lib/storage: do not change timestamps to constant rate if values are constant or have constant delta

This breaks the original timestamps, which results in issues like
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/120 and
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/141 .

.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.

.gitignore

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

.travis.yml

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

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

@@ -19,16 +19,20 @@ include deployment/*/Makefile
 clean:
 	rm -rf bin/*
 
+publish: publish-victoria-metrics
+
+package: package-victoria-metrics
+
 release: victoria-metrics-prod
 	cd bin && tar czf victoria-metrics-$(PKG_TAG).tar.gz victoria-metrics-prod
 
 fmt:
-	go fmt $(PKG_PREFIX)/lib/...
-	go fmt $(PKG_PREFIX)/app/...
+	GO111MODULE=on gofmt -l -w -s ./lib
+	GO111MODULE=on gofmt -l -w -s ./app
 
 vet:
-	go vet $(PKG_PREFIX)/lib/...
-	go vet $(PKG_PREFIX)/app/...
+	GO111MODULE=on go vet -mod=vendor ./lib/...
+	GO111MODULE=on go vet -mod=vendor ./app/...
 
 lint: install-golint
 	golint lib/...
@@ -46,19 +50,40 @@ errcheck: install-errcheck
 install-errcheck:
 	which errcheck || GO111MODULE=off go get -u github.com/kisielk/errcheck
 
+check_all: fmt vet lint errcheck golangci-lint
+
 test:
-	go test $(PKG_PREFIX)/lib/...
+	GO111MODULE=on go test -tags=integration -mod=vendor ./lib/... ./app/...
+
+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:
-	go test -bench=. $(PKG_PREFIX)/lib/...
+	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:
-	go get -u
-	go mod tidy
-	go mod vendor
+	GO111MODULE=on go get -u ./lib/...
+	GO111MODULE=on go get -u ./app/...
+	GO111MODULE=on go mod tidy
+	GO111MODULE=on go mod vendor
 
 quicktemplate-gen: install-qtc
 	qtc
 
 install-qtc:
 	which qtc || GO111MODULE=off go get -u github.com/valyala/quicktemplate/qtc
+
+
+golangci-lint: install-golangci-lint
+	golangci-lint run --exclude '(SA4003|SA1019):' -D errcheck
+
+install-golangci-lint:
+	which golangci-lint || GO111MODULE=off go get -u github.com/golangci/golangci-lint/cmd/golangci-lint

README.md

@@ -1,12 +1,17 @@
-<img  text-align="center" alt="Victoria Metrics" src="logo.png">
+[![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)
+[![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
 
-[![Latest Release](https://img.shields.io/github/release/VictoriaMetrics/VictoriaMetrics.svg?style=flat-square)](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/latest)
-
-VictoriaMetrics is 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/valyala/victoria-metrics/) and
+[docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) and
 in [source code](https://github.com/VictoriaMetrics/VictoriaMetrics).
 
 Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
@@ -16,15 +21,17 @@ 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.
 * 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).
+  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).
 * Easy operation:
   * VictoriaMetrics consists of a single executable without external dependencies.
   * All the configuration is done via explicit command-line flags with reasonable defaults.
@@ -37,7 +44,7 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   * [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 IoT sensors, connected car sensors and industrial sensors.
+* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars and industrial telemetry.
 * Has open source [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
 
 
@@ -46,72 +53,67 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
 
 ### Table of contents
 
-* [How to build from sources](#how-to-build-from-sources)
-* [How to start VictoriaMetrics](#how-to-start-victoriametrics)
-* [Prometheus setup](#prometheus-setup)
-* [Grafana setup](#grafana-setup)
-* [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)
-* [How to send data from OpenTSDB-compatible agents](#how-to-send-data-from-opentsdb-compatible-agents)
-* [How to apply new config / ugrade VictoriaMetrics](#how-to-apply-new-config--upgrade-victoriametrics)
-* [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)
-* [Federation](#federation)
-* [Capacity planning](#capacity-planning)
-* [High Availability](#high-availability)
-* [Multiple retentions](#multiple-retentions)
-* [Scalability and cluster version](#scalability-and-cluster-version)
-* [Security](#security)
-* [Tuning](#tuning)
-* [Monitoring](#monitoring)
-* [Troubleshooting](#troubleshooting)
-* [Community and contributions](#community-and-contributions)
-* [Reporting bugs](#reporting-bugs)
-
-
-### 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/valyala/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 `go build ./app/victoria-metrics` from the root folder of the repository.
-   It will build `victoria-metrics` binary in the root folder of the repository.
-
-#### Production build
-
-1. [Install docker](https://docs.docker.com/install/).
-2. Run `make victoria-metrics-prod` from the root folder of the respository.
-   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 `valyala/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`.
-
+  - [How to start VictoriaMetrics](#how-to-start-victoriametrics)
+  - [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)
+  - [Federation](#federation)
+  - [Capacity planning](#capacity-planning)
+  - [High availability](#high-availability)
+  - [Multiple retentions](#multiple-retentions)
+  - [Downsampling](#downsampling)
+  - [Multi-tenancy](#multi-tenancy)
+  - [Scalability and cluster version](#scalability-and-cluster-version)
+  - [Alerting](#alerting)
+  - [Security](#security)
+  - [Tuning](#tuning)
+  - [Monitoring](#monitoring)
+  - [Troubleshooting](#troubleshooting)
+- [Roadmap](#roadmap)
+- [Contacts](#contacts)
+- [Community and contributions](#community-and-contributions)
+- [Reporting bugs](#reporting-bugs)
+- [Victoria Metrics Logo](#victoria-metrics-logo)
+  - [Logo Usage Guidelines](#logo-usage-guidelines)
+    - [Font used:](#font-used)
+    - [Color Palette:](#color-palette)
+  - [We kindly ask:](#we-kindly-ask)
 
 
 ### 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. By default, it is disabled.
 
 Pass `-help` to see all the available flags with description and default values.
 
+It is recommended setting up [monitoring](#monitoring) for VictoriaMetrics.
+
 
 ### Prometheus setup
 
@@ -122,6 +124,7 @@ remote_write:
   - url: http://<victoriametrics-addr>:8428/api/v1/write
     queue_config:
       max_samples_per_send: 10000
+      max_shards: 100
 ```
 
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
@@ -131,11 +134,11 @@ 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
@@ -146,7 +149,11 @@ 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,
+since the previous versions may have issues with `remote_write`.
 
 
 ### Grafana setup
@@ -163,6 +170,28 @@ Then build graphs with the created datasource using [Prometheus query language](
 VictoriaMetrics supports native PromQL and [extends it with useful features](ExtendedPromQL).
 
 
+### How to upgrade VictoriaMetrics?
+
+It is safe upgrading VictoriaMetrics to new versions unless [release notes](https://github.com/VictoriaMetrics/VictoriaMetrics/releases)
+say otherwise. It is recommended performing regular upgrades to the latest version,
+since it may contain important bug fixes, performance optimizations or new features.
+
+Follow the following steps during the upgrade:
+
+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 the upgraded VictoriaMetrics.
+
+
+### How to apply new config to VictoriaMetrics?
+
+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 the new config.
+
+
 ### How to send data from InfluxDB-compatible agents such as [Telegraf](https://www.influxdata.com/time-series-platform/telegraf/)?
 
 Just use `http://<victoriametric-addr>:8428` url instead of InfluxDB url in agents' configs.
@@ -176,10 +205,46 @@ 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
-* Field names are mapped to time series names prefixed by `{measurement}.` value
-* Field values are mapped to time series values
-* Tags are mapped to Prometheus labels as-is
+* [`db` query arg](https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint) is mapped into `db` label value.
+* 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.
+* Field values are mapped to time series values.
+* Tags are mapped to Prometheus labels as-is.
+
+For example, the following Influx line:
+
+```
+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
+```
+
+Example for writing data with [Influx line protocol](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/)
+to local VictoriaMetrics using `curl`:
+
+```
+curl -d 'measurement,tag1=value1,tag2=value2 field1=123,field2=1.23' -X POST 'http://localhost:8428/write'
+```
+
+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__!=""}'
+```
+
+The `/api/v1/export` endpoint should return the following response:
+
+```
+{"metric":{"__name__":"measurement.field1","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560272508147]}
+{"metric":{"__name__":"measurement.field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
+```
 
 
 ### How to send data from Graphite-compatible agents such as [StatsD](https://github.com/etsy/statsd)?
@@ -188,36 +253,148 @@ VictoriaMetrics maps Influx data using the following rules:
 the following command will enable Graphite receiver in VictoriaMetrics on TCP and UDP port `2003`:
 
 ```
-/path/to/victoria-metrics-prod ... -graphiteListenAddr=:2003
+/path/to/victoria-metrics-prod -graphiteListenAddr=:2003
 ```
 
 2) Use the configured address in Graphite-compatible agents. For instance, set `graphiteHost`
 to the VictoriaMetrics host in `StatsD` configs.
 
 
+Example for writing data with Graphite plaintext protocol to local VictoriaMetrics using `nc`:
+
+```
+echo "foo.bar.baz;tag1=value1;tag2=value2 123 `date +%s`" | nc -N localhost 2003
+```
+
+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__!=""}'
+```
+
+The `/api/v1/export` endpoint should return the following response:
+
+```
+{"metric":{"__name__":"foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560277406000]}
+```
+
+
+### 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?
 
 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`:
 
 ```
-/path/to/victoria-metrics-prod ... -opentsdbListenAddr=:4242
+/path/to/victoria-metrics-prod -opentsdbListenAddr=:4242
 ```
 
 2) Send data to the given address from OpenTSDB-compatible agents.
 
 
-### How to apply new config / upgrade VictoriaMetrics?
+Example for writing data with OpenTSDB protocol to local VictoriaMetrics using `nc`:
 
-VictoriaMetrics must be restarted in order to upgrade or apply new config:
+```
+echo "put foo.bar.baz `date +%s` 123 tag1=value1 tag2=value2" | nc -N localhost 4242
+```
 
-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 the upgraded VictoriaMetrics with new config.
+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__!=""}'
+```
+
+The `/api/v1/export` endpoint should return the following response:
+
+```
+{"metric":{"__name__":"foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560277292000]}
+```
+
+
+### 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 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:
 
@@ -226,7 +403,7 @@ The page will return the following JSON response:
 ```
 
 Snapshots are created under `<-storageDataPath>/snapshots` directory, where `<-storageDataPath>`
-is the command-line flag value. Snapshots can be archived to backup storage via `rsync -L`, `scp -r`
+is the command-line flag value. Snapshots can be archived to backup storage via `cp -L`, `rsync -L`, `scp -r`
 or any similar tool that follows symlinks during copying.
 
 The `http://<victoriametrics-addr>:8428/snapshot/list` page contains the list of available snapshots.
@@ -236,6 +413,12 @@ to delete `<snapshot-name>` snapshot.
 
 Navigate to `http://<victoriametrics-addr>:8428/snapshot/delete_all` in order to delete all the snapshots.
 
+Steps for restoring from a snapshot:
+1. Stop VictoriaMetrics with `kill -INT`.
+2. Remove the entire contents of the directory pointed by `-storageDataPath` command-line flag.
+3. Copy snapshot contents to the directory pointed by `-storageDataPath`.
+4. Start VictoriaMetrics.
+
 
 ### How to delete time series?
 
@@ -267,31 +450,55 @@ 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 overriden.
+`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.
-  VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited with `-memory.allowedPercent` flag.
+  The number of active time series may be obtained from `vm_cache_entries{type="storage/hour_metric_ids"}` metric
+  exproted on the `/metrics` page.
+  VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited by `-memory.allowedPercent` flag.
+
 * CPU cores: a CPU core per 300K inserted data points per second. So, ~4 CPU cores are required for processing
-  the insert stream of 1M data points per second.
+  the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data.
+  See [this article](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) for details.
   If you see lower numbers per CPU core, then it is likely active time series info doesn't fit caches,
   so you need more RAM for lowering CPU usage.
-* 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
 
-1) Install multiple VictoriaMetrics instances in distinct datacenters.
+1) Install multiple VictoriaMetrics instances in distinct datacenters (availability zones).
 2) Add addresses of these instances to `remote_write` section in Prometheus config:
 
 ```yml
@@ -316,6 +523,10 @@ kill -HUP `pidof prometheus`
 6) Set up Prometheus datasource in Grafana that points to Promxy.
 
 
+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`.
+
+
 ### Multiple retentions
 
 Just start multiple VictoriaMetrics instances with distinct values for the following flags:
@@ -325,20 +536,43 @@ Just start multiple VictoriaMetrics instances with distinct values for the follo
 * `-httpListenAddr`, so clients may reach VictoriaMetrics instance with proper retention
 
 
+### Downsampling
+
+There is no downsampling support at the moment, but:
+- VictoriaMetrics is optimized for querying big amounts of raw data. See benchmark results for heavy queries
+  in [this article](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
+- VictoriaMetrics has good compression for on-disk data. See [this article](https://medium.com/@valyala/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932)
+  for details.
+
+These properties reduce the need in downsampling. We plan to implement downsampling in the future.
+See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/36) for details.
+
+
+### Multi-tenancy
+
+Single-node VictoriaMetrics doesn't support multi-tenancy. Use [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster) instead.
+
+
 ### 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
-such as Thanos, Uber M3, InfluxDB or TimescaleDB.
+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
 horizontally scalable long-term remote storage for really large Prometheus deployments.
 [Contact us](mailto:info@victoriametrics.com) for paid support.
 
 
+### Alerting
+
+VictoriaMetrics doesn't support rule evaluation and alerting yet, so these actions must be performed either
+on [Prometheus side](https://prometheus.io/docs/alerting/overview/) or on [Grafana side](https://grafana.com/docs/alerting/rules/).
+
+
 ### 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.
@@ -353,10 +587,10 @@ 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.
 
 
@@ -366,6 +600,21 @@ VictoriaMetrics exports internal metrics in Prometheus format on the `/metrics`
 Add this page to Prometheus' scrape config in order to collect VictoriaMetrics metrics.
 There is [an official Grafana dashboard for single-node VictoriaMetrics](https://grafana.com/dashboards/10229).
 
+The most interesting metrics are:
+
+* `vm_cache_entries{type="storage/hour_metric_ids"}` - the number of time series with new data points during the last hour
+  aka active time series.
+* `vm_rows{type="indexdb"}` - the number of rows in inverted index. Each label in each unique time series adds a single
+  row into the inverted index. An approximate number of time series in the database may be calculated as
+  `vm_rows{type="indexdb"} / (avg_labels_per_series + 1)`, where `avg_labels_per_series` is the average number of labels
+  per each time series.
+* Sum of `vm_rows{type="storage/big"}` and `vm_rows{type="storage/small"}` - total number of `(timestamp, value)` data points
+  in the database.
+* Sum of all the `vm_cache_size_bytes` metrics - the total size of all the caches in the database.
+* `vm_allowed_memory_bytes` - the maximum allowed size for caches in the database. It is calculated as `system_memory * <-memory.allowedPercent> / 100`,
+  where `system_memory` is the amount of system memory and `-memory.allowedPercent` is the corresponding flag value.
+* `vm_rows_inserted_total` - the total number of inserted rows since VictoriaMetrics start.
+
 
 ### Troubleshooting
 
@@ -376,10 +625,47 @@ There is [an official Grafana dashboard for single-node VictoriaMetrics](https:/
   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
+  for automatic recovering from such errors.
+
+## 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).
+
 
 ## Community and contributions
 
-Feel free asking any questions regarding VictoriaMetrics [here](https://groups.google.com/forum/#!forum/victorametrics-users).
+Feel free asking any questions regarding VictoriaMetrics:
+
+- [slack](http://slack.victoriametrics.com/)
+- [telergam-en](https://t.me/VictoriaMetrics_en)
+- [telergam-ru](https://t.me/VictoriaMetrics_ru1)
+- [google groups](https://groups.google.com/forum/#!forum/victorametrics-users)
+
+
+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):
 

app/victoria-metrics/Makefile

@@ -1,5 +1,8 @@
 # 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
+
 victoria-metrics-prod:
 	APP_NAME=victoria-metrics $(MAKE) app-via-docker
 
@@ -12,10 +15,58 @@ publish-victoria-metrics:
 
 run-victoria-metrics:
 	mkdir -p victoria-metrics-data
-	DOCKER_OPTS='-v $(shell pwd)/victoria-metrics-data:/victoria-metrics-data -p 8428:8428 -p 2003:2003 -p 2003:2003/udp' \
+	DOCKER_OPTS='-v $(shell pwd)/victoria-metrics-data:/victoria-metrics-data' \
 	APP_NAME=victoria-metrics \
 	ARGS='-graphiteListenAddr=:2003 -opentsdbListenAddr=:4242 -retentionPeriod=12 -search.maxUniqueTimeseries=1000000 -search.maxQueryDuration=10m' \
 	$(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' $(MAKE) app-via-docker
+
+victoria-metrics-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(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' $(MAKE) app-via-docker
+
+victoria-metrics-pure:
+	GO111MODULE=on CGO_ENABLED=0 go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-pure ./app/victoria-metrics
+
+victoria-metrics-pure-prod:
+	APP_NAME=victoria-metrics APP_SUFFIX='-pure' DOCKER_OPTS='--env CGO_ENABLED=0' $(MAKE) app-via-docker
+
+### 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_test.go

@@ -0,0 +1,281 @@
+// +build integration
+
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io/ioutil"
+	"log"
+	"net"
+	"net/http"
+	"os"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+const (
+	testFixturesDir        = "testdata"
+	testStorageSuffix      = "vm-test-storage"
+	testHTTPListenAddr     = ":7654"
+	testStatsDListenAddr   = ":2003"
+	testOpenTSDBListenAddr = ":4242"
+	testLogLevel           = "INFO"
+)
+
+const (
+	testReadHTTPPath   = "http://127.0.0.1" + testHTTPListenAddr
+	testWriteHTTPPath  = "http://127.0.0.1" + testHTTPListenAddr + "/write"
+	testHealthHTTPPath = "http://127.0.0.1" + testHTTPListenAddr + "/health"
+)
+
+const (
+	testStorageInitTimeout = 10 * time.Second
+)
+
+var (
+	storagePath   string
+	insertionTime = time.Now().UTC()
+)
+
+type test struct {
+	Name   string `json:"name"`
+	Data   string `json:"data"`
+	Query  string `json:"query"`
+	Result []Row  `json:"result"`
+}
+
+type Row struct {
+	Metric     map[string]string `json:"metric"`
+	Values     []float64         `json:"values"`
+	Timestamps []int64           `json:"timestamps"`
+}
+
+func TestMain(m *testing.M) {
+	setUp()
+	code := m.Run()
+	tearDown()
+	os.Exit(code)
+}
+
+func setUp() {
+	storagePath = filepath.Join(os.TempDir(), testStorageSuffix)
+	processFlags()
+	logger.Init()
+	vmstorage.InitWithoutMetrics()
+	vmselect.Init()
+	vminsert.Init()
+	go httpserver.Serve(*httpListenAddr, requestHandler)
+	readyStorageCheckFunc := func() bool {
+		resp, err := http.Get(testHealthHTTPPath)
+		if err != nil {
+			return false
+		}
+		resp.Body.Close()
+		return resp.StatusCode == 200
+	}
+	if err := waitFor(testStorageInitTimeout, readyStorageCheckFunc); err != nil {
+		log.Fatalf("http server can't start for %s seconds, err %s", testStorageInitTimeout, err)
+	}
+}
+
+func processFlags() {
+	flag.Parse()
+	for _, fs := range []struct {
+		flag  string
+		value string
+	}{
+		{flag: "storageDataPath", value: storagePath},
+		{flag: "httpListenAddr", value: testHTTPListenAddr},
+		{flag: "graphiteListenAddr", value: testStatsDListenAddr},
+		{flag: "opentsdbListenAddr", value: testOpenTSDBListenAddr},
+		{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)
+		}
+	}
+}
+
+func waitFor(timeout time.Duration, f func() bool) error {
+	fraction := timeout / 10
+	for i := fraction; i < timeout; i += fraction {
+		if f() {
+			return nil
+		}
+		time.Sleep(fraction)
+	}
+	return fmt.Errorf("timeout")
+}
+
+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)
+}
+
+func TestWriteRead(t *testing.T) {
+	t.Run("write", testWrite)
+	time.Sleep(1 * time.Second)
+	vmstorage.Stop()
+
+	// open storage after stop in write
+	vmstorage.InitWithoutMetrics()
+	t.Run("read", testRead)
+}
+
+func testWrite(t *testing.T) {
+	t.Run("influxdb", func(t *testing.T) {
+		for _, test := range readIn("influxdb", t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
+			t.Run(test.Name, func(t *testing.T) {
+				t.Parallel()
+				httpWrite(t, testWriteHTTPPath, test.Data)
+			})
+		}
+	})
+	t.Run("graphite", func(t *testing.T) {
+		for _, test := range readIn("graphite", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+			t.Run(test.Name, func(t *testing.T) {
+				t.Parallel()
+				tcpWrite(t, "127.0.0.1"+testStatsDListenAddr, test.Data)
+			})
+		}
+	})
+	t.Run("opentsdb", func(t *testing.T) {
+		for _, test := range readIn("opentsdb", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+			t.Run(test.Name, func(t *testing.T) {
+				t.Parallel()
+				tcpWrite(t, "127.0.0.1"+testOpenTSDBListenAddr, test.Data)
+			})
+		}
+	})
+}
+
+func testRead(t *testing.T) {
+	for _, engine := range []string{"graphite", "opentsdb", "influxdb"} {
+		t.Run(engine, func(t *testing.T) {
+			for _, test := range readIn(engine, t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
+				test := test
+				t.Run(test.Name, func(t *testing.T) {
+					t.Parallel()
+					rowContains(t, httpRead(t, testReadHTTPPath, test.Query), test.Result)
+				})
+			}
+		})
+	}
+}
+
+func readIn(readFor string, t *testing.T, timeStr string) []test {
+	t.Helper()
+	s := newSuite(t)
+	var tt []test
+	s.noError(filepath.Walk(filepath.Join(testFixturesDir, readFor), func(path string, info os.FileInfo, err error) error {
+		if filepath.Ext(path) != ".json" {
+			return nil
+		}
+		b, err := ioutil.ReadFile(path)
+		s.noError(err)
+		item := test{}
+		s.noError(json.Unmarshal(b, &item))
+		item.Data = strings.Replace(item.Data, "{TIME}", timeStr, 1)
+		tt = append(tt, item)
+		return nil
+	}))
+	if len(tt) == 0 {
+		t.Fatalf("no test found in %s", filepath.Join(testFixturesDir, readFor))
+	}
+	return tt
+}
+
+func httpWrite(t *testing.T, address string, data string) {
+	t.Helper()
+	s := newSuite(t)
+	resp, err := http.Post(address, "", bytes.NewBufferString(data))
+	s.noError(err)
+	s.noError(resp.Body.Close())
+	s.equalInt(resp.StatusCode, 204)
+}
+
+func tcpWrite(t *testing.T, address string, data string) {
+	t.Helper()
+	s := newSuite(t)
+	conn, err := net.Dial("tcp", address)
+	s.noError(err)
+	defer conn.Close()
+	n, err := conn.Write([]byte(data))
+	s.noError(err)
+	s.equalInt(n, len(data))
+}
+
+func httpRead(t *testing.T, address, query string) []Row {
+	t.Helper()
+	s := newSuite(t)
+	resp, err := http.Get(address + query)
+	s.noError(err)
+	defer resp.Body.Close()
+	s.equalInt(resp.StatusCode, 200)
+	var rows []Row
+	for dec := json.NewDecoder(resp.Body); dec.More(); {
+		var row Row
+		s.noError(dec.Decode(&row))
+		rows = append(rows, row)
+	}
+	return rows
+}
+
+func rowContains(t *testing.T, rows, contains []Row) {
+	t.Helper()
+	for _, r := range rows {
+		contains = removeIfFound(r, contains)
+	}
+	if len(contains) > 0 {
+		t.Fatalf("result rows %+v not found in %+v", contains, rows)
+	}
+}
+
+func removeIfFound(r Row, contains []Row) []Row {
+	for i, item := range contains {
+		// todo check time
+		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Values, item.Values) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+type suite struct{ t *testing.T }
+
+func newSuite(t *testing.T) *suite { return &suite{t: t} }
+
+func (s *suite) noError(err error) {
+	s.t.Helper()
+	if err != nil {
+		s.t.Errorf("unexpected error %v", err)
+		s.t.FailNow()
+	}
+}
+
+func (s *suite) equalInt(a, b int) {
+	s.t.Helper()
+	if a != b {
+		s.t.Errorf("%d not equal %d", a, b)
+		s.t.FailNow()
+	}
+}

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

@@ -0,0 +1,8 @@
+{
+  "name": "basic_insertion",
+  "data": "graphite.foo.bar.baz;tag1=value1;tag2=value2 123 {TIME}",
+  "query": "/api/v1/export?match={__name__!=\"\"}",
+  "result": [
+    {"metric":{"__name__":"graphite.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123]}
+  ]
+}

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

@@ -0,0 +1,9 @@
+{
+  "name": "basic_insertion",
+  "data": "measurement,tag1=value1,tag2=value2 field1=1.23,field2=123",
+  "query": "/api/v1/export?match={__name__!=\"\"}",
+  "result": [
+    {"metric":{"__name__":"measurement_field2","tag1":"value1","tag2":"value2"},"values":[123]},
+    {"metric":{"__name__":"measurement_field1","tag1":"value1","tag2":"value2"},"values":[1.23]}
+  ]
+}

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

@@ -0,0 +1,8 @@
+{
+  "name": "basic_insertion",
+  "data": "put openstdb.foo.bar.baz {TIME} 123 tag1=value1 tag2=value2",
+  "query": "/api/v1/export?match={__name__!=\"\"}",
+  "result": [
+    {"metric":{"__name__":"openstdb.foo.bar.baz","tag1":"value1","tag2":"value2"},"values":[123]}
+  ]
+}

app/vminsert/common/insert_ctx.go

@@ -30,11 +30,10 @@ func (ctx *InsertCtx) Reset(rowsLen int) {
 		mr.MetricNameRaw = nil
 	}
 	ctx.mrs = ctx.mrs[:0]
-
 	if n := rowsLen - cap(ctx.mrs); n > 0 {
 		ctx.mrs = append(ctx.mrs[:cap(ctx.mrs)], make([]storage.MetricRow, n)...)
 	}
-	ctx.mrs = ctx.mrs[:rowsLen]
+	ctx.mrs = ctx.mrs[:0]
 	ctx.metricNamesBuf = ctx.metricNamesBuf[:0]
 }
 

app/vminsert/common/lines_reader.go

@@ -0,0 +1,68 @@
+package common
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+)
+
+// The maximum size of a single line returned by ReadLinesBlock.
+const maxLineSize = 256 * 1024
+
+// Default size in bytes of a single block returned by ReadLinesBlock.
+const defaultBlockSize = 64 * 1024
+
+// ReadLinesBlock reads a block of lines delimited by '\n' from tailBuf and r into dstBuf.
+//
+// Trailing chars after the last newline are put into tailBuf.
+//
+// Returns (dstBuf, tailBuf).
+func ReadLinesBlock(r io.Reader, dstBuf, tailBuf []byte) ([]byte, []byte, error) {
+	if cap(dstBuf) < defaultBlockSize {
+		dstBuf = bytesutil.Resize(dstBuf, defaultBlockSize)
+	}
+	dstBuf = append(dstBuf[:0], tailBuf...)
+	tailBuf = tailBuf[:0]
+again:
+	n, err := r.Read(dstBuf[len(dstBuf):cap(dstBuf)])
+	// Check for error only if zero bytes read from r, i.e. no forward progress made.
+	// Otherwise process the read data.
+	if n == 0 {
+		if err == nil {
+			return dstBuf, tailBuf, fmt.Errorf("no forward progress made")
+		}
+		if err == io.EOF && len(dstBuf) > 0 {
+			// Missing newline in the end of stream. This is OK,
+			// so suppress io.EOF for now. It will be returned during the next
+			// call to ReadLinesBlock.
+			// This fixes https://github.com/VictoriaMetrics/VictoriaMetrics/issues/60 .
+			return dstBuf, tailBuf, nil
+		}
+		return dstBuf, tailBuf, err
+	}
+	dstBuf = dstBuf[:len(dstBuf)+n]
+
+	// Search for the last newline in dstBuf and put the rest into tailBuf.
+	nn := bytes.LastIndexByte(dstBuf[len(dstBuf)-n:], '\n')
+	if nn < 0 {
+		// Didn't found at least a single line.
+		if len(dstBuf) > maxLineSize {
+			return dstBuf, tailBuf, fmt.Errorf("too long line: more than %d bytes", maxLineSize)
+		}
+		if cap(dstBuf) < 2*len(dstBuf) {
+			// Increase dsbBuf capacity, so more data could be read into it.
+			dstBufLen := len(dstBuf)
+			dstBuf = bytesutil.Resize(dstBuf, 2*cap(dstBuf))
+			dstBuf = dstBuf[:dstBufLen]
+		}
+		goto again
+	}
+
+	// Found at least a single line. Return it.
+	nn += len(dstBuf) - n
+	tailBuf = append(tailBuf[:0], dstBuf[nn+1:]...)
+	dstBuf = dstBuf[:nn]
+	return dstBuf, tailBuf, nil
+}

app/vminsert/common/lines_reader_test.go

@@ -0,0 +1,213 @@
+package common
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"reflect"
+	"testing"
+)
+
+func TestReadLinesBlockFailure(t *testing.T) {
+	f := func(s string) {
+		t.Helper()
+		r := bytes.NewBufferString(s)
+		if _, _, err := ReadLinesBlock(r, nil, nil); err == nil {
+			t.Fatalf("expecting non-nil error")
+		}
+		sbr := &singleByteReader{
+			b: []byte(s),
+		}
+		if _, _, err := ReadLinesBlock(sbr, nil, nil); err == nil {
+			t.Fatalf("expecting non-nil error")
+		}
+		fr := &failureReader{}
+		if _, _, err := ReadLinesBlock(fr, nil, nil); err == nil {
+			t.Fatalf("expecting non-nil error")
+		}
+	}
+
+	// empty string
+	f("")
+
+	// too long string
+	b := make([]byte, maxLineSize+1)
+	f(string(b))
+}
+
+type failureReader struct{}
+
+func (fr *failureReader) Read(p []byte) (int, error) {
+	return 0, fmt.Errorf("some error")
+}
+
+func TestReadLinesBlockMultiLinesSingleByteReader(t *testing.T) {
+	f := func(s string, linesExpected []string) {
+		t.Helper()
+
+		r := &singleByteReader{
+			b: []byte(s),
+		}
+		var err error
+		var dstBuf, tailBuf []byte
+		var lines []string
+		for {
+			dstBuf, tailBuf, err = ReadLinesBlock(r, dstBuf, tailBuf)
+			if err != nil {
+				if err == io.EOF {
+					break
+				}
+				t.Fatalf("unexpected error in ReadLinesBlock(%q): %s", s, err)
+			}
+			lines = append(lines, string(dstBuf))
+		}
+		if !reflect.DeepEqual(lines, linesExpected) {
+			t.Fatalf("unexpected lines after reading %q: got %q; want %q", s, lines, linesExpected)
+		}
+	}
+
+	f("", nil)
+	f("foo", []string{"foo"})
+	f("foo\n", []string{"foo"})
+	f("foo\nbar", []string{"foo", "bar"})
+	f("\nfoo\nbar", []string{"", "foo", "bar"})
+	f("\nfoo\nbar\n", []string{"", "foo", "bar"})
+	f("\nfoo\nbar\n\n", []string{"", "foo", "bar", ""})
+}
+
+func TestReadLinesBlockMultiLinesBytesBuffer(t *testing.T) {
+	f := func(s string, linesExpected []string) {
+		t.Helper()
+
+		r := bytes.NewBufferString(s)
+		var err error
+		var dstBuf, tailBuf []byte
+		var lines []string
+		for {
+			dstBuf, tailBuf, err = ReadLinesBlock(r, dstBuf, tailBuf)
+			if err != nil {
+				if err == io.EOF {
+					break
+				}
+				t.Fatalf("unexpected error in ReadLinesBlock(%q): %s", s, err)
+			}
+			lines = append(lines, string(dstBuf))
+		}
+		if !reflect.DeepEqual(lines, linesExpected) {
+			t.Fatalf("unexpected lines after reading %q: got %q; want %q", s, lines, linesExpected)
+		}
+	}
+
+	f("", nil)
+	f("foo", []string{"foo"})
+	f("foo\n", []string{"foo"})
+	f("foo\nbar", []string{"foo", "bar"})
+	f("\nfoo\nbar", []string{"\nfoo", "bar"})
+	f("\nfoo\nbar\n", []string{"\nfoo\nbar"})
+	f("\nfoo\nbar\n\n", []string{"\nfoo\nbar\n"})
+}
+
+func TestReadLinesBlockSuccessSingleByteReader(t *testing.T) {
+	f := func(s, dstBufExpected, tailBufExpected string) {
+		t.Helper()
+
+		r := &singleByteReader{
+			b: []byte(s),
+		}
+		dstBuf, tailBuf, err := ReadLinesBlock(r, nil, nil)
+		if err != nil {
+			t.Fatalf("unexpected error: %s", err)
+		}
+		if string(dstBuf) != dstBufExpected {
+			t.Fatalf("unexpected dstBuf; got %q; want %q; tailBuf=%q", dstBuf, dstBufExpected, tailBuf)
+		}
+		if string(tailBuf) != tailBufExpected {
+			t.Fatalf("unexpected tailBuf; got %q; want %q; dstBuf=%q", tailBuf, tailBufExpected, dstBuf)
+		}
+
+		// Verify the same with non-empty dstBuf and tailBuf
+		r = &singleByteReader{
+			b: []byte(s),
+		}
+		dstBuf, tailBuf, err = ReadLinesBlock(r, dstBuf, tailBuf[:0])
+		if err != nil {
+			t.Fatalf("non-empty bufs: unexpected error: %s", err)
+		}
+		if string(dstBuf) != dstBufExpected {
+			t.Fatalf("non-empty bufs: unexpected dstBuf; got %q; want %q; tailBuf=%q", dstBuf, dstBufExpected, tailBuf)
+		}
+		if string(tailBuf) != tailBufExpected {
+			t.Fatalf("non-empty bufs: unexpected tailBuf; got %q; want %q; dstBuf=%q", tailBuf, tailBufExpected, dstBuf)
+		}
+	}
+
+	f("\n", "", "")
+	f("foo\n", "foo", "")
+	f("\nfoo", "", "")
+	f("foo\nbar", "foo", "")
+	f("foo\nbar\nbaz", "foo", "")
+	f("foo", "foo", "")
+
+	// The maximum line size
+	b := make([]byte, maxLineSize+10)
+	b[maxLineSize] = '\n'
+	f(string(b), string(b[:maxLineSize]), "")
+}
+
+func TestReadLinesBlockSuccessBytesBuffer(t *testing.T) {
+	f := func(s, dstBufExpected, tailBufExpected string) {
+		t.Helper()
+
+		r := bytes.NewBufferString(s)
+		dstBuf, tailBuf, err := ReadLinesBlock(r, nil, nil)
+		if err != nil {
+			t.Fatalf("unexpected error: %s", err)
+		}
+		if string(dstBuf) != dstBufExpected {
+			t.Fatalf("unexpected dstBuf; got %q; want %q; tailBuf=%q", dstBuf, dstBufExpected, tailBuf)
+		}
+		if string(tailBuf) != tailBufExpected {
+			t.Fatalf("unexpected tailBuf; got %q; want %q; dstBuf=%q", tailBuf, tailBufExpected, dstBuf)
+		}
+
+		// Verify the same with non-empty dstBuf and tailBuf
+		r = bytes.NewBufferString(s)
+		dstBuf, tailBuf, err = ReadLinesBlock(r, dstBuf, tailBuf[:0])
+		if err != nil {
+			t.Fatalf("non-empty bufs: unexpected error: %s", err)
+		}
+		if string(dstBuf) != dstBufExpected {
+			t.Fatalf("non-empty bufs: unexpected dstBuf; got %q; want %q; tailBuf=%q", dstBuf, dstBufExpected, tailBuf)
+		}
+		if string(tailBuf) != tailBufExpected {
+			t.Fatalf("non-empty bufs: unexpected tailBuf; got %q; want %q; dstBuf=%q", tailBuf, tailBufExpected, dstBuf)
+		}
+	}
+
+	f("\n", "", "")
+	f("foo\n", "foo", "")
+	f("\nfoo", "", "foo")
+	f("foo\nbar", "foo", "bar")
+	f("foo\nbar\nbaz", "foo\nbar", "baz")
+
+	// The maximum line size
+	b := make([]byte, maxLineSize+10)
+	b[maxLineSize] = '\n'
+	f(string(b), string(b[:maxLineSize]), string(b[maxLineSize+1:]))
+}
+
+type singleByteReader struct {
+	b []byte
+}
+
+func (sbr *singleByteReader) Read(p []byte) (int, error) {
+	if len(sbr.b) == 0 {
+		return 0, io.EOF
+	}
+	n := copy(p, sbr.b[:1])
+	sbr.b = sbr.b[n:]
+	if len(sbr.b) == 0 {
+		return n, io.EOF
+	}
+	return n, nil
+}

app/vminsert/concurrencylimiter/concurrencylimiter.go

@@ -1,34 +1,70 @@
 package concurrencylimiter
 
 import (
+	"flag"
 	"fmt"
 	"runtime"
 	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
+	"github.com/VictoriaMetrics/metrics"
 )
 
+var maxConcurrentInserts = flag.Int("maxConcurrentInserts", runtime.GOMAXPROCS(-1)*4, "The maximum number of concurrent inserts")
+
 var (
-	// ch is the channel for limiting concurrent inserts.
-	// Put an item into it before performing an insert and remove
-	// the item after the insert is complete.
-	ch = make(chan struct{}, runtime.GOMAXPROCS(-1)*2)
+	// ch is the channel for limiting concurrent calls to Do.
+	ch chan struct{}
 
 	// waitDuration is the amount of time to wait until at least a single
-	// concurrent insert out of cap(Ch) inserts is complete.
+	// concurrent Do call out of cap(ch) inserts is complete.
 	waitDuration = time.Second * 30
 )
 
+// Init initializes concurrencylimiter.
+//
+// Init must be called after flag.Parse call.
+func Init() {
+	ch = make(chan struct{}, *maxConcurrentInserts)
+}
+
 // Do calls f with the limited concurrency.
 func Do(f func() error) error {
-	// Limit the number of conurrent inserts in order to prevent from excess
+	// Limit the number of conurrent f calls in order to prevent from excess
 	// memory usage and CPU trashing.
-	t := time.NewTimer(waitDuration)
 	select {
 	case ch <- struct{}{}:
-		t.Stop()
+		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{}{}:
+		timerpool.Put(t)
 		err := f()
 		<-ch
 		return err
 	case <-t.C:
-		return fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase the number of CPUs or reduce the load", cap(ch))
+		timerpool.Put(t)
+		concurrencyLimitTimeout.Inc()
+		return fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase -maxConcurrentInserts or reduce the load", cap(ch))
 	}
 }
+
+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

@@ -86,7 +86,9 @@ func (r *Row) unmarshal(s string, tagsPool []Tag) ([]Tag, error) {
 
 	n = strings.IndexByte(tail, ' ')
 	if n < 0 {
-		return tagsPool, fmt.Errorf("cannot find whitespace between value and timestamp in %q", s)
+		// There is no timestamp. Use default timestamp instead.
+		r.Value = fastfloat.ParseBestEffort(tail)
+		return tagsPool, nil
 	}
 	r.Value = fastfloat.ParseBestEffort(tail[:n])
 	r.Timestamp = fastfloat.ParseInt64BestEffort(tail[n+1:])
@@ -112,6 +114,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
 			var err error
 			tagsPool, err = r.unmarshal(s, tagsPool)
 			if err != nil {
+				err = fmt.Errorf("cannot unmarshal Graphite line %q: %s", s, err)
 				return dst, tagsPool, err
 			}
 			return dst, tagsPool, nil
@@ -119,6 +122,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
 		var err error
 		tagsPool, err = r.unmarshal(s[:n], tagsPool)
 		if err != nil {
+			err = fmt.Errorf("cannot unmarshal Graphite line %q: %s", s[:n], err)
 			return dst, tagsPool, err
 		}
 		s = s[n+1:]

app/vminsert/graphite/parser_test.go

@@ -22,9 +22,6 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	// Missing value
 	f("aaa")
 
-	// Missing timestamp
-	f("aaa 1123")
-
 	// Invalid multiline
 	f("aaa\nbbb 123 34")
 
@@ -81,6 +78,14 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 		}},
 	})
 
+	// Missing timestamp
+	f("aaa 1123", &Rows{
+		Rows: []Row{{
+			Metric: "aaa",
+			Value:  1123,
+		}},
+	})
+
 	// Tags
 	f("foo;bar=baz 1 2", &Rows{
 		Rows: []Row{{
@@ -116,13 +121,17 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Multi lines
-	f("foo 0.3 2\nbar.baz 0.34 43\n", &Rows{
+	f("foo 0.3 2\naaa 3\nbar.baz 0.34 43\n", &Rows{
 		Rows: []Row{
 			{
 				Metric:    "foo",
 				Value:     0.3,
 				Timestamp: 2,
 			},
+			{
+				Metric: "aaa",
+				Value:  3,
+			},
 			{
 				Metric:    "bar.baz",
 				Value:     0.34,

app/vminsert/graphite/request_handler.go

@@ -1,7 +1,6 @@
 package graphite
 
 import (
-	"bytes"
 	"fmt"
 	"io"
 	"net"
@@ -15,7 +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.
 //
@@ -52,11 +54,10 @@ 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()
 }
 
-const maxReadPacketSize = 4 * 1024 * 1024
-
 const flushTimeout = 3 * time.Second
 
 func (ctx *pushCtx) Read(r io.Reader) bool {
@@ -71,40 +72,40 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			return false
 		}
 	}
-	lr := io.LimitReader(r, maxReadPacketSize)
-	ctx.reqBuf.Reset()
-	ctx.reqBuf.B = append(ctx.reqBuf.B[:0], ctx.tailBuf...)
-	n, err := io.CopyBuffer(&ctx.reqBuf, lr, ctx.copyBuf[:])
-	if err != nil {
-		if ne, ok := err.(net.Error); ok && ne.Timeout() {
+	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
+	if ctx.err != nil {
+		if ne, ok := ctx.err.(net.Error); ok && ne.Timeout() {
 			// Flush the read data on timeout and try reading again.
+			ctx.err = nil
 		} else {
-			graphiteReadErrors.Inc()
-			ctx.err = fmt.Errorf("cannot read graphite plaintext protocol data: %s", err)
+			if ctx.err != io.EOF {
+				graphiteReadErrors.Inc()
+				ctx.err = fmt.Errorf("cannot read graphite plaintext protocol data: %s", ctx.err)
+			}
 			return false
 		}
-	} else if n < maxReadPacketSize {
-		// Mark the end of stream.
-		ctx.err = io.EOF
 	}
-
-	// Parse all the rows until the last newline in ctx.reqBuf.B
-	nn := bytes.LastIndexByte(ctx.reqBuf.B, '\n')
-	ctx.tailBuf = ctx.tailBuf[:0]
-	if nn >= 0 {
-		ctx.tailBuf = append(ctx.tailBuf[:0], ctx.reqBuf.B[nn+1:]...)
-		ctx.reqBuf.B = ctx.reqBuf.B[:nn]
-	}
-	if err = ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf.B)); err != nil {
+	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.B), err)
+		ctx.err = fmt.Errorf("cannot unmarshal graphite plaintext protocol data with size %d: %s", len(ctx.reqBuf), err)
 		return false
 	}
 
-	// Convert timestamps from seconds to milliseconds
-	for i := range ctx.Rows.Rows {
-		ctx.Rows.Rows[i].Timestamp *= 1e3
+	// Fill missing timestamps with the current timestamp rounded to seconds.
+	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 rows {
+		rows[i].Timestamp *= 1e3
+	}
+
 	return true
 }
 
@@ -112,9 +113,8 @@ type pushCtx struct {
 	Rows   Rows
 	Common common.InsertCtx
 
-	reqBuf  bytesutil.ByteBuffer
+	reqBuf  []byte
 	tailBuf []byte
-	copyBuf [16 * 1024]byte
 
 	err error
 }
@@ -129,7 +129,7 @@ func (ctx *pushCtx) Error() error {
 func (ctx *pushCtx) reset() {
 	ctx.Rows.Reset()
 	ctx.Common.Reset(0)
-	ctx.reqBuf.Reset()
+	ctx.reqBuf = ctx.reqBuf[:0]
 	ctx.tailBuf = ctx.tailBuf[:0]
 
 	ctx.err = nil

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

@@ -196,6 +196,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]R
 			var err error
 			tagsPool, fieldsPool, err = r.unmarshal(s, tagsPool, fieldsPool)
 			if err != nil {
+				err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s, err)
 				return dst, tagsPool, fieldsPool, err
 			}
 			return dst, tagsPool, fieldsPool, nil
@@ -203,6 +204,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]R
 		var err error
 		tagsPool, fieldsPool, err = r.unmarshal(s[:n], tagsPool, fieldsPool)
 		if err != nil {
+			err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s[:n], err)
 			return dst, tagsPool, fieldsPool, err
 		}
 		s = s[n+1:]
@@ -293,8 +295,10 @@ func parseFieldValue(s string, hasQuotedFields bool) (float64, error) {
 		if len(s) < 2 || s[len(s)-1] != '"' {
 			return 0, fmt.Errorf("missing closing quote for quoted field value %s", s)
 		}
-		// Quoted string is translated to empty value.
-		return 0, nil
+		// Try converting quoted string to number, since sometimes Influx agents
+		// send numbers as strings.
+		s = s[1 : len(s)-1]
+		return fastfloat.ParseBestEffort(s), nil
 	}
 	ch := s[len(s)-1]
 	if ch == 'i' {

app/vminsert/influx/parser_test.go

@@ -241,17 +241,27 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Line with multiple tags, multiple fields and timestamp
-	f(`system,host=ip-172-16-10-144 uptime_format="3 days, 21:01" 1557761040000000000`, &Rows{
+	f(`system,host=ip-172-16-10-144 uptime_format="3 days, 21:01",quoted_float="-1.23",quoted_int="123" 1557761040000000000`, &Rows{
 		Rows: []Row{{
 			Measurement: "system",
 			Tags: []Tag{{
 				Key:   "host",
 				Value: "ip-172-16-10-144",
 			}},
-			Fields: []Field{{
-				Key:   "uptime_format",
-				Value: 0,
-			}},
+			Fields: []Field{
+				{
+					Key:   "uptime_format",
+					Value: 0,
+				},
+				{
+					Key:   "quoted_float",
+					Value: -1.23,
+				},
+				{
+					Key:   "quoted_int",
+					Value: 123,
+				},
+			},
 			Timestamp: 1557761040000000000,
 		}},
 	})
@@ -338,4 +348,30 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+	// No newline after the second line.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/82
+	f("foo,tag=xyz field=1.23 48934\n"+
+		"bar x=-1i", &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,
+				}},
+			},
+		},
+	})
 }

app/vminsert/influx/request_handler.go

@@ -1,8 +1,8 @@
 package influx
 
 import (
-	"bytes"
 	"compress/gzip"
+	"flag"
 	"fmt"
 	"io"
 	"net/http"
@@ -17,7 +17,15 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="influx"}`)
+var (
+	measurementFieldSeparator = flag.String("influxMeasurementFieldSeparator", "_", "Separator for `{measurement}{separator}{field_name}` metric name when inserted via Influx line protocol")
+	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"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="influx"}`)
+)
 
 // InsertHandler processes remote write for influx line protocol.
 //
@@ -79,6 +87,7 @@ 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]
@@ -89,18 +98,25 @@ func (ctx *pushCtx) InsertRows(db string) error {
 		}
 		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
 		ctx.metricGroupBuf = append(ctx.metricGroupBuf[:0], r.Measurement...)
-		ctx.metricGroupBuf = append(ctx.metricGroupBuf, '.')
+		skipFieldKey := len(r.Fields) == 1 && *skipSingleField
+		if !skipFieldKey {
+			ctx.metricGroupBuf = append(ctx.metricGroupBuf, *measurementFieldSeparator...)
+		}
 		metricGroupPrefixLen := len(ctx.metricGroupBuf)
 		for j := range r.Fields {
 			f := &r.Fields[j]
-			ctx.metricGroupBuf = append(ctx.metricGroupBuf[:metricGroupPrefixLen], f.Key...)
+			if !skipFieldKey {
+				ctx.metricGroupBuf = append(ctx.metricGroupBuf[:metricGroupPrefixLen], f.Key...)
+			}
 			metricGroup := bytesutil.ToUnsafeString(ctx.metricGroupBuf)
 			ic.Labels = ic.Labels[:0]
 			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()
 }
 
@@ -123,36 +139,21 @@ func putGzipReader(zr *gzip.Reader) {
 
 var gzipReaderPool sync.Pool
 
-const maxReadPacketSize = 4 * 1024 * 1024
-
 func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 	if ctx.err != nil {
 		return false
 	}
-	lr := io.LimitReader(r, maxReadPacketSize)
-	ctx.reqBuf.Reset()
-	ctx.reqBuf.B = append(ctx.reqBuf.B[:0], ctx.tailBuf...)
-	n, err := io.CopyBuffer(&ctx.reqBuf, lr, ctx.copyBuf[:])
-	if err != nil {
-		influxReadErrors.Inc()
-		ctx.err = fmt.Errorf("cannot read influx line protocol data: %s", err)
+	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
+	if ctx.err != nil {
+		if ctx.err != io.EOF {
+			influxReadErrors.Inc()
+			ctx.err = fmt.Errorf("cannot read influx line protocol data: %s", ctx.err)
+		}
 		return false
 	}
-	if n < maxReadPacketSize {
-		// Mark the end of stream.
-		ctx.err = io.EOF
-	}
-
-	// Parse all the rows until the last newline in ctx.reqBuf.B
-	nn := bytes.LastIndexByte(ctx.reqBuf.B, '\n')
-	ctx.tailBuf = ctx.tailBuf[:0]
-	if nn >= 0 {
-		ctx.tailBuf = append(ctx.tailBuf[:0], ctx.reqBuf.B[nn+1:]...)
-		ctx.reqBuf.B = ctx.reqBuf.B[:nn]
-	}
-	if err = ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf.B)); err != nil {
+	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.B), err)
+		ctx.err = fmt.Errorf("cannot unmarshal influx line protocol data with size %d: %s", len(ctx.reqBuf), err)
 		return false
 	}
 
@@ -169,6 +170,7 @@ 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 {
@@ -191,9 +193,8 @@ type pushCtx struct {
 	Rows   Rows
 	Common common.InsertCtx
 
-	reqBuf         bytesutil.ByteBuffer
+	reqBuf         []byte
 	tailBuf        []byte
-	copyBuf        [16 * 1024]byte
 	metricNameBuf  []byte
 	metricGroupBuf []byte
 
@@ -211,7 +212,7 @@ func (ctx *pushCtx) reset() {
 	ctx.Rows.Reset()
 	ctx.Common.Reset(0)
 
-	ctx.reqBuf.Reset()
+	ctx.reqBuf = ctx.reqBuf[:0]
 	ctx.tailBuf = ctx.tailBuf[:0]
 	ctx.metricNameBuf = ctx.metricNameBuf[:0]
 	ctx.metricGroupBuf = ctx.metricGroupBuf[:0]

app/vminsert/main.go

@@ -6,6 +6,7 @@ import (
 	"net/http"
 	"strings"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/graphite"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/influx"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdb"
@@ -22,6 +23,7 @@ var (
 
 // Init initializes vminsert.
 func Init() {
+	concurrencylimiter.Init()
 	if len(*graphiteListenAddr) > 0 {
 		go graphite.Serve(*graphiteListenAddr)
 	}
@@ -63,7 +65,8 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		w.WriteHeader(http.StatusNoContent)
 		return true
 	case "/query":
-		// Emulate fake response for influx query
+		// Emulate fake response for influx query.
+		// This is required for TSBS benchmark.
 		influxQueryRequests.Inc()
 		fmt.Fprintf(w, `{"results":[{"series":[{"values":[]}]}]}`)
 		return true

app/vminsert/opentsdb/parser.go

@@ -111,6 +111,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
 			var err error
 			tagsPool, err = r.unmarshal(s, tagsPool)
 			if err != nil {
+				err = fmt.Errorf("cannot unmarshal OpenTSDB line %q: %s", s, err)
 				return dst, tagsPool, err
 			}
 			return dst, tagsPool, nil
@@ -118,6 +119,7 @@ func unmarshalRows(dst []Row, s string, tagsPool []Tag) ([]Row, []Tag, error) {
 		var err error
 		tagsPool, err = r.unmarshal(s[:n], tagsPool)
 		if err != nil {
+			err = fmt.Errorf("cannot unmarshal OpenTSDB line %q: %s", s[:n], err)
 			return dst, tagsPool, err
 		}
 		s = s[n+1:]

app/vminsert/opentsdb/request_handler.go

@@ -1,7 +1,6 @@
 package opentsdb
 
 import (
-	"bytes"
 	"fmt"
 	"io"
 	"net"
@@ -15,7 +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.
 //
@@ -52,11 +54,10 @@ 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()
 }
 
-const maxReadPacketSize = 4 * 1024 * 1024
-
 const flushTimeout = 3 * time.Second
 
 func (ctx *pushCtx) Read(r io.Reader) bool {
@@ -71,33 +72,22 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			return false
 		}
 	}
-	lr := io.LimitReader(r, maxReadPacketSize)
-	ctx.reqBuf.Reset()
-	ctx.reqBuf.B = append(ctx.reqBuf.B[:0], ctx.tailBuf...)
-	n, err := io.CopyBuffer(&ctx.reqBuf, lr, ctx.copyBuf[:])
-	if err != nil {
-		if ne, ok := err.(net.Error); ok && ne.Timeout() {
+	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
+	if ctx.err != nil {
+		if ne, ok := ctx.err.(net.Error); ok && ne.Timeout() {
 			// Flush the read data on timeout and try reading again.
+			ctx.err = nil
 		} else {
-			opentsdbReadErrors.Inc()
-			ctx.err = fmt.Errorf("cannot read OpenTSDB put protocol data: %s", err)
+			if ctx.err != io.EOF {
+				opentsdbReadErrors.Inc()
+				ctx.err = fmt.Errorf("cannot read OpenTSDB put protocol data: %s", ctx.err)
+			}
 			return false
 		}
-	} else if n < maxReadPacketSize {
-		// Mark the end of stream.
-		ctx.err = io.EOF
 	}
-
-	// Parse all the rows until the last newline in ctx.reqBuf.B
-	nn := bytes.LastIndexByte(ctx.reqBuf.B, '\n')
-	ctx.tailBuf = ctx.tailBuf[:0]
-	if nn >= 0 {
-		ctx.tailBuf = append(ctx.tailBuf[:0], ctx.reqBuf.B[nn+1:]...)
-		ctx.reqBuf.B = ctx.reqBuf.B[:nn]
-	}
-	if err = ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf.B)); err != nil {
+	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.B), err)
+		ctx.err = fmt.Errorf("cannot unmarshal OpenTSDB put protocol data with size %d: %s", len(ctx.reqBuf), err)
 		return false
 	}
 
@@ -112,9 +102,8 @@ type pushCtx struct {
 	Rows   Rows
 	Common common.InsertCtx
 
-	reqBuf  bytesutil.ByteBuffer
+	reqBuf  []byte
 	tailBuf []byte
-	copyBuf [16 * 1024]byte
 
 	err error
 }
@@ -129,7 +118,7 @@ func (ctx *pushCtx) Error() error {
 func (ctx *pushCtx) reset() {
 	ctx.Rows.Reset()
 	ctx.Common.Reset(0)
-	ctx.reqBuf.Reset()
+	ctx.reqBuf = ctx.reqBuf[:0]
 	ctx.tailBuf = ctx.tailBuf[:0]
 
 	ctx.err = nil

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/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

@@ -14,6 +14,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
 	"github.com/VictoriaMetrics/metrics"
 )
 
@@ -29,28 +30,49 @@ 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 second until giving up. This should resolve short bursts in requests.
-	t := time.NewTimer(*maxQueueDuration)
 	select {
 	case concurrencyCh <- struct{}{}:
-		t.Stop()
 		defer func() { <-concurrencyCh }()
-	case <-t.C:
-		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()
+			httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
+			return true
+		}
 	}
 
 	path := strings.Replace(r.URL.Path, "//", "/", -1)
@@ -115,6 +137,15 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 			return true
 		}
 		return true
+	case "/api/v1/labels/count":
+		labelsCountRequests.Inc()
+		httpserver.EnableCORS(w, r)
+		if err := prometheus.LabelsCountHandler(w, r); err != nil {
+			labelsCountErrors.Inc()
+			sendPrometheusError(w, r, err)
+			return true
+		}
+		return true
 	case "/api/v1/export":
 		exportRequests.Inc()
 		if err := prometheus.ExportHandler(w, r); err != nil {
@@ -178,6 +209,9 @@ var (
 	labelsRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/labels"}`)
 	labelsErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/labels"}`)
 
+	labelsCountRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/labels/count"}`)
+	labelsCountErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/labels/count"}`)
+
 	deleteRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/admin/tsdb/delete_series"}`)
 	deleteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/admin/tsdb/delete_series"}`)
 

app/vmselect/netstorage/netstorage.go

@@ -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()
 
@@ -400,6 +404,36 @@ func GetLabelValues(labelName string, deadline Deadline) ([]string, error) {
 	return labelValues, nil
 }
 
+// GetLabelEntries returns all the label entries until the given deadline.
+func GetLabelEntries(deadline Deadline) ([]storage.TagEntry, error) {
+	labelEntries, err := vmstorage.SearchTagEntries(*maxTagKeysPerSearch, *maxTagValuesPerSearch)
+	if err != nil {
+		return nil, fmt.Errorf("error during label entries request: %s", err)
+	}
+
+	// Substitute "" with "__name__"
+	for i := range labelEntries {
+		e := &labelEntries[i]
+		if e.Key == "" {
+			e.Key = "__name__"
+		}
+	}
+
+	// Sort labelEntries by the number of label values in each entry.
+	sort.Slice(labelEntries, func(i, j int) bool {
+		a, b := labelEntries[i].Values, labelEntries[j].Values
+		if len(a) < len(b) {
+			return true
+		}
+		if len(a) > len(b) {
+			return false
+		}
+		return labelEntries[i].Key < labelEntries[j].Key
+	})
+
+	return labelEntries, nil
+}
+
 // GetSeriesCount returns the number of unique series.
 func GetSeriesCount(deadline Deadline) (uint64, error) {
 	n, err := vmstorage.GetSeriesCount()
@@ -429,7 +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 {
@@ -445,35 +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

@@ -22,9 +22,7 @@ func InitTmpBlocksDir(tmpDirPath string) {
 		tmpDirPath = os.TempDir()
 	}
 	tmpBlocksDir = tmpDirPath + "/searchResults"
-	if err := os.RemoveAll(tmpBlocksDir); err != nil {
-		logger.Panicf("FATAL: cannot remove %q: %s", tmpBlocksDir, err)
-	}
+	fs.MustRemoveAll(tmpBlocksDir)
 	if err := fs.MkdirAllIfNotExist(tmpBlocksDir); err != nil {
 		logger.Panicf("FATAL: cannot create %q: %s", tmpBlocksDir, err)
 	}

app/vmselect/prometheus/labels_count_response.qtpl

@@ -0,0 +1,17 @@
+{% import "github.com/VictoriaMetrics/VictoriaMetrics/lib/storage" %}
+
+{% stripspace %}
+LabelsCountResponse generates response for /api/v1/labels/count .
+{% func LabelsCountResponse(labelEntries []storage.TagEntry) %}
+{
+	"status":"success",
+	"data":{
+		{% for i, e := range labelEntries %}
+			{%q= e.Key %}:{%d= len(e.Values) %}
+			{% if i+1 < len(labelEntries) %},{% endif %}
+		{% endfor %}
+	}
+}
+{% endfunc %}
+
+{% endstripspace %}

app/vmselect/prometheus/labels_count_response.qtpl.go

@@ -0,0 +1,74 @@
+// Code generated by qtc from "labels_count_response.qtpl". DO NOT EDIT.
+// See https://github.com/valyala/quicktemplate for details.
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:1
+package prometheus
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:1
+import "github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+
+// LabelsCountResponse generates response for /api/v1/labels/count .
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:5
+import (
+	qtio422016 "io"
+
+	qt422016 "github.com/valyala/quicktemplate"
+)
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:5
+var (
+	_ = qtio422016.Copy
+	_ = qt422016.AcquireByteBuffer
+)
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:5
+func StreamLabelsCountResponse(qw422016 *qt422016.Writer, labelEntries []storage.TagEntry) {
+//line app/vmselect/prometheus/labels_count_response.qtpl:5
+	qw422016.N().S(`{"status":"success","data":{`)
+//line app/vmselect/prometheus/labels_count_response.qtpl:9
+	for i, e := range labelEntries {
+//line app/vmselect/prometheus/labels_count_response.qtpl:10
+		qw422016.N().Q(e.Key)
+//line app/vmselect/prometheus/labels_count_response.qtpl:10
+		qw422016.N().S(`:`)
+//line app/vmselect/prometheus/labels_count_response.qtpl:10
+		qw422016.N().D(len(e.Values))
+//line app/vmselect/prometheus/labels_count_response.qtpl:11
+		if i+1 < len(labelEntries) {
+//line app/vmselect/prometheus/labels_count_response.qtpl:11
+			qw422016.N().S(`,`)
+//line app/vmselect/prometheus/labels_count_response.qtpl:11
+		}
+//line app/vmselect/prometheus/labels_count_response.qtpl:12
+	}
+//line app/vmselect/prometheus/labels_count_response.qtpl:12
+	qw422016.N().S(`}}`)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+}
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+func WriteLabelsCountResponse(qq422016 qtio422016.Writer, labelEntries []storage.TagEntry) {
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	qw422016 := qt422016.AcquireWriter(qq422016)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	StreamLabelsCountResponse(qw422016, labelEntries)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	qt422016.ReleaseWriter(qw422016)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+}
+
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+func LabelsCountResponse(labelEntries []storage.TagEntry) string {
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	qb422016 := qt422016.AcquireByteBuffer()
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	WriteLabelsCountResponse(qb422016, labelEntries)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	qs422016 := string(qb422016.B)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	qt422016.ReleaseByteBuffer(qb422016)
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+	return qs422016
+//line app/vmselect/prometheus/labels_count_response.qtpl:15
+}

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"
@@ -37,9 +40,21 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 		return fmt.Errorf("cannot parse request form values: %s", err)
 	}
 	matches := r.Form["match[]"]
-	maxLookback := getDuration(r, "max_lookback", defaultStep)
-	start := getTime(r, "start", ct-maxLookback)
-	end := getTime(r, "end", ct)
+	if len(matches) == 0 {
+		return fmt.Errorf("missing `match[]` arg")
+	}
+	maxLookback, err := getDuration(r, "max_lookback", defaultStep)
+	if err != nil {
+		return err
+	}
+	start, err := getTime(r, "start", ct-maxLookback)
+	if err != nil {
+		return err
+	}
+	end, err := getTime(r, "end", ct)
+	if err != nil {
+		return err
+	}
 	deadline := getDeadline(r)
 	if start >= end {
 		start = end - defaultStep
@@ -53,7 +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)
 	}
@@ -61,7 +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
@@ -97,10 +112,19 @@ func ExportHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		// Maintain backwards compatibility
 		match := r.FormValue("match")
+		if len(match) == 0 {
+			return fmt.Errorf("missing `match[]` arg")
+		}
 		matches = []string{match}
 	}
-	start := getTime(r, "start", 0)
-	end := getTime(r, "end", ct)
+	start, err := getTime(r, "start", 0)
+	if err != nil {
+		return err
+	}
+	end, err := getTime(r, "end", ct)
+	if err != nil {
+		return err
+	}
 	format := r.FormValue("format")
 	deadline := getDeadline(r)
 	if start >= end {
@@ -136,7 +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)
 	}
@@ -144,7 +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
@@ -156,6 +180,11 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 	w.Header().Set("Content-Type", contentType)
 	writeResponseFunc(w, resultsCh)
 
+	// Consume all the data from resultsCh in the event writeResponseFunc
+	// fails to consume all the data.
+	for bb := range resultsCh {
+		quicktemplate.ReleaseByteBuffer(bb)
+	}
 	err = <-doneCh
 	if err != nil {
 		return fmt.Errorf("error during data fetching: %s", err)
@@ -175,6 +204,9 @@ func DeleteHandler(r *http.Request) error {
 		return fmt.Errorf("start and end aren't supported. Remove these args from the query in order to delete all the matching metrics")
 	}
 	matches := r.Form["match[]"]
+	if len(matches) == 0 {
+		return fmt.Errorf("missing `match[]` arg")
+	}
 	tagFilterss, err := getTagFilterssFromMatches(matches)
 	if err != nil {
 		return err
@@ -201,9 +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")
@@ -212,8 +274,69 @@ 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.
+func LabelsCountHandler(w http.ResponseWriter, r *http.Request) error {
+	startTime := time.Now()
+	deadline := getDeadline(r)
+	labelEntries, err := netstorage.GetLabelEntries(deadline)
+	if err != nil {
+		return fmt.Errorf(`cannot obtain label entries: %s`, err)
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	WriteLabelsCountResponse(w, labelEntries)
+	labelsCountDuration.UpdateDuration(startTime)
+	return nil
+}
+
+var labelsCountDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/labels/count"}`)
+
 // LabelsHandler processes /api/v1/labels request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#getting-label-names
@@ -260,8 +383,22 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 		return fmt.Errorf("cannot parse form values: %s", err)
 	}
 	matches := r.Form["match[]"]
-	start := getTime(r, "start", ct-defaultStep)
-	end := getTime(r, "end", ct)
+	if len(matches) == 0 {
+		return fmt.Errorf("missing `match[]` arg")
+	}
+	end, err := getTime(r, "end", ct)
+	if err != nil {
+		return err
+	}
+	// 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
+	}
 	deadline := getDeadline(r)
 
 	tagFilterss, err := getTagFilterssFromMatches(matches)
@@ -276,7 +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)
 	}
@@ -284,7 +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
@@ -297,11 +434,10 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	WriteSeriesResponse(w, resultsCh)
 
 	// Consume all the data from resultsCh in the event WriteSeriesResponse
-	// fail to consume all the data.
+	// fails to consume all the data.
 	for bb := range resultsCh {
 		quicktemplate.ReleaseByteBuffer(bb)
 	}
-
 	err = <-doneCh
 	if err != nil {
 		return fmt.Errorf("error during data fetching: %s", err)
@@ -320,8 +456,17 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 
 	query := r.FormValue("query")
-	start := getTime(r, "time", ct)
-	step := getDuration(r, "step", latencyOffset)
+	if len(query) == 0 {
+		return fmt.Errorf("missing `query` arg")
+	}
+	start, err := getTime(r, "time", ct)
+	if err != nil {
+		return err
+	}
+	step, err := getDuration(r, "step", latencyOffset)
+	if err != nil {
+		return err
+	}
 	deadline := getDeadline(r)
 
 	if len(query) > *maxQueryLen {
@@ -363,7 +508,7 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 		Step:     step,
 		Deadline: deadline,
 	}
-	result, err := promql.Exec(&ec, query)
+	result, err := promql.Exec(&ec, query, true)
 	if err != nil {
 		return fmt.Errorf("cannot execute %q: %s", query, err)
 	}
@@ -384,9 +529,21 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 
 	query := r.FormValue("query")
-	start := getTime(r, "start", ct-defaultStep)
-	end := getTime(r, "end", ct)
-	step := getDuration(r, "step", defaultStep)
+	if len(query) == 0 {
+		return fmt.Errorf("missing `query` arg")
+	}
+	start, err := getTime(r, "start", ct-defaultStep)
+	if err != nil {
+		return err
+	}
+	end, err := getTime(r, "end", ct)
+	if err != nil {
+		return err
+	}
+	step, err := getDuration(r, "step", defaultStep)
+	if err != nil {
+		return err
+	}
 	deadline := getDeadline(r)
 	mayCache := !getBool(r, "nocache")
 
@@ -409,12 +566,12 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 		Deadline: deadline,
 		MayCache: mayCache,
 	}
-	result, err := promql.Exec(&ec, query)
+	result, err := promql.Exec(&ec, query, false)
 	if err != nil {
 		return fmt.Errorf("cannot execute %q: %s", query, err)
 	}
 	if ct-end < latencyOffset {
-		adjustLastPoints(result)
+		result = adjustLastPoints(result)
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -427,17 +584,17 @@ var queryRangeDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/
 
 // adjustLastPoints substitutes the last point values with the previous
 // point values, since the last points may contain garbage.
-func adjustLastPoints(tss []netstorage.Result) {
+func adjustLastPoints(tss []netstorage.Result) []netstorage.Result {
 	if len(tss) == 0 {
-		return
+		return nil
 	}
 
 	// Search for the last non-NaN value across all the timeseries.
 	lastNonNaNIdx := -1
 	for i := range tss {
-		r := &tss[i]
-		j := len(r.Values) - 1
-		for j >= 0 && math.IsNaN(r.Values[j]) {
+		values := tss[i].Values
+		j := len(values) - 1
+		for j >= 0 && math.IsNaN(values[j]) {
 			j--
 		}
 		if j > lastNonNaNIdx {
@@ -446,75 +603,97 @@ func adjustLastPoints(tss []netstorage.Result) {
 	}
 	if lastNonNaNIdx == -1 {
 		// All timeseries contain only NaNs.
-		return
+		return nil
 	}
 
-	// Substitute last three values starting from lastNonNaNIdx
+	// Substitute the last two values starting from lastNonNaNIdx
 	// with the previous values for each timeseries.
 	for i := range tss {
-		r := &tss[i]
-		for j := 0; j < 3; j++ {
+		values := tss[i].Values
+		for j := 0; j < 2; j++ {
 			idx := lastNonNaNIdx + j
-			if idx <= 0 || idx >= len(r.Values) {
+			if idx <= 0 || idx >= len(values) || math.IsNaN(values[idx-1]) {
 				continue
 			}
-			r.Values[idx] = r.Values[idx-1]
+			values[idx] = values[idx-1]
 		}
 	}
+	return tss
 }
 
-func getTime(r *http.Request, argKey string, defaultValue int64) int64 {
+func getTime(r *http.Request, argKey string, defaultValue int64) (int64, error) {
 	argValue := r.FormValue(argKey)
 	if len(argValue) == 0 {
-		return defaultValue
+		return defaultValue, nil
 	}
 	secs, err := strconv.ParseFloat(argValue, 64)
 	if err != nil {
 		// Try parsing string format
 		t, err := time.Parse(time.RFC3339, argValue)
 		if err != nil {
-			return defaultValue
+			// Handle Prometheus'-provided minTime and maxTime.
+			// See https://github.com/prometheus/client_golang/issues/614
+			switch argValue {
+			case prometheusMinTimeFormatted:
+				return minTimeMsecs, nil
+			case prometheusMaxTimeFormatted:
+				return maxTimeMsecs, nil
+			}
+			return 0, fmt.Errorf("cannot parse %q=%q: %s", argKey, argValue, err)
 		}
 		secs = float64(t.UnixNano()) / 1e9
 	}
 	msecs := int64(secs * 1e3)
-	if msecs < minTimeMsecs || msecs > maxTimeMsecs {
-		return defaultValue
+	if msecs < minTimeMsecs {
+		msecs = 0
 	}
-	return msecs
+	if msecs > maxTimeMsecs {
+		msecs = maxTimeMsecs
+	}
+	return msecs, nil
 }
 
+var (
+	// These constants were obtained from https://github.com/prometheus/prometheus/blob/91d7175eaac18b00e370965f3a8186cc40bf9f55/web/api/v1/api.go#L442
+	// See https://github.com/prometheus/client_golang/issues/614 for details.
+	prometheusMinTimeFormatted = time.Unix(math.MinInt64/1000+62135596801, 0).UTC().Format(time.RFC3339Nano)
+	prometheusMaxTimeFormatted = time.Unix(math.MaxInt64/1000-62135596801, 999999999).UTC().Format(time.RFC3339Nano)
+)
+
 const (
 	// These values prevent from overflow when storing msec-precision time in int64.
-	minTimeMsecs = int64(-1<<63) / 1e6
+	minTimeMsecs = 0 // use 0 instead of `int64(-1<<63) / 1e6` because the storage engine doesn't actually support negative time
 	maxTimeMsecs = int64(1<<63-1) / 1e6
 )
 
-func getDuration(r *http.Request, argKey string, defaultValue int64) int64 {
+func getDuration(r *http.Request, argKey string, defaultValue int64) (int64, error) {
 	argValue := r.FormValue(argKey)
 	if len(argValue) == 0 {
-		return defaultValue
+		return defaultValue, nil
 	}
 	secs, err := strconv.ParseFloat(argValue, 64)
 	if err != nil {
 		// Try parsing string format
 		d, err := time.ParseDuration(argValue)
 		if err != nil {
-			return defaultValue
+			return 0, fmt.Errorf("cannot parse %q=%q: %s", argKey, argValue, err)
 		}
 		secs = d.Seconds()
 	}
 	msecs := int64(secs * 1e3)
 	if msecs <= 0 || msecs > maxDurationMsecs {
-		return defaultValue
+		return 0, fmt.Errorf("%q=%dms is out of allowed range [%d ... %d]", argKey, msecs, 0, int64(maxDurationMsecs))
 	}
-	return msecs
+	return msecs, nil
 }
 
 const maxDurationMsecs = 100 * 365 * 24 * 3600 * 1000
 
 func getDeadline(r *http.Request) netstorage.Deadline {
-	d := getDuration(r, "timeout", 0)
+	d, err := getDuration(r, "timeout", 0)
+	if err != nil {
+		d = 0
+	}
 	dMax := int64(maxQueryDuration.Seconds() * 1e3)
 	if d <= 0 || d > dMax {
 		d = dMax

app/vmselect/prometheus/prometheus_test.go

@@ -0,0 +1,78 @@
+package prometheus
+
+import (
+	"fmt"
+	"net/http"
+	"net/url"
+	"testing"
+)
+
+func TestGetTimeSuccess(t *testing.T) {
+	f := func(s string, timestampExpected int64) {
+		t.Helper()
+		urlStr := fmt.Sprintf("http://foo.bar/baz?s=%s", url.QueryEscape(s))
+		r, err := http.NewRequest("GET", urlStr, nil)
+		if err != nil {
+			t.Fatalf("unexpected error in NewRequest: %s", err)
+		}
+
+		// Verify defaultValue
+		ts, err := getTime(r, "foo", 123)
+		if err != nil {
+			t.Fatalf("unexpected error when obtaining default time from getTime(%q): %s", s, err)
+		}
+		if ts != 123 {
+			t.Fatalf("unexpected default value for getTime(%q); got %d; want %d", s, ts, 123)
+		}
+
+		// Verify timestampExpected
+		ts, err = getTime(r, "s", 123)
+		if err != nil {
+			t.Fatalf("unexpected error in getTime(%q): %s", s, err)
+		}
+		if ts != timestampExpected {
+			t.Fatalf("unexpected timestamp for getTime(%q); got %d; want %d", s, ts, timestampExpected)
+		}
+	}
+
+	f("2019-07-07T20:01:02Z", 1562529662000)
+	f("2019-07-07T20:47:40+03:00", 1562521660000)
+	f("-292273086-05-16T16:47:06Z", minTimeMsecs)
+	f("292277025-08-18T07:12:54.999999999Z", maxTimeMsecs)
+	f("1562529662.324", 1562529662324)
+	f("-9223372036.854", minTimeMsecs)
+	f("-9223372036.855", minTimeMsecs)
+	f("9223372036.855", maxTimeMsecs)
+}
+
+func TestGetTimeError(t *testing.T) {
+	f := func(s string) {
+		t.Helper()
+		urlStr := fmt.Sprintf("http://foo.bar/baz?s=%s", url.QueryEscape(s))
+		r, err := http.NewRequest("GET", urlStr, nil)
+		if err != nil {
+			t.Fatalf("unexpected error in NewRequest: %s", err)
+		}
+
+		// Verify defaultValue
+		ts, err := getTime(r, "foo", 123)
+		if err != nil {
+			t.Fatalf("unexpected error when obtaining default time from getTime(%q): %s", s, err)
+		}
+		if ts != 123 {
+			t.Fatalf("unexpected default value for getTime(%q); got %d; want %d", s, ts, 123)
+		}
+
+		// Verify timestampExpected
+		_, err = getTime(r, "s", 123)
+		if err == nil {
+			t.Fatalf("expecting non-nil error in getTime(%q)", s)
+		}
+	}
+
+	f("foo")
+	f("2019-07-07T20:01:02Zisdf")
+	f("2019-07-07T20:47:40+03:00123")
+	f("-292273086-05-16T16:47:07Z")
+	f("292277025-08-18T07:12:54.999999998Z")
+}

app/vmselect/promql/aggr.go

@@ -6,6 +6,9 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
 
 var aggrFuncs = map[string]aggrFunc{
@@ -26,6 +29,8 @@ var aggrFuncs = map[string]aggrFunc{
 	"median":   aggrFuncMedian,
 	"limitk":   aggrFuncLimitK,
 	"distinct": newAggrFunc(aggrFuncDistinct),
+	"sum2":     newAggrFunc(aggrFuncSum2),
+	"geomean":  newAggrFunc(aggrFuncGeomean),
 }
 
 type aggrFunc func(afa *aggrFuncArg) ([]*timeseries, error)
@@ -65,33 +70,26 @@ func newAggrFunc(afe func(tss []*timeseries) []*timeseries) aggrFunc {
 	}
 }
 
+func removeGroupTags(metricName *storage.MetricName, modifier *modifierExpr) {
+	groupOp := strings.ToLower(modifier.Op)
+	switch groupOp {
+	case "", "by":
+		metricName.RemoveTagsOn(modifier.Args)
+	case "without":
+		metricName.RemoveTagsIgnoring(modifier.Args)
+	default:
+		logger.Panicf("BUG: unknown group modifier: %q", groupOp)
+	}
+}
+
 func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeseries, modifier *modifierExpr, keepOriginal bool) ([]*timeseries, error) {
 	arg := copyTimeseriesMetricNames(argOrig)
 
-	// Filter out superflouos tags.
-	var groupTags []string
-	groupOp := "by"
-	if modifier.Op != "" {
-		groupTags = modifier.Args
-		groupOp = strings.ToLower(modifier.Op)
-	}
-	switch groupOp {
-	case "by":
-		for _, ts := range arg {
-			ts.MetricName.RemoveTagsOn(groupTags)
-		}
-	case "without":
-		for _, ts := range arg {
-			ts.MetricName.RemoveTagsIgnoring(groupTags)
-		}
-	default:
-		return nil, fmt.Errorf(`unknown modifier: %q`, groupOp)
-	}
-
 	// Perform grouping.
 	m := make(map[string][]*timeseries)
 	bb := bbPool.Get()
 	for i, ts := range arg {
+		removeGroupTags(&ts.MetricName, modifier)
 		bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
 		if keepOriginal {
 			ts = argOrig[i]
@@ -100,10 +98,18 @@ func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeserie
 	}
 	bbPool.Put(bb)
 
+	srcTssCount := 0
+	dstTssCount := 0
 	rvs := make([]*timeseries, 0, len(m))
 	for _, tss := range m {
 		rv := afe(tss)
 		rvs = append(rvs, rv...)
+		srcTssCount += len(tss)
+		dstTssCount += len(rv)
+		if dstTssCount > 2000 && dstTssCount > 16*srcTssCount {
+			// This looks like count_values explosion.
+			return nil, fmt.Errorf(`too many timeseries after aggragation; got %d; want less than %d`, dstTssCount, 16*srcTssCount)
+		}
 	}
 	return rvs, nil
 }
@@ -132,6 +138,52 @@ func aggrFuncSum(tss []*timeseries) []*timeseries {
 	return tss[:1]
 }
 
+func aggrFuncSum2(tss []*timeseries) []*timeseries {
+	dst := tss[0]
+	for i := range dst.Values {
+		sum2 := float64(0)
+		count := 0
+		for _, ts := range tss {
+			v := ts.Values[i]
+			if math.IsNaN(v) {
+				continue
+			}
+			sum2 += v * v
+			count++
+		}
+		if count == 0 {
+			sum2 = nan
+		}
+		dst.Values[i] = sum2
+	}
+	return tss[:1]
+}
+
+func aggrFuncGeomean(tss []*timeseries) []*timeseries {
+	if len(tss) == 1 {
+		// Fast path - nothing to geomean.
+		return tss
+	}
+	dst := tss[0]
+	for i := range dst.Values {
+		p := 1.0
+		count := 0
+		for _, ts := range tss {
+			v := ts.Values[i]
+			if math.IsNaN(v) {
+				continue
+			}
+			p *= v
+			count++
+		}
+		if count == 0 {
+			p = nan
+		}
+		dst.Values[i] = math.Pow(p, 1/float64(count))
+	}
+	return tss[:1]
+}
+
 func aggrFuncMin(tss []*timeseries) []*timeseries {
 	if len(tss) == 1 {
 		// Fast path - nothing to min.
@@ -260,7 +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]
 }
@@ -301,6 +357,9 @@ func aggrFuncCountValues(afa *aggrFuncArg) ([]*timeseries, error) {
 		m := make(map[float64]bool)
 		for _, ts := range tss {
 			for _, v := range ts.Values {
+				if math.IsNaN(v) {
+					continue
+				}
 				m[v] = true
 			}
 		}
@@ -313,7 +372,7 @@ func aggrFuncCountValues(afa *aggrFuncArg) ([]*timeseries, error) {
 		var rvs []*timeseries
 		for _, v := range values {
 			var dst timeseries
-			dst.CopyFrom(tss[0])
+			dst.CopyFromShallowTimestamps(tss[0])
 			dst.MetricName.RemoveTag(dstLabel)
 			dst.MetricName.AddTag(dstLabel, strconv.FormatFloat(v, 'g', -1, 64))
 			for i := range dst.Values {
@@ -372,7 +431,7 @@ func newAggrFuncTopK(isReverse bool) aggrFunc {
 					ts.Values[n] = nan
 				}
 			}
-			return rvs
+			return removeNaNs(rvs)
 		}
 		return aggrFuncExt(afe, args[1], &afa.ae.Modifier, true)
 	}
@@ -457,6 +516,7 @@ func newAggrQuantileFunc(phis []float64) func(tss []*timeseries) []*timeseries {
 			idx := int(math.Round(float64(len(tss)-1) * phi))
 			dst.Values[n] = tss[idx].Values[n]
 		}
+		tss[0] = dst
 		return tss[:1]
 	}
 }

app/vmselect/promql/aggr_incremental.go

@@ -0,0 +1,450 @@
+package promql
+
+import (
+	"math"
+	"strings"
+	"sync"
+)
+
+// callbacks for optimized incremental calculations for aggregate functions
+// over rollups over metricExpr.
+//
+// These calculations save RAM for aggregates over big number of time series.
+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,
+	},
+}
+
+type incrementalAggrFuncContext struct {
+	ae *aggrFuncExpr
+
+	mLock sync.Mutex
+	m     map[uint]map[string]*incrementalAggrContext
+
+	callbacks *incrementalAggrFuncCallbacks
+}
+
+func newIncrementalAggrFuncContext(ae *aggrFuncExpr, callbacks *incrementalAggrFuncCallbacks) *incrementalAggrFuncContext {
+	return &incrementalAggrFuncContext{
+		ae:        ae,
+		m:         make(map[uint]map[string]*incrementalAggrContext),
+		callbacks: callbacks,
+	}
+}
+
+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)
+	iac := m[string(bb.B)]
+	if iac == nil {
+		tsAggr := &timeseries{
+			Values:     make([]float64, len(ts.Values)),
+			Timestamps: ts.Timestamps,
+			denyReuse:  true,
+		}
+		tsAggr.MetricName.CopyFrom(&ts.MetricName)
+		iac = &incrementalAggrContext{
+			ts:     tsAggr,
+			values: make([]float64, len(ts.Values)),
+		}
+		m[string(bb.B)] = iac
+	}
+	bbPool.Put(bb)
+	iafc.callbacks.updateAggrFunc(iac, ts.Values)
+}
+
+func (iafc *incrementalAggrFuncContext) finalizeTimeseries() []*timeseries {
+	// There is no need in iafc.mLock.Lock here, since finalizeTimeseries must be called
+	// without concurrent goroutines touching iafc.
+	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 mGlobal {
+		finalizeAggrFunc(iac)
+		tss = append(tss, iac.ts)
+	}
+	return tss
+}
+
+type incrementalAggrFuncCallbacks struct {
+	updateAggrFunc   func(iac *incrementalAggrContext, values []float64)
+	mergeAggrFunc    func(dst, src *incrementalAggrContext)
+	finalizeAggrFunc func(iac *incrementalAggrContext)
+}
+
+func getIncrementalAggrFuncCallbacks(name string) *incrementalAggrFuncCallbacks {
+	name = strings.ToLower(name)
+	return incrementalAggrFuncCallbacksMap[name]
+}
+
+type incrementalAggrContext struct {
+	ts     *timeseries
+	values []float64
+}
+
+func finalizeAggrCommon(iac *incrementalAggrContext) {
+	counts := iac.values
+	dstValues := iac.ts.Values
+	_ = dstValues[len(counts)-1]
+	for i, v := range counts {
+		if v == 0 {
+			dstValues[i] = nan
+		}
+	}
+}
+
+func updateAggrSum(iac *incrementalAggrContext, values []float64) {
+	dstValues := iac.ts.Values
+	dstCounts := iac.values
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		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
+	}
+}
+
+func updateAggrMin(iac *incrementalAggrContext, values []float64) {
+	dstValues := iac.ts.Values
+	dstCounts := iac.values
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v < dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
+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
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v > dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
+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.
+	dstValues := iac.ts.Values
+	dstCounts := iac.values
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
+		dstCounts[i]++
+	}
+}
+
+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
+	_ = dstValues[len(counts)-1]
+	for i, v := range counts {
+		if v == 0 {
+			dstValues[i] = nan
+			continue
+		}
+		dstValues[i] /= v
+	}
+}
+
+func updateAggrCount(iac *incrementalAggrContext, values []float64) {
+	dstValues := iac.ts.Values
+	_ = dstValues[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		dstValues[i]++
+	}
+}
+
+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) {
+	dstValues := iac.ts.Values
+	for i, v := range dstValues {
+		if v == 0 {
+			dstValues[i] = nan
+		}
+	}
+}
+
+func updateAggrSum2(iac *incrementalAggrContext, values []float64) {
+	dstValues := iac.ts.Values
+	dstCounts := iac.values
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v * v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v * v
+	}
+}
+
+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
+	}
+}
+
+func updateAggrGeomean(iac *incrementalAggrContext, values []float64) {
+	dstValues := iac.ts.Values
+	dstCounts := iac.values
+	_ = dstValues[len(values)-1]
+	_ = dstCounts[len(values)-1]
+	for i, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] *= v
+		dstCounts[i]++
+	}
+}
+
+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
+	_ = dstValues[len(counts)-1]
+	for i, v := range counts {
+		if v == 0 {
+			dstValues[i] = nan
+			continue
+		}
+		dstValues[i] = math.Pow(dstValues[i], 1/v)
+	}
+}

app/vmselect/promql/aggr_incremental_test.go

@@ -0,0 +1,187 @@
+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
+		}
+		if v1 != v2 {
+			return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
+		}
+	}
+	return nil
+}

app/vmselect/promql/arch_arm64.go

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

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
 	}
 }
@@ -271,7 +274,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 			rvsLeft := make([]*timeseries, len(right))
 			tsLeft := left[0]
 			for i, tsRight := range right {
-				tsRight.MetricName.ResetMetricGroup()
+				resetMetricGroupIfRequired(be, tsRight)
 				rvsLeft[i] = tsLeft
 			}
 			return rvsLeft, right, right, nil
@@ -281,7 +284,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 			rvsRight := make([]*timeseries, len(left))
 			tsRight := right[0]
 			for i, tsLeft := range left {
-				tsLeft.MetricName.ResetMetricGroup()
+				resetMetricGroupIfRequired(be, tsLeft)
 				rvsRight[i] = tsRight
 			}
 			return left, rvsRight, left, nil
@@ -296,8 +299,13 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 		if len(tss) == 1 {
 			return nil
 		}
-		return fmt.Errorf(`duplicate timeseries on the %s side of %q: %s %s`, side, be.Op, stringMetricTags(&tss[0].MetricName), be.GroupModifier.AppendString(nil))
+		if mergeNonOverlappingTimeseries(tss) {
+			return nil
+		}
+		return fmt.Errorf(`duplicate timeseries on the %s side of %s %s: %s and %s`, side, be.Op, be.GroupModifier.AppendString(nil),
+			stringMetricTags(&tss[0].MetricName), stringMetricTags(&tss[1].MetricName))
 	}
+
 	var rvsLeft, rvsRight []*timeseries
 	mLeft, mRight := createTimeseriesMapByTagSet(be, left, right)
 	joinOp := strings.ToLower(be.JoinModifier.Op)
@@ -335,7 +343,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 			if err := ensureOneX("right", tssRight); err != nil {
 				return nil, nil, nil, err
 			}
-			tssLeft[0].MetricName.ResetMetricGroup()
+			resetMetricGroupIfRequired(be, tssLeft[0])
 			rvsLeft = append(rvsLeft, tssLeft[0])
 			rvsRight = append(rvsRight, tssRight[0])
 		default:
@@ -349,6 +357,19 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	return rvsLeft, rvsRight, dst, nil
 }
 
+func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
+	if isBinaryOpCmp(be.Op) && !be.Bool {
+		// Do not reset MetricGroup for non-boolean `compare` binary ops like Prometheus does.
+		return
+	}
+	switch be.Op {
+	case "default", "if", "ifnot":
+		// Do not reset MetricGroup for these ops.
+		return
+	}
+	ts.MetricName.ResetMetricGroup()
+}
+
 func binaryOpPlus(left, right float64) float64 {
 	return left + right
 }
@@ -497,3 +518,26 @@ func isScalar(arg []*timeseries) bool {
 	}
 	return len(mn.Tags) == 0
 }
+
+func mergeNonOverlappingTimeseries(tss []*timeseries) bool {
+	if len(tss) < 2 {
+		logger.Panicf("BUG: expecting at least two timeseries. Got %d", len(tss))
+	}
+
+	// Check whether time series in tss overlap.
+	var dst timeseries
+	dst.CopyFromShallowTimestamps(tss[0])
+	dstValues := dst.Values
+	for _, ts := range tss[1:] {
+		for i, value := range ts.Values {
+			if math.IsNaN(dstValues[i]) {
+				dstValues[i] = value
+			} else if !math.IsNaN(value) {
+				// Time series overlap.
+				return false
+			}
+		}
+	}
+	tss[0].CopyFromShallowTimestamps(&dst)
+	return true
+}

app/vmselect/promql/eval.go

@@ -16,10 +16,10 @@ import (
 )
 
 var (
-	maxPointsPerTimeseries = flag.Int("search.maxPointsPerTimeseries", 10e3, "The maximum points per a single timeseries returned from the search")
+	maxPointsPerTimeseries = flag.Int("search.maxPointsPerTimeseries", 30e3, "The maximum points per a single timeseries returned from the search")
 )
 
-// The minumum number of points per timeseries for enabling time rounding.
+// The minimum number of points per timeseries for enabling time rounding.
 // This improves cache hit ratio for frequently requested queries over
 // big time ranges.
 const minTimeseriesPointsForTimeRounding = 50
@@ -31,7 +31,7 @@ const minTimeseriesPointsForTimeRounding = 50
 func ValidateMaxPointsPerTimeseries(start, end, step int64) error {
 	points := (end-start)/step + 1
 	if uint64(points) > uint64(*maxPointsPerTimeseries) {
-		return fmt.Errorf(`too many points for the given step=%d, start=%d and end=%d: %d; cannot exceed %d points`,
+		return fmt.Errorf(`too many points for the given step=%d, start=%d and end=%d: %d; cannot exceed -search.maxPointsPerTimeseries=%d`,
 			step, start, end, uint64(points), *maxPointsPerTimeseries)
 	}
 	return nil
@@ -145,14 +145,14 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		re := &rollupExpr{
 			Expr: me,
 		}
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re)
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, me.AppendString(nil), err)
 		}
 		return rv, nil
 	}
 	if re, ok := e.(*rollupExpr); ok {
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re)
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, re.AppendString(nil), err)
 		}
@@ -188,13 +188,30 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		if err != nil {
 			return nil, err
 		}
-		rv, err := evalRollupFunc(ec, fe.Name, rf, re)
+		rv, err := evalRollupFunc(ec, fe.Name, rf, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, fe.AppendString(nil), err)
 		}
 		return rv, nil
 	}
 	if ae, ok := e.(*aggrFuncExpr); ok {
+		if callbacks := getIncrementalAggrFuncCallbacks(ae.Name); callbacks != nil {
+			fe, nrf := tryGetArgRollupFuncWithMetricExpr(ae)
+			if fe != nil {
+				// There is an optimized path for calculating aggrFuncExpr over rollupFunc over metricExpr.
+				// The optimized path saves RAM for aggregates over big number of time series.
+				args, re, err := evalRollupFuncArgs(ec, fe)
+				if err != nil {
+					return nil, err
+				}
+				rf, err := nrf(args)
+				if err != nil {
+					return nil, err
+				}
+				iafc := newIncrementalAggrFuncContext(ae, callbacks)
+				return evalRollupFunc(ec, fe.Name, rf, re, iafc)
+			}
+		}
 		args, err := evalExprs(ec, ae.Args)
 		if err != nil {
 			return nil, err
@@ -249,6 +266,69 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 	return nil, fmt.Errorf("unexpected expression %q", e.AppendString(nil))
 }
 
+func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFunc) {
+	if len(ae.Args) != 1 {
+		return nil, nil
+	}
+	e := ae.Args[0]
+	// Make sure e contains one of the following:
+	// - metricExpr
+	// - metricExpr[d]
+	// - rollupFunc(metricExpr)
+	// - rollupFunc(metricExpr[d])
+
+	if me, ok := e.(*metricExpr); ok {
+		// e = metricExpr
+		if me.IsEmpty() {
+			return nil, nil
+		}
+		fe := &funcExpr{
+			Name: "default_rollup",
+			Args: []expr{me},
+		}
+		nrf := getRollupFunc(fe.Name)
+		return fe, nrf
+	}
+	if re, ok := e.(*rollupExpr); ok {
+		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+			return nil, nil
+		}
+		// e = rollupExpr(metricExpr)
+		fe := &funcExpr{
+			Name: "default_rollup",
+			Args: []expr{re},
+		}
+		nrf := getRollupFunc(fe.Name)
+		return fe, nrf
+	}
+	fe, ok := e.(*funcExpr)
+	if !ok {
+		return nil, nil
+	}
+	nrf := getRollupFunc(fe.Name)
+	if nrf == nil {
+		return nil, nil
+	}
+	rollupArgIdx := getRollupArgIdx(fe.Name)
+	arg := fe.Args[rollupArgIdx]
+	if me, ok := arg.(*metricExpr); ok {
+		if me.IsEmpty() {
+			return nil, nil
+		}
+		return &funcExpr{
+			Name: fe.Name,
+			Args: []expr{me},
+		}, nrf
+	}
+	if re, ok := arg.(*rollupExpr); ok {
+		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+			return nil, nil
+		}
+		return fe, nrf
+	}
+	return nil, nil
+}
+
 func evalExprs(ec *EvalConfig, es []expr) ([][]*timeseries, error) {
 	var rvs [][]*timeseries
 	for _, e := range es {
@@ -308,7 +388,7 @@ func getRollupExprArg(arg expr) *rollupExpr {
 	return &reNew
 }
 
-func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr) ([]*timeseries, error) {
+func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr, iafc *incrementalAggrFuncContext) ([]*timeseries, error) {
 	ecNew := ec
 	var offset int64
 	if len(re.Offset) > 0 {
@@ -325,19 +405,11 @@ func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr)
 	var rvs []*timeseries
 	var err error
 	if me, ok := re.Expr.(*metricExpr); ok {
-		if me.IsEmpty() {
-			rvs = evalNumber(ecNew, nan)
-		} else {
-			var window int64
-			if len(re.Window) > 0 {
-				window, err = DurationValue(re.Window, ec.Step)
-				if err != nil {
-					return nil, err
-				}
-			}
-			rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, me, window)
-		}
+		rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, me, iafc, re.Window)
 	} else {
+		if iafc != nil {
+			logger.Panicf("BUG: iafc must be nil for rollup %q over subquery %q", name, re.AppendString(nil))
+		}
 		rvs, err = evalRollupFuncWithSubquery(ecNew, name, rf, re)
 	}
 	if err != nil {
@@ -379,8 +451,7 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	}
 
 	ecSQ := newEvalConfig(ec)
-	ecSQ.Start -= window + maxSilenceInterval
-	ecSQ.End += step
+	ecSQ.Start -= window + maxSilenceInterval + step
 	ecSQ.Step = step
 	if err := ValidateMaxPointsPerTimeseries(ecSQ.Start, ecSQ.End, ecSQ.Step); err != nil {
 		return nil, err
@@ -395,30 +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
 }
 
@@ -472,31 +532,27 @@ func removeNanValues(dstValues []float64, dstTimestamps []int64, values []float6
 	return dstValues, dstTimestamps
 }
 
-func getMaxPointsPerRollup() int {
-	maxPointsPerRollupOnce.Do(func() {
-		n := memory.Allowed() / 16 / 8
-		if n <= 16 {
-			n = 16
-		}
-		maxPointsPerRollup = n
-	})
-	return maxPointsPerRollup
-}
-
-var (
-	maxPointsPerRollup     int
-	maxPointsPerRollupOnce sync.Once
-)
-
 var (
 	rollupResultCacheFullHits    = metrics.NewCounter(`vm_rollup_result_cache_full_hits_total`)
 	rollupResultCachePartialHits = metrics.NewCounter(`vm_rollup_result_cache_partial_hits_total`)
 	rollupResultCacheMiss        = metrics.NewCounter(`vm_rollup_result_cache_miss_total`)
 )
 
-func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me *metricExpr, window int64) ([]*timeseries, error) {
+func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me *metricExpr, iafc *incrementalAggrFuncContext, windowStr string) ([]*timeseries, error) {
+	if me.IsEmpty() {
+		return evalNumber(ec, nan), nil
+	}
+	var window int64
+	if len(windowStr) > 0 {
+		var err error
+		window, err = DurationValue(windowStr, ec.Step)
+		if err != nil {
+			return nil, err
+		}
+	}
+
 	// Search for partial results in cache.
-	tssCached, start := rollupResultCacheV.Get(name, ec, me, window)
+	tssCached, start := rollupResultCacheV.Get(name, ec, me, iafc, window)
 	if start > ec.End {
 		// The result is fully cached.
 		rollupResultCacheFullHits.Inc()
@@ -514,7 +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
 	}
@@ -533,28 +589,79 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	// Verify timeseries fit available memory after the rollup.
 	// Take into account points from tssCached.
 	pointsPerTimeseries := 1 + (ec.End-ec.Start)/ec.Step
-	if uint64(pointsPerTimeseries) > uint64(getMaxPointsPerRollup()/rssLen/len(rcs)) {
+	rollupPoints := mulNoOverflow(pointsPerTimeseries, int64(rssLen*len(rcs)))
+	rollupMemorySize := mulNoOverflow(rollupPoints, 16)
+	rml := getRollupMemoryLimiter()
+	if !rml.Get(uint64(rollupMemorySize)) {
 		rss.Cancel()
-		return nil, fmt.Errorf("cannot process more than %d data points for %d time series with %d points in each time series; "+
-			"possible solutions are: reducing the number of matching time series; switching to node with more RAM; increasing `step` query arg (%gs)",
-			getMaxPointsPerRollup(), rssLen*len(rcs), pointsPerTimeseries, float64(ec.Step)/1e3)
+		return nil, fmt.Errorf("not enough memory for processing %d data points across %d time series with %d points in each time series; "+
+			"possible solutions are: reducing the number of matching time series; switching to node with more RAM; "+
+			"increasing -memory.allowedPercent; increasing `step` query arg (%gs)",
+			rollupPoints, rssLen*len(rcs), pointsPerTimeseries, float64(ec.Step)/1e3)
 	}
+	defer rml.Put(uint64(rollupMemorySize))
 
 	// Evaluate rollup
-	tss := make([]*timeseries, 0, rssLen*len(rcs))
+	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
+	var tss []*timeseries
+	if iafc != nil {
+		tss, err = evalRollupWithIncrementalAggregate(iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+	} else {
+		tss, err = evalRollupNoIncrementalAggregate(rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+	}
+	if err != nil {
+		return nil, err
+	}
+	tss = mergeTimeseries(tssCached, tss, start, ec)
+	rollupResultCacheV.Put(name, ec, me, iafc, window, tss)
+
+	return tss, nil
+}
+
+var (
+	rollupMemoryLimiter     memoryLimiter
+	rollupMemoryLimiterOnce sync.Once
+)
+
+func getRollupMemoryLimiter() *memoryLimiter {
+	rollupMemoryLimiterOnce.Do(func() {
+		rollupMemoryLimiter.MaxSize = uint64(memory.Allowed()) / 4
+	})
+	return &rollupMemoryLimiter
+}
+
+func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
+	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
+	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
+		preFunc(rs.Values, rs.Timestamps)
+		ts := getTimeseries()
+		defer putTimeseries(ts)
+		for _, rc := range rcs {
+			ts.Reset()
+			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 {
+		return nil, err
+	}
+	tss := iafc.finalizeTimeseries()
+	return tss, nil
+}
+
+func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConfig,
+	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
+	tss := make([]*timeseries, 0, rss.Len()*len(rcs))
 	var tssLock sync.Mutex
-	err = rss.RunParallel(func(rs *netstorage.Result) {
+	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()
@@ -563,18 +670,23 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	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, window, tss)
-
 	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] {
@@ -584,13 +696,14 @@ func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64
 	}
 	newRollupConfig := func(rf rollupFunc, tagValue string) *rollupConfig {
 		return &rollupConfig{
-			TagValue:   tagValue,
-			Func:       rf,
-			Start:      start,
-			End:        end,
-			Step:       step,
-			Window:     window,
-			Timestamps: sharedTimestamps,
+			TagValue:        tagValue,
+			Func:            rf,
+			Start:           start,
+			End:             end,
+			Step:            step,
+			Window:          window,
+			MayAdjustWindow: rollupFuncsMayAdjustWindow[name],
+			Timestamps:      sharedTimestamps,
 		}
 	}
 	appendRollupConfigs := func(dst []*rollupConfig) []*rollupConfig {
@@ -617,6 +730,11 @@ func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64
 			deltaValues(values)
 		}
 		rcs = appendRollupConfigs(rcs)
+	case "rollup_candlestick":
+		rcs = append(rcs, newRollupConfig(rollupFirst, "open"))
+		rcs = append(rcs, newRollupConfig(rollupLast, "close"))
+		rcs = append(rcs, newRollupConfig(rollupMin, "low"))
+		rcs = append(rcs, newRollupConfig(rollupMax, "high"))
 	default:
 		rcs = append(rcs, newRollupConfig(rf, ""))
 	}
@@ -653,3 +771,11 @@ func evalTime(ec *EvalConfig) []*timeseries {
 	}
 	return rv
 }
+
+func mulNoOverflow(a, b int64) int64 {
+	if math.MaxInt64/b < a {
+		// Overflow
+		return math.MaxInt64
+	}
+	return a * b
+}

app/vmselect/promql/exec.go

@@ -1,16 +1,23 @@
 package promql
 
 import (
+	"flag"
 	"fmt"
 	"math"
 	"sort"
 	"sync"
 	"sync/atomic"
+	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/metrics"
 )
 
+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) {
@@ -22,8 +29,20 @@ func ExpandWithExprs(q string) (string, error) {
 	return string(buf), nil
 }
 
-// Exec executes q for the given ec until the deadline.
-func Exec(ec *EvalConfig, q string) ([]netstorage.Result, error) {
+// Exec executes q for the given ec.
+func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result, error) {
+	if *logSlowQueryDuration > 0 {
+		startTime := time.Now()
+		defer func() {
+			d := time.Since(startTime)
+			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()
+			}
+		}()
+	}
+
 	ec.validate()
 
 	e, err := parsePromQLWithCache(q)
@@ -50,6 +69,14 @@ func Exec(ec *EvalConfig, q string) ([]netstorage.Result, error) {
 	}
 	ec.End -= ec.Step
 
+	if isFirstPointOnly {
+		// Remove all the points except the first one from every time series.
+		for _, ts := range rv {
+			ts.Values = ts.Values[:1]
+			ts.Timestamps = ts.Timestamps[:1]
+		}
+	}
+
 	maySort := maySortResults(e, rv)
 	result, err := timeseriesToResult(rv, maySort)
 	if err != nil {
@@ -107,18 +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

@@ -62,7 +62,7 @@ func TestExecSuccess(t *testing.T) {
 			Deadline: netstorage.NewDeadline(time.Minute),
 		}
 		for i := 0; i < 5; i++ {
-			result, err := Exec(ec, q)
+			result, err := Exec(ec, q, false)
 			if err != nil {
 				t.Fatalf(`unexpected error when executing %q: %s`, q, err)
 			}
@@ -286,7 +286,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `time()[300s:100s] offset 100s`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{800, 1000, 1200, 1400, 1600, 1800},
+			Values:     []float64{900, 1100, 1300, 1500, 1700, 1900},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -297,7 +297,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `time()[1.5i:0.5i] offset 0.5i`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{800, 1000, 1200, 1400, 1600, 1800},
+			Values:     []float64{900, 1100, 1300, 1500, 1700, 1900},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -308,7 +308,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `time()[300s] offset 100s`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{700, 900, 1100, 1300, 1500, 1700},
+			Values:     []float64{900, 1100, 1300, 1500, 1700, 1900},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -319,7 +319,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `time()[300s]`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{800, 1000, 1200, 1400, 1600, 1800},
+			Values:     []float64{1000, 1200, 1400, 1600, 1800, 2000},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -574,6 +574,30 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`clamp_max(alias(time(),"foobar"), 1400)`, func(t *testing.T) {
+		t.Parallel()
+		q := `clamp_max(alias(time(), "foobar"), 1400)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1000, 1200, 1400, 1400, 1400, 1400},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.MetricGroup = []byte("foobar")
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`CLAmp_MAx(alias(time(),"foobar"), 1400)`, func(t *testing.T) {
+		t.Parallel()
+		q := `CLAmp_MAx(alias(time(), "foobar"), 1400)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1000, 1200, 1400, 1400, 1400, 1400},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.MetricGroup = []byte("foobar")
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run("clamp_min(time(), -time()+3000)", func(t *testing.T) {
 		t.Parallel()
 		q := `clamp_min(time(), -time()+2500)`
@@ -789,6 +813,18 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{}
 		f(q, resultExpected)
 	})
+	t.Run(`alias()`, func(t *testing.T) {
+		t.Parallel()
+		q := `alias(time(), "foobar")`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1000, 1200, 1400, 1600, 1800, 2000},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.MetricGroup = []byte("foobar")
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`label_set(tag)`, func(t *testing.T) {
 		t.Parallel()
 		q := `label_set(time(), "tagname", "tagvalue")`
@@ -1266,6 +1302,34 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`label_transform(mismatch)`, func(t *testing.T) {
+		t.Parallel()
+		q := `label_transform(time(), "__name__", "foobar", "xx")`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1000, 1200, 1400, 1600, 1800, 2000},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`label_transform(match)`, func(t *testing.T) {
+		t.Parallel()
+		q := `label_transform(
+			label_set(time(), "foo", "a.bar.baz"),
+			"foo", "\\.", "-")`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1000, 1200, 1400, 1600, 1800, 2000},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("a-bar-baz"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`label_replace(mismatch)`, func(t *testing.T) {
 		t.Parallel()
 		q := `label_replace(time(), "__name__", "x${1}y", "foo", ".+")`
@@ -1410,6 +1474,62 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r1, r2}
 		f(q, resultExpected)
 	})
+	t.Run(`a cmp scalar (leave MetricGroup)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort_desc((
+			label_set(time(), "__name__", "foo", "a", "x"),
+			label_set(time()+200, "__name__", "bar", "a", "x"),
+		) > 1300)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, 1400, 1600, 1800, 2000, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.MetricGroup = []byte("bar")
+		r1.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("a"),
+			Value: []byte("x"),
+		}}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, 1400, 1600, 1800, 2000},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.MetricGroup = []byte("foo")
+		r2.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("a"),
+			Value: []byte("x"),
+		}}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
+	t.Run(`a cmp bool scalar (drop MetricGroup)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort_desc((
+			label_set(time(), "__name__", "foo", "a", "x"),
+			label_set(time()+200, "__name__", "bar", "a", "y"),
+		) >= bool 1200)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, 1, 1, 1, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("a"),
+			Value: []byte("y"),
+		}}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{0, 1, 1, 1, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("a"),
+			Value: []byte("x"),
+		}}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
 	t.Run(`1 > 2`, func(t *testing.T) {
 		t.Parallel()
 		q := `1 > 2`
@@ -1512,13 +1632,14 @@ func TestExecSuccess(t *testing.T) {
 	t.Run(`vector default scalar`, func(t *testing.T) {
 		t.Parallel()
 		q := `sort_desc(union(
-			label_set(time() > 1400, "foo", "bar"),
-			label_set(time() < 1700, "foo", "baz")) default 123)`
+			label_set(time() > 1400, "__name__", "x", "foo", "bar"),
+			label_set(time() < 1700, "__name__", "y", "foo", "baz")) default 123)`
 		r1 := netstorage.Result{
 			MetricName: metricNameExpected,
 			Values:     []float64{123, 123, 123, 1600, 1800, 2000},
 			Timestamps: timestampsExpected,
 		}
+		r1.MetricName.MetricGroup = []byte("x")
 		r1.MetricName.Tags = []storage.Tag{{
 			Key:   []byte("foo"),
 			Value: []byte("bar"),
@@ -1528,6 +1649,7 @@ func TestExecSuccess(t *testing.T) {
 			Values:     []float64{1000, 1200, 1400, 1600, 123, 123},
 			Timestamps: timestampsExpected,
 		}
+		r2.MetricName.MetricGroup = []byte("y")
 		r2.MetricName.Tags = []storage.Tag{{
 			Key:   []byte("foo"),
 			Value: []byte("baz"),
@@ -1699,6 +1821,24 @@ 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.Parallel()
+		q := `label_set(time()/10, "foo", "bar") + on(foo) group_left() (
+			label_set(time() < 1400, "foo", "bar", "op", "le"),
+			label_set(time() >= 1400, "foo", "bar", "op", "ge"),
+		)`
+		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"),
+		}}
+		resultExpected := []netstorage.Result{r1}
+		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)`
@@ -2018,6 +2158,26 @@ 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 := `sort(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")
+		))`
+		resultExpected := []netstorage.Result{}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort(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")
+		))`
+		resultExpected := []netstorage.Result{}
+		f(q, resultExpected)
+	})
 	t.Run(`median_over_time()`, func(t *testing.T) {
 		t.Parallel()
 		q := `median_over_time({})`
@@ -2068,6 +2228,17 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`avg(scalar) wiTHout (xx, yy)`, func(t *testing.T) {
+		t.Parallel()
+		q := `avg wiTHout (xx, yy) (123)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{123, 123, 123, 123, 123, 123},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`sum(time)`, func(t *testing.T) {
 		t.Parallel()
 		q := `sum(time()/100)`
@@ -2079,6 +2250,51 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`geomean(time)`, func(t *testing.T) {
+		t.Parallel()
+		q := `geomean(time()/100)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{10, 12, 14, 16, 18, 20},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`geomean_over_time(time)`, func(t *testing.T) {
+		t.Parallel()
+		q := `round(geomean_over_time(alias(time()/100, "foobar")[3i]), 0.1)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{6.8, 8.8, 10.9, 12.9, 14.9, 16.9},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.MetricGroup = []byte("foobar")
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`sum2(time)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sum2(time()/100)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{100, 144, 196, 256, 324, 400},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`sum2_over_time(time)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sum2_over_time(alias(time()/100, "foobar")[3i])`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{155, 251, 371, 515, 683, 875},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`sum(multi-vector)`, func(t *testing.T) {
 		t.Parallel()
 		q := `sum(label_set(10, "foo", "bar") or label_set(time()/100, "baz", "sss"))`
@@ -2090,6 +2306,39 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`geomean(multi-vector)`, func(t *testing.T) {
+		t.Parallel()
+		q := `round(geomean(label_set(10, "foo", "bar") or label_set(time()/100, "baz", "sss")), 0.1)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{10, 11, 11.8, 12.6, 13.4, 14.1},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`sum2(multi-vector)`, func(t *testing.T) {
+		t.Parallel()
+		q := `sum2(label_set(10, "foo", "bar") or label_set(time()/100, "baz", "sss"))`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{200, 244, 296, 356, 424, 500},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`sqrt(sum2(multi-vector))`, func(t *testing.T) {
+		t.Parallel()
+		q := `round(sqrt(sum2(label_set(10, "foo", "bar") or label_set(time()/100, "baz", "sss"))))`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{14, 16, 17, 19, 21, 22},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`avg(multi-vector)`, func(t *testing.T) {
 		t.Parallel()
 		q := `avg(label_set(10, "foo", "bar") or label_set(time()/100, "baz", "sss"))`
@@ -2114,10 +2363,10 @@ 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}
@@ -2279,6 +2528,21 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r1, r2}
 		f(q, resultExpected)
 	})
+	t.Run(`topk(1, nan_timeseries)`, func(t *testing.T) {
+		t.Parallel()
+		q := `topk(1, label_set(NaN, "foo", "bar") or label_set(time()/150, "baz", "sss")) default 0`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{6.666666666666667, 8, 9.333333333333334, 10.666666666666666, 12, 13.333333333333334},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("baz"),
+			Value: []byte("sss"),
+		}}
+		resultExpected := []netstorage.Result{r1}
+		f(q, resultExpected)
+	})
 	t.Run(`topk(2)`, func(t *testing.T) {
 		t.Parallel()
 		q := `sort(topk(2, label_set(10, "foo", "bar") or label_set(time()/150, "baz", "sss")))`
@@ -2378,7 +2642,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `distinct_over_time((time() < 1700)[500s])`
 		r1 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{3, 3, 3, 2, 1, nan},
+			Values:     []float64{3, 3, 3, 3, 2, 1},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r1}
@@ -2389,7 +2653,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `distinct_over_time((time() < 1700)[2.5i])`
 		r1 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{3, 3, 3, 2, 1, nan},
+			Values:     []float64{3, 3, 3, 3, 2, 1},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r1}
@@ -2645,7 +2909,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `integrate(time()*1e-3)`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{200, 240.00000000000003, 280, 320, 360, 400},
+			Values:     []float64{160, 200, 240.00000000000003, 280, 320, 360},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2667,7 +2931,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate(2000-time())`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{4.5, 3.5, 2.5, 1.5, 0.5, -0.5},
+			Values:     []float64{5.5, 4.5, 3.5, 2.5, 1.5, 0.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2678,7 +2942,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate((2000-time())[100s])`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{4.5, 3.5, 2.5, 1.5, 0.5, -0.5},
+			Values:     []float64{5.5, 4.5, 3.5, 2.5, 1.5, 0.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2689,7 +2953,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate((2000-time())[100s:])`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{4.5, 3.5, 2.5, 1.5, 0.5, -0.5},
+			Values:     []float64{5.5, 4.5, 3.5, 2.5, 1.5, 0.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2700,7 +2964,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate((2000-time())[100s:100s])`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{4, 6.5, 4.5, 2.5, 0.5, -1.5},
+			Values:     []float64{5.5, 4.5, 6.5, 4.5, 2.5, 0.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2711,7 +2975,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate((2000-time())[100s:100s] offset 100s)`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{4.5, 3.5, 5.5, 3.5, 1.5, -0.5},
+			Values:     []float64{6, 5, 7.5, 5.5, 3.5, 1.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2722,7 +2986,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `rate((2000-time())[100s:100s] offset 100s)[:] offset 100s`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{6, 5, 7.5, 5.5, 3.5, 1.5},
+			Values:     []float64{7, 6, 5, 7.5, 5.5, 3.5},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -2744,7 +3008,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `increase(2000-time())`
 		r := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{900, 700, 500, 300, 100, -100},
+			Values:     []float64{1100, 900, 700, 500, 300, 100},
 			Timestamps: timestampsExpected,
 		}
 		resultExpected := []netstorage.Result{r}
@@ -3009,6 +3273,48 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`rollup_candlestick()`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort(rollup_candlestick(round(rand(0),0.01)[:10s]))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{0.02, 0.02, 0.03, 0, 0.03, 0.02},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("rollup"),
+			Value: []byte("low"),
+		}}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{0.32, 0.82, 0.13, 0.28, 0.86, 0.57},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("rollup"),
+			Value: []byte("close"),
+		}}
+		r3 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{0.9, 0.32, 0.82, 0.13, 0.28, 0.86},
+			Timestamps: timestampsExpected,
+		}
+		r3.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("rollup"),
+			Value: []byte("open"),
+		}}
+		r4 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{0.85, 0.94, 0.97, 0.93, 0.98, 0.92},
+			Timestamps: timestampsExpected,
+		}
+		r4.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("rollup"),
+			Value: []byte("high"),
+		}}
+		resultExpected := []netstorage.Result{r1, r2, r3, r4}
+		f(q, resultExpected)
+	})
 	t.Run(`rollup_increase()`, func(t *testing.T) {
 		t.Parallel()
 		q := `sort(rollup_increase(time()))`
@@ -3047,7 +3353,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `sort(rollup(time()[:50s]))`
 		r1 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{1050, 1250, 1450, 1650, 1850, 2050},
+			Values:     []float64{850, 1050, 1250, 1450, 1650, 1850},
 			Timestamps: timestampsExpected,
 		}
 		r1.MetricName.Tags = []storage.Tag{{
@@ -3056,21 +3362,21 @@ func TestExecSuccess(t *testing.T) {
 		}}
 		r2 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{1200, 1400, 1600, 1800, 2000, 2200},
+			Values:     []float64{925, 1125, 1325, 1525, 1725, 1925},
 			Timestamps: timestampsExpected,
 		}
 		r2.MetricName.Tags = []storage.Tag{{
 			Key:   []byte("rollup"),
-			Value: []byte("max"),
+			Value: []byte("avg"),
 		}}
 		r3 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{1125, 1325, 1525, 1725, 1925, 2125},
+			Values:     []float64{1000, 1200, 1400, 1600, 1800, 2000},
 			Timestamps: timestampsExpected,
 		}
 		r3.MetricName.Tags = []storage.Tag{{
 			Key:   []byte("rollup"),
-			Value: []byte("avg"),
+			Value: []byte("max"),
 		}}
 		resultExpected := []netstorage.Result{r1, r2, r3}
 		f(q, resultExpected)
@@ -3429,7 +3735,14 @@ func TestExecError(t *testing.T) {
 			Deadline: netstorage.NewDeadline(time.Minute),
 		}
 		for i := 0; i < 4; i++ {
-			rv, err := Exec(ec, q)
+			rv, err := Exec(ec, q, false)
+			if err == nil {
+				t.Fatalf(`expecting non-nil error on %q`, q)
+			}
+			if rv != nil {
+				t.Fatalf(`expecting nil rv`)
+			}
+			rv, err = Exec(ec, q, true)
 			if err == nil {
 				t.Fatalf(`expecting non-nil error on %q`, q)
 			}
@@ -3461,6 +3774,7 @@ func TestExecError(t *testing.T) {
 	f(`hour(1,2)`)
 	f(`label_join()`)
 	f(`label_replace(1)`)
+	f(`label_transform(1)`)
 	f(`label_set()`)
 	f(`label_set(1, "foo")`)
 	f(`label_del()`)
@@ -3506,6 +3820,9 @@ func TestExecError(t *testing.T) {
 	f(`keep_last_value()`)
 	f(`distinct_over_time()`)
 	f(`distinct()`)
+	f(`alias()`)
+	f(`alias(1)`)
+	f(`alias(1, "foo", "bar")`)
 
 	// Invalid argument type
 	f(`median_over_time({}, 2)`)
@@ -3535,6 +3852,11 @@ func TestExecError(t *testing.T) {
 	f(`label_replace(1, "foo", "bar", 4, 5)`)
 	f(`label_replace(1, "foo", "bar", "baz", 5)`)
 	f(`label_replace(1, "foo", "bar", "baz", "invalid(regexp")`)
+	f(`label_transform(1, 2, 3, 4)`)
+	f(`label_transform(1, "foo", 3, 4)`)
+	f(`label_transform(1, "foo", "bar", 4)`)
+	f(`label_transform(1, "foo", "invalid(regexp", "baz`)
+	f(`alias(1, 2)`)
 
 	// Duplicate timeseries
 	f(`(label_set(1, "foo", "bar") or label_set(2, "foo", "baz"))
@@ -3545,12 +3867,31 @@ func TestExecError(t *testing.T) {
 	f(`1 + group_left() (label_set(1, "foo", bar"), label_set(2, "foo", "baz"))`)
 	f(`1 + on() group_left() (label_set(1, "foo", bar"), label_set(2, "foo", "baz"))`)
 	f(`1 + on(a) group_left(b) (label_set(1, "foo", bar"), label_set(2, "foo", "baz"))`)
+	f(`label_set(1, "foo", "bar") + on(foo) group_left() (label_set(1, "foo", "bar", "a", "b"), label_set(1, "foo", "bar", "a", "c"))`)
 	f(`(label_set(1, "foo", bar"), label_set(2, "foo", "baz")) + group_right 1`)
 	f(`(label_set(1, "foo", bar"), label_set(2, "foo", "baz")) + on() group_right 1`)
 	f(`(label_set(1, "foo", bar"), label_set(2, "foo", "baz")) + on(a) group_right(b,c) 1`)
 	f(`(label_set(1, "foo", bar"), label_set(2, "foo", "baz")) + on() 1`)
+	f(`(label_set(1, "foo", "bar", "a", "b"), label_set(1, "foo", "bar", "a", "c")) + on(foo) group_right() label_set(1, "foo", "bar")`)
 	f(`1 + on() (label_set(1, "foo", bar"), label_set(2, "foo", "baz"))`)
 
+	// duplicate metrics after binary op
+	f(`(
+		label_set(time(), "__name__", "foo", "a", "x"),
+		label_set(time()+200, "__name__", "bar", "a", "x"),
+	) > bool 1300`)
+	f(`(
+		label_set(time(), "__name__", "foo", "a", "x"),
+		label_set(time()+200, "__name__", "bar", "a", "x"),
+	) + 10`)
+
+	// Invalid aggregates
+	f(`sum(1, 2)`)
+	f(`sum(1) foo (bar)`)
+	f(`sum foo () (bar)`)
+	f(`sum(foo) by (1)`)
+	f(`count(foo) without ("bar")`)
+
 	// With expressions
 	f(`ttf()`)
 	f(`ttf(1, 2)`)

app/vmselect/promql/lexer.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"strconv"
 	"strings"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
 type lexer struct {
@@ -85,10 +87,7 @@ again:
 		goto tokenFoundLabel
 	}
 	if isIdentPrefix(s) {
-		token, err = scanIdent(s)
-		if err != nil {
-			return "", err
-		}
+		token = scanIdent(s)
 		goto tokenFoundLabel
 	}
 	if isStringPrefix(s) {
@@ -210,15 +209,103 @@ func scanPositiveNumber(s string) (string, error) {
 	return s[:j], nil
 }
 
-func scanIdent(s string) (string, error) {
-	if len(s) == 0 {
-		return "", fmt.Errorf("ident cannot be empty")
-	}
+func scanIdent(s string) string {
 	i := 0
-	for i < len(s) && isIdentChar(s[i]) {
-		i++
+	for i < len(s) {
+		if isIdentChar(s[i]) {
+			i++
+			continue
+		}
+		if s[i] != '\\' {
+			break
+		}
+
+		// Do not verify the next char, since it is escaped.
+		i += 2
+		if i > len(s) {
+			i--
+			break
+		}
 	}
-	return s[:i], nil
+	if i == 0 {
+		logger.Panicf("BUG: scanIdent couldn't find a single ident char; make sure isIdentPrefix called before scanIdent")
+	}
+	return s[:i]
+}
+
+func unescapeIdent(s string) string {
+	n := strings.IndexByte(s, '\\')
+	if n < 0 {
+		return s
+	}
+	dst := make([]byte, 0, len(s))
+	for {
+		dst = append(dst, s[:n]...)
+		s = s[n+1:]
+		if len(s) == 0 {
+			return string(dst)
+		}
+		if s[0] == 'x' && len(s) >= 3 {
+			h1 := fromHex(s[1])
+			h2 := fromHex(s[2])
+			if h1 >= 0 && h2 >= 0 {
+				dst = append(dst, byte((h1<<4)|h2))
+				s = s[3:]
+			} else {
+				dst = append(dst, s[0])
+				s = s[1:]
+			}
+		} else {
+			dst = append(dst, s[0])
+			s = s[1:]
+		}
+		n = strings.IndexByte(s, '\\')
+		if n < 0 {
+			dst = append(dst, s...)
+			return string(dst)
+		}
+	}
+}
+
+func fromHex(ch byte) int {
+	if ch >= '0' && ch <= '9' {
+		return int(ch - '0')
+	}
+	if ch >= 'a' && ch <= 'f' {
+		return int((ch - 'a') + 10)
+	}
+	if ch >= 'A' && ch <= 'F' {
+		return int((ch - 'A') + 10)
+	}
+	return -1
+}
+
+func toHex(n byte) byte {
+	if n < 10 {
+		return '0' + n
+	}
+	return 'a' + (n - 10)
+}
+
+func appendEscapedIdent(dst, s []byte) []byte {
+	for i := 0; i < len(s); i++ {
+		ch := s[i]
+		if isIdentChar(ch) {
+			if i == 0 && !isFirstIdentChar(ch) {
+				// hex-encode the first char
+				dst = append(dst, '\\', 'x', toHex(ch>>4), toHex(ch&0xf))
+			} else {
+				dst = append(dst, ch)
+			}
+		} else if ch >= 0x20 && ch < 0x7f {
+			// Leave ASCII printable chars as is
+			dst = append(dst, '\\', ch)
+		} else {
+			// hex-encode non-printable chars
+			dst = append(dst, '\\', 'x', toHex(ch>>4), toHex(ch&0xf))
+		}
+	}
+	return dst
 }
 
 func (lex *lexer) Prev() {
@@ -353,6 +440,10 @@ func isIdentPrefix(s string) bool {
 	if len(s) == 0 {
 		return false
 	}
+	if s[0] == '\\' {
+		// Assume this is an escape char for the next char.
+		return true
+	}
 	return isFirstIdentChar(s[0])
 }
 
@@ -367,7 +458,7 @@ func isIdentChar(ch byte) bool {
 	if isFirstIdentChar(ch) {
 		return true
 	}
-	return isDecimalChar(ch) || ch == ':' || ch == '.'
+	return isDecimalChar(ch) || ch == '.'
 }
 
 func isSpaceChar(ch byte) bool {

app/vmselect/promql/lexer_test.go

@@ -5,6 +5,57 @@ import (
 	"testing"
 )
 
+func TestUnescapeIdent(t *testing.T) {
+	f := func(s, resultExpected string) {
+		t.Helper()
+		result := unescapeIdent(s)
+		if result != resultExpected {
+			t.Fatalf("unexpected result for unescapeIdent(%q); got %q; want %q", s, result, resultExpected)
+		}
+	}
+	f("", "")
+	f("a", "a")
+	f("\\", "")
+	f(`\\`, `\`)
+	f(`\foo\-bar`, `foo-bar`)
+	f(`a\\\\b\"c\d`, `a\\b"cd`)
+	f(`foo.bar:baz_123`, `foo.bar:baz_123`)
+	f(`foo\ bar`, `foo bar`)
+	f(`\x21`, `!`)
+	f(`\xeDfoo\x2Fbar\-\xqw\x`, "\xedfoo\x2fbar-xqwx")
+}
+
+func TestAppendEscapedIdent(t *testing.T) {
+	f := func(s, resultExpected string) {
+		t.Helper()
+		result := appendEscapedIdent(nil, []byte(s))
+		if string(result) != resultExpected {
+			t.Fatalf("unexpected result for appendEscapedIdent(%q); got %q; want %q", s, result, resultExpected)
+		}
+	}
+	f(`a`, `a`)
+	f(`a.b:c_23`, `a.b:c_23`)
+	f(`a b-cd+dd\`, `a\ b\-cd\+dd\\`)
+	f("a\x1E\x20\xee", `a\x1e\ \xee`)
+	f("\x2e\x2e", `\x2e.`)
+}
+
+func TestScanIdent(t *testing.T) {
+	f := func(s, resultExpected string) {
+		t.Helper()
+		result := scanIdent(s)
+		if result != resultExpected {
+			t.Fatalf("unexpected result for scanIdent(%q): got %q; want %q", s, result, resultExpected)
+		}
+	}
+	f("a", "a")
+	f("foo.bar:baz_123", "foo.bar:baz_123")
+	f("a+b", "a")
+	f("foo()", "foo")
+	f(`a\-b+c`, `a\-b`)
+	f(`a\ b\\\ c\`, `a\ b\\\ c\`)
+}
+
 func TestLexerNextPrev(t *testing.T) {
 	var lex lexer
 	lex.Init("foo bar baz")

app/vmselect/promql/memory_limiter.go

@@ -0,0 +1,33 @@
+package promql
+
+import (
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+type memoryLimiter struct {
+	MaxSize uint64
+
+	mu    sync.Mutex
+	usage uint64
+}
+
+func (ml *memoryLimiter) Get(n uint64) bool {
+	ml.mu.Lock()
+	ok := n <= ml.MaxSize && ml.MaxSize-n >= ml.usage
+	if ok {
+		ml.usage += n
+	}
+	ml.mu.Unlock()
+	return ok
+}
+
+func (ml *memoryLimiter) Put(n uint64) {
+	ml.mu.Lock()
+	if n > ml.usage {
+		logger.Panicf("BUG: n=%d cannot exceed %d", n, ml.usage)
+	}
+	ml.usage -= n
+	ml.mu.Unlock()
+}

app/vmselect/promql/memory_limiter_test.go

@@ -0,0 +1,56 @@
+package promql
+
+import (
+	"testing"
+)
+
+func TestMemoryLimiter(t *testing.T) {
+	var ml memoryLimiter
+	ml.MaxSize = 100
+
+	// Allocate memory
+	if !ml.Get(10) {
+		t.Fatalf("cannot get 10 out of %d bytes", ml.MaxSize)
+	}
+	if ml.usage != 10 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 10)
+	}
+	if !ml.Get(20) {
+		t.Fatalf("cannot get 20 out of 90 bytes")
+	}
+	if ml.usage != 30 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 30)
+	}
+	if ml.Get(1000) {
+		t.Fatalf("unexpected get for 1000 bytes")
+	}
+	if ml.usage != 30 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 30)
+	}
+	if ml.Get(71) {
+		t.Fatalf("unexpected get for 71 bytes")
+	}
+	if ml.usage != 30 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 30)
+	}
+	if !ml.Get(70) {
+		t.Fatalf("cannot get 70 bytes")
+	}
+	if ml.usage != 100 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 100)
+	}
+
+	// Return memory back
+	ml.Put(10)
+	ml.Put(70)
+	if ml.usage != 20 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 20)
+	}
+	if !ml.Get(30) {
+		t.Fatalf("cannot get 30 bytes")
+	}
+	ml.Put(50)
+	if ml.usage != 0 {
+		t.Fatalf("unexpected usage; got %d; want %d", ml.usage, 0)
+	}
+}

app/vmselect/promql/parser.go

@@ -6,7 +6,6 @@ import (
 	"strings"
 	"sync"
 
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
@@ -19,12 +18,13 @@ func getDefaultWithArgExprs() []*withArgExpr {
 
 			// ttf - time to fuckup
 			`ttf(freev) = smooth_exponential(
-				clamp_max(clamp_min(freev, 0) / clamp_min(deriv(-freev), 0), 365*24*3600),
+				clamp_max(clamp_max(-freev, 0) / clamp_max(deriv_fast(freev), 0), 365*24*3600),
 				clamp_max(step()/300, 1)
 			)`,
 
 			`median_over_time(m) = quantile_over_time(0.5, m)`,
 			`range_median(q) = range_quantile(0.5, q)`,
+			`alias(q, name) = label_set(q, "__name__", name)`,
 		})
 	})
 	return defaultWithArgExprs
@@ -744,7 +744,7 @@ func expandWithExpr(was []*withArgExpr, e expr) (expr, error) {
 		if !t.HasNonEmptyMetricGroup() {
 			return t, nil
 		}
-		k := bytesutil.ToUnsafeString(t.TagFilters[0].Value)
+		k := string(appendEscapedIdent(nil, t.TagFilters[0].Value))
 		wa := getWithArgExpr(was, k)
 		if wa == nil {
 			return t, nil
@@ -811,7 +811,9 @@ func expandModifierArgs(was []*withArgExpr, args []string) ([]string, error) {
 			continue
 		}
 		if len(wa.Args) > 0 {
-			return nil, fmt.Errorf("cannot use func %q instead of %q in %s", wa.Name, arg, args)
+			// Template funcs cannot be used inside modifier list. Leave the arg as is.
+			dstArgs = append(dstArgs, arg)
+			continue
 		}
 		me, ok := wa.Expr.(*metricExpr)
 		if ok {
@@ -851,6 +853,10 @@ func expandModifierArgs(was []*withArgExpr, args []string) ([]string, error) {
 
 func expandWithExprExt(was []*withArgExpr, wa *withArgExpr, args []expr) (expr, error) {
 	if len(wa.Args) != len(args) {
+		if args == nil {
+			// Just return metricExpr with the wa.Name name.
+			return newMetricExpr(wa.Name), nil
+		}
 		return nil, fmt.Errorf("invalid number of args for %q; got %d; want %d", wa.Name, len(args), len(wa.Args))
 	}
 	wasNew := make([]*withArgExpr, 0, len(was)+len(args))
@@ -869,6 +875,14 @@ func expandWithExprExt(was []*withArgExpr, wa *withArgExpr, args []expr) (expr,
 	return expandWithExpr(wasNew, wa.Expr)
 }
 
+func newMetricExpr(name string) *metricExpr {
+	return &metricExpr{
+		TagFilters: []storage.TagFilter{{
+			Value: []byte(name),
+		}},
+	}
+}
+
 func extractStringValue(token string) (string, error) {
 	if !isStringPrefix(token) {
 		return "", fmt.Errorf(`stringExpr must contain only string literals; got %q`, token)
@@ -1074,9 +1088,6 @@ func (p *parser) parseTagFilterExpr() (*tagFilterExpr, error) {
 	}
 	var tfe tagFilterExpr
 	tfe.Key = p.lex.Token
-	if tfe.Key == "__name__" {
-		tfe.Key = ""
-	}
 	if err := p.lex.Next(); err != nil {
 		return nil, err
 	}
@@ -1125,8 +1136,16 @@ func (tfe *tagFilterExpr) toTagFilter() (*storage.TagFilter, error) {
 	}
 
 	var tf storage.TagFilter
-	tf.Key = []byte(tfe.Key)
-	tf.Value = []byte(tfe.Value.S)
+	tf.Key = []byte(unescapeIdent(tfe.Key))
+	if len(tfe.Key) == 0 {
+		tf.Value = []byte(unescapeIdent(tfe.Value.S))
+	} else {
+		tf.Value = []byte(tfe.Value.S)
+	}
+	if string(tf.Key) == "__name__" {
+		// This is required for storage.Search
+		tf.Key = nil
+	}
 	tf.IsRegexp = tfe.IsRegexp
 	tf.IsNegative = tfe.IsNegative
 	if !tf.IsRegexp {
@@ -1507,7 +1526,7 @@ func (wa *withArgExpr) AppendString(dst []byte) []byte {
 }
 
 type rollupExpr struct {
-	// The expression for the rollup. Usually it is metricExpr, but may be arbitary expr
+	// The expression for the rollup. Usually it is metricExpr, but may be arbitrary expr
 	// if subquery is used. https://prometheus.io/blog/2019/01/28/subquery-support/
 	Expr expr
 
@@ -1585,7 +1604,7 @@ func (me *metricExpr) AppendString(dst []byte) []byte {
 	if len(tfs) > 0 {
 		tf := &tfs[0]
 		if len(tf.Key) == 0 && !tf.IsNegative && !tf.IsRegexp {
-			dst = append(dst, tf.Value...)
+			dst = appendEscapedIdent(dst, tf.Value)
 			tfs = tfs[1:]
 		}
 	}
@@ -1627,7 +1646,7 @@ func appendStringTagFilter(dst []byte, tf *storage.TagFilter) []byte {
 	if len(tf.Key) == 0 {
 		dst = append(dst, "__name__"...)
 	} else {
-		dst = append(dst, tf.Key...)
+		dst = appendEscapedIdent(dst, tf.Key)
 	}
 	var op string
 	if tf.IsNegative {

app/vmselect/promql/parser_test.go

@@ -118,6 +118,17 @@ func TestParsePromQLSuccess(t *testing.T) {
 	same("with")
 	same("WITH")
 	same("With")
+	same("alias")
+	same(`alias{foo="bar"}`)
+	same(`aLIas{alias="aa"}`)
+	another(`al\ias`, `alias`)
+	// identifiers with with escape chars
+	same(`foo\ bar`)
+	same(`foo\-bar\{{baz\+bar="aa"}`)
+	another(`\x2E\x2ef\oo{b\xEF\ar="aa"}`, `\x2e.foo{b\xefar="aa"}`)
+	// Duplicate filters
+	same(`foo{__name__="bar"}`)
+	same(`foo{a="b", a="c", __name__="aaa", b="d"}`)
 	// Metric filters ending with comma
 	another(`m{foo="bar",}`, `m{foo="bar"}`)
 	// String concat in tag value
@@ -251,6 +262,8 @@ func TestParsePromQLSuccess(t *testing.T) {
 	same(`rate(rate(m[5m]))`)
 	same(`rate(rate(m[5m])[1h:])`)
 	same(`rate(rate(m[5m])[1h:3s])`)
+	// funcName with escape chars
+	same(`foo\(ba\-r()`)
 
 	// aggrFuncExpr
 	same(`sum(http_server_request) by ()`)
@@ -295,10 +308,14 @@ func TestParsePromQLSuccess(t *testing.T) {
 	another(`with (ct={job="test", i="bar"}) ct + {ct, x="d"} + foo{ct, ct} + ctx(1)`,
 		`(({job="test", i="bar"} + {job="test", i="bar", x="d"}) + foo{job="test", i="bar"}) + ctx(1)`)
 	another(`with (foo = bar) {__name__=~"foo"}`, `{__name__=~"foo"}`)
-	another(`with (foo = bar) {__name__="foo"}`, `bar`)
+	another(`with (foo = bar) foo{__name__="foo"}`, `bar`)
 	another(`with (foo = bar) {__name__="foo", x="y"}`, `bar{x="y"}`)
 	another(`with (foo(bar) = {__name__!="bar"}) foo(x)`, `{__name__!="bar"}`)
-	another(`with (foo(bar) = {__name__="bar"}) foo(x)`, `x`)
+	another(`with (foo(bar) = bar{__name__="bar"}) foo(x)`, `x`)
+	another(`with (foo\-bar(baz) = baz + baz) foo\-bar((x,y))`, `(x, y) + (x, y)`)
+	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x*y)`, `(x * y) + (x * y)`)
+	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x\*y)`, `x\*y + x\*y`)
+	another(`with (foo\-bar(b\ az) = b\ az + b\ az) foo\-bar(x\*y)`, `x\*y + x\*y`)
 	// override ttf to something new.
 	another(`with (ttf = a) ttf + b`, `a + b`)
 	// override ttf to ru
@@ -332,8 +349,11 @@ func TestParsePromQLSuccess(t *testing.T) {
 	another(`with (x="a", y=x) y+"bc"`, `"abc"`)
 	another(`with (x="a", y="b"+x) "we"+y+"z"+f()`, `"webaz" + f()`)
 	another(`with (f(x) = m{foo=x+"y", bar="y"+x, baz=x} + x) f("qwe")`, `m{foo="qwey", bar="yqwe", baz="qwe"} + "qwe"`)
+	another(`with (f(a)=a) f`, `f`)
+	another(`with (f\q(a)=a) f\q`, `fq`)
 
 	// Verify withExpr for aggr func modifiers
+	another(`with (f(x) = x, y = sum(m) by (f)) y`, `sum(m) by (f)`)
 	another(`with (f(x) = sum(m) by (x)) f(foo)`, `sum(m) by (foo)`)
 	another(`with (f(x) = sum(m) by (x)) f((foo, bar, foo))`, `sum(m) by (foo, bar)`)
 	another(`with (f(x) = sum(m) without (x,y)) f((a, b))`, `sum(m) without (a, b, y)`)
@@ -658,7 +678,7 @@ func TestParsePromQLError(t *testing.T) {
 	f(`with (x=m) f(b, a{x})`)
 	f(`with (x=m) sum(a{x})`)
 	f(`with (x=m) (a{x})`)
-	f(`with (f(a)=a) f`)
+	f(`with (f(a)=a) f(1, 2)`)
 	f(`with (f(x)=x{foo="bar"}) f(1)`)
 	f(`with (f(x)=x{foo="bar"}) f(m + n)`)
 	f(`with (f = with`)
@@ -668,8 +688,7 @@ func TestParsePromQLError(t *testing.T) {
 	f(`with (f(,)=x) x`)
 	f(`with (x(a) = {b="c"}) foo{x}`)
 	f(`with (f(x) = m{foo=xx}) f("qwe")`)
-	f(`a + with(f(x)=x) f`)
-	f(`with (f(x) = x, y = sum(m) by (f)) y`)
+	f(`a + with(f(x)=x) f(1,2)`)
 	f(`with (f(x) = sum(m) by (x)) f({foo="bar"})`)
 	f(`with (f(x) = sum(m) by (x)) f((xx(), {foo="bar"}))`)
 	f(`with (f(x) = m + on (x) n) f(xx())`)

app/vmselect/promql/regexp_cache.go

@@ -1,7 +1,6 @@
 package promql
 
 import (
-	"fmt"
 	"regexp"
 	"sync"
 	"sync/atomic"
@@ -10,12 +9,16 @@ import (
 )
 
 func compileRegexpAnchored(re string) (*regexp.Regexp, error) {
+	reAnchored := "^(?:" + re + ")$"
+	return compileRegexp(reAnchored)
+}
+
+func compileRegexp(re string) (*regexp.Regexp, error) {
 	rcv := regexpCacheV.Get(re)
 	if rcv != nil {
 		return rcv.r, rcv.err
 	}
-	regexAnchored := fmt.Sprintf("^(?:%s)$", re)
-	r, err := regexp.Compile(regexAnchored)
+	r, err := regexp.Compile(re)
 	rcv = &regexpCacheValue{
 		r:   r,
 		err: err,
@@ -77,7 +80,7 @@ func (rc *regexpCache) Get(regexp string) *regexpCacheValue {
 	rcv := rc.m[regexp]
 	rc.mu.RUnlock()
 
-	if rc == nil {
+	if rcv == nil {
 		atomic.AddUint64(&rc.misses, 1)
 	}
 	return rcv

app/vmselect/promql/rollup.go

@@ -19,13 +19,14 @@ var rollupFuncs = map[string]newRollupFunc{
 	// See funcs accepting range-vector on https://prometheus.io/docs/prometheus/latest/querying/functions/ .
 	"changes":            newRollupFuncOneArg(rollupChanges),
 	"delta":              newRollupFuncOneArg(rollupDelta),
-	"deriv":              newRollupFuncOneArg(rollupDeriv),
+	"deriv":              newRollupFuncOneArg(rollupDerivSlow),
+	"deriv_fast":         newRollupFuncOneArg(rollupDerivFast),
 	"holt_winters":       newRollupHoltWinters,
 	"idelta":             newRollupFuncOneArg(rollupIdelta),
 	"increase":           newRollupFuncOneArg(rollupDelta),  // + rollupFuncsRemoveCounterResets
 	"irate":              newRollupFuncOneArg(rollupIderiv), // + rollupFuncsRemoveCounterResets
 	"predict_linear":     newRollupPredictLinear,
-	"rate":               newRollupFuncOneArg(rollupDeriv), // + rollupFuncsRemoveCounterResets
+	"rate":               newRollupFuncOneArg(rollupDerivFast), // + rollupFuncsRemoveCounterResets
 	"resets":             newRollupFuncOneArg(rollupResets),
 	"avg_over_time":      newRollupFuncOneArg(rollupAvg),
 	"min_over_time":      newRollupFuncOneArg(rollupMin),
@@ -37,6 +38,8 @@ var rollupFuncs = map[string]newRollupFunc{
 	"stdvar_over_time":   newRollupFuncOneArg(rollupStdvar),
 
 	// Additional rollup funcs.
+	"sum2_over_time":     newRollupFuncOneArg(rollupSum2),
+	"geomean_over_time":  newRollupFuncOneArg(rollupGeomean),
 	"first_over_time":    newRollupFuncOneArg(rollupFirst),
 	"last_over_time":     newRollupFuncOneArg(rollupLast),
 	"distinct_over_time": newRollupFuncOneArg(rollupDistinct),
@@ -47,6 +50,17 @@ var rollupFuncs = map[string]newRollupFunc{
 	"rollup_deriv":       newRollupFuncOneArg(rollupFake),
 	"rollup_delta":       newRollupFuncOneArg(rollupFake),
 	"rollup_increase":    newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
+	"rollup_candlestick": newRollupFuncOneArg(rollupFake),
+}
+
+var rollupFuncsMayAdjustWindow = map[string]bool{
+	"default_rollup":  true,
+	"first_over_time": true,
+	"last_over_time":  true,
+	"deriv":           true,
+	"deriv_fast":      true,
+	"irate":           true,
+	"rate":            true,
 }
 
 var rollupFuncsRemoveCounterResets = map[string]bool{
@@ -64,6 +78,7 @@ var rollupFuncsKeepMetricGroup = map[string]bool{
 	"max_over_time":      true,
 	"quantile_over_time": true,
 	"rollup":             true,
+	"geomean_over_time":  true,
 }
 
 func getRollupArgIdx(funcName string) int {
@@ -120,6 +135,13 @@ type rollupConfig struct {
 	Step   int64
 	Window int64
 
+	// Whether window may be adjusted to 2 x interval between data points.
+	// This is needed for functions which have dt in the denominator
+	// such as rate, deriv, etc.
+	// Without the adjustement their value would jump in unexpected directions
+	// when using window smaller than 2 x scrape_interval.
+	MayAdjustWindow bool
+
 	Timestamps []int64
 }
 
@@ -162,7 +184,7 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	if window <= 0 {
 		window = rc.Step
 	}
-	if window < maxPrevInterval {
+	if rc.MayAdjustWindow && window < maxPrevInterval {
 		window = maxPrevInterval
 	}
 	rfa := getRollupFuncArg()
@@ -171,8 +193,7 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 
 	i := 0
 	j := 0
-	for _, ts := range rc.Timestamps {
-		tEnd := ts + rc.Step
+	for _, tEnd := range rc.Timestamps {
 		tStart := tEnd - window
 		n := sort.Search(len(timestamps)-i, func(n int) bool {
 			return timestamps[i+n] > tStart
@@ -210,10 +231,27 @@ func getMaxPrevInterval(timestamps []int64) int64 {
 	}
 	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) {
@@ -246,12 +284,14 @@ func deltaValues(values []float64) {
 	if len(values) == 0 {
 		return
 	}
+	prevDelta := float64(0)
 	prevValue := values[0]
 	for i, v := range values[1:] {
-		values[i] = v - prevValue
+		prevDelta = v - prevValue
+		values[i] = prevDelta
 		prevValue = v
 	}
-	values[len(values)-1] = nan
+	values[len(values)-1] = prevDelta
 }
 
 func derivValues(values []float64, timestamps []int64) {
@@ -260,16 +300,23 @@ func derivValues(values []float64, timestamps []int64) {
 	if len(values) == 0 {
 		return
 	}
+	prevDeriv := float64(0)
 	prevValue := values[0]
 	prevTs := timestamps[0]
 	for i, v := range values[1:] {
 		ts := timestamps[i+1]
+		if ts == prevTs {
+			// Use the previous value for duplicate timestamps.
+			values[i] = prevDeriv
+			continue
+		}
 		dt := float64(ts-prevTs) * 1e-3
-		values[i] = (v - prevValue) / dt
+		prevDeriv = (v - prevValue) / dt
+		values[i] = prevDeriv
 		prevValue = v
 		prevTs = ts
 	}
-	values[len(values)-1] = nan
+	values[len(values)-1] = prevDeriv
 }
 
 type newRollupFunc func(args []interface{}) (rollupFunc, error)
@@ -296,11 +343,11 @@ func newRollupHoltWinters(args []interface{}) (rollupFunc, error) {
 		return nil, err
 	}
 	rf := func(rfa *rollupFuncArg) float64 {
-		// There is no need in handling NaNs here, since they must be cleanup up
+		// There is no need in handling NaNs here, since they must be cleaned up
 		// before calling rollup funcs.
 		values := rfa.values
 		if len(values) == 0 {
-			return nan
+			return rfa.prevValue
 		}
 		sf := sfs[rfa.idx]
 		if sf <= 0 || sf >= 1 {
@@ -342,41 +389,55 @@ func newRollupPredictLinear(args []interface{}) (rollupFunc, error) {
 		return nil, err
 	}
 	rf := func(rfa *rollupFuncArg) float64 {
-		// There is no need in handling NaNs here, since they must be cleanup up
-		// before calling rollup funcs.
-		values := rfa.values
-		timestamps := rfa.timestamps
-		if len(values) == 0 {
+		v, k := linearRegression(rfa)
+		if math.IsNaN(v) {
 			return nan
 		}
-
-		// See https://en.wikipedia.org/wiki/Simple_linear_regression#Numerical_example
-		// TODO: determine whether this shit really works.
-		tFirst := rfa.prevTimestamp
-		vSum := rfa.prevValue
-		if math.IsNaN(rfa.prevValue) {
-			tFirst = timestamps[0]
-			vSum = 0
-		}
-		tSum := float64(0)
-		tvSum := float64(0)
-		ttSum := float64(0)
-		for i, v := range values {
-			dt := float64(timestamps[i]-tFirst) * 1e-3
-			vSum += v
-			tSum += dt
-			tvSum += dt * v
-			ttSum += dt * dt
-		}
-		n := float64(len(values))
-		k := (n*tvSum - tSum*vSum) / (n*ttSum - tSum*tSum)
-		v := (vSum - k*tSum) / n
 		sec := secs[rfa.idx]
 		return v + k*sec
 	}
 	return rf, nil
 }
 
+func linearRegression(rfa *rollupFuncArg) (float64, float64) {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	timestamps := rfa.timestamps
+	if len(values) == 0 {
+		return rfa.prevValue, 0
+	}
+
+	// See https://en.wikipedia.org/wiki/Simple_linear_regression#Numerical_example
+	tFirst := rfa.prevTimestamp
+	vSum := rfa.prevValue
+	tSum := float64(0)
+	tvSum := float64(0)
+	ttSum := float64(0)
+	n := 1.0
+	if math.IsNaN(rfa.prevValue) {
+		tFirst = timestamps[0]
+		vSum = 0
+		n = 0
+	}
+	for i, v := range values {
+		dt := float64(timestamps[i]-tFirst) * 1e-3
+		vSum += v
+		tSum += dt
+		tvSum += dt * v
+		ttSum += dt * dt
+	}
+	n += float64(len(values))
+	if n == 1 {
+		return vSum, 0
+	}
+	k := (n*tvSum - tSum*vSum) / (n*ttSum - tSum*tSum)
+	v := (vSum - k*tSum) / n
+	// Adjust v to the last timestamp on the given time range.
+	v += k * (float64(timestamps[len(timestamps)-1]-tFirst) * 1e-3)
+	return v, k
+}
+
 func newRollupQuantile(args []interface{}) (rollupFunc, error) {
 	if err := expectRollupArgsNum(args, 2); err != nil {
 		return nil, err
@@ -386,11 +447,15 @@ func newRollupQuantile(args []interface{}) (rollupFunc, error) {
 		return nil, err
 	}
 	rf := func(rfa *rollupFuncArg) float64 {
-		// There is no need in handling NaNs here, since they must be cleanup up
+		// There is no need in handling NaNs here, since they must be cleaned up
 		// before calling rollup funcs.
 		values := rfa.values
 		if len(values) == 0 {
-			return nan
+			return rfa.prevValue
+		}
+		if len(values) == 1 {
+			// Fast path - only a single value.
+			return values[0]
 		}
 		hf := histogram.GetFast()
 		for _, v := range values {
@@ -408,11 +473,11 @@ func rollupAvg(rfa *rollupFuncArg) float64 {
 	// Do not use `Rapid calculation methods` at https://en.wikipedia.org/wiki/Standard_deviation,
 	// since it is slower and has no significant benefits in precision.
 
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue
 	}
 	var sum float64
 	for _, v := range values {
@@ -422,11 +487,11 @@ func rollupAvg(rfa *rollupFuncArg) float64 {
 }
 
 func rollupMin(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue
 	}
 	minValue := values[0]
 	for _, v := range values {
@@ -438,11 +503,11 @@ func rollupMin(rfa *rollupFuncArg) float64 {
 }
 
 func rollupMax(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue
 	}
 	maxValue := values[0]
 	for _, v := range values {
@@ -454,11 +519,11 @@ func rollupMax(rfa *rollupFuncArg) float64 {
 }
 
 func rollupSum(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue
 	}
 	var sum float64
 	for _, v := range values {
@@ -467,12 +532,43 @@ func rollupSum(rfa *rollupFuncArg) float64 {
 	return sum
 }
 
-func rollupCount(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+func rollupSum2(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue * rfa.prevValue
+	}
+	var sum2 float64
+	for _, v := range values {
+		sum2 += v * v
+	}
+	return sum2
+}
+
+func rollupGeomean(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	if len(values) == 0 {
+		return rfa.prevValue
+	}
+	p := 1.0
+	for _, v := range values {
+		p *= v
+	}
+	return math.Pow(p, 1/float64(len(values)))
+}
+
+func rollupCount(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	if len(values) == 0 {
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	return float64(len(values))
 }
@@ -485,11 +581,18 @@ func rollupStddev(rfa *rollupFuncArg) float64 {
 func rollupStdvar(rfa *rollupFuncArg) float64 {
 	// See `Rapid calculation methods` at https://en.wikipedia.org/wiki/Standard_deviation
 
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
+	}
+	if len(values) == 1 {
+		// Fast path.
+		return values[0]
 	}
 	var avg float64
 	var count float64
@@ -504,7 +607,7 @@ func rollupStdvar(rfa *rollupFuncArg) float64 {
 }
 
 func rollupDelta(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	prevValue := rfa.prevValue
@@ -516,32 +619,42 @@ func rollupDelta(rfa *rollupFuncArg) float64 {
 		values = values[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		return 0
 	}
 	return values[len(values)-1] - prevValue
 }
 
 func rollupIdelta(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	lastValue := values[len(values)-1]
 	values = values[:len(values)-1]
 	if len(values) == 0 {
 		prevValue := rfa.prevValue
 		if math.IsNaN(prevValue) {
-			return nan
+			return 0
 		}
 		return lastValue - prevValue
 	}
 	return lastValue - values[len(values)-1]
 }
 
-func rollupDeriv(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+func rollupDerivSlow(rfa *rollupFuncArg) float64 {
+	// Use linear regression like Prometheus does.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/73
+	_, k := linearRegression(rfa)
+	return k
+}
+
+func rollupDerivFast(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	timestamps := rfa.timestamps
@@ -557,7 +670,7 @@ func rollupDeriv(rfa *rollupFuncArg) float64 {
 		timestamps = timestamps[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		return 0
 	}
 	vEnd := values[len(values)-1]
 	tEnd := timestamps[len(timestamps)-1]
@@ -567,43 +680,54 @@ func rollupDeriv(rfa *rollupFuncArg) float64 {
 }
 
 func rollupIderiv(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	timestamps := rfa.timestamps
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	vEnd := values[len(values)-1]
 	tEnd := timestamps[len(timestamps)-1]
 	values = values[:len(values)-1]
 	timestamps = timestamps[:len(timestamps)-1]
-	prevValue := rfa.prevValue
-	prevTimestamp := rfa.prevTimestamp
-	if len(values) == 0 {
-		if math.IsNaN(prevValue) {
-			return nan
-		}
-	} else {
-		prevValue = values[len(values)-1]
-		prevTimestamp = timestamps[len(timestamps)-1]
+	// Skip data points with duplicate timestamps.
+	for len(timestamps) > 0 && timestamps[len(timestamps)-1] >= tEnd {
+		timestamps = timestamps[:len(timestamps)-1]
 	}
-	dv := vEnd - prevValue
-	dt := tEnd - prevTimestamp
+	var tStart int64
+	var vStart float64
+	if len(timestamps) == 0 {
+		if math.IsNaN(rfa.prevValue) {
+			return 0
+		}
+		tStart = rfa.prevTimestamp
+		vStart = rfa.prevValue
+	} else {
+		tStart = timestamps[len(timestamps)-1]
+		vStart = values[len(timestamps)-1]
+	}
+	dv := vEnd - vStart
+	dt := tEnd - tStart
 	return dv / (float64(dt) / 1000)
 }
 
 func rollupChanges(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
-	if len(values) == 0 {
-		return nan
-	}
-	n := 0
 	prevValue := rfa.prevValue
+	n := 0
 	if math.IsNaN(prevValue) {
+		if len(values) == 0 {
+			return nan
+		}
 		prevValue = values[0]
+		values = values[1:]
+		n++
 	}
 	for _, v := range values {
 		if v != prevValue {
@@ -615,11 +739,14 @@ func rollupChanges(rfa *rollupFuncArg) float64 {
 }
 
 func rollupResets(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	prevValue := rfa.prevValue
 	if math.IsNaN(prevValue) {
@@ -627,7 +754,7 @@ func rollupResets(rfa *rollupFuncArg) float64 {
 		values = values[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		return 0
 	}
 	n := 0
 	for _, v := range values {
@@ -646,7 +773,7 @@ func rollupFirst(rfa *rollupFuncArg) float64 {
 		return v
 	}
 
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
@@ -655,24 +782,27 @@ func rollupFirst(rfa *rollupFuncArg) float64 {
 	return values[0]
 }
 
-var rollupDefault = rollupFirst
+var rollupDefault = rollupLast
 
 func rollupLast(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		return rfa.prevValue
 	}
 	return values[len(values)-1]
 }
 
 func rollupDistinct(rfa *rollupFuncArg) float64 {
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	m := make(map[float64]struct{})
 	for _, v := range values {
@@ -684,12 +814,15 @@ func rollupDistinct(rfa *rollupFuncArg) float64 {
 func rollupIntegrate(rfa *rollupFuncArg) float64 {
 	prevTimestamp := rfa.prevTimestamp
 
-	// There is no need in handling NaNs here, since they must be cleanup up
+	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
 	values := rfa.values
 	timestamps := rfa.timestamps
 	if len(values) == 0 {
-		return nan
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	prevValue := rfa.prevValue
 	if math.IsNaN(prevValue) {
@@ -699,7 +832,7 @@ func rollupIntegrate(rfa *rollupFuncArg) float64 {
 		timestamps = timestamps[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		return 0
 	}
 
 	var sum float64

app/vmselect/promql/rollup_result_cache.go

@@ -2,6 +2,7 @@ package promql
 
 import (
 	"crypto/rand"
+	"flag"
 	"fmt"
 	"runtime"
 	"sync"
@@ -15,6 +16,8 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
+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.
 }
@@ -47,6 +50,10 @@ func InitRollupResultCache(cachePath string) {
 	} else {
 		c = fastcache.New(getRollupResultCacheSize())
 	}
+	if *disableCache {
+		c.Reset()
+	}
+
 	stats := &fastcache.Stats{}
 	var statsLock sync.Mutex
 	var statsLastUpdate time.Time
@@ -64,7 +71,7 @@ func InitRollupResultCache(cachePath string) {
 		return stats
 	}
 	if len(rollupResultCachePath) > 0 {
-		logger.Infof("loaded rollupResult cache from %q in %s; entriesCount: %d, bytesSize: %d",
+		logger.Infof("loaded rollupResult cache from %q in %s; entriesCount: %d, sizeBytes: %d",
 			rollupResultCachePath, time.Since(startTime), fcs().EntriesCount, fcs().BytesSize)
 	}
 
@@ -101,7 +108,7 @@ func StopRollupResultCache() {
 		var fcs fastcache.Stats
 		rollupResultCacheV.c.UpdateStats(&fcs)
 		rollupResultCacheV.c.Reset()
-		logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, bytesSize: %d",
+		logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, sizeBytes: %d",
 			rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
 	}
 }
@@ -118,8 +125,8 @@ func ResetRollupResultCache() {
 	rollupResultCacheV.c.Reset()
 }
 
-func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExpr, window int64) (tss []*timeseries, newStart int64) {
-	if !ec.mayCache() {
+func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64) (tss []*timeseries, newStart int64) {
+	if *disableCache || !ec.mayCache() {
 		return nil, ec.Start
 	}
 
@@ -127,7 +134,7 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	bb := bbPool.Get()
 	defer bbPool.Put(bb)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	if len(metainfoBuf) == 0 {
 		return nil, ec.Start
@@ -145,7 +152,7 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	if len(resultBuf) == 0 {
 		mi.RemoveKey(key)
 		metainfoBuf = mi.Marshal(metainfoBuf[:0])
-		bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, window, ec.Step)
+		bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
 		rrc.c.Set(bb.B, metainfoBuf)
 		return nil, ec.Start
 	}
@@ -189,8 +196,8 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	return tss, newStart
 }
 
-func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, window int64, tss []*timeseries) {
-	if len(tss) == 0 || !ec.mayCache() {
+func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64, tss []*timeseries) {
+	if *disableCache || len(tss) == 0 || !ec.mayCache() {
 		return
 	}
 
@@ -235,7 +242,7 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 	bb.B = key.Marshal(bb.B[:0])
 	rrc.c.SetBig(bb.B, tssMarshaled)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	var mi rollupResultCacheMetainfo
 	if len(metainfoBuf) > 0 {
@@ -263,10 +270,16 @@ var (
 var tooBigRollupResults = metrics.NewCounter("vm_too_big_rollup_results_total")
 
 // Increment this value every time the format of the cache changes.
-const rollupResultCacheVersion = 4
+const rollupResultCacheVersion = 5
 
-func marshalRollupResultCacheKey(dst []byte, funcName string, me *metricExpr, window, step int64) []byte {
+func marshalRollupResultCacheKey(dst []byte, funcName string, me *metricExpr, iafc *incrementalAggrFuncContext, window, step int64) []byte {
 	dst = append(dst, rollupResultCacheVersion)
+	if iafc == nil {
+		dst = append(dst, 0)
+	} else {
+		dst = append(dst, 1)
+		dst = iafc.ae.AppendString(dst)
+	}
 	dst = encoding.MarshalUint64(dst, uint64(len(funcName)))
 	dst = append(dst, funcName...)
 	dst = encoding.MarshalInt64(dst, window)

app/vmselect/promql/rollup_result_cache_test.go

@@ -23,10 +23,15 @@ func TestRollupResultCache(t *testing.T) {
 			Value: []byte("xxx"),
 		}},
 	}
+	iafc := &incrementalAggrFuncContext{
+		ae: &aggrFuncExpr{
+			Name: "foobar",
+		},
+	}
 
 	// Try obtaining an empty value.
 	t.Run("empty", func(t *testing.T) {
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != ec.Start {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, ec.Start)
 		}
@@ -36,21 +41,42 @@ func TestRollupResultCache(t *testing.T) {
 	})
 
 	// Store timeseries overlapping with start
-	t.Run("start-overlap", func(t *testing.T) {
+	t.Run("start-overlap-no-iafc", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{800, 1000, 1200},
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1400 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
 		}
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
+				Timestamps: []int64{1000, 1200},
+				Values:     []float64{1, 2},
+			},
+		}
+		testTimeseriesEqual(t, tss, tssExpected)
+	})
+	t.Run("start-overlap-with-iafc", func(t *testing.T) {
+		ResetRollupResultCache()
+		tss := []*timeseries{
+			{
+				Timestamps: []int64{800, 1000, 1200},
+				Values:     []float64{0, 1, 2},
+			},
+		}
+		rollupResultCacheV.Put(funcName, ec, me, iafc, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, iafc, window)
+		if newStart != 1400 {
+			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
+		}
+		tssExpected := []*timeseries{
+			{
 				Timestamps: []int64{1000, 1200},
 				Values:     []float64{1, 2},
 			},
@@ -62,13 +88,13 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("end-overlap", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1800, 2000, 2200, 2400},
 				Values:     []float64{333, 0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -81,13 +107,13 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("full-cover", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1200, 1400, 1600},
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -100,13 +126,13 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("before-start", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{200, 400, 600},
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -119,13 +145,13 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("after-end", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{2200, 2400, 2600},
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1000 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1000)
 		}
@@ -138,18 +164,18 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("bigger-than-start-end", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{800, 1000, 1200, 1400, 1600, 1800, 2000, 2200},
 				Values:     []float64{0, 1, 2, 3, 4, 5, 6, 7},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
@@ -161,18 +187,18 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("start-end-match", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
@@ -191,8 +217,8 @@ func TestRollupResultCache(t *testing.T) {
 			}
 			tss = append(tss, ts)
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss)
-		tssResult, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss)
+		tssResult, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 2200 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 2200)
 		}
@@ -203,32 +229,32 @@ func TestRollupResultCache(t *testing.T) {
 	t.Run("multi-timeseries", func(t *testing.T) {
 		ResetRollupResultCache()
 		tss1 := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{800, 1000, 1200},
 				Values:     []float64{0, 1, 2},
 			},
 		}
 		tss2 := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1800, 2000, 2200, 2400},
 				Values:     []float64{333, 0, 1, 2},
 			},
 		}
 		tss3 := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1200, 1400, 1600},
 				Values:     []float64{0, 1, 2},
 			},
 		}
-		rollupResultCacheV.Put(funcName, ec, me, window, tss1)
-		rollupResultCacheV.Put(funcName, ec, me, window, tss2)
-		rollupResultCacheV.Put(funcName, ec, me, window, tss3)
-		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, window)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss1)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss2)
+		rollupResultCacheV.Put(funcName, ec, me, nil, window, tss3)
+		tss, newStart := rollupResultCacheV.Get(funcName, ec, me, nil, window)
 		if newStart != 1400 {
 			t.Fatalf("unexpected newStart; got %d; want %d", newStart, 1400)
 		}
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200},
 				Values:     []float64{1, 2},
 			},
@@ -249,14 +275,14 @@ func TestMergeTimeseries(t *testing.T) {
 	t.Run("bStart=ec.Start", func(t *testing.T) {
 		a := []*timeseries{}
 		b := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
 		}
 		tss := mergeTimeseries(a, b, 1000, ec)
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{1, 2, 3, 4, 5, 6},
 			},
@@ -266,14 +292,14 @@ func TestMergeTimeseries(t *testing.T) {
 	t.Run("a-empty", func(t *testing.T) {
 		a := []*timeseries{}
 		b := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1400, 1600, 1800, 2000},
 				Values:     []float64{3, 4, 5, 6},
 			},
 		}
 		tss := mergeTimeseries(a, b, bStart, ec)
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{nan, nan, 3, 4, 5, 6},
 			},
@@ -282,7 +308,7 @@ func TestMergeTimeseries(t *testing.T) {
 	})
 	t.Run("b-empty", func(t *testing.T) {
 		a := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200},
 				Values:     []float64{2, 1},
 			},
@@ -290,7 +316,7 @@ func TestMergeTimeseries(t *testing.T) {
 		b := []*timeseries{}
 		tss := mergeTimeseries(a, b, bStart, ec)
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{2, 1, nan, nan, nan, nan},
 			},
@@ -299,20 +325,20 @@ func TestMergeTimeseries(t *testing.T) {
 	})
 	t.Run("non-empty", func(t *testing.T) {
 		a := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200},
 				Values:     []float64{2, 1},
 			},
 		}
 		b := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1400, 1600, 1800, 2000},
 				Values:     []float64{3, 4, 5, 6},
 			},
 		}
 		tss := mergeTimeseries(a, b, bStart, ec)
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{2, 1, 3, 4, 5, 6},
 			},
@@ -321,14 +347,14 @@ func TestMergeTimeseries(t *testing.T) {
 	})
 	t.Run("non-empty-distinct-metric-names", func(t *testing.T) {
 		a := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1000, 1200},
 				Values:     []float64{2, 1},
 			},
 		}
 		a[0].MetricName.MetricGroup = []byte("bar")
 		b := []*timeseries{
-			&timeseries{
+			{
 				Timestamps: []int64{1400, 1600, 1800, 2000},
 				Values:     []float64{3, 4, 5, 6},
 			},
@@ -336,14 +362,14 @@ func TestMergeTimeseries(t *testing.T) {
 		b[0].MetricName.MetricGroup = []byte("foo")
 		tss := mergeTimeseries(a, b, bStart, ec)
 		tssExpected := []*timeseries{
-			&timeseries{
+			{
 				MetricName: storage.MetricName{
 					MetricGroup: []byte("foo"),
 				},
 				Timestamps: []int64{1000, 1200, 1400, 1600, 1800, 2000},
 				Values:     []float64{nan, nan, 3, 4, 5, 6},
 			},
-			&timeseries{
+			{
 				MetricName: storage.MetricName{
 					MetricGroup: []byte("bar"),
 				},

app/vmselect/promql/rollup_test.go

@@ -10,6 +10,68 @@ var (
 	testTimestamps = []int64{5, 15, 24, 36, 49, 60, 78, 80, 97, 115, 120, 130}
 )
 
+func TestRollupIderivDuplicateTimestamps(t *testing.T) {
+	rfa := &rollupFuncArg{
+		values:     []float64{1, 2, 3, 4, 5},
+		timestamps: []int64{100, 100, 200, 300, 300},
+	}
+	n := rollupIderiv(rfa)
+	if n != 20 {
+		t.Fatalf("unexpected value; got %v; want %v", n, 20)
+	}
+
+	rfa = &rollupFuncArg{
+		values:     []float64{1, 2, 3, 4, 5},
+		timestamps: []int64{100, 100, 300, 300, 300},
+	}
+	n = rollupIderiv(rfa)
+	if n != 15 {
+		t.Fatalf("unexpected value; got %v; want %v", n, 15)
+	}
+
+	rfa = &rollupFuncArg{
+		prevValue:  nan,
+		values:     []float64{},
+		timestamps: []int64{},
+	}
+	n = rollupIderiv(rfa)
+	if !math.IsNaN(n) {
+		t.Fatalf("unexpected value; got %v; want %v", n, nan)
+	}
+
+	rfa = &rollupFuncArg{
+		prevValue:  nan,
+		values:     []float64{15},
+		timestamps: []int64{100},
+	}
+	n = rollupIderiv(rfa)
+	if n != 0 {
+		t.Fatalf("unexpected value; got %v; want %v", n, 0)
+	}
+
+	rfa = &rollupFuncArg{
+		prevTimestamp: 100,
+		prevValue:     10,
+		values:        []float64{15},
+		timestamps:    []int64{100},
+	}
+	n = rollupIderiv(rfa)
+	if n != inf {
+		t.Fatalf("unexpected value; got %v; want %v", n, inf)
+	}
+
+	rfa = &rollupFuncArg{
+		prevTimestamp: 100,
+		prevValue:     10,
+		values:        []float64{15, 20},
+		timestamps:    []int64{100, 100},
+	}
+	n = rollupIderiv(rfa)
+	if n != inf {
+		t.Fatalf("unexpected value; got %v; want %v", n, inf)
+	}
+}
+
 func TestRemoveCounterResets(t *testing.T) {
 	removeCounterResets(nil)
 
@@ -38,19 +100,19 @@ func TestDeltaValues(t *testing.T) {
 
 	values := []float64{123}
 	deltaValues(values)
-	valuesExpected := []float64{nan}
+	valuesExpected := []float64{0}
 	testRowsEqual(t, values, testTimestamps[:1], valuesExpected, testTimestamps[:1])
 
 	values = append([]float64{}, testValues...)
 	deltaValues(values)
-	valuesExpected = []float64{-89, 10, -23, 33, -20, 65, -87, 32, -12, 2, 0, nan}
+	valuesExpected = []float64{-89, 10, -23, 33, -20, 65, -87, 32, -12, 2, 0, 0}
 	testRowsEqual(t, values, testTimestamps, valuesExpected, testTimestamps)
 
 	// remove counter resets
 	values = append([]float64{}, testValues...)
 	removeCounterResets(values)
 	deltaValues(values)
-	valuesExpected = []float64{34, 10, 21, 33, 34, 65, 12, 32, 32, 2, 0, nan}
+	valuesExpected = []float64{34, 10, 21, 33, 34, 65, 12, 32, 32, 2, 0, 0}
 	testRowsEqual(t, values, testTimestamps, valuesExpected, testTimestamps)
 }
 
@@ -59,13 +121,13 @@ func TestDerivValues(t *testing.T) {
 
 	values := []float64{123}
 	derivValues(values, testTimestamps[:1])
-	valuesExpected := []float64{nan}
+	valuesExpected := []float64{0}
 	testRowsEqual(t, values, testTimestamps[:1], valuesExpected, testTimestamps[:1])
 
 	values = append([]float64{}, testValues...)
 	derivValues(values, testTimestamps)
 	valuesExpected = []float64{-8900, 1111.111111111111, -1916.6666666666665, 2538.461538461538, -1818.1818181818182, 3611.111111111111,
-		-43500, 1882.3529411764705, -666.6666666666666, 400, 0, nan}
+		-43500, 1882.3529411764705, -666.6666666666666, 400, 0, 0}
 	testRowsEqual(t, values, testTimestamps, valuesExpected, testTimestamps)
 
 	// remove counter resets
@@ -73,8 +135,15 @@ func TestDerivValues(t *testing.T) {
 	removeCounterResets(values)
 	derivValues(values, testTimestamps)
 	valuesExpected = []float64{3400, 1111.111111111111, 1750, 2538.461538461538, 3090.909090909091, 3611.111111111111,
-		6000, 1882.3529411764705, 1777.7777777777776, 400, 0, nan}
+		6000, 1882.3529411764705, 1777.7777777777776, 400, 0, 0}
 	testRowsEqual(t, values, testTimestamps, valuesExpected, testTimestamps)
+
+	// duplicate timestamps
+	values = []float64{1, 2, 3, 4, 5, 6, 7}
+	timestamps := []int64{100, 100, 200, 200, 300, 400, 400}
+	derivValues(values, timestamps)
+	valuesExpected = []float64{0, 20, 20, 20, 10, 10, 10}
+	testRowsEqual(t, values, timestamps, valuesExpected, timestamps)
 }
 
 func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpected *metricExpr, vExpected float64) {
@@ -143,10 +212,10 @@ func TestRollupPredictLinear(t *testing.T) {
 		testRollupFunc(t, "predict_linear", args, &me, vExpected)
 	}
 
-	f(0e-3, 63.739757761102624)
-	f(50e-3, 50.39682764539959)
-	f(100e-3, 37.053897529696556)
-	f(200e-3, 10.368037298290488)
+	f(0e-3, 30.382432471845043)
+	f(50e-3, 17.03950235614201)
+	f(100e-3, 3.696572240438975)
+	f(200e-3, -22.989287990967092)
 }
 
 func TestRollupHoltWinters(t *testing.T) {
@@ -189,10 +258,11 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
 		testRollupFunc(t, funcName, args, &me, vExpected)
 	}
 
-	f("default_rollup", 123)
-	f("changes", 10)
+	f("default_rollup", 34)
+	f("changes", 11)
 	f("delta", -89)
-	f("deriv", -712)
+	f("deriv", -266.85860231406065)
+	f("deriv_fast", -712)
 	f("idelta", 0)
 	f("increase", 275)
 	f("irate", 0)
@@ -202,12 +272,16 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f("min_over_time", 12)
 	f("max_over_time", 123)
 	f("sum_over_time", 565)
+	f("sum2_over_time", 37951)
+	f("geomean_over_time", 39.33466603189148)
 	f("count_over_time", 12)
 	f("stddev_over_time", 30.752935722554287)
 	f("stdvar_over_time", 945.7430555555555)
 	f("first_over_time", 123)
 	f("last_over_time", 34)
 	f("integrate", 61.0275)
+	f("distinct_over_time", 8)
+	f("ideriv", 0)
 }
 
 func TestRollupNewRollupFuncError(t *testing.T) {
@@ -259,7 +333,7 @@ func TestRollupNoWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, nan, nan, nan, 123}
+		valuesExpected := []float64{nan, nan, nan, nan, nan}
 		timestampsExpected := []int64{0, 1, 2, 3, 4}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -267,14 +341,14 @@ func TestRollupNoWindowNoPoints(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupDelta,
 			Start:  120,
-			End:    144,
+			End:    148,
 			Step:   4,
 			Window: 0,
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{2, 2, 0, 0, 0, nan, nan}
-		timestampsExpected := []int64{120, 124, 128, 132, 136, 140, 144}
+		valuesExpected := []float64{2, 0, 0, 0, 0, 0, 0, 0}
+		timestampsExpected := []int64{120, 124, 128, 132, 136, 140, 144, 148}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 }
@@ -290,22 +364,22 @@ func TestRollupWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, nan, nan, nan, 123}
+		valuesExpected := []float64{nan, nan, nan, nan, nan}
 		timestampsExpected := []int64{0, 1, 2, 3, 4}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 	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{34, 34, 34, nan}
+		timestampsExpected := []int64{161, 171, 181, 191}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 }
@@ -315,14 +389,14 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupFirst,
 			Start:  0,
-			End:    20,
+			End:    25,
 			Step:   5,
 			Window: 0,
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{123, 123, 123, 123, 123}
-		timestampsExpected := []int64{0, 5, 10, 15, 20}
+		valuesExpected := []float64{nan, 123, 123, 123, 34, 34}
+		timestampsExpected := []int64{0, 5, 10, 15, 20, 25}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 	t.Run("afterEnd", func(t *testing.T) {
@@ -335,7 +409,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{44, 34, 34, nan}
+		valuesExpected := []float64{12, 44, 34, 34}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -349,7 +423,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 123, 54, 44, nan}
+		valuesExpected := []float64{nan, nan, 123, 54, 44}
 		timestampsExpected := []int64{-50, 0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -366,7 +440,7 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{123, 123, 34, 34, 44}
+		valuesExpected := []float64{nan, 123, 123, 34, 34}
 		timestampsExpected := []int64{0, 5, 10, 15, 20}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -380,7 +454,7 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{34, 34, nan, nan}
+		valuesExpected := []float64{44, 34, 34, 34}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -394,7 +468,7 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{54, 44, nan, nan}
+		valuesExpected := []float64{nan, 54, 44, 34}
 		timestampsExpected := []int64{0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -411,7 +485,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{123, 21, 12, 34, nan}
+		valuesExpected := []float64{nan, 123, 21, 12, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -425,7 +499,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{4, 4, 3, 1, nan}
+		valuesExpected := []float64{nan, 4, 4, 3, 1}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -439,7 +513,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{21, 12, 32, 34, nan}
+		valuesExpected := []float64{nan, 21, 12, 32, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -453,7 +527,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{123, 99, 44, 34, nan}
+		valuesExpected := []float64{nan, 123, 99, 44, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -467,7 +541,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{222, 199, 110, 34, nan}
+		valuesExpected := []float64{nan, 222, 199, 110, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -481,7 +555,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{-102, -9, 22, 0, nan}
+		valuesExpected := []float64{nan, -102, -9, 22, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -495,7 +569,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{33, -87, 0, nan}
+		valuesExpected := []float64{0, 33, -87, 0}
 		timestampsExpected := []int64{10, 50, 90, 130}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -509,10 +583,24 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{3, 4, 3, 0, nan}
+		valuesExpected := []float64{nan, 4, 4, 3, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("changes_small_window", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupChanges,
+			Start:  0,
+			End:    45,
+			Step:   9,
+			Window: 9,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 1, 1, 1, 1, 0}
+		timestampsExpected := []int64{0, 9, 18, 27, 36, 45}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("resets", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupResets,
@@ -523,7 +611,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{2, 2, 1, 0, nan}
+		valuesExpected := []float64{nan, 2, 2, 1, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -537,13 +625,13 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{55.5, 49.75, 36.666666666666664, 34, nan}
+		valuesExpected := []float64{nan, 55.5, 49.75, 36.666666666666664, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 	t.Run("deriv", func(t *testing.T) {
 		rc := rollupConfig{
-			Func:   rollupDeriv,
+			Func:   rollupDerivSlow,
 			Start:  0,
 			End:    160,
 			Step:   40,
@@ -551,7 +639,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{-3290.3225806451615, -204.54545454545456, 550, 0, nan}
+		valuesExpected := []float64{0, -2879.310344827587, 558.0608793686592, 422.84569138276544, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -565,7 +653,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{-1916.6666666666665, -43500, 400, 0, nan}
+		valuesExpected := []float64{nan, -1916.6666666666665, -43500, 400, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -579,7 +667,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{39.81519810323691, 32.080952292598795, 5.2493385826745405, 0, nan}
+		valuesExpected := []float64{nan, 39.81519810323691, 32.080952292598795, 5.2493385826745405, 5.830951894845301}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -593,7 +681,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{4.6035, 4.3934999999999995, 2.166, 0.34, nan}
+		valuesExpected := []float64{nan, 4.6035, 4.3934999999999995, 2.166, 0.34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -607,7 +695,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{4, 4, 3, 1, nan}
+		valuesExpected := []float64{nan, 4, 4, 3, 1}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})

app/vmselect/promql/timeseries.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"sort"
 	"strconv"
+	"sync"
 	"unsafe"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
@@ -38,11 +39,13 @@ func (ts *timeseries) String() string {
 	return fmt.Sprintf("MetricName=%s, Values=%g, Timestamps=%d", &ts.MetricName, ts.Values, ts.Timestamps)
 }
 
-func (ts *timeseries) CopyFrom(src *timeseries) {
+func (ts *timeseries) CopyFromShallowTimestamps(src *timeseries) {
 	ts.Reset()
 	ts.MetricName.CopyFrom(&src.MetricName)
 	ts.Values = append(ts.Values[:0], src.Values...)
-	ts.Timestamps = append(ts.Timestamps[:0], src.Timestamps...)
+	ts.Timestamps = src.Timestamps
+
+	ts.denyReuse = true
 }
 
 func (ts *timeseries) CopyFromMetricNames(src *timeseries) {
@@ -59,6 +62,20 @@ func (ts *timeseries) CopyShallow(src *timeseries) {
 	ts.denyReuse = true
 }
 
+func getTimeseries() *timeseries {
+	if v := timeseriesPool.Get(); v != nil {
+		return v.(*timeseries)
+	}
+	return &timeseries{}
+}
+
+func putTimeseries(ts *timeseries) {
+	ts.Reset()
+	timeseriesPool.Put(ts)
+}
+
+var timeseriesPool sync.Pool
+
 func marshalTimeseriesFast(tss []*timeseries, maxSize int, step int64) []byte {
 	if len(tss) == 0 {
 		logger.Panicf("BUG: tss cannot be empty")

app/vmselect/promql/transform.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"math"
 	"math/rand"
+	"regexp"
 	"sort"
 	"strconv"
 	"strings"
@@ -61,6 +62,7 @@ var transformFuncs = map[string]transformFunc{
 	"label_keep":         transformLabelKeep,
 	"label_copy":         transformLabelCopy,
 	"label_move":         transformLabelMove,
+	"label_transform":    transformLabelTransform,
 	"union":              transformUnion,
 	"":                   transformUnion, // empty func is a synonim to union
 	"keep_last_value":    transformKeepLastValue,
@@ -123,7 +125,8 @@ func newTransformFuncOneArg(tf func(v float64) float64) transformFunc {
 }
 
 func doTransformValues(arg []*timeseries, tf func(values []float64), fe *funcExpr) ([]*timeseries, error) {
-	keepMetricGroup := transformFuncsKeepMetricGroup[fe.Name]
+	name := strings.ToLower(fe.Name)
+	keepMetricGroup := transformFuncsKeepMetricGroup[name]
 	for _, ts := range arg {
 		if !keepMetricGroup {
 			ts.MetricName.ResetMetricGroup()
@@ -294,14 +297,12 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 		if err != nil {
 			continue
 		}
-		var dst timeseries
-		dst.CopyFrom(ts)
-		dst.MetricName.ResetMetricGroup()
-		dst.MetricName.RemoveTag("le")
-		bb.B = marshalMetricTagsSorted(bb.B[:0], &dst.MetricName)
+		ts.MetricName.ResetMetricGroup()
+		ts.MetricName.RemoveTag("le")
+		bb.B = marshalMetricTagsSorted(bb.B[:0], &ts.MetricName)
 		m[string(bb.B)] = append(m[string(bb.B)], x{
 			le: le,
-			ts: &dst,
+			ts: ts,
 		})
 	}
 	bbPool.Put(bb)
@@ -323,6 +324,14 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 		if math.IsNaN(phi) {
 			return nan
 		}
+		// Verify for broken buckets with NaN or negative values.
+		for _, xs := range xss {
+			v := xs.ts.Values[i]
+			if math.IsNaN(v) || v < 0 {
+				// Broken bucket.
+				return nan
+			}
+		}
 		if phi < 0 {
 			return -inf
 		}
@@ -333,10 +342,6 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 		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
@@ -394,13 +399,6 @@ func runningAvg(a, b float64, idx int) float64 {
 	return a + (b-a)/float64(idx+1)
 }
 
-func keepLastValue(a, b float64, idx int) float64 {
-	if math.IsNaN(b) {
-		return a
-	}
-	return b
-}
-
 func skipLeadingNaNs(values []float64) []float64 {
 	i := 0
 	for i < len(values) && math.IsNaN(values[i]) {
@@ -641,9 +639,7 @@ func transformUnion(tfa *transformFuncArg) ([]*timeseries, error) {
 				continue
 			}
 			m[string(bb.B)] = true
-			var dst timeseries
-			dst.CopyFrom(ts)
-			rvs = append(rvs, &dst)
+			rvs = append(rvs, ts)
 		}
 	}
 	bbPool.Put(bb)
@@ -816,6 +812,31 @@ func transformLabelJoin(tfa *transformFuncArg) ([]*timeseries, error) {
 	return rvs, nil
 }
 
+func transformLabelTransform(tfa *transformFuncArg) ([]*timeseries, error) {
+	args := tfa.args
+	if err := expectTransformArgsNum(args, 4); err != nil {
+		return nil, err
+	}
+	label, err := getString(args[1], 1)
+	if err != nil {
+		return nil, err
+	}
+	regex, err := getString(args[2], 2)
+	if err != nil {
+		return nil, err
+	}
+	replacement, err := getString(args[3], 3)
+	if err != nil {
+		return nil, err
+	}
+
+	r, err := compileRegexp(regex)
+	if err != nil {
+		return nil, fmt.Errorf(`cannot compile regex %q: %s`, regex, err)
+	}
+	return labelReplace(args[0], label, r, label, replacement)
+}
+
 func transformLabelReplace(tfa *transformFuncArg) ([]*timeseries, error) {
 	args := tfa.args
 	if err := expectTransformArgsNum(args, 5); err != nil {
@@ -842,11 +863,12 @@ func transformLabelReplace(tfa *transformFuncArg) ([]*timeseries, error) {
 	if err != nil {
 		return nil, fmt.Errorf(`cannot compile regex %q: %s`, regex, err)
 	}
+	return labelReplace(args[0], srcLabel, r, dstLabel, replacement)
+}
 
+func labelReplace(tss []*timeseries, srcLabel string, r *regexp.Regexp, dstLabel, replacement string) ([]*timeseries, error) {
 	replacementBytes := []byte(replacement)
-
-	rvs := args[0]
-	for _, ts := range rvs {
+	for _, ts := range tss {
 		mn := &ts.MetricName
 		dstValue := getDstValue(mn, dstLabel)
 		srcValue := mn.GetTagValue(srcLabel)
@@ -856,7 +878,7 @@ func transformLabelReplace(tfa *transformFuncArg) ([]*timeseries, error) {
 			mn.RemoveTag(dstLabel)
 		}
 	}
-	return rvs, nil
+	return tss, nil
 }
 
 func transformLn(v float64) float64 {

app/vmstorage/main.go

@@ -28,11 +28,20 @@ var (
 
 // Init initializes vmstorage.
 func Init() {
+	InitWithoutMetrics()
+	registerStorageMetrics()
+}
+
+// InitWithoutMetrics must be called instead of Init inside tests.
+//
+// This allows multiple Init / Stop cycles.
+func InitWithoutMetrics() {
 	if err := encoding.CheckPrecisionBits(uint8(*precisionBits)); err != nil {
 		logger.Fatalf("invalid `-precisionBits`: %s", err)
 	}
 	logger.Infof("opening storage at %q with retention period %d months", *DataPath, *retentionPeriod)
 	startTime := time.Now()
+	WG = syncwg.WaitGroup{}
 	strg, err := storage.OpenStorage(*DataPath, *retentionPeriod)
 	if err != nil {
 		logger.Fatalf("cannot open a storage at %s with retention period %d months: %s", *DataPath, *retentionPeriod, err)
@@ -45,10 +54,9 @@ func Init() {
 	partsCount := tm.SmallPartsCount + tm.BigPartsCount
 	blocksCount := tm.SmallBlocksCount + tm.BigBlocksCount
 	rowsCount := tm.SmallRowsCount + tm.BigRowsCount
-	logger.Infof("successfully opened storage %q in %s; partsCount: %d; blocksCount: %d; rowsCount: %d",
-		*DataPath, time.Since(startTime), partsCount, blocksCount, rowsCount)
-
-	registerStorageMetrics(Storage)
+	sizeBytes := tm.SmallSizeBytes + tm.BigSizeBytes
+	logger.Infof("successfully opened storage %q in %s; partsCount: %d; blocksCount: %d; rowsCount: %d; sizeBytes: %d",
+		*DataPath, time.Since(startTime), partsCount, blocksCount, rowsCount, sizeBytes)
 }
 
 // Storage is a storage.
@@ -96,6 +104,14 @@ func SearchTagValues(tagKey []byte, maxTagValues int) ([]string, error) {
 	return values, err
 }
 
+// SearchTagEntries searches for tag entries.
+func SearchTagEntries(maxTagKeys, maxTagValues int) ([]storage.TagEntry, error) {
+	WG.Add(1)
+	tagEntries, err := Storage.SearchTagEntries(maxTagKeys, maxTagValues)
+	WG.Done()
+	return tagEntries, err
+}
+
 // GetSeriesCount returns the number of time series in the storage.
 func GetSeriesCount() (uint64, error) {
 	WG.Add(1)
@@ -203,7 +219,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	}
 }
 
-func registerStorageMetrics(strg *storage.Storage) {
+func registerStorageMetrics() {
 	mCache := &storage.Metrics{}
 	var mCacheLock sync.Mutex
 	var lastUpdateTime time.Time
@@ -215,7 +231,7 @@ func registerStorageMetrics(strg *storage.Storage) {
 			return mCache
 		}
 		var mc storage.Metrics
-		strg.UpdateMetrics(&mc)
+		Storage.UpdateMetrics(&mc)
 		mCache = &mc
 		lastUpdateTime = time.Now()
 		return mCache
@@ -285,6 +301,18 @@ func registerStorageMetrics(strg *storage.Storage) {
 	metrics.NewGauge(`vm_missing_tsids_for_metric_id_total`, func() float64 {
 		return float64(idbm().MissingTSIDsForMetricID)
 	})
+	metrics.NewGauge(`vm_recent_hour_metric_ids_search_calls_total`, func() float64 {
+		return float64(idbm().RecentHourMetricIDsSearchCalls)
+	})
+	metrics.NewGauge(`vm_recent_hour_metric_ids_search_hits_total`, func() float64 {
+		return float64(idbm().RecentHourMetricIDsSearchHits)
+	})
+	metrics.NewGauge(`vm_date_metric_ids_search_calls_total`, func() float64 {
+		return float64(idbm().DateMetricIDsSearchCalls)
+	})
+	metrics.NewGauge(`vm_date_metric_ids_search_hits_total`, func() float64 {
+		return float64(idbm().DateMetricIDsSearchHits)
+	})
 
 	metrics.NewGauge(`vm_assisted_merges_total{type="storage/small"}`, func() float64 {
 		return float64(tm().SmallAssistedMerges)
@@ -320,6 +348,39 @@ func registerStorageMetrics(strg *storage.Storage) {
 		return float64(idbm().BlocksCount)
 	})
 
+	metrics.NewGauge(`vm_data_size_bytes{type="storage/big"}`, func() float64 {
+		return float64(tm().BigSizeBytes)
+	})
+	metrics.NewGauge(`vm_data_size_bytes{type="storage/small"}`, func() float64 {
+		return float64(tm().SmallSizeBytes)
+	})
+	metrics.NewGauge(`vm_data_size_bytes{type="indexdb"}`, func() float64 {
+		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)
 	})
@@ -342,6 +403,9 @@ func registerStorageMetrics(strg *storage.Storage) {
 	metrics.NewGauge(`vm_cache_entries{type="storage/date_metricID"}`, func() float64 {
 		return float64(m().DateMetricIDCacheSize)
 	})
+	metrics.NewGauge(`vm_cache_entries{type="storage/hour_metric_ids"}`, func() float64 {
+		return float64(m().HourMetricIDCacheSize)
+	})
 	metrics.NewGauge(`vm_cache_entries{type="storage/bigIndexBlocks"}`, func() float64 {
 		return float64(tm().BigIndexBlocksCacheSize)
 	})
@@ -357,24 +421,30 @@ func registerStorageMetrics(strg *storage.Storage) {
 	metrics.NewGauge(`vm_cache_entries{type="indexdb/tagFilters"}`, func() float64 {
 		return float64(idbm().TagCacheSize)
 	})
+	metrics.NewGauge(`vm_cache_entries{type="indexdb/uselessTagFilters"}`, func() float64 {
+		return float64(idbm().UselessTagFiltersCacheSize)
+	})
 	metrics.NewGauge(`vm_cache_entries{type="storage/regexps"}`, func() float64 {
 		return float64(storage.RegexpCacheSize())
 	})
 
 	metrics.NewGauge(`vm_cache_size_bytes{type="storage/tsid"}`, func() float64 {
-		return float64(m().TSIDCacheBytesSize)
+		return float64(m().TSIDCacheSizeBytes)
 	})
 	metrics.NewGauge(`vm_cache_size_bytes{type="storage/metricIDs"}`, func() float64 {
-		return float64(m().MetricIDCacheBytesSize)
+		return float64(m().MetricIDCacheSizeBytes)
 	})
 	metrics.NewGauge(`vm_cache_size_bytes{type="storage/metricName"}`, func() float64 {
-		return float64(m().MetricNameCacheBytesSize)
+		return float64(m().MetricNameCacheSizeBytes)
 	})
 	metrics.NewGauge(`vm_cache_size_bytes{type="storage/date_metricID"}`, func() float64 {
-		return float64(m().DateMetricIDCacheBytesSize)
+		return float64(m().DateMetricIDCacheSizeBytes)
 	})
 	metrics.NewGauge(`vm_cache_size_bytes{type="indexdb/tagFilters"}`, func() float64 {
-		return float64(idbm().TagCacheBytesSize)
+		return float64(idbm().TagCacheSizeBytes)
+	})
+	metrics.NewGauge(`vm_cache_size_bytes{type="indexdb/uselessTagFilters"}`, func() float64 {
+		return float64(idbm().UselessTagFiltersCacheSizeBytes)
 	})
 
 	metrics.NewGauge(`vm_cache_requests_total{type="storage/tsid"}`, func() float64 {
@@ -404,6 +474,9 @@ func registerStorageMetrics(strg *storage.Storage) {
 	metrics.NewGauge(`vm_cache_requests_total{type="indexdb/tagFilters"}`, func() float64 {
 		return float64(idbm().TagCacheRequests)
 	})
+	metrics.NewGauge(`vm_cache_requests_total{type="indexdb/uselessTagFilters"}`, func() float64 {
+		return float64(idbm().UselessTagFiltersCacheRequests)
+	})
 	metrics.NewGauge(`vm_cache_requests_total{type="storage/regexps"}`, func() float64 {
 		return float64(storage.RegexpCacheRequests())
 	})
@@ -435,6 +508,9 @@ func registerStorageMetrics(strg *storage.Storage) {
 	metrics.NewGauge(`vm_cache_misses_total{type="indexdb/tagFilters"}`, func() float64 {
 		return float64(idbm().TagCacheMisses)
 	})
+	metrics.NewGauge(`vm_cache_misses_total{type="indexdb/uselessTagFilters"}`, func() float64 {
+		return float64(idbm().UselessTagFiltersCacheMisses)
+	})
 	metrics.NewGauge(`vm_cache_misses_total{type="storage/regexps"}`, func() float64 {
 		return float64(storage.RegexpCacheMisses())
 	})

deployment/docker/Makefile

@@ -1,5 +1,5 @@
-DOCKER_NAMESPACE := valyala
-BUILDER_IMAGE := local/builder:go1.12.5
+DOCKER_NAMESPACE := victoriametrics
+BUILDER_IMAGE := local/builder:go1.12.7
 CERTS_IMAGE := local/certs:1.0.2
 
 package-certs:
@@ -18,8 +18,10 @@ app-via-docker: package-certs package-builder
 		-w /VictoriaMetrics \
 		--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/README.md

@@ -0,0 +1,33 @@
+### Folder contains basic images and tools for building and running Victoria Metrics in docker
+
+#### Docker compose
+
+To spin-up setup of VictoriaMetrics, Prometheus and Grafana run following command:
+
+`docker-compose up`
+
+##### VictoriaMetrics
+
+VictoriaMetrics opens following ports:
+* `--graphiteListenAddr=:2003`
+* `--opentsdbListenAddr=:4242`
+* `--httpListenAddr=:8428`
+
+##### Prometheus
+
+To access service open following [link](http://localhost:9090).
+
+Prometheus is already configured to use VictoriaMetrics as remote storage.
+
+##### Grafana
+
+To access service open following [link](http://localhost:3000).
+
+Default creds:
+* login - `admin`
+* password - `admin`
+
+Grafana is provisioned by default with following entities:
+* VictoriaMetrics datasource
+* Prometheus datasource
+* VictoriaMetrics overview dashboard

deployment/docker/builder/Dockerfile

@@ -1 +1,2 @@
-FROM golang:1.12.5
+FROM golang:1.12.7
+STOPSIGNAL SIGINT

deployment/docker/docker-compose.yml

@@ -0,0 +1,61 @@
+version: '3.5'
+services:
+  prometheus:
+    container_name: prometheus
+    image: prom/prometheus:v2.10.0
+    depends_on:
+      - "victoriametrics"
+    ports:
+      - 9090:9090
+    volumes:
+      - promdata:/prometheus
+      - ./prometheus.yml:/etc/prometheus/prometheus.yml
+    command:
+      - '--config.file=/etc/prometheus/prometheus.yml'
+      - '--storage.tsdb.path=/prometheus'
+    networks:
+      - vm_net
+    restart: always
+  victoriametrics:
+    container_name: victoriametrics
+    image: victoriametrics/victoria-metrics
+    ports:
+      - 8428:8428
+      - 2003:2003
+      - 4242:4242
+    volumes:
+      - vmdata:/storage
+    command:
+      - '--storageDataPath=/storage'
+      - '--graphiteListenAddr=:2003'
+      - '--opentsdbListenAddr=:4242'
+      - '--httpListenAddr=:8428'
+    networks:
+      - vm_net
+    restart: always
+  grafana:
+    container_name: grafana
+    image: grafana/grafana:6.2.1
+    entrypoint: >
+      /bin/sh -c "
+      cd /var/lib/grafana &&
+      mkdir -p dashboards &&
+      sed 's/$${DS_PROMETHEUS}/Prometheus/g' vm.json > dashboards/vm.json &&
+      /run.sh"
+    depends_on:
+      - "victoriametrics"
+    ports:
+      - 3000:3000
+    volumes:
+      - grafanadata:/var/lib/grafana
+      - ./provisioning/:/etc/grafana/provisioning/
+      - ./../../dashboards/victoriametrics.json:/var/lib/grafana/vm.json
+    networks:
+      - vm_net
+    restart: always
+volumes:
+  promdata: {}
+  vmdata: {}
+  grafanadata: {}
+networks:
+  vm_net:

deployment/docker/prometheus.yml

@@ -0,0 +1,16 @@
+global:
+  scrape_interval:     10s
+  evaluation_interval: 10s
+
+remote_write:
+  - url: "http://victoriametrics:8428/api/v1/write"
+    queue_config:
+      max_samples_per_send: 10000
+
+scrape_configs:
+  - job_name: 'prometheus'
+    static_configs:
+      - targets: ['prometheus:9090']
+  - job_name: 'victoriametrics'
+    static_configs:
+      - targets: ['victoriametrics:8428']

deployment/docker/provisioning/dashboards/dashboard.yml

@@ -0,0 +1,9 @@
+apiVersion: 1
+
+providers:
+- name: Prometheus
+  orgId: 1
+  folder: ''
+  type: file
+  options:
+    path: /var/lib/grafana/dashboards

deployment/docker/provisioning/datasources/datasource.yml

@@ -0,0 +1,14 @@
+apiVersion: 1
+
+datasources:
+    - name: Prometheus
+      type: prometheus
+      access: proxy
+      url: http://prometheus:9090
+      isDefault: false
+
+    - name: VictoriaMetrics
+      type: prometheus
+      access: proxy
+      url: http://victoriametrics:8428
+      isDefault: true

go.mod

@@ -1,16 +1,18 @@
 module github.com/VictoriaMetrics/VictoriaMetrics
 
 require (
-	github.com/VictoriaMetrics/fastcache v1.5.0
-	github.com/VictoriaMetrics/metrics v1.4.0
+	github.com/VictoriaMetrics/fastcache v1.5.1
+	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 v1.7.5
 	github.com/spaolacci/murmur3 v1.1.0 // indirect
 	github.com/valyala/fastjson v1.4.1
-	github.com/valyala/gozstd v1.5.0
+	github.com/valyala/gozstd v1.5.1
 	github.com/valyala/histogram v1.0.1
 	github.com/valyala/quicktemplate v1.1.1
-	golang.org/x/sys v0.0.0-20190502175342-a43fa875dd82
+	golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa
 )
 
 go 1.12

go.sum

@@ -1,10 +1,10 @@
 github.com/OneOfOne/xxhash v1.2.2/go.mod h1:HSdplMjZKSmBqAxg5vPj2TmRDmfkzw+cTzAElWljhcU=
 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 v1.5.0 h1:z8t2QV/CDXWVJ9vy9yRtGGDoOvk9W2aXQBijbLk0KCc=
-github.com/VictoriaMetrics/fastcache v1.5.0/go.mod h1:+jv9Ckb+za/P1ZRg/sulP5Ni1v49daAVERr0H3CuscE=
-github.com/VictoriaMetrics/metrics v1.4.0 h1:3+XdciC4E8sywx+0PStXhtIdWxXP2bdJ06Whw0mViQE=
-github.com/VictoriaMetrics/metrics v1.4.0/go.mod h1:QZAL5yLaXvhSPeib0ahluGo9VK0HXDZHovKaKlpuWvs=
+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 v1.7.1 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 v1.7.5 h1:NMapGoDIKPKpk2hpcgAU6XHfsREHG2p8PIg7C3f/jpI=
+github.com/klauspost/compress v1.7.5/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 v1.5.0 h1:OI7Z2e+GkvmmdRFiJeyuByhT/WMTLuRLa43Z2Tjzenw=
-github.com/valyala/gozstd v1.5.0/go.mod h1:oYOS+oJovjw9ewtrwEYb9+ybolEXd6pHyLMuAWN5zts=
+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/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/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 v0.0.0-20190502175342-a43fa875dd82 h1:vsphBvatvfbhlb4PO1BYSr9dzugGxJ/SQHoNufZJq1w=
-golang.org/x/sys v0.0.0-20190502175342-a43fa875dd82/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa h1:KIDDMLT1O0Nr7TSxp8xM5tJcdn8tgyAONntO829og1M=
+golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

lib/bytesutil/bytebuffer.go

@@ -13,6 +13,7 @@ var (
 	// Verify ByteBuffer implements the given interfaces.
 	_ io.Writer       = &ByteBuffer{}
 	_ fs.ReadAtCloser = &ByteBuffer{}
+	_ io.ReaderFrom   = &ByteBuffer{}
 
 	// Verify reader implement filestream.ReadCloser interface.
 	_ filestream.ReadCloser = &reader{}
@@ -44,7 +45,31 @@ func (bb *ByteBuffer) ReadAt(p []byte, offset int64) {
 		logger.Panicf("BUG: too big offset=%d; cannot exceed len(bb.B)=%d", offset, len(bb.B))
 	}
 	if n := copy(p, bb.B[offset:]); n < len(p) {
-		logger.Panicf("BUG: EOF occured after reading %d bytes out of %d bytes at offset %d", n, len(p), offset)
+		logger.Panicf("BUG: EOF occurred after reading %d bytes out of %d bytes at offset %d", n, len(p), offset)
+	}
+}
+
+// ReadFrom reads all the data from r to bb until EOF.
+func (bb *ByteBuffer) ReadFrom(r io.Reader) (int64, error) {
+	b := bb.B
+	bLen := len(b)
+	b = Resize(b, 4*1024)
+	b = b[:cap(b)]
+	offset := bLen
+	for {
+		if free := len(b) - offset; free < offset {
+			n := len(b)
+			b = append(b, make([]byte, n)...)
+		}
+		n, err := r.Read(b[offset:])
+		offset += n
+		if err != nil {
+			bb.B = b[:offset]
+			if err == io.EOF {
+				err = nil
+			}
+			return int64(offset - bLen), err
+		}
 	}
 }
 

lib/bytesutil/bytebuffer_test.go

@@ -1,6 +1,7 @@
 package bytesutil
 
 import (
+	"bytes"
 	"fmt"
 	"io"
 	"testing"
@@ -66,6 +67,92 @@ func TestByteBuffer(t *testing.T) {
 	}
 }
 
+func TestByteBufferReadFrom(t *testing.T) {
+	var bbPool ByteBufferPool
+
+	t.Run("zero_bytes", func(t *testing.T) {
+		t.Parallel()
+		bb := bbPool.Get()
+		defer bbPool.Put(bb)
+		src := bytes.NewBufferString("")
+		n, err := bb.ReadFrom(src)
+		if err != nil {
+			t.Fatalf("error when reading empty string: %s", err)
+		}
+		if n != 0 {
+			t.Fatalf("unexpected number of bytes read; got %d; want %d", n, 0)
+		}
+		if len(bb.B) != 0 {
+			t.Fatalf("unexpejcted len(bb.B); got %d; want %d", len(bb.B), 0)
+		}
+	})
+
+	t.Run("non_zero_bytes", func(t *testing.T) {
+		t.Parallel()
+		bb := bbPool.Get()
+		defer bbPool.Put(bb)
+		s := "foobarbaz"
+		src := bytes.NewBufferString(s)
+		n, err := bb.ReadFrom(src)
+		if err != nil {
+			t.Fatalf("error when reading non-empty string: %s", err)
+		}
+		if n != int64(len(s)) {
+			t.Fatalf("unexpected number of bytes read; got %d; want %d", n, len(s))
+		}
+		if string(bb.B) != s {
+			t.Fatalf("unexpected value read; got %q; want %q", bb.B, s)
+		}
+	})
+
+	t.Run("big_number_of_bytes", func(t *testing.T) {
+		t.Parallel()
+		bb := bbPool.Get()
+		defer bbPool.Put(bb)
+		b := make([]byte, 1024*1024+234)
+		for i := range b {
+			b[i] = byte(i)
+		}
+		src := bytes.NewBuffer(b)
+		n, err := bb.ReadFrom(src)
+		if err != nil {
+			t.Fatalf("cannot read big value: %s", err)
+		}
+		if n != int64(len(b)) {
+			t.Fatalf("unexpected number of bytes read; got %d; want %d", n, len(b))
+		}
+		if string(bb.B) != string(b) {
+			t.Fatalf("unexpected value read; got %q; want %q", bb.B, b)
+		}
+	})
+
+	t.Run("non_empty_bb", func(t *testing.T) {
+		t.Parallel()
+		bb := bbPool.Get()
+		defer bbPool.Put(bb)
+		prefix := []byte("prefix")
+		bb.B = append(bb.B[:0], prefix...)
+		s := "aosdfdsafdjsf"
+		src := bytes.NewBufferString(s)
+		n, err := bb.ReadFrom(src)
+		if err != nil {
+			t.Fatalf("cannot read to non-empty bb: %s", err)
+		}
+		if n != int64(len(s)) {
+			t.Fatalf("unexpected number of bytes read; got %d; want %d", n, len(s))
+		}
+		if len(bb.B) != len(prefix)+len(s) {
+			t.Fatalf("unexpected bb.B len; got %d; want %d", len(bb.B), len(prefix)+len(s))
+		}
+		if string(bb.B[:len(prefix)]) != string(prefix) {
+			t.Fatalf("unexpected prefix; got %q; want %q", bb.B[:len(prefix)], prefix)
+		}
+		if string(bb.B[len(prefix):]) != s {
+			t.Fatalf("unexpected data read; got %q; want %q", bb.B[len(prefix):], s)
+		}
+	})
+}
+
 func TestByteBufferRead(t *testing.T) {
 	var bb ByteBuffer
 

lib/decimal/decimal.go

@@ -186,6 +186,7 @@ func maxUpExponent(v int64) int16 {
 		v = -v
 	}
 	if v < 0 {
+		// Handle corner case for v=-1<<63
 		return 0
 	}
 

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

@@ -1,27 +1,34 @@
 package encoding
 
 import (
-	"github.com/valyala/gozstd"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding/zstd"
+	"github.com/VictoriaMetrics/metrics"
 )
 
-// CompressZSTD compresses src, appends the result to dst and returns
-// the appended dst.
-//
-// src must be non-empty.
-func CompressZSTD(dst, src []byte) []byte {
-	return gozstd.CompressLevel(dst, src, 5)
-}
-
 // CompressZSTDLevel appends compressed src to dst and returns
 // the appended dst.
 //
 // The given compressLevel is used for the compression.
 func CompressZSTDLevel(dst, src []byte, compressLevel int) []byte {
-	return gozstd.CompressLevel(dst, src, compressLevel)
+	compressCalls.Inc()
+	originalBytes.Add(len(src))
+	dstLen := len(dst)
+	dst = zstd.CompressLevel(dst, src, compressLevel)
+	compressedBytes.Add(len(dst) - dstLen)
+	return dst
 }
 
 // DecompressZSTD decompresses src, appends the result to dst and returns
 // the appended dst.
 func DecompressZSTD(dst, src []byte) ([]byte, error) {
-	return gozstd.Decompress(dst, src)
+	decompressCalls.Inc()
+	return zstd.Decompress(dst, src)
 }
+
+var (
+	compressCalls   = metrics.NewCounter(`vm_zstd_block_compress_calls_total`)
+	decompressCalls = metrics.NewCounter(`vm_zstd_block_decompress_calls_total`)
+
+	originalBytes   = metrics.NewCounter(`vm_zstd_block_original_bytes_total`)
+	compressedBytes = metrics.NewCounter(`vm_zstd_block_compressed_bytes_total`)
+)

lib/encoding/compress_test.go

@@ -17,7 +17,7 @@ func TestCompressDecompressZSTD(t *testing.T) {
 }
 
 func testCompressDecompressZSTD(t *testing.T, b []byte) {
-	bc := CompressZSTD(nil, b)
+	bc := CompressZSTDLevel(nil, b, 5)
 	bNew, err := DecompressZSTD(nil, bc)
 	if err != nil {
 		t.Fatalf("unexpected error when decompressing b=%x from bc=%x: %s", b, bc, err)
@@ -27,7 +27,7 @@ func testCompressDecompressZSTD(t *testing.T, b []byte) {
 	}
 
 	prefix := []byte{1, 2, 33}
-	bcNew := CompressZSTD(prefix, b)
+	bcNew := CompressZSTDLevel(prefix, b, 5)
 	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)
 	}

lib/encoding/encoding.go

@@ -117,7 +117,7 @@ func marshalInt64Array(dst []byte, a []int64, precisionBits uint8) (result []byt
 
 	bb := bbPool.Get()
 	if isGauge(a) {
-		// Guage values are better compressed with delta encoding.
+		// Gauge values are better compressed with delta encoding.
 		mt = MarshalTypeZSTDNearestDelta
 		pb := precisionBits
 		if pb < 6 {
@@ -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_windows.go

@@ -0,0 +1,9 @@
+package filestream
+
+func (st *streamTracker) adviseDontNeed(n int, fdatasync bool) error {
+	return nil
+}
+
+func (st *streamTracker) close() error {
+	return nil
+}

lib/flagutil/array.go

@@ -1,18 +0,0 @@
-package flagutil
-
-import "strings"
-
-// Array holds an array of flag values
-type Array []string
-
-// String implements flag.Value interface
-func (a *Array) String() string {
-	return strings.Join(*a, ",")
-}
-
-// Set implements flag.Value interface
-func (a *Array) Set(value string) error {
-	values := strings.Split(value, ",")
-	*a = append(*a, values...)
-	return nil
-}

lib/flagutil/array_test.go

@@ -1,34 +0,0 @@
-package flagutil
-
-import (
-	"flag"
-	"os"
-	"testing"
-)
-
-var fooFlag Array
-
-func init() {
-	os.Args = append(os.Args, "--fooFlag=foo", "--fooFlag=bar")
-	flag.Var(&fooFlag, "fooFlag", "test")
-}
-
-func TestMain(m *testing.M) {
-	flag.Parse()
-	os.Exit(m.Run())
-}
-
-func TestArray(t *testing.T) {
-	expected := map[string]struct{}{
-		"foo": {},
-		"bar": {},
-	}
-	if len(expected) != len(fooFlag) {
-		t.Errorf("len array flag (%d) is not equal to %d", len(fooFlag), len(expected))
-	}
-	for _, i := range fooFlag {
-		if _, ok := expected[i]; !ok {
-			t.Errorf("unexpected item in array %v", i)
-		}
-	}
-}

lib/fs/fs.go

@@ -5,6 +5,8 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"strings"
+	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
@@ -70,8 +72,8 @@ var (
 	readersCount = metrics.NewCounter(`vm_fs_readers`)
 )
 
-// SyncPath syncs contents of the given path.
-func SyncPath(path string) {
+// MustSyncPath syncs contents of the given path.
+func MustSyncPath(path string) {
 	d, err := os.Open(path)
 	if err != nil {
 		logger.Panicf("FATAL: cannot open %q: %s", path, err)
@@ -111,8 +113,8 @@ func WriteFile(path string, data []byte) error {
 	if err != nil {
 		return fmt.Errorf("cannot obtain absolute path to %q: %s", path, err)
 	}
-	parentDirPath, _ := filepath.Split(absPath)
-	SyncPath(parentDirPath)
+	parentDirPath := filepath.Dir(absPath)
+	MustSyncPath(parentDirPath)
 
 	return nil
 }
@@ -122,7 +124,7 @@ func MkdirAllIfNotExist(path string) error {
 	if IsPathExist(path) {
 		return nil
 	}
-	return os.MkdirAll(path, 0755)
+	return mkdirSync(path)
 }
 
 // MkdirAllFailIfExist creates the given path dir if it isn't exist.
@@ -132,7 +134,18 @@ func MkdirAllFailIfExist(path string) error {
 	if IsPathExist(path) {
 		return fmt.Errorf("the %q already exists", path)
 	}
-	return os.MkdirAll(path, 0755)
+	return mkdirSync(path)
+}
+
+func mkdirSync(path string) error {
+	if err := os.MkdirAll(path, 0755); err != nil {
+		return err
+	}
+	// Sync the parent directory, so the created directory becomes visible
+	// in the fs after power loss.
+	parentDirPath := filepath.Dir(path)
+	MustSyncPath(parentDirPath)
+	return nil
 }
 
 // RemoveDirContents removes all the contents of the given dir it it exists.
@@ -159,11 +172,9 @@ func RemoveDirContents(dir string) {
 			continue
 		}
 		fullPath := dir + "/" + name
-		if err := os.RemoveAll(fullPath); err != nil {
-			logger.Panicf("FATAL: cannot remove %q: %s", fullPath, err)
-		}
+		MustRemoveAll(fullPath)
 	}
-	SyncPath(dir)
+	MustSyncPath(dir)
 }
 
 // MustClose must close the given file f.
@@ -174,6 +185,18 @@ func MustClose(f *os.File) {
 	}
 }
 
+// MustFileSize returns file size for the given path.
+func MustFileSize(path string) uint64 {
+	fi, err := os.Stat(path)
+	if err != nil {
+		logger.Panicf("FATAL: cannot stat %q: %s", path, err)
+	}
+	if fi.IsDir() {
+		logger.Panicf("FATAL: %q must be a file, not a directory", path)
+	}
+	return uint64(fi.Size())
+}
+
 // IsPathExist returns whether the given path exists.
 func IsPathExist(path string) bool {
 	if _, err := os.Stat(path); err != nil {
@@ -185,18 +208,79 @@ func IsPathExist(path string) bool {
 	return true
 }
 
-// MustRemoveAllSynced removes path with all the contents
-// and syncs the parent directory, so it no longer contains the path.
-func MustRemoveAllSynced(path string) {
-	if err := os.RemoveAll(path); err != nil {
+func mustSyncParentDirIfExists(path string) {
+	parentDirPath := filepath.Dir(path)
+	if !IsPathExist(parentDirPath) {
+		return
+	}
+	MustSyncPath(parentDirPath)
+}
+
+// MustRemoveAll removes path with all the contents.
+//
+// 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)
 	}
-	SyncPath(filepath.Dir(path))
+	// 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()
 }
 
 // HardLinkFiles makes hard links for all the files from srcDir in dstDir.
 func HardLinkFiles(srcDir, dstDir string) error {
-	if err := os.MkdirAll(dstDir, 0755); err != nil {
+	if err := mkdirSync(dstDir); err != nil {
 		return fmt.Errorf("cannot create dstDir=%q: %s", dstDir, err)
 	}
 
@@ -227,7 +311,7 @@ func HardLinkFiles(srcDir, dstDir string) error {
 		}
 	}
 
-	SyncPath(dstDir)
+	MustSyncPath(dstDir)
 	return nil
 }
 

lib/httpserver/httpserver.go

@@ -31,6 +31,8 @@ var (
 	httpAuthPassword = flag.String("httpAuth.password", "", "Password for HTTP Basic Auth. The authentication is disabled -httpAuth.username is empty")
 	metricsAuthKey   = flag.String("metricsAuthKey", "", "Auth key for /metrics. It overrides httpAuth settings")
 	pprofAuthKey     = flag.String("pprofAuthKey", "", "Auth key for /debug/pprof. It overrides httpAuth settings")
+
+	disableResponseCompression = flag.Bool("http.disableResponseCompression", false, "Disable compression of HTTP responses for saving CPU resources. By default compression is enabled to save network bandwidth")
 )
 
 var (
@@ -51,6 +53,8 @@ type RequestHandler func(w http.ResponseWriter, r *http.Request) bool
 // By default all the responses are transparently compressed, since Google
 // charges a lot for the egress traffic. The compression may be disabled
 // by calling DisableResponseCompression before writing the first byte to w.
+//
+// The compression is also disabled if -http.disableResponseCompression flag is set.
 func Serve(addr string, rh RequestHandler) {
 	scheme := "http"
 	if *tlsEnable {
@@ -224,6 +228,9 @@ func checkBasicAuth(w http.ResponseWriter, r *http.Request) bool {
 }
 
 func maybeGzipResponseWriter(w http.ResponseWriter, r *http.Request) http.ResponseWriter {
+	if *disableResponseCompression {
+		return w
+	}
 	ae := r.Header.Get("Accept-Encoding")
 	if ae == "" {
 		return w

lib/memory/memory_linux.go

@@ -2,6 +2,7 @@ package memory
 
 import (
 	"io/ioutil"
+	"os/exec"
 	"strconv"
 	"syscall"
 
@@ -19,17 +20,57 @@ func sysTotalMemory() int {
 	// See https://stackoverflow.com/questions/42187085/check-mem-limit-within-a-docker-container .
 	data, err := ioutil.ReadFile("/sys/fs/cgroup/memory/memory.limit_in_bytes")
 	if err != nil {
-		return totalMem
+		// Try determining the amount of memory inside lxc container.
+		mem, err := readLXCMemoryLimit(totalMem)
+		if err != nil {
+			return totalMem
+		}
+		return mem
 	}
-	for len(data) > 0 && data[len(data)-1] == '\n' {
-		data = data[:len(data)-1]
-	}
-	mem, err := strconv.Atoi(string(data))
+	mem, err := readPositiveInt(data, totalMem)
 	if err != nil {
 		return totalMem
 	}
-	if mem > totalMem {
-		mem = totalMem
+	if mem != totalMem {
+		return mem
+	}
+
+	// Try reading LXC memory limit, since it looks like the cgroup limit doesn't work
+	mem, err = readLXCMemoryLimit(totalMem)
+	if err != nil {
+		return totalMem
 	}
 	return mem
 }
+
+func readLXCMemoryLimit(totalMem int) (int, error) {
+	// Read memory limit according to https://unix.stackexchange.com/questions/242718/how-to-find-out-how-much-memory-lxc-container-is-allowed-to-consume
+	// This should properly determine the limit inside lxc container.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/84
+	cmd := exec.Command("/bin/sh", "-c",
+		`cat /sys/fs/cgroup/memory$(cat /proc/self/cgroup | grep memory | cut -d: -f3)/memory.limit_in_bytes`)
+	data, err := cmd.Output()
+	if err != nil {
+		return 0, err
+	}
+	return readPositiveInt(data, totalMem)
+}
+
+func readPositiveInt(data []byte, maxN int) (int, error) {
+	for len(data) > 0 && data[len(data)-1] == '\n' {
+		data = data[:len(data)-1]
+	}
+	n, err := strconv.ParseUint(string(data), 10, 64)
+	if err != nil {
+		return 0, err
+	}
+	if int64(n) < 0 || int64(int(n)) != int64(n) {
+		// Int overflow.
+		return maxN, nil
+	}
+	ni := int(n)
+	if ni > maxN {
+		return maxN, nil
+	}
+	return ni, nil
+}

lib/memory/memory_windows.go

@@ -0,0 +1,40 @@
+package memory
+
+import (
+	"syscall"
+	"unsafe"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+// This has been adapted from https://github.com/pbnjay/memory.
+
+type memStatusEx struct {
+	dwLength     uint32
+	dwMemoryLoad uint32
+	ullTotalPhys uint64
+	unused       [6]uint64
+}
+
+func sysTotalMemory() int {
+	kernel32, err := syscall.LoadDLL("kernel32.dll")
+	if err != nil {
+		logger.Panicf("FATAL: cannot load kernel32.dll: %s", err)
+	}
+	globalMemoryStatusEx, err := kernel32.FindProc("GlobalMemoryStatusEx")
+	if err != nil {
+		logger.Panicf("FATAL: cannot find GlobalMemoryStatusEx: %s", err)
+	}
+	msx := &memStatusEx{
+		dwLength: uint32(unsafe.Sizeof(memStatusEx{})),
+	}
+	r, _, err := globalMemoryStatusEx.Call(uintptr(unsafe.Pointer(msx)))
+	if r == 0 {
+		logger.Panicf("FATAL: error in GlobalMemoryStatusEx: %s", err)
+	}
+	n := int(msx.ullTotalPhys)
+	if uint64(n) != msx.ullTotalPhys {
+		logger.Panicf("FATAL: int overflow for msx.ullTotalPhys=%d", msx.ullTotalPhys)
+	}
+	return n
+}

lib/mergeset/block_stream_reader.go

@@ -91,6 +91,13 @@ func (bsr *blockStreamReader) reset() {
 	bsr.err = nil
 }
 
+func (bsr *blockStreamReader) String() string {
+	if len(bsr.path) > 0 {
+		return bsr.path
+	}
+	return bsr.ph.String()
+}
+
 // InitFromInmemoryPart initializes bsr from the given ip.
 func (bsr *blockStreamReader) InitFromInmemoryPart(ip *inmemoryPart) {
 	bsr.reset()

lib/mergeset/block_stream_writer.go

@@ -2,7 +2,6 @@ package mergeset
 
 import (
 	"fmt"
-	"os"
 	"path/filepath"
 	"sync"
 
@@ -92,7 +91,7 @@ func (bsw *blockStreamWriter) InitFromFilePart(path string, nocache bool, compre
 	metaindexPath := path + "/metaindex.bin"
 	metaindexFile, err := filestream.Create(metaindexPath, false)
 	if err != nil {
-		_ = os.RemoveAll(path)
+		fs.MustRemoveAll(path)
 		return fmt.Errorf("cannot create metaindex file: %s", err)
 	}
 
@@ -100,7 +99,7 @@ func (bsw *blockStreamWriter) InitFromFilePart(path string, nocache bool, compre
 	indexFile, err := filestream.Create(indexPath, nocache)
 	if err != nil {
 		metaindexFile.MustClose()
-		_ = os.RemoveAll(path)
+		fs.MustRemoveAll(path)
 		return fmt.Errorf("cannot create index file: %s", err)
 	}
 
@@ -109,7 +108,7 @@ func (bsw *blockStreamWriter) InitFromFilePart(path string, nocache bool, compre
 	if err != nil {
 		metaindexFile.MustClose()
 		indexFile.MustClose()
-		_ = os.RemoveAll(path)
+		fs.MustRemoveAll(path)
 		return fmt.Errorf("cannot create items file: %s", err)
 	}
 
@@ -119,7 +118,7 @@ func (bsw *blockStreamWriter) InitFromFilePart(path string, nocache bool, compre
 		metaindexFile.MustClose()
 		indexFile.MustClose()
 		itemsFile.MustClose()
-		_ = os.RemoveAll(path)
+		fs.MustRemoveAll(path)
 		return fmt.Errorf("cannot create lens file: %s", err)
 	}
 
@@ -152,6 +151,12 @@ func (bsw *blockStreamWriter) MustClose() {
 	bsw.itemsWriter.MustClose()
 	bsw.lensWriter.MustClose()
 
+	// Sync bsw.path contents to make sure it doesn't disappear
+	// after system crash or power loss.
+	if bsw.path != "" {
+		fs.MustSyncPath(bsw.path)
+	}
+
 	bsw.reset()
 }
 

lib/mergeset/inmemory_part.go

@@ -84,13 +84,18 @@ func (ip *inmemoryPart) Init(ib *inmemoryBlock) {
 // It is unsafe re-using ip while the returned part is in use.
 func (ip *inmemoryPart) NewPart() *part {
 	ph := ip.ph
-	p, err := newPart(&ph, "", ip.metaindexData.NewReader(), &ip.indexData, &ip.itemsData, &ip.lensData)
+	size := ip.size()
+	p, err := newPart(&ph, "", size, ip.metaindexData.NewReader(), &ip.indexData, &ip.itemsData, &ip.lensData)
 	if err != nil {
 		logger.Panicf("BUG: cannot create a part from inmemoryPart: %s", err)
 	}
 	return p
 }
 
+func (ip *inmemoryPart) size() uint64 {
+	return uint64(len(ip.metaindexData.B) + len(ip.indexData.B) + len(ip.itemsData.B) + len(ip.lensData.B))
+}
+
 func getInmemoryPart() *inmemoryPart {
 	v := ipPool.Get()
 	if v == nil {

lib/mergeset/merge.go

@@ -29,7 +29,7 @@ func mergeBlockStreams(ph *partHeader, bsw *blockStreamWriter, bsrs []*blockStre
 	if err == errForciblyStopped {
 		return err
 	}
-	return fmt.Errorf("cannot merge block streams: %s", err)
+	return fmt.Errorf("cannot merge %d block streams: %s: %s", len(bsrs), bsrs, err)
 }
 
 var bsmPool = &sync.Pool{

lib/mergeset/part.go

@@ -48,6 +48,8 @@ type part struct {
 
 	path string
 
+	size uint64
+
 	mrs []metaindexRow
 
 	indexFile fs.ReadAtCloser
@@ -71,6 +73,7 @@ func openFilePart(path string) (*part, error) {
 	if err != nil {
 		return nil, fmt.Errorf("cannot open %q: %s", metaindexPath, err)
 	}
+	metaindexSize := fs.MustFileSize(metaindexPath)
 
 	indexPath := path + "/index.bin"
 	indexFile, err := fs.OpenReaderAt(indexPath)
@@ -78,6 +81,7 @@ func openFilePart(path string) (*part, error) {
 		metaindexFile.MustClose()
 		return nil, fmt.Errorf("cannot open %q: %s", indexPath, err)
 	}
+	indexSize := fs.MustFileSize(indexPath)
 
 	itemsPath := path + "/items.bin"
 	itemsFile, err := fs.OpenReaderAt(itemsPath)
@@ -86,6 +90,7 @@ func openFilePart(path string) (*part, error) {
 		indexFile.MustClose()
 		return nil, fmt.Errorf("cannot open %q: %s", itemsPath, err)
 	}
+	itemsSize := fs.MustFileSize(itemsPath)
 
 	lensPath := path + "/lens.bin"
 	lensFile, err := fs.OpenReaderAt(lensPath)
@@ -95,11 +100,13 @@ func openFilePart(path string) (*part, error) {
 		itemsFile.MustClose()
 		return nil, fmt.Errorf("cannot open %q: %s", lensPath, err)
 	}
+	lensSize := fs.MustFileSize(lensPath)
 
-	return newPart(&ph, path, metaindexFile, indexFile, itemsFile, lensFile)
+	size := metaindexSize + indexSize + itemsSize + lensSize
+	return newPart(&ph, path, size, metaindexFile, indexFile, itemsFile, lensFile)
 }
 
-func newPart(ph *partHeader, path string, metaindexReader filestream.ReadCloser, indexFile, itemsFile, lensFile fs.ReadAtCloser) (*part, error) {
+func newPart(ph *partHeader, path string, size uint64, metaindexReader filestream.ReadCloser, indexFile, itemsFile, lensFile fs.ReadAtCloser) (*part, error) {
 	var errors []error
 	mrs, err := unmarshalMetaindexRows(nil, metaindexReader)
 	if err != nil {
@@ -109,6 +116,7 @@ func newPart(ph *partHeader, path string, metaindexReader filestream.ReadCloser,
 
 	p := &part{
 		path: path,
+		size: size,
 		mrs:  mrs,
 
 		indexFile: indexFile,

lib/mergeset/part_search.go

@@ -356,16 +356,15 @@ func (ps *partSearch) readInmemoryBlock(bh *blockHeader) (*inmemoryBlock, error)
 }
 
 func binarySearchKey(items [][]byte, key []byte) int {
-	if len(items) > 0 {
-		if string(key) <= string(items[0]) {
-			// Fast path - the item is the first.
-			return 0
-		}
-		if len(items) > 1 && string(key) <= string(items[1]) {
-			// Fast path - the item is the second.
-			return 1
-		}
+	if len(items) == 0 {
+		return 0
 	}
+	if string(key) <= string(items[0]) {
+		// Fast path - the item is the first.
+		return 0
+	}
+	items = items[1:]
+	offset := uint(1)
 
 	// This has been copy-pasted from https://golang.org/src/sort/search.go
 	n := uint(len(items))
@@ -378,5 +377,5 @@ func binarySearchKey(items [][]byte, key []byte) int {
 			j = h
 		}
 	}
-	return int(i)
+	return int(i + offset)
 }

lib/mergeset/part_search_test.go

@@ -156,7 +156,8 @@ func newTestPart(blocksCount, maxItemsPerBlock int) (*part, []string, error) {
 	if itemsMerged != uint64(len(items)) {
 		return nil, nil, fmt.Errorf("unexpected itemsMerged; got %d; want %d", itemsMerged, len(items))
 	}
-	p, err := newPart(&ip.ph, "partName", ip.metaindexData.NewReader(), &ip.indexData, &ip.itemsData, &ip.lensData)
+	size := ip.size()
+	p, err := newPart(&ip.ph, "partName", size, ip.metaindexData.NewReader(), &ip.indexData, &ip.itemsData, &ip.lensData)
 	if err != nil {
 		return nil, nil, fmt.Errorf("cannot create part: %s", err)
 	}

lib/mergeset/table.go

@@ -161,8 +161,8 @@ func OpenTable(path string) (*Table, error) {
 
 	var m TableMetrics
 	tb.UpdateMetrics(&m)
-	logger.Infof("table %q has been opened in %s; partsCount: %d; blocksCount: %d, itemsCount: %d",
-		path, time.Since(startTime), m.PartsCount, m.BlocksCount, m.ItemsCount)
+	logger.Infof("table %q has been opened in %s; partsCount: %d; blocksCount: %d, itemsCount: %d; sizeBytes: %d",
+		path, time.Since(startTime), m.PartsCount, m.BlocksCount, m.ItemsCount, m.SizeBytes)
 
 	return tb, nil
 }
@@ -208,7 +208,7 @@ func (tb *Table) MustClose() {
 	logger.Infof("%d inmemory parts have been flushed to files in %s on %q", len(pws), time.Since(startTime), tb.path)
 
 	// Remove references to parts from the tb, so they may be eventually closed
-	// after all the seraches are done.
+	// after all the searches are done.
 	tb.partsLock.Lock()
 	parts := tb.parts
 	tb.parts = nil
@@ -242,6 +242,7 @@ type TableMetrics struct {
 
 	BlocksCount uint64
 	ItemsCount  uint64
+	SizeBytes   uint64
 
 	DataBlocksCacheSize     uint64
 	DataBlocksCacheRequests uint64
@@ -274,6 +275,7 @@ func (tb *Table) UpdateMetrics(m *TableMetrics) {
 
 		m.BlocksCount += p.ph.blocksCount
 		m.ItemsCount += p.ph.itemsCount
+		m.SizeBytes += p.size
 
 		m.DataBlocksCacheSize += p.ibCache.Len()
 		m.DataBlocksCacheRequests += p.ibCache.Requests()
@@ -727,6 +729,7 @@ func (tb *Table) mergeParts(pws []*partWrapper, stopCh <-chan struct{}, isOuterP
 	if err != nil {
 		return fmt.Errorf("cannot open merged part %q: %s", dstPartPath, err)
 	}
+	newPSize := newP.size
 	newPW := &partWrapper{
 		p:        newP,
 		refCount: 1,
@@ -761,7 +764,7 @@ func (tb *Table) mergeParts(pws []*partWrapper, stopCh <-chan struct{}, isOuterP
 
 	d := time.Since(startTime)
 	if d > 10*time.Second {
-		logger.Infof("merged %d items in %s at %d items/sec to %q", outItemsCount, d, int(float64(outItemsCount)/d.Seconds()), dstPartPath)
+		logger.Infof("merged %d items in %s at %d items/sec to %q; sizeBytes: %d", outItemsCount, d, int(float64(outItemsCount)/d.Seconds()), dstPartPath, newPSize)
 	}
 
 	return nil
@@ -846,22 +849,18 @@ func openParts(path string) ([]*partWrapper, error) {
 	}
 
 	txnDir := path + "/txn"
-	if err := os.RemoveAll(txnDir); err != nil {
-		return nil, fmt.Errorf("cannot remove %q: %s", txnDir, err)
-	}
+	fs.MustRemoveAll(txnDir)
 	if err := fs.MkdirAllFailIfExist(txnDir); err != nil {
 		return nil, fmt.Errorf("cannot create %q: %s", txnDir, err)
 	}
 
 	tmpDir := path + "/tmp"
-	if err := os.RemoveAll(tmpDir); err != nil {
-		return nil, fmt.Errorf("cannot remove %q: %s", tmpDir, err)
-	}
+	fs.MustRemoveAll(tmpDir)
 	if err := fs.MkdirAllFailIfExist(tmpDir); err != nil {
 		return nil, fmt.Errorf("cannot create %q: %s", tmpDir, err)
 	}
 
-	fs.SyncPath(path)
+	fs.MustSyncPath(path)
 
 	// Open parts.
 	fis, err := d.Readdir(-1)
@@ -965,9 +964,9 @@ func (tb *Table) CreateSnapshotAt(dstDir string) error {
 		}
 	}
 
-	fs.SyncPath(dstDir)
+	fs.MustSyncPath(dstDir)
 	parentDir := filepath.Dir(dstDir)
-	fs.SyncPath(parentDir)
+	fs.MustSyncPath(parentDir)
 
 	logger.Infof("created Table snapshot of %q at %q in %s", srcDir, dstDir, time.Since(startTime))
 	return nil
@@ -1033,9 +1032,7 @@ func runTransaction(txnLock *sync.RWMutex, pathPrefix, txnPath string) error {
 		if err != nil {
 			return fmt.Errorf("invalid path to remove: %s", err)
 		}
-		if err := os.RemoveAll(path); err != nil {
-			return fmt.Errorf("cannot remove %q: %s", path, err)
-		}
+		fs.MustRemoveAll(path)
 	}
 
 	// Move the new part to new directory.
@@ -1061,7 +1058,7 @@ func runTransaction(txnLock *sync.RWMutex, pathPrefix, txnPath string) error {
 	}
 
 	// Flush pathPrefix directory metadata to the underying storage.
-	fs.SyncPath(pathPrefix)
+	fs.MustSyncPath(pathPrefix)
 
 	// Remove the transaction file.
 	if err := os.Remove(txnPath); err != nil {

lib/mergeset/table_search_timing_test.go

@@ -33,6 +33,9 @@ func benchmarkTableSearch(b *testing.B, itemsCount int) {
 	// Force finishing pending merges
 	tb.MustClose()
 	tb, err = OpenTable(path)
+	if err != nil {
+		b.Fatalf("unexpected error when re-opening table %q: %s", path, err)
+	}
 	defer tb.MustClose()
 
 	keys := make([][]byte, len(items))

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/netutil/tcplistener.go

@@ -47,8 +47,6 @@ type TCPListener struct {
 
 	net.Listener
 
-	name string
-
 	accepts      *metrics.Counter
 	acceptErrors *metrics.Counter
 

lib/procutil/signal.go

@@ -6,7 +6,7 @@ import (
 	"syscall"
 )
 
-// WaitForSigterm waits fro either SIGTERM or SIGINT
+// WaitForSigterm waits for either SIGTERM or SIGINT
 //
 // Returns the caught signal.
 func WaitForSigterm() os.Signal {