".tmp", false)
+	f("tmp.123", false)
+	f(".tmp.123.xx", false)
+	f(".tmp.1", true)
+	f("asdf.dff.tmp.123", true)
+	f("asdf.sdfds.tmp.dfd", false)
+	f("dfd.sdfds.dfds.1232", false)
+}

lib/httpserver/httpserver.go

@@ -423,7 +423,29 @@ var (
 func Errorf(w http.ResponseWriter, format string, args ...interface{}) {
 	errStr := fmt.Sprintf(format, args...)
 	logger.Errorf("%s", errStr)
-	http.Error(w, errStr, http.StatusBadRequest)
+
+	// Extract statusCode from args
+	statusCode := http.StatusBadRequest
+	for _, arg := range args {
+		if esc, ok := arg.(*ErrorWithStatusCode); ok {
+			statusCode = esc.StatusCode
+			break
+		}
+	}
+	http.Error(w, errStr, statusCode)
+}
+
+// ErrorWithStatusCode is error with HTTP status code.
+//
+// The given StatusCode is sent to client when the error is passed to Errorf.
+type ErrorWithStatusCode struct {
+	Err        error
+	StatusCode int
+}
+
+// Error implements error interface.
+func (e *ErrorWithStatusCode) Error() string {
+	return e.Err.Error()
 }
 
 func isTrivialNetworkError(err error) bool {

lib/memory/memory.go

@@ -10,27 +10,41 @@ import (
 
 var allowedMemPercent = flag.Float64("memory.allowedPercent", 60, "Allowed percent of system memory VictoriaMetrics caches may occupy")
 
-var allowedMemory int
+var (
+	allowedMemory   int
+	remainingMemory int
+)
 
 var once sync.Once
 
+func initOnce() {
+	if !flag.Parsed() {
+		// Do not use logger.Panicf here, since logger may be uninitialized yet.
+		panic(fmt.Errorf("BUG: memory.Allowed must be called only after flag.Parse call"))
+	}
+	if *allowedMemPercent < 10 || *allowedMemPercent > 200 {
+		logger.Panicf("FATAL: -memory.allowedPercent must be in the range [10...200]; got %f", *allowedMemPercent)
+	}
+	percent := *allowedMemPercent / 100
+
+	mem := sysTotalMemory()
+	allowedMemory = int(float64(mem) * percent)
+	remainingMemory = mem - allowedMemory
+	logger.Infof("limiting caches to %d bytes, leaving %d bytes to the OS according to -memory.allowedPercent=%g", allowedMemory, remainingMemory, *allowedMemPercent)
+}
+
 // Allowed returns the amount of system memory allowed to use by the app.
 //
 // The function must be called only after flag.Parse is called.
 func Allowed() int {
-	once.Do(func() {
-		if !flag.Parsed() {
-			// Do not use logger.Panicf here, since logger may be uninitialized yet.
-			panic(fmt.Errorf("BUG: memory.Allowed must be called only after flag.Parse call"))
-		}
-		if *allowedMemPercent < 10 || *allowedMemPercent > 200 {
-			logger.Panicf("FATAL: -memory.allowedPercent must be in the range [10...200]; got %f", *allowedMemPercent)
-		}
-		percent := *allowedMemPercent / 100
-
-		mem := sysTotalMemory()
-		allowedMemory = int(float64(mem) * percent)
-		logger.Infof("limiting caches to %d bytes of RAM according to -memory.allowedPercent=%g", allowedMemory, *allowedMemPercent)
-	})
+	once.Do(initOnce)
 	return allowedMemory
 }
+
+// Remaining returns the amount of memory remaining to the OS.
+//
+// This function must be called only after flag.Parse is called.
+func Remaining() int {
+	once.Do(initOnce)
+	return remainingMemory
+}

lib/memory/memory_bsd.go

@@ -0,0 +1,17 @@
+// +build freebsd openbsd dragonfly netbsd
+
+package memory
+
+import (
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+// This code has been adopted from https://github.com/pbnjay/memory
+
+func sysTotalMemory() int {
+	s, err := sysctlUint64("hw.physmem")
+	if err != nil {
+		logger.Panicf("FATAL: cannot determine system memory: %s", err)
+	}
+	return int(s)
+}

lib/memory/memory_darwin.go

@@ -1,9 +1,6 @@
 package memory
 
 import (
-	"syscall"
-	"unsafe"
-
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
@@ -15,16 +12,3 @@ func sysTotalMemory() int {
 	}
 	return int(s)
 }
-
-func sysctlUint64(name string) (uint64, error) {
-	s, err := syscall.Sysctl(name)
-	if err != nil {
-		return 0, err
-	}
-	// hack because the string conversion above drops a \0
-	b := []byte(s)
-	if len(b) < 8 {
-		b = append(b, 0)
-	}
-	return *(*uint64)(unsafe.Pointer(&b[0])), nil
-}

lib/memory/sysctl.go

@@ -0,0 +1,22 @@
+// +build darwin freebsd openbsd dragonfly netbsd
+
+package memory
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+// This has been adapted from github.com/pbnjay/memory.
+func sysctlUint64(name string) (uint64, error) {
+	s, err := syscall.Sysctl(name)
+	if err != nil {
+		return 0, err
+	}
+	// hack because the string conversion above drops a \0
+	b := []byte(s)
+	if len(b) < 8 {
+		b = append(b, 0)
+	}
+	return *(*uint64)(unsafe.Pointer(&b[0])), nil
+}

lib/mergeset/part_header.go

@@ -164,7 +164,7 @@ func (ph *partHeader) WriteMetadata(partPath string) error {
 		return fmt.Errorf("cannot marshal metadata: %s", err)
 	}
 	metadataPath := partPath + "/metadata.json"
-	if err := fs.WriteFile(metadataPath, metadata); err != nil {
+	if err := fs.WriteFileAtomically(metadataPath, metadata); err != nil {
 		return fmt.Errorf("cannot create %q: %s", metadataPath, err)
 	}
 	return nil

lib/mergeset/table.go

@@ -16,7 +16,6 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/syncwg"
-	"golang.org/x/sys/unix"
 )
 
 // maxParts is the maximum number of parts in the table.
@@ -60,6 +59,8 @@ const rawItemsFlushInterval = time.Second
 type Table struct {
 	path string
 
+	flushCallback func()
+
 	partsLock sync.Mutex
 	parts     []*partWrapper
 
@@ -122,8 +123,11 @@ func (pw *partWrapper) decRef() {
 
 // OpenTable opens a table on the given path.
 //
+// Optional flushCallback is called every time new data batch is flushed
+// to the underlying storage and becomes visible to search.
+//
 // The table is created if it doesn't exist yet.
-func OpenTable(path string) (*Table, error) {
+func OpenTable(path string, flushCallback func()) (*Table, error) {
 	path = filepath.Clean(path)
 	logger.Infof("opening table %q...", path)
 	startTime := time.Now()
@@ -134,13 +138,9 @@ func OpenTable(path string) (*Table, error) {
 	}
 
 	// Protect from concurrent opens.
-	flockFile := path + "/flock.lock"
-	flockF, err := os.Create(flockFile)
+	flockF, err := fs.CreateFlockFile(path)
 	if err != nil {
-		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
-	}
-	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
-		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+		return nil, err
 	}
 
 	// Open table parts.
@@ -150,11 +150,12 @@ func OpenTable(path string) (*Table, error) {
 	}
 
 	tb := &Table{
-		path:     path,
-		parts:    pws,
-		mergeIdx: uint64(time.Now().UnixNano()),
-		flockF:   flockF,
-		stopCh:   make(chan struct{}),
+		path:          path,
+		flushCallback: flushCallback,
+		parts:         pws,
+		mergeIdx:      uint64(time.Now().UnixNano()),
+		flockF:        flockF,
+		stopCh:        make(chan struct{}),
 	}
 	tb.startPartMergers()
 	tb.startRawItemsFlusher()
@@ -449,6 +450,9 @@ func (tb *Table) mergeRawItemsBlocks(blocksToMerge []*inmemoryBlock) {
 		if err := tb.mergeParts(pws, nil, true); err != nil {
 			logger.Panicf("FATAL: cannot merge raw parts: %s", err)
 		}
+		if tb.flushCallback != nil {
+			tb.flushCallback()
+		}
 	}
 
 	for {
@@ -715,7 +719,7 @@ func (tb *Table) mergeParts(pws []*partWrapper, stopCh <-chan struct{}, isOuterP
 	dstPartPath := ph.Path(tb.path, mergeIdx)
 	fmt.Fprintf(&bb, "%s -> %s\n", tmpPartPath, dstPartPath)
 	txnPath := fmt.Sprintf("%s/txn/%016X", tb.path, mergeIdx)
-	if err := fs.WriteFile(txnPath, bb.B); err != nil {
+	if err := fs.WriteFileAtomically(txnPath, bb.B); err != nil {
 		return fmt.Errorf("cannot create transaction file %q: %s", txnPath, err)
 	}
 
@@ -808,19 +812,7 @@ func (tb *Table) maxOutPartItems() uint64 {
 }
 
 func (tb *Table) maxOutPartItemsSlow() uint64 {
-	// Determine the amount of free space on tb.path.
-	d, err := os.Open(tb.path)
-	if err != nil {
-		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", tb.path, err)
-	}
-	defer fs.MustClose(d)
-
-	fd := d.Fd()
-	var stat unix.Statfs_t
-	if err := unix.Fstatfs(int(fd), &stat); err != nil {
-		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", tb.path, err)
-	}
-	freeSpace := stat.Bavail * uint64(stat.Bsize)
+	freeSpace := fs.MustGetFreeSpace(tb.path)
 
 	// Calculate the maximum number of items in the output merge part
 	// by dividing the freeSpace by 4 and by the number of concurrent
@@ -994,7 +986,12 @@ func runTransactions(txnLock *sync.RWMutex, path string) error {
 	})
 
 	for _, fi := range fis {
-		txnPath := txnDir + "/" + fi.Name()
+		fn := fi.Name()
+		if fs.IsTemporaryFileName(fn) {
+			// Skip temporary files, which could be left after unclean shutdown.
+			continue
+		}
+		txnPath := txnDir + "/" + fn
 		if err := runTransaction(txnLock, path, txnPath); err != nil {
 			return fmt.Errorf("cannot run transaction from %q: %s", txnPath, err)
 		}

lib/mergeset/table_search_test.go

@@ -5,6 +5,7 @@ import (
 	"math/rand"
 	"os"
 	"sort"
+	"sync/atomic"
 	"testing"
 	"time"
 )
@@ -39,7 +40,7 @@ func TestTableSearchSerial(t *testing.T) {
 
 	func() {
 		// Re-open the table and verify the search works.
-		tb, err := OpenTable(path)
+		tb, err := OpenTable(path, nil)
 		if err != nil {
 			t.Fatalf("cannot open table: %s", err)
 		}
@@ -74,7 +75,7 @@ func TestTableSearchConcurrent(t *testing.T) {
 
 	// Re-open the table and verify the search works.
 	func() {
-		tb, err := OpenTable(path)
+		tb, err := OpenTable(path, nil)
 		if err != nil {
 			t.Fatalf("cannot open table: %s", err)
 		}
@@ -146,7 +147,11 @@ func testTableSearchSerial(tb *Table, items []string) error {
 }
 
 func newTestTable(path string, itemsCount int) (*Table, []string, error) {
-	tb, err := OpenTable(path)
+	var flushes uint64
+	flushCallback := func() {
+		atomic.AddUint64(&flushes, 1)
+	}
+	tb, err := OpenTable(path, flushCallback)
 	if err != nil {
 		return nil, nil, fmt.Errorf("cannot open table: %s", err)
 	}
@@ -159,6 +164,9 @@ func newTestTable(path string, itemsCount int) (*Table, []string, error) {
 		items[i] = item
 	}
 	tb.DebugFlush()
+	if itemsCount > 0 && atomic.LoadUint64(&flushes) == 0 {
+		return nil, nil, fmt.Errorf("unexpeted zero flushes for itemsCount=%d", itemsCount)
+	}
 
 	sort.Strings(items)
 	return tb, items, nil

lib/mergeset/table_search_timing_test.go

@@ -32,7 +32,7 @@ func benchmarkTableSearch(b *testing.B, itemsCount int) {
 
 	// Force finishing pending merges
 	tb.MustClose()
-	tb, err = OpenTable(path)
+	tb, err = OpenTable(path, nil)
 	if err != nil {
 		b.Fatalf("unexpected error when re-opening table %q: %s", path, err)
 	}

lib/mergeset/table_test.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"sync"
+	"sync/atomic"
 	"testing"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
@@ -20,7 +21,7 @@ func TestTableOpenClose(t *testing.T) {
 	}()
 
 	// Create a new table
-	tb, err := OpenTable(path)
+	tb, err := OpenTable(path, nil)
 	if err != nil {
 		t.Fatalf("cannot create new table: %s", err)
 	}
@@ -30,7 +31,7 @@ func TestTableOpenClose(t *testing.T) {
 
 	// Re-open created table multiple times.
 	for i := 0; i < 10; i++ {
-		tb, err := OpenTable(path)
+		tb, err := OpenTable(path, nil)
 		if err != nil {
 			t.Fatalf("cannot open created table: %s", err)
 		}
@@ -44,14 +45,14 @@ func TestTableOpenMultipleTimes(t *testing.T) {
 		_ = os.RemoveAll(path)
 	}()
 
-	tb1, err := OpenTable(path)
+	tb1, err := OpenTable(path, nil)
 	if err != nil {
 		t.Fatalf("cannot open table: %s", err)
 	}
 	defer tb1.MustClose()
 
 	for i := 0; i < 10; i++ {
-		tb2, err := OpenTable(path)
+		tb2, err := OpenTable(path, nil)
 		if err == nil {
 			tb2.MustClose()
 			t.Fatalf("expecting non-nil error when opening already opened table")
@@ -68,7 +69,11 @@ func TestTableAddItemSerial(t *testing.T) {
 		_ = os.RemoveAll(path)
 	}()
 
-	tb, err := OpenTable(path)
+	var flushes uint64
+	flushCallback := func() {
+		atomic.AddUint64(&flushes, 1)
+	}
+	tb, err := OpenTable(path, flushCallback)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -78,6 +83,9 @@ func TestTableAddItemSerial(t *testing.T) {
 
 	// Verify items count after pending items flush.
 	tb.DebugFlush()
+	if atomic.LoadUint64(&flushes) == 0 {
+		t.Fatalf("unexpected zero flushes")
+	}
 
 	var m TableMetrics
 	tb.UpdateMetrics(&m)
@@ -91,7 +99,7 @@ func TestTableAddItemSerial(t *testing.T) {
 	testReopenTable(t, path, itemsCount)
 
 	// Add more items in order to verify merge between inmemory parts and file-based parts.
-	tb, err = OpenTable(path)
+	tb, err = OpenTable(path, nil)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -124,7 +132,7 @@ func TestTableCreateSnapshotAt(t *testing.T) {
 		_ = os.RemoveAll(path)
 	}()
 
-	tb, err := OpenTable(path)
+	tb, err := OpenTable(path, nil)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -155,13 +163,13 @@ func TestTableCreateSnapshotAt(t *testing.T) {
 	}()
 
 	// Verify snapshots contain all the data.
-	tb1, err := OpenTable(snapshot1)
+	tb1, err := OpenTable(snapshot1, nil)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
 	defer tb1.MustClose()
 
-	tb2, err := OpenTable(snapshot2)
+	tb2, err := OpenTable(snapshot2, nil)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -205,7 +213,11 @@ func TestTableAddItemsConcurrent(t *testing.T) {
 		_ = os.RemoveAll(path)
 	}()
 
-	tb, err := OpenTable(path)
+	var flushes uint64
+	flushCallback := func() {
+		atomic.AddUint64(&flushes, 1)
+	}
+	tb, err := OpenTable(path, flushCallback)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -215,6 +227,10 @@ func TestTableAddItemsConcurrent(t *testing.T) {
 
 	// Verify items count after pending items flush.
 	tb.DebugFlush()
+	if atomic.LoadUint64(&flushes) == 0 {
+		t.Fatalf("unexpected zero flushes")
+	}
+
 	var m TableMetrics
 	tb.UpdateMetrics(&m)
 	if m.ItemsCount != itemsCount {
@@ -227,7 +243,7 @@ func TestTableAddItemsConcurrent(t *testing.T) {
 	testReopenTable(t, path, itemsCount)
 
 	// Add more items in order to verify merge between inmemory parts and file-based parts.
-	tb, err = OpenTable(path)
+	tb, err = OpenTable(path, nil)
 	if err != nil {
 		t.Fatalf("cannot open %q: %s", path, err)
 	}
@@ -269,7 +285,7 @@ func testReopenTable(t *testing.T, path string, itemsCount int) {
 	t.Helper()
 
 	for i := 0; i < 10; i++ {
-		tb, err := OpenTable(path)
+		tb, err := OpenTable(path, nil)
 		if err != nil {
 			t.Fatalf("cannot re-open %q: %s", path, err)
 		}

lib/netutil/conn.go

@@ -72,9 +72,10 @@ func (sc *statConn) Read(p []byte) (int, error) {
 	sc.cm.readCalls.Inc()
 	sc.cm.readBytes.Add(n)
 	if err != nil && err != io.EOF {
-		sc.cm.readErrors.Inc()
 		if ne, ok := err.(net.Error); ok && ne.Timeout() {
 			sc.cm.readTimeouts.Inc()
+		} else {
+			sc.cm.readErrors.Inc()
 		}
 	}
 	return n, err
@@ -96,9 +97,10 @@ func (sc *statConn) Write(p []byte) (int, error) {
 	sc.cm.writeCalls.Inc()
 	sc.cm.writtenBytes.Add(n)
 	if err != nil {
-		sc.cm.writeErrors.Inc()
 		if ne, ok := err.(net.Error); ok && ne.Timeout() {
 			sc.cm.writeTimeouts.Inc()
+		} else {
+			sc.cm.writeErrors.Inc()
 		}
 	}
 	return n, err

lib/prompb/remote.pb.go

@@ -31,7 +31,7 @@ func (m *WriteRequest) Unmarshal(dAtA []byte) error {
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -118,7 +118,7 @@ func skipRemote(dAtA []byte) (n int, err error) {
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -176,7 +176,7 @@ func skipRemote(dAtA []byte) (n int, err error) {
 					}
 					b := dAtA[iNdEx]
 					iNdEx++
-					innerWire |= (uint64(b) & 0x7F) << shift
+					innerWire |= uint64(b&0x7F) << shift
 					if b < 0x80 {
 						break
 					}

lib/prompb/types.pb.go

@@ -43,7 +43,7 @@ func (m *Sample) Unmarshal(dAtA []byte) error {
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -82,7 +82,7 @@ func (m *Sample) Unmarshal(dAtA []byte) error {
 				}
 				b := dAtA[iNdEx]
 				iNdEx++
-				m.Timestamp |= (int64(b) & 0x7F) << shift
+				m.Timestamp |= int64(b&0x7F) << shift
 				if b < 0x80 {
 					break
 				}
@@ -128,7 +128,7 @@ func (m *TimeSeries) Unmarshal(dAtA []byte, dstLabels []Label, dstSamples []Samp
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -255,7 +255,7 @@ func (m *Label) Unmarshal(dAtA []byte) error {
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -283,7 +283,7 @@ func (m *Label) Unmarshal(dAtA []byte) error {
 				}
 				b := dAtA[iNdEx]
 				iNdEx++
-				stringLen |= (uint64(b) & 0x7F) << shift
+				stringLen |= uint64(b&0x7F) << shift
 				if b < 0x80 {
 					break
 				}
@@ -312,7 +312,7 @@ func (m *Label) Unmarshal(dAtA []byte) error {
 				}
 				b := dAtA[iNdEx]
 				iNdEx++
-				stringLen |= (uint64(b) & 0x7F) << shift
+				stringLen |= uint64(b&0x7F) << shift
 				if b < 0x80 {
 					break
 				}
@@ -363,7 +363,7 @@ func skipTypes(dAtA []byte) (n int, err error) {
 			}
 			b := dAtA[iNdEx]
 			iNdEx++
-			wire |= (uint64(b) & 0x7F) << shift
+			wire |= uint64(b&0x7F) << shift
 			if b < 0x80 {
 				break
 			}
@@ -421,7 +421,7 @@ func skipTypes(dAtA []byte) (n int, err error) {
 					}
 					b := dAtA[iNdEx]
 					iNdEx++
-					innerWire |= (uint64(b) & 0x7F) << shift
+					innerWire |= uint64(b&0x7F) << shift
 					if b < 0x80 {
 						break
 					}

lib/storage/block.go

@@ -205,19 +205,6 @@ func (b *Block) MarshalData(timestampsBlockOffset, valuesBlockOffset uint64) ([]
 	b.bh.ValuesBlockSize = uint32(len(b.valuesData))
 	b.values = b.values[:0]
 
-	if len(timestamps) > 1 && (b.bh.ValuesMarshalType == encoding.MarshalTypeConst || b.bh.ValuesMarshalType == encoding.MarshalTypeDeltaConst) {
-		// Special case - values are constant or are changed with constant rate.
-		// In this case we may 'cheat' by assuming timestamps are changed
-		// at ideal constant rate. This improves timestamps' compression rate.
-		minTimestamp := timestamps[0]
-		maxTimestamp := timestamps[len(timestamps)-1]
-		delta := (maxTimestamp - minTimestamp) / int64(len(timestamps)-1)
-		ts := minTimestamp
-		for i := 1; i < len(timestamps); i++ {
-			ts += delta
-			timestamps[i] = ts
-		}
-	}
 	b.timestampsData, b.bh.TimestampsMarshalType, b.bh.MinTimestamp = encoding.MarshalTimestamps(b.timestampsData[:0], timestamps, b.bh.PrecisionBits)
 	b.bh.TimestampsBlockOffset = timestampsBlockOffset
 	b.bh.TimestampsBlockSize = uint32(len(b.timestampsData))

lib/storage/block_header.go

@@ -181,6 +181,10 @@ func unmarshalBlockHeaders(dst []blockHeader, src []byte, blockHeadersCount int)
 		logger.Panicf("BUG: blockHeadersCount must be greater than zero; got %d", blockHeadersCount)
 	}
 	dstLen := len(dst)
+	if n := dstLen + blockHeadersCount - cap(dst); n > 0 {
+		dst = append(dst[:cap(dst)], make([]blockHeader, n)...)
+		dst = dst[:dstLen]
+	}
 	var bh blockHeader
 	for len(src) > 0 {
 		tmp, err := bh.Unmarshal(src)

lib/storage/index_db.go

@@ -18,6 +18,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/mergeset"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	xxhash "github.com/cespare/xxhash/v2"
 )
@@ -52,17 +53,17 @@ type indexDB struct {
 	extDBLock sync.Mutex
 
 	// Cache for fast TagFilters -> TSIDs lookup.
-	tagCache *fastcache.Cache
+	tagCache *workingsetcache.Cache
 
 	// Cache for fast MetricID -> TSID lookup.
-	metricIDCache *fastcache.Cache
+	metricIDCache *workingsetcache.Cache
 
 	// Cache for fast MetricID -> MetricName lookup.
-	metricNameCache *fastcache.Cache
+	metricNameCache *workingsetcache.Cache
 
 	// Cache holding useless TagFilters entries, which have no tag filters
 	// matching low number of metrics.
-	uselessTagFiltersCache *fastcache.Cache
+	uselessTagFiltersCache *workingsetcache.Cache
 
 	indexSearchPool sync.Pool
 
@@ -101,7 +102,7 @@ type indexDB struct {
 }
 
 // openIndexDB opens index db from the given path with the given caches.
-func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (*indexDB, error) {
+func openIndexDB(path string, metricIDCache, metricNameCache *workingsetcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (*indexDB, error) {
 	if metricIDCache == nil {
 		logger.Panicf("BUG: metricIDCache must be non-nil")
 	}
@@ -115,7 +116,7 @@ func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, c
 		logger.Panicf("BUG: prevHourMetricIDs must be non-nil")
 	}
 
-	tb, err := mergeset.OpenTable(path)
+	tb, err := mergeset.OpenTable(path, invalidateTagCache)
 	if err != nil {
 		return nil, fmt.Errorf("cannot open indexDB %q: %s", path, err)
 	}
@@ -130,10 +131,10 @@ func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, c
 		tb:       tb,
 		name:     name,
 
-		tagCache:               fastcache.New(mem / 32),
+		tagCache:               workingsetcache.New(mem/32, time.Hour),
 		metricIDCache:          metricIDCache,
 		metricNameCache:        metricNameCache,
-		uselessTagFiltersCache: fastcache.New(mem / 128),
+		uselessTagFiltersCache: workingsetcache.New(mem/128, time.Hour),
 
 		currHourMetricIDs: currHourMetricIDs,
 		prevHourMetricIDs: prevHourMetricIDs,
@@ -273,8 +274,8 @@ func (db *indexDB) decRef() {
 	db.SetExtDB(nil)
 
 	// Free space occupied by caches owned by db.
-	db.tagCache.Reset()
-	db.uselessTagFiltersCache.Reset()
+	db.tagCache.Stop()
+	db.uselessTagFiltersCache.Stop()
 
 	db.tagCache = nil
 	db.metricIDCache = nil
@@ -291,20 +292,36 @@ func (db *indexDB) decRef() {
 }
 
 func (db *indexDB) getFromTagCache(key []byte) ([]TSID, bool) {
-	value := db.tagCache.GetBig(nil, key)
-	if len(value) == 0 {
+	compressedBuf := tagBufPool.Get()
+	defer tagBufPool.Put(compressedBuf)
+	compressedBuf.B = db.tagCache.GetBig(compressedBuf.B[:0], key)
+	if len(compressedBuf.B) == 0 {
 		return nil, false
 	}
-	tsids, err := unmarshalTSIDs(nil, value)
+	buf := tagBufPool.Get()
+	defer tagBufPool.Put(buf)
+	var err error
+	buf.B, err = encoding.DecompressZSTD(buf.B[:0], compressedBuf.B)
+	if err != nil {
+		logger.Panicf("FATAL: cannot decompress tsids from tagCache: %s", err)
+	}
+	tsids, err := unmarshalTSIDs(nil, buf.B)
 	if err != nil {
 		logger.Panicf("FATAL: cannot unmarshal tsids from tagCache: %s", err)
 	}
 	return tsids, true
 }
 
+var tagBufPool bytesutil.ByteBufferPool
+
 func (db *indexDB) putToTagCache(tsids []TSID, key []byte) {
-	value := marshalTSIDs(nil, tsids)
-	db.tagCache.SetBig(key, value)
+	buf := tagBufPool.Get()
+	buf.B = marshalTSIDs(buf.B[:0], tsids)
+	compressedBuf := tagBufPool.Get()
+	compressedBuf.B = encoding.CompressZSTDLevel(compressedBuf.B[:0], buf.B, 1)
+	tagBufPool.Put(buf)
+	db.tagCache.SetBig(key, compressedBuf.B)
+	tagBufPool.Put(compressedBuf)
 }
 
 func (db *indexDB) getFromMetricIDCache(dst *TSID, metricID uint64) error {
@@ -388,7 +405,7 @@ func unmarshalTSIDs(dst []TSID, src []byte) ([]TSID, error) {
 	return dst, nil
 }
 
-func (db *indexDB) invalidateTagCache() {
+func invalidateTagCache() {
 	// This function must be fast, since it is called each
 	// time new timeseries is added.
 	atomic.AddUint64(&tagFiltersKeyGen, 1)
@@ -496,8 +513,8 @@ func (db *indexDB) createTSIDByName(dst *TSID, metricName []byte) error {
 		return fmt.Errorf("cannot create indexes: %s", err)
 	}
 
-	// Invalidate tag cache, since it doesn't contain tags for the created mn -> TSID mapping.
-	db.invalidateTagCache()
+	// There is no need in invalidating tag cache, since it is invalidated
+	// on db.tb flush via invalidateTagCache flushCallback passed to OpenTable.
 
 	return nil
 }
@@ -873,7 +890,10 @@ func (db *indexDB) DeleteTSIDs(tfss []*TagFilters) (int, error) {
 	db.updateDeletedMetricIDs(metricIDs)
 
 	// Reset TagFilters -> TSIDS cache, since it may contain deleted TSIDs.
-	db.invalidateTagCache()
+	invalidateTagCache()
+
+	// Do not reset uselessTagFiltersCache, since the found metricIDs
+	// on cache miss are filtered out later with deletedMetricIDs.
 
 	// Delete TSIDs in the extDB.
 	if db.doExtDB(func(extDB *indexDB) {
@@ -974,7 +994,8 @@ func (db *indexDB) searchTSIDs(tfss []*TagFilters, tr TimeRange, maxMetrics int)
 		extTSIDs, err = is.searchTSIDs(tfss, tr, maxMetrics)
 		extDB.putIndexSearch(is)
 
-		db.putToTagCache(tsids, tfKeyExtBuf.B)
+		sort.Slice(extTSIDs, func(i, j int) bool { return extTSIDs[i].Less(&extTSIDs[j]) })
+		extDB.putToTagCache(extTSIDs, tfKeyExtBuf.B)
 	}) {
 		if err != nil {
 			return nil, err
@@ -1218,6 +1239,82 @@ func (is *indexSearch) updateMetricIDsByMetricNameMatch(metricIDs, srcMetricIDs
 	return nil
 }
 
+func (is *indexSearch) getTagFilterWithMinMetricIDsCountOptimized(tfs *TagFilters, tr TimeRange, maxMetrics int) (*tagFilter, map[uint64]struct{}, error) {
+	// Try fast path with the minimized number of maxMetrics.
+	maxMetricsAdjusted := is.adjustMaxMetricsAdaptive(tr, maxMetrics)
+	minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetricsAdjusted)
+	if err == nil {
+		return minTf, minMetricIDs, nil
+	}
+	if err != errTooManyMetrics {
+		return nil, nil, err
+	}
+
+	// All the tag filters match too many metrics.
+
+	// Slow path: try filtering the matching metrics by time range.
+	// This should work well for cases when old metrics are constantly substituted
+	// by big number of new metrics. For example, prometheus-operator creates many new
+	// metrics for each new deployment.
+	//
+	// Allow fetching up to 20*maxMetrics metrics for the given time range
+	// in the hope these metricIDs will be filtered out by other filters later.
+	maxTimeRangeMetrics := 20 * maxMetrics
+	metricIDsForTimeRange, err := is.getMetricIDsForTimeRange(tr, maxTimeRangeMetrics+1)
+	if err == errMissingMetricIDsForDate {
+		// Slow path: try to select find the tag filter without maxMetrics adjustement.
+		minTf, minMetricIDs, err = is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetrics)
+		if err == nil {
+			return minTf, minMetricIDs, nil
+		}
+		if err != errTooManyMetrics {
+			return nil, nil, err
+		}
+		return nil, nil, fmt.Errorf("cannot find tag filter matching less than %d time series; "+
+			"either increase -search.maxUniqueTimeseries or use more specific tag filters", maxMetrics)
+	}
+	if err != nil {
+		return nil, nil, err
+	}
+	if len(metricIDsForTimeRange) <= maxTimeRangeMetrics {
+		return nil, metricIDsForTimeRange, nil
+	}
+
+	// Slow path: try to select the tag filter without maxMetrics adjustement.
+	minTf, minMetricIDs, err = is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetrics)
+	if err == nil {
+		return minTf, minMetricIDs, nil
+	}
+	if err != errTooManyMetrics {
+		return nil, nil, err
+	}
+	return nil, nil, fmt.Errorf("more than %d time series found on the time range %s; either increase -search.maxUniqueTimeseries or shrink the time range",
+		maxMetrics, tr.String())
+}
+
+const maxDaysForDateMetricIDs = 40
+
+func (is *indexSearch) adjustMaxMetricsAdaptive(tr TimeRange, maxMetrics int) int {
+	minDate := uint64(tr.MinTimestamp) / msecPerDay
+	maxDate := uint64(tr.MaxTimestamp) / msecPerDay
+	if maxDate-minDate > maxDaysForDateMetricIDs {
+		// Cannot reduce maxMetrics for the given time range,
+		// since it is expensive extracting metricIDs for the given tr.
+		return maxMetrics
+	}
+	hmPrev := is.db.prevHourMetricIDs.Load().(*hourMetricIDs)
+	if !hmPrev.isFull {
+		return maxMetrics
+	}
+	hourMetrics := len(hmPrev.m)
+	if hourMetrics >= 256 && maxMetrics > hourMetrics/4 {
+		// It is cheaper to filter on the hour or day metrics if the minimum
+		// number of matching metrics across tfs exceeds hourMetrics / 4.
+		return hourMetrics / 4
+	}
+	return maxMetrics
+}
+
 func (is *indexSearch) getTagFilterWithMinMetricIDsCountAdaptive(tfs *TagFilters, maxMetrics int) (*tagFilter, map[uint64]struct{}, error) {
 	kb := &is.kb
 	kb.B = append(kb.B[:0], uselessMultiTagFiltersKeyPrefix)
@@ -1266,29 +1363,6 @@ func (is *indexSearch) getTagFilterWithMinMetricIDsCountAdaptive(tfs *TagFilters
 
 var errTooManyMetrics = errors.New("all the tag filters match too many metrics")
 
-const maxDaysForDateMetricIDs = 40
-
-func (is *indexSearch) adjustMaxMetricsAdaptive(tr TimeRange, maxMetrics int) int {
-	minDate := uint64(tr.MinTimestamp) / msecPerDay
-	maxDate := uint64(tr.MaxTimestamp) / msecPerDay
-	if maxDate-minDate > maxDaysForDateMetricIDs {
-		// Cannot reduce maxMetrics for the given time range,
-		// since the it is expensive extracting metricIDs for the given tr.
-		return maxMetrics
-	}
-	hmPrev := is.db.prevHourMetricIDs.Load().(*hourMetricIDs)
-	if !hmPrev.isFull {
-		return maxMetrics
-	}
-	hourMetrics := len(hmPrev.m)
-	if hourMetrics >= 256 && maxMetrics > hourMetrics/4 {
-		// It is cheaper to filter on the hour or day metrics if the minimum
-		// number of matching metrics across tfs exceeds hourMetrics / 4.
-		return hourMetrics / 4
-	}
-	return maxMetrics
-}
-
 func (is *indexSearch) getTagFilterWithMinMetricIDsCount(tfs *TagFilters, maxMetrics int) (*tagFilter, map[uint64]struct{}, error) {
 	var minMetricIDs map[uint64]struct{}
 	var minTf *tagFilter
@@ -1381,7 +1455,7 @@ func matchTagFilters(mn *MetricName, tfs []*tagFilter, kb *bytesutil.ByteBuffer)
 				continue
 			}
 
-			// Found the matching tag name. Match for the value.
+			// Found the matching tag name. Match the value.
 			b := tag.Marshal(kb.B)
 			kb.B = b[:len(kb.B)]
 			ok, err := matchTagFilter(b, tf)
@@ -1394,7 +1468,7 @@ func matchTagFilters(mn *MetricName, tfs []*tagFilter, kb *bytesutil.ByteBuffer)
 			tagMatched = true
 			break
 		}
-		if !tagMatched {
+		if !tagMatched && !tf.isNegative {
 			// Matching tag name wasn't found.
 			return false, nil
 		}
@@ -1463,37 +1537,9 @@ func (is *indexSearch) updateMetricIDsForTagFilters(metricIDs map[uint64]struct{
 	// Sort tag filters for faster ts.Seek below.
 	sort.Slice(tfs.tfs, func(i, j int) bool { return bytes.Compare(tfs.tfs[i].prefix, tfs.tfs[j].prefix) < 0 })
 
-	maxMetricsAdjusted := is.adjustMaxMetricsAdaptive(tr, maxMetrics)
-	minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetricsAdjusted)
+	minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCountOptimized(tfs, tr, maxMetrics)
 	if err != nil {
-		if err != errTooManyMetrics {
-			return err
-		}
-
-		// All the tag filters match too many metrics.
-
-		// Slow path: try filtering the matching metrics by time range.
-		// This should work well for cases when old metrics are constantly substituted
-		// by big number of new metrics. For example, prometheus-operator creates many new
-		// metrics for each new deployment.
-		//
-		// Allow fetching up to 20*maxMetrics metrics for the given time range
-		// in the hope these metricIDs will be filtered out by other filters below.
-		maxTimeRangeMetrics := 20 * maxMetrics
-		metricIDsForTimeRange, err := is.getMetricIDsForTimeRange(tr, maxTimeRangeMetrics+1)
-		if err == errMissingMetricIDsForDate {
-			return fmt.Errorf("cannot find tag filter matching less than %d time series; either increase -search.maxUniqueTimeseries or use more specific tag filters",
-				maxMetrics)
-		}
-		if err != nil {
-			return err
-		}
-		if len(metricIDsForTimeRange) > maxTimeRangeMetrics {
-			return fmt.Errorf("more than %d time series found on the time range %s; either increase -search.maxUniqueTimeseries or shrink the time range",
-				maxTimeRangeMetrics, tr.String())
-		}
-		minMetricIDs = metricIDsForTimeRange
-		minTf = nil
+		return err
 	}
 
 	// Find intersection of minTf with other tfs.

lib/storage/index_db_test.go

@@ -12,7 +12,7 @@ import (
 	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
-	"github.com/VictoriaMetrics/fastcache"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 )
 
 func TestMarshalUnmarshalTSIDs(t *testing.T) {
@@ -57,10 +57,10 @@ func TestMarshalUnmarshalTSIDs(t *testing.T) {
 }
 
 func TestIndexDBOpenClose(t *testing.T) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
 
 	var hmCurr atomic.Value
 	hmCurr.Store(&hourMetricIDs{})
@@ -85,10 +85,10 @@ func TestIndexDB(t *testing.T) {
 	const metricGroups = 10
 
 	t.Run("serial", func(t *testing.T) {
-		metricIDCache := fastcache.New(1234)
-		metricNameCache := fastcache.New(1234)
-		defer metricIDCache.Reset()
-		defer metricNameCache.Reset()
+		metricIDCache := workingsetcache.New(1234, time.Hour)
+		metricNameCache := workingsetcache.New(1234, time.Hour)
+		defer metricIDCache.Stop()
+		defer metricNameCache.Stop()
 
 		var hmCurr atomic.Value
 		hmCurr.Store(&hourMetricIDs{})
@@ -142,10 +142,10 @@ func TestIndexDB(t *testing.T) {
 	})
 
 	t.Run("concurrent", func(t *testing.T) {
-		metricIDCache := fastcache.New(1234)
-		metricNameCache := fastcache.New(1234)
-		defer metricIDCache.Reset()
-		defer metricNameCache.Reset()
+		metricIDCache := workingsetcache.New(1234, time.Hour)
+		metricNameCache := workingsetcache.New(1234, time.Hour)
+		defer metricIDCache.Stop()
+		defer metricNameCache.Stop()
 
 		var hmCurr atomic.Value
 		hmCurr.Store(&hourMetricIDs{})
@@ -753,8 +753,8 @@ func TestMatchTagFilters(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %s", err)
 	}
-	if ok {
-		t.Fatalf("Shouldn't match")
+	if !ok {
+		t.Fatalf("Should match")
 	}
 	tfs.Reset()
 	if err := tfs.Add([]byte("non-existing-tag"), []byte("foob.+metric"), true, true); err != nil {
@@ -764,8 +764,19 @@ func TestMatchTagFilters(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %s", err)
 	}
-	if ok {
-		t.Fatalf("Shouldn't match")
+	if !ok {
+		t.Fatalf("Should match")
+	}
+	tfs.Reset()
+	if err := tfs.Add([]byte("non-existing-tag"), []byte(".+"), true, true); err != nil {
+		t.Fatalf("cannot add regexp, negative filter: %s", err)
+	}
+	ok, err = matchTagFilters(&mn, toTFPointers(tfs.tfs), &bb)
+	if err != nil {
+		t.Fatalf("unexpected error: %s", err)
+	}
+	if !ok {
+		t.Fatalf("Should match")
 	}
 
 	// Negative match by existing tag
@@ -859,6 +870,17 @@ func TestMatchTagFilters(t *testing.T) {
 	if !ok {
 		t.Fatalf("Should match")
 	}
+	tfs.Reset()
+	if err := tfs.Add([]byte("key 3"), []byte(""), true, false); err != nil {
+		t.Fatalf("cannot add regexp, negative filter: %s", err)
+	}
+	ok, err = matchTagFilters(&mn, toTFPointers(tfs.tfs), &bb)
+	if err != nil {
+		t.Fatalf("unexpected error: %s", err)
+	}
+	if !ok {
+		t.Fatalf("Should match")
+	}
 
 	// Positive match by multiple tags and MetricGroup
 	tfs.Reset()

lib/storage/index_db_timing_test.go

@@ -3,20 +3,50 @@ package storage
 import (
 	"fmt"
 	"os"
+	"regexp"
 	"strconv"
 	"sync/atomic"
 	"testing"
+	"time"
 
-	"github.com/VictoriaMetrics/fastcache"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 )
 
+func BenchmarkRegexpFilterMatch(b *testing.B) {
+	b.ReportAllocs()
+	b.RunParallel(func(pb *testing.PB) {
+		re := regexp.MustCompile(`.*foo-bar-baz.*`)
+		b := []byte("fdsffd foo-bar-baz assd fdsfad dasf dsa")
+		for pb.Next() {
+			if !re.Match(b) {
+				panic("BUG: regexp must match!")
+			}
+			b[0]++
+		}
+	})
+}
+
+func BenchmarkRegexpFilterMismatch(b *testing.B) {
+	b.ReportAllocs()
+	b.RunParallel(func(pb *testing.PB) {
+		re := regexp.MustCompile(`.*foo-bar-baz.*`)
+		b := []byte("fdsffd foo-bar sfddsf assd nmn,mfdsdsakj")
+		for pb.Next() {
+			if re.Match(b) {
+				panic("BUG: regexp mustn't match!")
+			}
+			b[0]++
+		}
+	})
+}
+
 func BenchmarkIndexDBAddTSIDs(b *testing.B) {
 	const recordsPerLoop = 1e3
 
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
 
 	var hmCurr atomic.Value
 	hmCurr.Store(&hourMetricIDs{})
@@ -78,86 +108,11 @@ func benchmarkIndexDBAddTSIDs(db *indexDB, tsid *TSID, mn *MetricName, startOffs
 	}
 }
 
-func BenchmarkIndexDBSearchTSIDs(b *testing.B) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
-
-	var hmCurr atomic.Value
-	hmCurr.Store(&hourMetricIDs{})
-	var hmPrev atomic.Value
-	hmPrev.Store(&hourMetricIDs{})
-
-	const dbName = "bench-index-db-search-tsids"
-	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
-	if err != nil {
-		b.Fatalf("cannot open indexDB: %s", err)
-	}
-	defer func() {
-		db.MustClose()
-		if err := os.RemoveAll(dbName); err != nil {
-			b.Fatalf("cannot remove indexDB: %s", err)
-		}
-	}()
-
-	const recordsCount = 1e5
-
-	// Fill the db with recordsCount records.
-	var mn MetricName
-	mn.MetricGroup = []byte("rps")
-	for i := 0; i < 2; i++ {
-		key := fmt.Sprintf("key_%d", i)
-		value := fmt.Sprintf("value_%d", i)
-		mn.AddTag(key, value)
-	}
-	var tsid TSID
-	var metricName []byte
-	is := db.getIndexSearch()
-	defer db.putIndexSearch(is)
-	for i := 0; i < recordsCount; i++ {
-		mn.sortTags()
-		metricName = mn.Marshal(metricName[:0])
-		if err := is.GetOrCreateTSIDByName(&tsid, metricName); err != nil {
-			b.Fatalf("cannot insert record: %s", err)
-		}
-	}
-
-	b.SetBytes(1)
-	b.ReportAllocs()
-	b.ResetTimer()
-	b.RunParallel(func(pb *testing.PB) {
-		tags := []Tag{
-			{[]byte("key_0"), []byte("value_0")},
-			{[]byte("key_1"), []byte("value_1")},
-		}
-		var tfs TagFilters
-		tfss := []*TagFilters{&tfs}
-		i := 0
-		for pb.Next() {
-			tfs.Reset()
-			for j := range tags {
-				if err := tfs.Add(tags[j].Key, tags[j].Value, false, false); err != nil {
-					panic(fmt.Errorf("BUG: unexpected error: %s", err))
-				}
-			}
-			tsids, err := db.searchTSIDs(tfss, TimeRange{}, 1e5)
-			if err != nil {
-				panic(fmt.Errorf("unexpected error in search for tfs=%s: %s", &tfs, err))
-			}
-			if len(tsids) == 0 && i < recordsCount {
-				panic(fmt.Errorf("zero tsids found for tfs=%s", &tfs))
-			}
-			i++
-		}
-	})
-}
-
 func BenchmarkIndexDBGetTSIDs(b *testing.B) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
 
 	var hmCurr atomic.Value
 	hmCurr.Store(&hourMetricIDs{})

lib/storage/metric_name.go

@@ -9,6 +9,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
 )
 
@@ -366,17 +367,8 @@ func (mn *MetricName) Unmarshal(src []byte) error {
 		}
 	}
 
-	// Verify no identical tag keys.
-	if len(mn.Tags) > 0 {
-		prevKey := mn.Tags[0].Key
-		for i := range mn.Tags[1:] {
-			t := &mn.Tags[1+i]
-			if bytes.Equal(t.Key, prevKey) {
-				return fmt.Errorf("found duplicate key %q", prevKey)
-			}
-			prevKey = t.Key
-		}
-	}
+	// There is no need in verifying for identical tag keys,
+	// since they must be handled in MetricName.Marshal inside marshalTags.
 
 	return nil
 }
@@ -392,9 +384,18 @@ const maxLabelNameLen = 256
 const maxLabelValueLen = 16 * 1024
 
 // The maximum number of labels per each timeseries.
+var maxLabelsPerTimeseries = 30
+
+// SetMaxLabelsPerTimeseries sets the limit on the number of labels
+// per each time series.
 //
-// Superflouos lables are dropped.
-const maxLabelsPerTimeseries = 30
+// Superfouos labels are dropped.
+func SetMaxLabelsPerTimeseries(maxLabels int) {
+	if maxLabels <= 0 {
+		logger.Panicf("BUG: maxLabels must be positive; got %d", maxLabels)
+	}
+	maxLabelsPerTimeseries = maxLabels
+}
 
 // MarshalMetricNameRaw marshals labels to dst and returns the result.
 //
@@ -574,8 +575,15 @@ func (ts *canonicalTagsSort) Swap(i, j int) {
 }
 
 func marshalTags(dst []byte, tags []Tag) []byte {
+	var prevKey []byte
 	for i := range tags {
-		dst = tags[i].Marshal(dst)
+		t := &tags[i]
+		if string(prevKey) == string(t.Key) {
+			// Skip duplicate keys, since they aren't allowed in Prometheus data model.
+			continue
+		}
+		prevKey = t.Key
+		dst = t.Marshal(dst)
 	}
 	return dst
 }

lib/storage/metric_name_test.go

@@ -34,6 +34,32 @@ func testMetricNameSortTags(t *testing.T, tags, expectedTags []string) {
 	}
 }
 
+func TestMetricNameMarshalDuplicateKeys(t *testing.T) {
+	var mn MetricName
+	mn.MetricGroup = []byte("xxx")
+	mn.AddTag("foo", "bar")
+	mn.AddTag("duplicate", "tag")
+	mn.AddTag("duplicate", "tag")
+	mn.AddTag("tt", "xx")
+	mn.AddTag("duplicate", "tag2")
+
+	var mnExpected MetricName
+	mnExpected.MetricGroup = []byte("xxx")
+	mnExpected.AddTag("duplicate", "tag")
+	mnExpected.AddTag("foo", "bar")
+	mnExpected.AddTag("tt", "xx")
+
+	mn.sortTags()
+	data := mn.Marshal(nil)
+	var mn1 MetricName
+	if err := mn1.Unmarshal(data); err != nil {
+		t.Fatalf("cannot unmarshal mn %s: %s", &mn, err)
+	}
+	if !reflect.DeepEqual(&mnExpected, &mn1) {
+		t.Fatalf("unexpected mn unmarshaled;\ngot\n%+v\nwant\n%+v", &mn1, &mnExpected)
+	}
+}
+
 func TestMetricNameMarshalUnmarshal(t *testing.T) {
 	for i := 0; i < 10; i++ {
 		for tagsCount := 0; tagsCount < 10; tagsCount++ {

lib/storage/part.go

@@ -144,7 +144,7 @@ func (p *part) MustClose() {
 	p.valuesFile.MustClose()
 	p.indexFile.MustClose()
 
-	isBig := p.ph.RowsCount > maxRowsPerSmallPart
+	isBig := p.ph.RowsCount > maxRowsPerSmallPart()
 	p.ibCache.Reset(isBig)
 }
 

lib/storage/part_search.go

@@ -28,6 +28,9 @@ type partSearch struct {
 	// tr is a time range to search.
 	tr TimeRange
 
+	// Skip populating timestampsData and valuesData in Block if fetchData=false.
+	fetchData bool
+
 	metaindex []metaindexRow
 
 	ibCache *indexBlockCache
@@ -48,6 +51,7 @@ func (ps *partSearch) reset() {
 	ps.p = nil
 	ps.tsids = ps.tsids[:0]
 	ps.tsidIdx = 0
+	ps.fetchData = true
 	ps.metaindex = nil
 	ps.ibCache = nil
 	ps.bhs = nil
@@ -61,7 +65,7 @@ func (ps *partSearch) reset() {
 }
 
 // Init initializes the ps with the given p, tsids and tr.
-func (ps *partSearch) Init(p *part, tsids []TSID, tr TimeRange) {
+func (ps *partSearch) Init(p *part, tsids []TSID, tr TimeRange, fetchData bool) {
 	ps.reset()
 	ps.p = p
 
@@ -72,6 +76,7 @@ func (ps *partSearch) Init(p *part, tsids []TSID, tr TimeRange) {
 		ps.tsids = append(ps.tsids[:0], tsids...)
 	}
 	ps.tr = tr
+	ps.fetchData = fetchData
 	ps.metaindex = p.metaindex
 	ps.ibCache = &p.ibCache
 
@@ -281,11 +286,14 @@ func (ps *partSearch) searchBHS() bool {
 
 func (ps *partSearch) readBlock(bh *blockHeader) {
 	ps.Block.Reset()
+	ps.Block.bh = *bh
+	if !ps.fetchData {
+		return
+	}
+
 	ps.Block.timestampsData = bytesutil.Resize(ps.Block.timestampsData[:0], int(bh.TimestampsBlockSize))
 	ps.p.timestampsFile.ReadAt(ps.Block.timestampsData, int64(bh.TimestampsBlockOffset))
 
 	ps.Block.valuesData = bytesutil.Resize(ps.Block.valuesData[:0], int(bh.ValuesBlockSize))
 	ps.p.valuesFile.ReadAt(ps.Block.valuesData, int64(bh.ValuesBlockOffset))
-
-	ps.Block.bh = *bh
 }