deployment/docker: switch Go builder from go1.12.8 to go1.12.9

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150

.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

@@ -13,10 +13,14 @@ before_install:
   - GO111MODULE=off go get -u github.com/kisielk/errcheck
 
 script:
-  - make check_all
+  - make check-all
   - git diff --exit-code
-  - make test_full
+  - 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)
+  - 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

@@ -50,19 +50,25 @@ errcheck: install-errcheck
 install-errcheck:
 	which errcheck || GO111MODULE=off go get -u github.com/kisielk/errcheck
 
-check_all: fmt vet lint errcheck golangci-lint
+check-all: fmt vet lint errcheck golangci-lint
 
 test:
-	GO111MODULE=on go test -mod=vendor ./lib/...
-	GO111MODULE=on go test -mod=vendor ./app/...
+	GO111MODULE=on go test -tags=integration -mod=vendor ./lib/... ./app/...
 
-test_full:
-	GO111MODULE=on go test -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./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:
 	GO111MODULE=on go test -mod=vendor -bench=. ./lib/...
 	GO111MODULE=on go test -mod=vendor -bench=. ./app/...
 
+benchmark-pure:
+	GO111MODULE=on CGO_ENABLED=0 go test -mod=vendor -bench=. ./lib/...
+	GO111MODULE=on CGO_ENABLED=0 go test -mod=vendor -bench=. ./app/...
+
 vendor-update:
 	GO111MODULE=on go get -u ./lib/...
 	GO111MODULE=on go get -u ./app/...

README.md

@@ -1,14 +1,15 @@
 [![Latest Release](https://img.shields.io/github/release/VictoriaMetrics/VictoriaMetrics.svg?style=flat-square)](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/latest)
+[![Slack](https://img.shields.io/badge/join%20slack-%23victoriametrics-brightgreen.svg)](http://slack.victoriametrics.com/)
 [![GitHub license](https://img.shields.io/github/license/VictoriaMetrics/VictoriaMetrics.svg)](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/LICENSE)
 [![Go Report](https://goreportcard.com/badge/github.com/VictoriaMetrics/VictoriaMetrics)](https://goreportcard.com/report/github.com/VictoriaMetrics/VictoriaMetrics)
 [![Build Status](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics.svg?branch=master)](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics)
 [![codecov](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics/branch/master/graph/badge.svg)](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics)
 
-<img text-align="center" alt="Victoria Metrics" src="logo.png">
+<img alt="Victoria Metrics" src="logo.png">
 
 ## Single-node VictoriaMetrics
 
-VictoriaMetrics is fast, cost-effective and scalable time series database. It can be used as a long-term remote storage for Prometheus.
+VictoriaMetrics is fast, cost-effective and scalable time-series database. It can be used as long-term remote storage for Prometheus.
 It is available in [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases),
 [docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) and
 in [source code](https://github.com/VictoriaMetrics/VictoriaMetrics).
@@ -20,15 +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.
@@ -50,23 +53,24 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
 
 ### Table of contents
 
-- [Single-node VictoriaMetrics](#single-node-victoriametrics)
-- [Prominent features](#prominent-features)
-- [Operation](#operation)
-  - [Table of contents](#table-of-contents)
+  - [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)
-  - [How to start VictoriaMetrics](#how-to-start-victoriametrics)
+  - [Start with docker-compose](#start-with-docker-compose)
   - [Setting up service](#setting-up-service)
   - [Third-party contributions](#third-party-contributions)
-  - [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 / upgrade 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)
@@ -77,10 +81,12 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [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)
@@ -91,56 +97,22 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [We kindly ask:](#we-kindly-ask)
 
 
-### How to build from sources
-
-We recommend using either [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) or
-[docker images](https://hub.docker.com/r/victoriametrics/victoria-metrics/) instead of building VictoriaMetrics
-from sources. Building from sources is reasonable when developing an additional features specific
-to your needs.
-
-
-#### Development build
-
-1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
-2. Run `make victoria-metrics` from the root folder of the repository.
-   It will build `victoria-metrics` binary and put it into the `bin` folder.
-
-#### Production build
-
-1. [Install docker](https://docs.docker.com/install/).
-2. Run `make victoria-metrics-prod` from the root folder of the repository.
-   It will build `victoria-metrics-prod` binary and put it into the `bin` folder.
-
-#### Building docker images
-
-Run `make package-victoria-metrics`. It will build `victoriametrics/victoria-metrics:<PKG_TAG>` docker image locally.
-`<PKG_TAG>` is auto-generated image tag, which depends on source code in the repository.
-The `<PKG_TAG>` may be manually set via `PKG_TAG=foobar make package-victoria-metrics`.
-
-
 ### How to start VictoriaMetrics
 
-Just start VictoriaMetrics executable or docker image with the desired command-line flags.
+Just start VictoriaMetrics [executable](https://github.com/VictoriaMetrics/VictoriaMetrics/releases)
+or [docker image](https://hub.docker.com/r/victoriametrics/victoria-metrics/) with the desired command-line flags.
 
-The following command line flags are used the most:
+The following command-line flags are used the most:
 
 * `-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory.
 * `-retentionPeriod` - retention period in months for the data. Older data is automatically deleted.
-* `-httpListenAddr` - TCP address to listen to for http requests. By default it listens port `8428` on all the network interfaces.
-* `-graphiteListenAddr` - TCP and UDP address to listen to for Graphite data. By default it is disabled.
-* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data. By default it is disabled.
+* `-httpListenAddr` - TCP address to listen to for http requests. By default, it listens port `8428` on all the network interfaces.
+* `-graphiteListenAddr` - TCP and UDP address to listen to for Graphite data. By default, it is disabled.
+* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data. By default, it is disabled.
 
 Pass `-help` to see all the available flags with description and default values.
 
-
-### Setting up service
-
-Read [these instructions](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/43) on how to set up VictoriaMetrics as a service in your OS.
-
-
-### Third-party contributions
-
-* [Unofficial yum repository](https://copr.fedorainfracloud.org/coprs/antonpatsev/VictoriaMetrics/) ([source code](https://github.com/patsevanton/victoriametrics-rpm))
+It is recommended setting up [monitoring](#monitoring) for VictoriaMetrics.
 
 
 ### Prometheus setup
@@ -152,7 +124,7 @@ remote_write:
   - url: http://<victoriametrics-addr>:8428/api/v1/write
     queue_config:
       max_samples_per_send: 10000
-      max_shards: 100
+      max_shards: 30
 ```
 
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
@@ -162,11 +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
@@ -177,7 +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
@@ -194,6 +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.
@@ -208,7 +206,9 @@ Do not forget substituting `<victoriametrics-addr>` with the real address where
 
 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 with `{measurement}{separator}` value. `{separator}` equals to `.` by default, but can be changed with `-influxMeasurementFieldSeparator` command-line flag.
+* 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.
 
@@ -221,17 +221,18 @@ foo,tag1=value1,tag2=value2 field1=12,field2=40
 is converted into the following Prometheus data points:
 
 ```
-foo.field1{tag1="value1", tag2="value2"} 12
-foo.field2{tag1="value1", tag2="value2"} 40
+foo_field1{tag1="value1", tag2="value2"} 12
+foo_field2{tag1="value1", tag2="value2"} 40
 ```
 
-Example for writing data with Influx line protocol to local VictoriaMetrics using `curl`:
+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'
 ```
 
-Arbitrary number of lines delimited by '\n' may be sent in a single request.
+An arbitrary number of lines delimited by '\n' may be sent in a single request.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
@@ -245,6 +246,9 @@ The `/api/v1/export` endpoint should return the following response:
 {"metric":{"__name__":"measurement.field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
 ```
 
+Note that Influx line protocol expects [timestamps in *nanoseconds* by default](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/#timestamp),
+while VictoriaMetrics stores them with *milliseconds* precision.
+
 
 ### How to send data from Graphite-compatible agents such as [StatsD](https://github.com/etsy/statsd)?
 
@@ -265,8 +269,8 @@ Example for writing data with Graphite plaintext protocol to local VictoriaMetri
 echo "foo.bar.baz;tag1=value1;tag2=value2 123 `date +%s`" | nc -N localhost 2003
 ```
 
-VictoriaMetrics sets the current time if timestamp is omitted.
-Arbitrary number of lines delimited by `\n` may be sent in one go.
+VictoriaMetrics sets the current time if the timestamp is omitted.
+An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
@@ -280,6 +284,14 @@ The `/api/v1/export` endpoint should return the following response:
 ```
 
 
+### Querying Graphite data
+
+Data sent to VictoriaMetrics via `Graphite plaintext protocol` may be read either via
+[Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/)
+or via [go-graphite/carbonapi](https://github.com/go-graphite/carbonapi/blob/master/cmd/carbonapi/carbonapi.example.prometheus.yaml).
+
+
+
 ### How to send data from OpenTSDB-compatible agents?
 
 1) Enable OpenTSDB receiver in VictoriaMetrics by setting `-opentsdbListenAddr` command line flag. For instance,
@@ -298,7 +310,7 @@ Example for writing data with OpenTSDB protocol to local VictoriaMetrics using `
 echo "put foo.bar.baz `date +%s` 123 tag1=value1 tag2=value2" | nc -N localhost 4242
 ```
 
-Arbitrary number of lines delimited by `\n` may be sent in one go.
+An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
@@ -312,18 +324,79 @@ The `/api/v1/export` endpoint should return the following response:
 ```
 
 
-### How to apply new config / upgrade VictoriaMetrics?
+### How to build from sources
 
-VictoriaMetrics must be restarted in order to upgrade or apply new config:
+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.
 
-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.
+
+#### Development build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics` from the root folder of the repository.
+   It builds `victoria-metrics` binary and puts it into the `bin` folder.
+
+#### Production build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make victoria-metrics-prod` from the root folder of the repository.
+   It builds `victoria-metrics-prod` binary and puts it into the `bin` folder.
+
+#### ARM build
+
+ARM build may run on Raspberry Pi or on [energy-efficient ARM servers](https://blog.cloudflare.com/arm-takes-wing/).
+
+#### Development ARM build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics-arm` or `make victoria-metrics-arm64` from the root folder of the repository.
+   It builds `victoria-metrics-arm` or `victoria-metrics-arm64` binary respectively and puts it into the `bin` folder.
+
+#### Production ARM build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make victoria-metrics-arm-prod` or `make victoria-metrics-arm64-prod` from the root folder of the repository.
+   It builds `victoria-metrics-arm-prod` or `victoria-metrics-arm64-prod` binary respectively and puts it into the `bin` folder.
+
+#### Pure Go build (CGO_ENABLED=0)
+
+`Pure Go` mode builds only Go code without [cgo](https://golang.org/cmd/cgo/) dependencies.
+This is an experimental mode, which may result in a lower compression ratio and slower decompression performance.
+Use it with caution!
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make victoria-metrics-pure` from the root folder of the repository.
+   It builds `victoria-metrics-pure` binary and puts it into the `bin` folder.
+
+#### Building docker images
+
+Run `make package-victoria-metrics`. It builds `victoriametrics/victoria-metrics:<PKG_TAG>` docker image locally.
+`<PKG_TAG>` is auto-generated image tag, which depends on source code in the repository.
+The `<PKG_TAG>` may be manually set via `PKG_TAG=foobar make package-victoria-metrics`.
+
+
+### Start with docker-compose
+
+[Docker-compose](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/docker-compose.yml)
+helps to spin up VictoriaMetrics, Prometheus and Grafana with one command.
+More details may be found [here](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/master/deployment/docker#folder-contains-basic-images-and-tools-for-building-and-running-victoria-metrics-in-docker).
+
+
+### Setting up service
+
+Read [these instructions](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/43) on how to set up VictoriaMetrics as a service in your OS.
+
+
+### Third-party contributions
+
+* [Unofficial yum repository](https://copr.fedorainfracloud.org/coprs/antonpatsev/VictoriaMetrics/) ([source code](https://github.com/patsevanton/victoriametrics-rpm))
 
 
 ### How to work with snapshots?
 
-VictoriaMetrics is able to create [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
+VictoriaMetrics can create [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
 for all the data stored under `-storageDataPath` directory.
 Navigate to `http://<victoriametrics-addr>:8428/snapshot/create` in order to create an instant snapshot.
 The page will return the following JSON response:
@@ -380,31 +453,55 @@ 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 overridden.
+`start` and `end` may contain either unix timestamp in seconds or [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) values. By default, the last point
+on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but can be overridden.
 For instance, `/federate?match[]=up&max_lookback=1h` would return last points on the `[now - 1h ... now]` interval. This may be useful for time series federation
 with scrape intervals exceeding `5m`.
 
 
 ### Capacity planning
 
-Rough estimation of the required resources:
+A rough estimation of the required resources for ingestion path:
 
 * RAM size: less than 1KB per active time series. So, ~1GB of RAM is required for 1M active time series.
   Time series is considered active if new data points have been added to it recently or if it has been recently queried.
-  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
@@ -429,6 +526,10 @@ 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:
@@ -446,7 +547,7 @@ There is no downsampling support at the moment, but:
 - VictoriaMetrics has good compression for on-disk data. See [this article](https://medium.com/@valyala/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932)
   for details.
 
-These properties reduce the need in downsampling. We plan implementing downsampling in the future.
+These properties reduce the need in downsampling. We plan to implement downsampling in the future.
 See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/36) for details.
 
 
@@ -458,17 +559,23 @@ Single-node VictoriaMetrics doesn't support multi-tenancy. Use [cluster version]
 ### Scalability and cluster version
 
 Though single-node VictoriaMetrics cannot scale to multiple nodes, it is optimized for resource usage - storage size / bandwidth / IOPS, RAM, CPU.
-This means that a single-node VictoriaMetrics may scale vertically and substitute moderately sized cluster built with competing solutions
-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.
@@ -483,10 +590,10 @@ 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.
 
 
@@ -496,6 +603,21 @@ 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
 
@@ -506,6 +628,26 @@ 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
 
@@ -516,10 +658,18 @@ Contact us with any questions regarding VictoriaMetrics at [info@victoriametrics
 
 Feel free asking any questions regarding VictoriaMetrics:
 
+- [slack](http://slack.victoriametrics.com/)
 - [telergam-en](https://t.me/VictoriaMetrics_en)
 - [telergam-ru](https://t.me/VictoriaMetrics_ru1)
 - [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):
 
 - Prefer simple code and architecture.

app/victoria-metrics/Makefile

@@ -21,7 +21,52 @@ run-victoria-metrics:
 	$(MAKE) run-via-docker
 
 victoria-metrics-arm:
-	CC=arm-linux-gnueabi-gcc CGO_ENABLED=1 GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm ./app/victoria-metrics
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm ./app/victoria-metrics
+
+victoria-metrics-arm-prod:
+	APP_NAME=victoria-metrics APP_SUFFIX='-arm' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm' $(MAKE) app-via-docker
 
 victoria-metrics-arm64:
-	CC=aarch64-linux-gnu-gcc CGO_ENABLED=1 GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm64 ./app/victoria-metrics
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm64 ./app/victoria-metrics
+
+victoria-metrics-arm64-prod:
+	APP_NAME=victoria-metrics APP_SUFFIX='-arm64' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm64' $(MAKE) app-via-docker
+
+victoria-metrics-pure:
+	GO111MODULE=on CGO_ENABLED=0 go build -mod=vendor -ldflags "$(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/lines_reader.go

@@ -24,6 +24,7 @@ func ReadLinesBlock(r io.Reader, dstBuf, tailBuf []byte) ([]byte, []byte, error)
 		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.
@@ -34,7 +35,7 @@ again:
 		}
 		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
+			// 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

app/vminsert/common/lines_reader_test.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"fmt"
 	"io"
+	"reflect"
 	"testing"
 )
 
@@ -40,7 +41,73 @@ func (fr *failureReader) Read(p []byte) (int, error) {
 	return 0, fmt.Errorf("some error")
 }
 
-func TestReadLineBlockSuccessSingleByteReader(t *testing.T) {
+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()
 
@@ -87,7 +154,7 @@ func TestReadLineBlockSuccessSingleByteReader(t *testing.T) {
 	f(string(b), string(b[:maxLineSize]), "")
 }
 
-func TestReadLineBlockSuccessBytesBuffer(t *testing.T) {
+func TestReadLinesBlockSuccessBytesBuffer(t *testing.T) {
 	f := func(s, dstBufExpected, tailBufExpected string) {
 		t.Helper()
 

app/vminsert/concurrencylimiter/concurrencylimiter.go

@@ -32,6 +32,17 @@ func Init() {
 func Do(f func() error) error {
 	// Limit the number of conurrent f calls in order to prevent from excess
 	// memory usage and CPU trashing.
+	select {
+	case ch <- struct{}{}:
+		err := f()
+		<-ch
+		return err
+	default:
+	}
+
+	// All the workers are busy.
+	// Sleep for up to waitDuration.
+	concurrencyLimitReached.Inc()
 	t := timerpool.Get(waitDuration)
 	select {
 	case ch <- struct{}{}:
@@ -41,9 +52,19 @@ func Do(f func() error) error {
 		return err
 	case <-t.C:
 		timerpool.Put(t)
-		concurrencyLimitErrors.Inc()
+		concurrencyLimitTimeout.Inc()
 		return fmt.Errorf("the server is overloaded with %d concurrent inserts; either increase -maxConcurrentInserts or reduce the load", cap(ch))
 	}
 }
 
-var concurrencyLimitErrors = metrics.NewCounter(`vm_concurrency_limit_errors_total`)
+var (
+	concurrencyLimitReached = metrics.NewCounter(`vm_concurrent_insert_limit_reached_total`)
+	concurrencyLimitTimeout = metrics.NewCounter(`vm_concurrent_insert_limit_timeout_total`)
+
+	_ = metrics.NewGauge(`vm_concurrent_insert_capacity`, func() float64 {
+		return float64(cap(ch))
+	})
+	_ = metrics.NewGauge(`vm_concurrent_insert_current`, func() float64 {
+		return float64(len(ch))
+	})
+)

app/vminsert/graphite/parser.go

@@ -114,6 +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
@@ -121,6 +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/request_handler.go

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

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:]

app/vminsert/influx/parser_test.go

@@ -348,4 +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

@@ -18,11 +18,14 @@ import (
 )
 
 var (
-	measurementFieldSeparator = flag.String("influxMeasurementFieldSeparator", ".", "Separator for `{measurement}{separator}{field_name}` metric name when inserted via Influx line protocol")
+	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"}`)
+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.
 //
@@ -84,6 +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]
@@ -109,8 +113,10 @@ func (ctx *pushCtx) InsertRows(db string) error {
 			ic.AddLabel("", metricGroup)
 			ic.WriteDataPoint(ctx.metricNameBuf, ic.Labels[:1], r.Timestamp, f.Value)
 		}
-		rowsInserted.Add(len(r.Fields))
+		rowsTotal += len(r.Fields)
 	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
 	return ic.FlushBufs()
 }
 
@@ -164,6 +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 {

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

@@ -14,7 +14,10 @@ import (
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var rowsInserted = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb"}`)
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb"}`)
+	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="opentsdb"}`)
+)
 
 // insertHandler processes remote write for OpenTSDB put protocol.
 //
@@ -51,6 +54,7 @@ func (ctx *pushCtx) InsertRows() error {
 		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
 	return ic.FlushBufs()
 }
 

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

@@ -30,29 +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 while until giving up. This should resolve short bursts in requests.
-	t := timerpool.Get(*maxQueueDuration)
 	select {
 	case concurrencyCh <- struct{}{}:
-		timerpool.Put(t)
 		defer func() { <-concurrencyCh }()
-	case <-t.C:
-		timerpool.Put(t)
-		httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
-		return true
+	default:
+		// Sleep for a while until giving up. This should resolve short bursts in requests.
+		concurrencyLimitReached.Inc()
+		t := timerpool.Get(*maxQueueDuration)
+		select {
+		case concurrencyCh <- struct{}{}:
+			timerpool.Put(t)
+			defer func() { <-concurrencyCh }()
+		case <-t.C:
+			timerpool.Put(t)
+			concurrencyLimitTimeout.Inc()
+			httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
+			return true
+		}
 	}
 
 	path := strings.Replace(r.URL.Path, "//", "/", -1)

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()
 
@@ -459,7 +463,7 @@ var ssPool sync.Pool
 var missingMetricNamesForMetricID = metrics.NewCounter(`vm_missing_metric_names_for_metric_id_total`)
 
 // ProcessSearchQuery performs sq on storage nodes until the given deadline.
-func ProcessSearchQuery(sq *storage.SearchQuery, deadline Deadline) (*Results, error) {
+func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadline) (*Results, error) {
 	// Setup search.
 	tfss, err := setupTfss(sq.TagFilterss)
 	if err != nil {
@@ -475,35 +479,38 @@ func ProcessSearchQuery(sq *storage.SearchQuery, deadline Deadline) (*Results, e
 
 	sr := getStorageSearch()
 	defer putStorageSearch(sr)
-	sr.Init(vmstorage.Storage, tfss, tr, *maxMetricsPerSearch)
+	sr.Init(vmstorage.Storage, tfss, tr, fetchData, *maxMetricsPerSearch)
 
 	tbf := getTmpBlocksFile()
 	m := make(map[string][]tmpBlockAddr)
+	blocksRead := 0
 	for sr.NextMetricBlock() {
+		blocksRead++
 		addr, err := tbf.WriteBlock(sr.MetricBlock.Block)
 		if err != nil {
 			putTmpBlocksFile(tbf)
-			return nil, fmt.Errorf("cannot write data to temporary blocks file: %s", err)
+			return nil, fmt.Errorf("cannot write data block #%d to temporary blocks file: %s", blocksRead, err)
 		}
 		if time.Until(deadline.Deadline) < 0 {
 			putTmpBlocksFile(tbf)
-			return nil, fmt.Errorf("timeout exceeded while fetching data from storage: %s", deadline.Timeout)
+			return nil, fmt.Errorf("timeout exceeded while fetching data block #%d from storage: %s", blocksRead, deadline.Timeout)
 		}
 		metricName := sr.MetricBlock.MetricName
 		m[string(metricName)] = append(m[string(metricName)], addr)
 	}
 	if err := sr.Error(); err != nil {
 		putTmpBlocksFile(tbf)
-		return nil, fmt.Errorf("search error: %s", err)
+		return nil, fmt.Errorf("search error after reading %d data blocks: %s", blocksRead, err)
 	}
 	if err := tbf.Finalize(); err != nil {
 		putTmpBlocksFile(tbf)
-		return nil, fmt.Errorf("cannot finalize temporary blocks file: %s", err)
+		return nil, fmt.Errorf("cannot finalize temporary blocks file with %d blocks: %s", blocksRead, err)
 	}
 
 	var rss Results
 	rss.packedTimeseries = make([]packedTimeseries, len(m))
 	rss.tr = tr
+	rss.fetchData = fetchData
 	rss.deadline = deadline
 	rss.tbf = tbf
 	i := 0

app/vmselect/prometheus/prometheus.go

@@ -6,12 +6,15 @@ import (
 	"math"
 	"net/http"
 	"runtime"
+	"sort"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/promql"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/quicktemplate"
@@ -65,7 +68,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 		MaxTimestamp: end,
 		TagFilterss:  tagFilterss,
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, true, deadline)
 	if err != nil {
 		return fmt.Errorf("cannot fetch data for %q: %s", sq, err)
 	}
@@ -73,7 +76,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	resultsCh := make(chan *quicktemplate.ByteBuffer)
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			WriteFederate(bb, rs)
 			resultsCh <- bb
@@ -157,7 +160,7 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 		MaxTimestamp: end,
 		TagFilterss:  tagFilterss,
 	}
-	rss, err := netstorage.ProcessSearchQuery(sq, deadline)
+	rss, err := netstorage.ProcessSearchQuery(sq, true, deadline)
 	if err != nil {
 		return fmt.Errorf("cannot fetch data for %q: %s", sq, err)
 	}
@@ -165,7 +168,7 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 	resultsCh := make(chan *quicktemplate.ByteBuffer, runtime.GOMAXPROCS(-1))
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			writeLineFunc(bb, rs)
 			resultsCh <- bb
@@ -230,9 +233,39 @@ var deleteDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
 	deadline := getDeadline(r)
-	labelValues, err := netstorage.GetLabelValues(labelName, deadline)
-	if err != nil {
-		return fmt.Errorf(`cannot obtain label values for %q: %s`, labelName, err)
+
+	if err := r.ParseForm(); err != nil {
+		return fmt.Errorf("cannot parse form values: %s", err)
+	}
+	var labelValues []string
+	if len(r.Form["match[]"]) == 0 && len(r.Form["start"]) == 0 && len(r.Form["end"]) == 0 {
+		var err error
+		labelValues, err = netstorage.GetLabelValues(labelName, deadline)
+		if err != nil {
+			return fmt.Errorf(`cannot obtain label values for %q: %s`, labelName, err)
+		}
+	} else {
+		// Extended functionality that allows filtering by label filters and time range
+		// i.e. /api/v1/label/foo/values?match[]=foobar{baz="abc"}&start=...&end=...
+		// is equivalent to `label_values(foobar{baz="abc"}, foo)` call on the selected
+		// time range in Grafana templating.
+		matches := r.Form["match[]"]
+		if len(matches) == 0 {
+			matches = []string{fmt.Sprintf("{%s!=''}", labelName)}
+		}
+		ct := currentTime()
+		end, err := getTime(r, "end", ct)
+		if err != nil {
+			return err
+		}
+		start, err := getTime(r, "start", end-defaultStep)
+		if err != nil {
+			return err
+		}
+		labelValues, err = labelValuesWithMatches(labelName, matches, start, end, deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain label values for %q, match[]=%q, start=%d, end=%d: %s", labelName, matches, start, end, err)
+		}
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -241,6 +274,50 @@ func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request
 	return nil
 }
 
+func labelValuesWithMatches(labelName string, matches []string, start, end int64, deadline netstorage.Deadline) ([]string, error) {
+	if len(matches) == 0 {
+		logger.Panicf("BUG: matches must be non-empty")
+	}
+	tagFilterss, err := getTagFilterssFromMatches(matches)
+	if err != nil {
+		return nil, err
+	}
+	if start >= end {
+		start = end - defaultStep
+	}
+	sq := &storage.SearchQuery{
+		MinTimestamp: start,
+		MaxTimestamp: end,
+		TagFilterss:  tagFilterss,
+	}
+	rss, err := netstorage.ProcessSearchQuery(sq, false, deadline)
+	if err != nil {
+		return nil, fmt.Errorf("cannot fetch data for %q: %s", sq, err)
+	}
+
+	m := make(map[string]struct{})
+	var mLock sync.Mutex
+	err = rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
+		labelValue := rs.MetricName.GetTagValue(labelName)
+		if len(labelValue) == 0 {
+			return
+		}
+		mLock.Lock()
+		m[string(labelValue)] = struct{}{}
+		mLock.Unlock()
+	})
+	if err != nil {
+		return nil, fmt.Errorf("error when data fetching: %s", err)
+	}
+
+	labelValues := make([]string, 0, len(m))
+	for labelValue := range m {
+		labelValues = append(labelValues, labelValue)
+	}
+	sort.Strings(labelValues)
+	return labelValues, nil
+}
+
 var labelValuesDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/label/{}/values"}`)
 
 // LabelsCountHandler processes /api/v1/labels/count request.
@@ -309,11 +386,16 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		return fmt.Errorf("missing `match[]` arg")
 	}
-	start, err := getTime(r, "start", ct-defaultStep)
+	end, err := getTime(r, "end", ct)
 	if err != nil {
 		return err
 	}
-	end, err := getTime(r, "end", ct)
+	// Do not set start to minTimeMsecs by default as Prometheus does,
+	// since this leads to fetching and scanning all the data from the storage,
+	// which can take a lot of time for big storages.
+	// It is better setting start as end-defaultStep by default.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/91
+	start, err := getTime(r, "start", end-defaultStep)
 	if err != nil {
 		return err
 	}
@@ -331,7 +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)
 	}
@@ -339,7 +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
@@ -426,7 +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)
 	}
@@ -484,12 +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")
@@ -502,17 +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 {
@@ -521,21 +603,22 @@ 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, error) {
@@ -548,20 +631,38 @@ func getTime(r *http.Request, argKey string, defaultValue int64) (int64, error)
 		// Try parsing string format
 		t, err := time.Parse(time.RFC3339, argValue)
 		if err != nil {
+			// 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 0, fmt.Errorf("%q=%dms is out of allowed range [%d ... %d]", argKey, msecs, minTimeMsecs, maxTimeMsecs)
+	if msecs < minTimeMsecs {
+		msecs = 0
+	}
+	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
 )
 
@@ -581,7 +682,7 @@ func getDuration(r *http.Request, argKey string, defaultValue int64) (int64, err
 	}
 	msecs := int64(secs * 1e3)
 	if msecs <= 0 || msecs > maxDurationMsecs {
-		return 0, fmt.Errorf("%q=%dms is out of allowed range [%d ... %d]", argKey, msecs, 0, maxDurationMsecs)
+		return 0, fmt.Errorf("%q=%dms is out of allowed range [%d ... %d]", argKey, msecs, 0, int64(maxDurationMsecs))
 	}
 	return msecs, nil
 }

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]
@@ -140,6 +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.
@@ -268,7 +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]
 }
@@ -383,7 +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)
 	}

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/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
 	}
 }
@@ -413,10 +416,25 @@ func binaryOpIfnot(left, right float64) float64 {
 }
 
 func binaryOpEq(left, right float64) bool {
+	// Special handling for nan == nan.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
+	if math.IsNaN(left) {
+		return math.IsNaN(right)
+	}
+
 	return left == right
 }
 
 func binaryOpNeq(left, right float64) bool {
+	// Special handling for comparison with nan.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
+	if math.IsNaN(left) {
+		return !math.IsNaN(right)
+	}
+	if math.IsNaN(right) {
+		return true
+	}
+
 	return left != right
 }
 

app/vmselect/promql/eval.go

@@ -16,7 +16,7 @@ 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 minimum number of points per timeseries for enabling time rounding.
@@ -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
 }
 
@@ -478,9 +538,21 @@ var (
 	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()
@@ -498,7 +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
 	}
@@ -530,36 +602,18 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	defer rml.Put(uint64(rollupMemorySize))
 
 	// Evaluate rollup
-	tss := make([]*timeseries, 0, rssLen*len(rcs))
-	var tssLock sync.Mutex
-	err = rss.RunParallel(func(rs *netstorage.Result) {
-		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
-
-			tssLock.Lock()
-			tss = append(tss, &ts)
-			tssLock.Unlock()
-		}
-	})
+	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
 	}
-	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)
+	rollupResultCacheV.Put(name, ec, me, iafc, window, tss)
 
 	return tss, nil
 }
@@ -576,6 +630,63 @@ func getRollupMemoryLimiter() *memoryLimiter {
 	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, workerID uint) {
+		preFunc(rs.Values, rs.Timestamps)
+		for _, rc := range rcs {
+			var ts timeseries
+			doRollupForTimeseries(rc, &ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
+			tssLock.Lock()
+			tss = append(tss, &ts)
+			tssLock.Unlock()
+		}
+	})
+	if err != nil {
+		return nil, err
+	}
+	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] {
@@ -585,13 +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 {
@@ -618,6 +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, ""))
 	}

app/vmselect/promql/exec.go

@@ -16,6 +16,8 @@ import (
 
 var logSlowQueryDuration = flag.Duration("search.logSlowQueryDuration", 5*time.Second, "Log queries with execution time exceeding this value. Zero disables slow query logging")
 
+var slowQueries = metrics.NewCounter(`vm_slow_queries_total`)
+
 // ExpandWithExprs expands WITH expressions inside q and returns the resulting
 // PromQL without WITH expressions.
 func ExpandWithExprs(q string) (string, error) {
@@ -27,14 +29,16 @@ 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: duration=%s, start=%d, end=%d, step=%d, query=%q", d, ec.Start/1000, ec.End/1000, ec.Step/1000, q)
+				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()
 			}
 		}()
 	}
@@ -65,6 +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 {
@@ -122,18 +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)`
@@ -2134,6 +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({})`
@@ -2184,6 +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)`
@@ -2195,6 +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"))`
@@ -2206,6 +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"))`
@@ -2230,10 +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}
@@ -2395,6 +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")))`
@@ -2494,7 +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}
@@ -2505,7 +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}
@@ -2761,7 +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}
@@ -2783,7 +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}
@@ -2794,7 +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}
@@ -2805,7 +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}
@@ -2816,7 +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}
@@ -2827,7 +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}
@@ -2838,7 +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}
@@ -2860,7 +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}
@@ -3125,6 +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()))`
@@ -3163,7 +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{{
@@ -3172,21 +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)
@@ -3545,7 +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)
 			}
@@ -3688,6 +3885,13 @@ func TestExecError(t *testing.T) {
 		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

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

app/vmselect/promql/parser.go

@@ -18,7 +18,7 @@ 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)
 			)`,
 
@@ -373,7 +373,7 @@ func (p *parser) parseSingleExpr() (expr, error) {
 }
 
 func (p *parser) parseSingleExprWithoutRollupSuffix() (expr, error) {
-	if isPositiveNumberPrefix(p.lex.Token) {
+	if isPositiveNumberPrefix(p.lex.Token) || isInfOrNaN(p.lex.Token) {
 		return p.parsePositiveNumberExpr()
 	}
 	if isStringPrefix(p.lex.Token) {
@@ -417,7 +417,7 @@ func (p *parser) parseSingleExprWithoutRollupSuffix() (expr, error) {
 }
 
 func (p *parser) parsePositiveNumberExpr() (*numberExpr, error) {
-	if !isPositiveNumberPrefix(p.lex.Token) {
+	if !isPositiveNumberPrefix(p.lex.Token) && !isInfOrNaN(p.lex.Token) {
 		return nil, fmt.Errorf(`positiveNumberExpr: unexpected token %q; want "number"`, p.lex.Token)
 	}
 
@@ -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)

app/vmselect/promql/parser_test.go

@@ -118,6 +118,10 @@ 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"}`)
@@ -166,14 +170,34 @@ func TestParsePromQLSuccess(t *testing.T) {
 	another(`-.2`, `-0.2`)
 	another(`-.2E-2`, `-0.002`)
 	same(`NaN`)
+	another(`nan`, `NaN`)
+	another(`NAN`, `NaN`)
+	another(`nAN`, `NaN`)
 	another(`Inf`, `+Inf`)
+	another(`INF`, `+Inf`)
+	another(`inf`, `+Inf`)
 	another(`+Inf`, `+Inf`)
 	another(`-Inf`, `-Inf`)
+	another(`-inF`, `-Inf`)
 
 	// binaryOpExpr
-	another(`NaN + 2 *3 * Inf`, `NaN`)
-	another(`Inf - Inf`, `NaN`)
-	another(`Inf + Inf`, `+Inf`)
+	another(`nan == nan`, `NaN`)
+	another(`nan ==bool nan`, `1`)
+	another(`nan !=bool nan`, `0`)
+	another(`nan !=bool 2`, `1`)
+	another(`2 !=bool nan`, `1`)
+	another(`nan >bool nan`, `0`)
+	another(`nan <bool nan`, `0`)
+	another(`1 ==bool nan`, `0`)
+	another(`NaN !=bool 1`, `1`)
+	another(`inf >=bool 2`, `1`)
+	another(`-1 >bool -inf`, `1`)
+	another(`-1 <bool -inf`, `0`)
+	another(`nan + 2 *3 * inf`, `NaN`)
+	another(`INF - Inf`, `NaN`)
+	another(`Inf + inf`, `+Inf`)
+	another(`1/0`, `+Inf`)
+	another(`0/0`, `NaN`)
 	another(`-m`, `0 - m`)
 	same(`m + ignoring () n[5m]`)
 	another(`M + IGNORING () N[5m]`, `M + ignoring () N[5m]`)
@@ -345,8 +369,11 @@ 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)`)
@@ -671,7 +698,7 @@ 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`)
@@ -681,8 +708,7 @@ 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/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),
@@ -37,16 +38,33 @@ 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),
 	"integrate":          newRollupFuncOneArg(rollupIntegrate),
 	"ideriv":             newRollupFuncOneArg(rollupIderiv),
+	"lifetime":           newRollupFuncOneArg(rollupLifetime),
+	"scrape_interval":    newRollupFuncOneArg(rollupScrapeInterval),
 	"rollup":             newRollupFuncOneArg(rollupFake),
 	"rollup_rate":        newRollupFuncOneArg(rollupFake), // + rollupFuncsRemoveCounterResets
 	"rollup_deriv":       newRollupFuncOneArg(rollupFake),
 	"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,
+	"lifetime":        true,
+	"scrape_interval": true,
 }
 
 var rollupFuncsRemoveCounterResets = map[string]bool{
@@ -64,6 +82,7 @@ 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 +139,13 @@ 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 +188,7 @@ 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 +197,7 @@ 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 +235,27 @@ 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 +288,14 @@ 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 +304,23 @@ 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 +347,11 @@ 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 +393,55 @@ 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 +451,15 @@ 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 +477,11 @@ 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 +491,11 @@ 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 +507,11 @@ 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 +523,11 @@ 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 +536,43 @@ 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 +585,18 @@ 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 +611,7 @@ 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
@@ -512,43 +619,61 @@ func rollupDelta(rfa *rollupFuncArg) float64 {
 		if len(values) == 0 {
 			return nan
 		}
+		if len(values) == 1 {
+			// Assume that the previous non-existing value was 0.
+			return values[0]
+		}
 		prevValue = values[0]
 		values = values[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		// Assume that the value didn't change on the given interval.
+		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
+		}
+		// Assume that the value didn't change on the given interval.
+		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
+			// Assume that the previous non-existing value was 0.
+			return lastValue
 		}
 		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
 	prevValue := rfa.prevValue
 	prevTimestamp := rfa.prevTimestamp
 	if math.IsNaN(prevValue) {
-		if len(values) == 0 {
+		if len(values) < 2 {
+			// It is impossible to calculate derivative on 0 or 1 values.
 			return nan
 		}
 		prevValue = values[0]
@@ -557,7 +682,8 @@ func rollupDeriv(rfa *rollupFuncArg) float64 {
 		timestamps = timestamps[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		// Assume that the value didn't change on the given interval.
+		return 0
 	}
 	vEnd := values[len(values)-1]
 	tEnd := timestamps[len(timestamps)-1]
@@ -567,43 +693,85 @@ 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 len(values) < 2 {
+		if len(values) == 0 || math.IsNaN(rfa.prevValue) {
+			// It is impossible to calculate derivative on 0 or 1 values.
+			return nan
+		}
+		return (values[0] - rfa.prevValue) / (float64(timestamps[0]-rfa.prevTimestamp) * 1e-3)
 	}
 	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) {
+	// Skip data points with duplicate timestamps.
+	for len(timestamps) > 0 && timestamps[len(timestamps)-1] >= tEnd {
+		timestamps = timestamps[:len(timestamps)-1]
+	}
+	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) * 1e-3)
+}
+
+func rollupLifetime(rfa *rollupFuncArg) float64 {
+	// Calculate the duration between the first and the last data points.
+	timestamps := rfa.timestamps
+	if math.IsNaN(rfa.prevValue) {
+		if len(timestamps) < 2 {
 			return nan
 		}
-	} else {
-		prevValue = values[len(values)-1]
-		prevTimestamp = timestamps[len(timestamps)-1]
+		return float64(timestamps[len(timestamps)-1]-timestamps[0]) * 1e-3
 	}
-	dv := vEnd - prevValue
-	dt := tEnd - prevTimestamp
-	return dv / (float64(dt) / 1000)
+	if len(timestamps) == 0 {
+		return nan
+	}
+	return float64(timestamps[len(timestamps)-1]-rfa.prevTimestamp) * 1e-3
+}
+
+func rollupScrapeInterval(rfa *rollupFuncArg) float64 {
+	// Calculate the average interval between data points.
+	timestamps := rfa.timestamps
+	if math.IsNaN(rfa.prevValue) {
+		if len(timestamps) < 2 {
+			return nan
+		}
+		return float64(timestamps[len(timestamps)-1]-timestamps[0]) * 1e-3 / float64(len(timestamps)-1)
+	}
+	if len(timestamps) == 0 {
+		return nan
+	}
+	return (float64(timestamps[len(timestamps)-1]-rfa.prevTimestamp) * 1e-3) / float64(len(timestamps))
 }
 
 func rollupChanges(rfa *rollupFuncArg) float64 {
-	// 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 +783,14 @@ 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 +798,7 @@ func rollupResets(rfa *rollupFuncArg) float64 {
 		values = values[1:]
 	}
 	if len(values) == 0 {
-		return nan
+		return 0
 	}
 	n := 0
 	for _, v := range values {
@@ -646,7 +817,7 @@ 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 +826,27 @@ 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 +858,15 @@ 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 +876,7 @@ 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

@@ -4,14 +4,15 @@ import (
 	"crypto/rand"
 	"flag"
 	"fmt"
-	"runtime"
 	"sync"
 	"sync/atomic"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -19,7 +20,7 @@ import (
 var disableCache = flag.Bool("search.disableCache", false, "Whether to disable response caching. This may be useful during data backfilling")
 
 var rollupResultCacheV = &rollupResultCache{
-	fastcache.New(1024 * 1024), // This is a cache for testing.
+	c: workingsetcache.New(1024*1024, time.Hour), // This is a cache for testing.
 }
 var rollupResultCachePath string
 
@@ -43,15 +44,17 @@ var (
 func InitRollupResultCache(cachePath string) {
 	rollupResultCachePath = cachePath
 	startTime := time.Now()
-	var c *fastcache.Cache
+	cacheSize := getRollupResultCacheSize()
+	var c *workingsetcache.Cache
 	if len(rollupResultCachePath) > 0 {
 		logger.Infof("loading rollupResult cache from %q...", rollupResultCachePath)
-		c = fastcache.LoadFromFileOrNew(rollupResultCachePath, getRollupResultCacheSize())
+		c = workingsetcache.Load(rollupResultCachePath, cacheSize, time.Hour)
 	} else {
-		c = fastcache.New(getRollupResultCacheSize())
+		c = workingsetcache.New(cacheSize, time.Hour)
 	}
 	if *disableCache {
-		c.Reset()
+		c.Stop()
+		c = nil
 	}
 
 	stats := &fastcache.Stats{}
@@ -71,7 +74,7 @@ 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)
 	}
 
@@ -96,25 +99,28 @@ func InitRollupResultCache(cachePath string) {
 // StopRollupResultCache closes the rollupResult cache.
 func StopRollupResultCache() {
 	if len(rollupResultCachePath) == 0 {
-		rollupResultCacheV.c.Reset()
+		if !*disableCache {
+			rollupResultCacheV.c.Stop()
+			rollupResultCacheV.c = nil
+		}
 		return
 	}
-	gomaxprocs := runtime.GOMAXPROCS(-1)
 	logger.Infof("saving rollupResult cache to %q...", rollupResultCachePath)
 	startTime := time.Now()
-	if err := rollupResultCacheV.c.SaveToFileConcurrent(rollupResultCachePath, gomaxprocs); err != nil {
+	if err := rollupResultCacheV.c.Save(rollupResultCachePath); err != nil {
 		logger.Errorf("cannot close rollupResult cache at %q: %s", rollupResultCachePath, err)
-	} else {
-		var fcs fastcache.Stats
-		rollupResultCacheV.c.UpdateStats(&fcs)
-		rollupResultCacheV.c.Reset()
-		logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, bytesSize: %d",
-			rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
+		return
 	}
+	var fcs fastcache.Stats
+	rollupResultCacheV.c.UpdateStats(&fcs)
+	rollupResultCacheV.c.Stop()
+	rollupResultCacheV.c = nil
+	logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, sizeBytes: %d",
+		rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
 }
 
 type rollupResultCache struct {
-	c *fastcache.Cache
+	c *workingsetcache.Cache
 }
 
 var rollupResultCacheResets = metrics.NewCounter(`vm_cache_resets_total{type="promql/rollupResult"}`)
@@ -125,7 +131,7 @@ func ResetRollupResultCache() {
 	rollupResultCacheV.c.Reset()
 }
 
-func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExpr, window int64) (tss []*timeseries, newStart int64) {
+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
 	}
@@ -134,7 +140,7 @@ 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
@@ -148,15 +154,23 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 		return nil, ec.Start
 	}
 	bb.B = key.Marshal(bb.B[:0])
-	resultBuf := rrc.c.GetBig(nil, bb.B)
-	if len(resultBuf) == 0 {
+	compressedResultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(compressedResultBuf)
+	compressedResultBuf.B = rrc.c.GetBig(compressedResultBuf.B[:0], bb.B)
+	if len(compressedResultBuf.B) == 0 {
 		mi.RemoveKey(key)
 		metainfoBuf = mi.Marshal(metainfoBuf[:0])
-		bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, 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
 	}
-	tss, err := unmarshalTimeseriesFast(resultBuf)
+	// Decompress into newly allocated byte slice, since tss returned from unmarshalTimeseriesFast
+	// refers to the byte slice, so it cannot be returned to the resultBufPool.
+	resultBuf, err := encoding.DecompressZSTD(nil, compressedResultBuf.B)
+	if err != nil {
+		logger.Panicf("BUG: cannot decompress resultBuf from rollupResultCache: %s; it looks like it was improperly saved", err)
+	}
+	tss, err = unmarshalTimeseriesFast(resultBuf)
 	if err != nil {
 		logger.Panicf("BUG: cannot unmarshal timeseries from rollupResultCache: %s; it looks like it was improperly saved", err)
 	}
@@ -196,7 +210,9 @@ 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) {
+var resultBufPool bytesutil.ByteBufferPool
+
+func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64, tss []*timeseries) {
 	if *disableCache || len(tss) == 0 || !ec.mayCache() {
 		return
 	}
@@ -227,11 +243,16 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 
 	// Store tss in the cache.
 	maxMarshaledSize := getRollupResultCacheSize() / 4
-	tssMarshaled := marshalTimeseriesFast(tss, maxMarshaledSize, ec.Step)
-	if tssMarshaled == nil {
+	resultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(resultBuf)
+	resultBuf.B = marshalTimeseriesFast(resultBuf.B[:0], tss, maxMarshaledSize, ec.Step)
+	if len(resultBuf.B) == 0 {
 		tooBigRollupResults.Inc()
 		return
 	}
+	compressedResultBuf := resultBufPool.Get()
+	defer resultBufPool.Put(compressedResultBuf)
+	compressedResultBuf.B = encoding.CompressZSTDLevel(compressedResultBuf.B[:0], resultBuf.B, 1)
 
 	bb := bbPool.Get()
 	defer bbPool.Put(bb)
@@ -240,9 +261,9 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 	key.prefix = rollupResultCacheKeyPrefix
 	key.suffix = atomic.AddUint64(&rollupResultCacheKeySuffix, 1)
 	bb.B = key.Marshal(bb.B[:0])
-	rrc.c.SetBig(bb.B, tssMarshaled)
+	rrc.c.SetBig(bb.B, compressedResultBuf.B)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, 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 {
@@ -270,10 +291,16 @@ 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 = 6
 
-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)
 		}
@@ -36,7 +41,7 @@ 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{
 			{
@@ -44,8 +49,29 @@ func TestRollupResultCache(t *testing.T) {
 				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{
+			{
+				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)
 		}
@@ -67,8 +93,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -86,8 +112,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -105,8 +131,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -124,8 +150,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -143,8 +169,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -166,8 +192,8 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}
@@ -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)
 		}
@@ -220,10 +246,10 @@ func TestRollupResultCache(t *testing.T) {
 				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)
 		}

app/vmselect/promql/rollup_test.go

@@ -10,6 +10,79 @@ 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 !math.IsNaN(n) {
+		t.Fatalf("unexpected value; got %v; want %v", n, nan)
+	}
+
+	rfa = &rollupFuncArg{
+		prevTimestamp: 90,
+		prevValue:     10,
+		values:        []float64{15},
+		timestamps:    []int64{100},
+	}
+	n = rollupIderiv(rfa)
+	if n != 500 {
+		t.Fatalf("unexpected value; got %v; want %v", n, 0.5)
+	}
+
+	rfa = &rollupFuncArg{
+		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 +111,19 @@ 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 +132,13 @@ 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 +146,15 @@ 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 +223,10 @@ 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 +269,11 @@ 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 +283,16 @@ 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 +344,7 @@ 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 +352,14 @@ 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 +375,22 @@ 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 +400,14 @@ 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 +420,7 @@ 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 +434,7 @@ 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 +451,7 @@ 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 +465,7 @@ 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 +479,7 @@ 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 +496,7 @@ 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 +510,7 @@ 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 +524,7 @@ 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 +538,7 @@ 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 +552,7 @@ 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 +566,7 @@ 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)
 	})
@@ -495,10 +580,66 @@ 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{123, 33, -87, 0}
 		timestampsExpected := []int64{10, 50, 90, 130}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("lifetime", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLifetime,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.031, 0.044, 0.04, 0.01}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("lifetime", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLifetime,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 200,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.031, 0.075, 0.115, 0.125}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("scrape_interval", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupScrapeInterval,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.010333333333333333, 0.011, 0.013333333333333334, 0.01}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("scrape_interval", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupScrapeInterval,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 80,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.010333333333333333, 0.010714285714285714, 0.012, 0.0125}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("changes", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupChanges,
@@ -509,10 +650,24 @@ 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 +678,7 @@ 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 +692,13 @@ 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 +706,7 @@ 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 +720,7 @@ 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 +734,7 @@ 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 +748,7 @@ 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 +762,7 @@ 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"
@@ -61,7 +62,21 @@ func (ts *timeseries) CopyShallow(src *timeseries) {
 	ts.denyReuse = true
 }
 
-func marshalTimeseriesFast(tss []*timeseries, maxSize int, step int64) []byte {
+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(dst []byte, tss []*timeseries, maxSize int, step int64) []byte {
 	if len(tss) == 0 {
 		logger.Panicf("BUG: tss cannot be empty")
 	}
@@ -77,13 +92,13 @@ func marshalTimeseriesFast(tss []*timeseries, maxSize int, step int64) []byte {
 
 	if size > maxSize {
 		// Do not marshal tss, since it would occupy too much space
-		return nil
+		return dst
 	}
 
 	// Allocate the buffer for the marshaled tss before its' marshaling.
 	// This should reduce memory fragmentation and memory usage.
-	dst := make([]byte, 0, size)
-	dst = marshalFastTimestamps(dst, tss[0].Timestamps)
+	dst = bytesutil.Resize(dst, size)
+	dst = marshalFastTimestamps(dst[:0], tss[0].Timestamps)
 	for _, ts := range tss {
 		dst = ts.marshalFastNoTimestamps(dst)
 	}

app/vmselect/promql/timeseries_test.go

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

app/vmselect/promql/transform.go

@@ -125,7 +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()
@@ -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

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.
@@ -211,7 +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
@@ -223,7 +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
@@ -340,6 +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)
 	})
@@ -388,22 +429,22 @@ func registerStorageMetrics(strg *storage.Storage) {
 	})
 
 	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().UselessTagFiltersCacheBytesSize)
+		return float64(idbm().UselessTagFiltersCacheSizeBytes)
 	})
 
 	metrics.NewGauge(`vm_cache_requests_total{type="storage/tsid"}`, func() float64 {

dashboards/victoriametrics.json

@@ -60,12 +60,12 @@
       }
     ]
   },
-  "description": "Overview for single node VictoriaMetrics",
+  "description": "Overview for single node VictoriaMetrics v1.22.2 or higher",
   "editable": true,
   "gnetId": 10229,
   "graphTooltip": 0,
   "id": null,
-  "iteration": 1559857416920,
+  "iteration": 1563651131627,
   "links": [
     {
       "icon": "doc",
@@ -668,7 +668,7 @@
       "dashLength": 10,
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
-      "description": "* `*` - unsupported query path\n* `/write` - insert into VM\n* `/metrics` - query VM system metrics\n* `/query` - query instant values\n* `/query_range` - query over a range of time\n* `/series` - match a certain label set\n* `/label/{}/values` - query a list of label values (variables mostly)",
+      "description": "Shows the number of active time series with new data points inserted during the last hour. High value may result in ingestion slowdown. \n\nSee following link for details:",
       "fill": 1,
       "gridPos": {
         "h": 8,
@@ -676,23 +676,27 @@
         "x": 0,
         "y": 19
       },
-      "id": 35,
+      "id": 51,
       "legend": {
-        "alignAsTable": true,
-        "avg": true,
-        "current": true,
+        "avg": false,
+        "current": false,
         "max": false,
         "min": false,
         "show": true,
-        "sort": "current",
-        "sortDesc": true,
         "total": false,
-        "values": true
+        "values": false
       },
       "lines": true,
       "linewidth": 1,
-      "links": [],
-      "nullPointMode": "null as zero",
+      "links": [
+        {
+          "targetBlank": true,
+          "title": "troubleshooting",
+          "type": "absolute",
+          "url": "https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#troubleshooting"
+        }
+      ],
+      "nullPointMode": "null",
       "options": {},
       "percentage": false,
       "pointradius": 2,
@@ -704,10 +708,10 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(rate(vm_http_request_errors_total{job=\"$job\"}[$__interval])) by (path) > 0",
+          "expr": "vm_cache_entries{job=\"$job\", type=\"storage/hour_metric_ids\"}",
           "format": "time_series",
           "intervalFactor": 1,
-          "legendFormat": "{{path}}",
+          "legendFormat": "Active time series",
           "refId": "A"
         }
       ],
@@ -715,10 +719,10 @@
       "timeFrom": null,
       "timeRegions": [],
       "timeShift": null,
-      "title": "Requests error rate",
+      "title": "Active time series",
       "tooltip": {
         "shared": true,
-        "sort": 2,
+        "sort": 0,
         "value_type": "individual"
       },
       "type": "graph",
@@ -859,7 +863,7 @@
       "dashLength": 10,
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
-      "description": "",
+      "description": "* `*` - unsupported query path\n* `/write` - insert into VM\n* `/metrics` - query VM system metrics\n* `/query` - query instant values\n* `/query_range` - query over a range of time\n* `/series` - match a certain label set\n* `/label/{}/values` - query a list of label values (variables mostly)",
       "fill": 1,
       "gridPos": {
         "h": 8,
@@ -867,20 +871,23 @@
         "x": 0,
         "y": 27
       },
-      "id": 37,
+      "id": 35,
       "legend": {
-        "avg": false,
-        "current": false,
+        "alignAsTable": true,
+        "avg": true,
+        "current": true,
         "max": false,
         "min": false,
-        "show": false,
+        "show": true,
+        "sort": "current",
+        "sortDesc": true,
         "total": false,
-        "values": false
+        "values": true
       },
       "lines": true,
       "linewidth": 1,
       "links": [],
-      "nullPointMode": "null",
+      "nullPointMode": "null as zero",
       "options": {},
       "percentage": false,
       "pointradius": 2,
@@ -892,11 +899,10 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(vm_tcplistener_conns{job=\"$job\"})",
+          "expr": "sum(rate(vm_http_request_errors_total{job=\"$job\"}[$__interval])) by (path) > 0",
           "format": "time_series",
-          "hide": false,
           "intervalFactor": 1,
-          "legendFormat": "connections",
+          "legendFormat": "{{path}}",
           "refId": "A"
         }
       ],
@@ -904,10 +910,10 @@
       "timeFrom": null,
       "timeRegions": [],
       "timeShift": null,
-      "title": "TCP connections",
+      "title": "Requests error rate",
       "tooltip": {
         "shared": true,
-        "sort": 0,
+        "sort": 2,
         "value_type": "individual"
       },
       "type": "graph",
@@ -920,7 +926,6 @@
       },
       "yaxes": [
         {
-          "decimals": null,
           "format": "short",
           "label": null,
           "logBase": 1,
@@ -1031,13 +1036,102 @@
         "alignLevel": null
       }
     },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "description": "",
+      "fill": 1,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 0,
+        "y": 35
+      },
+      "id": 37,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": false,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {},
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(vm_tcplistener_conns{job=\"$job\"})",
+          "format": "time_series",
+          "hide": false,
+          "intervalFactor": 1,
+          "legendFormat": "connections",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "TCP connections",
+      "tooltip": {
+        "shared": true,
+        "sort": 0,
+        "value_type": "individual"
+      },
+      "type": "graph",
+      "xaxis": {
+        "buckets": null,
+        "mode": "time",
+        "name": null,
+        "show": true,
+        "values": []
+      },
+      "yaxes": [
+        {
+          "decimals": null,
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        },
+        {
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        }
+      ],
+      "yaxis": {
+        "align": false,
+        "alignLevel": null
+      }
+    },
     {
       "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 35
+        "y": 43
       },
       "id": 14,
       "panels": [],
@@ -1056,7 +1150,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 36
+        "y": 44
       },
       "id": 10,
       "legend": {
@@ -1152,7 +1246,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 36
+        "y": 44
       },
       "id": 34,
       "legend": {
@@ -1254,7 +1348,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 44
+        "y": 52
       },
       "id": 30,
       "legend": {
@@ -1341,7 +1435,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 44
+        "y": 52
       },
       "id": 36,
       "legend": {
@@ -1416,17 +1510,190 @@
         "alignLevel": null
       }
     },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "description": "Shows amount of on-disk space occupied by data points.",
+      "fill": 1,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 0,
+        "y": 60
+      },
+      "id": 53,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": true,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {},
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(vm_data_size_bytes{job=\"$job\", type!=\"indexdb\"})",
+          "format": "time_series",
+          "intervalFactor": 1,
+          "legendFormat": "Size",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "Disk space usage (datapoints)",
+      "tooltip": {
+        "shared": true,
+        "sort": 0,
+        "value_type": "individual"
+      },
+      "type": "graph",
+      "xaxis": {
+        "buckets": null,
+        "mode": "time",
+        "name": null,
+        "show": true,
+        "values": []
+      },
+      "yaxes": [
+        {
+          "format": "bytes",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        },
+        {
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        }
+      ],
+      "yaxis": {
+        "align": false,
+        "alignLevel": null
+      }
+    },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "description": "Shows amount of on-disk space occupied by inverted index.",
+      "fill": 1,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 12,
+        "y": 60
+      },
+      "id": 55,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": true,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {},
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "vm_data_size_bytes{job=\"$job\", type=\"indexdb\"}",
+          "format": "time_series",
+          "intervalFactor": 1,
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "Disk space usage (index)",
+      "tooltip": {
+        "shared": true,
+        "sort": 0,
+        "value_type": "individual"
+      },
+      "type": "graph",
+      "xaxis": {
+        "buckets": null,
+        "mode": "time",
+        "name": null,
+        "show": true,
+        "values": []
+      },
+      "yaxes": [
+        {
+          "format": "bytes",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        },
+        {
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        }
+      ],
+      "yaxis": {
+        "align": false,
+        "alignLevel": null
+      }
+    },
     {
       "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 52
+        "y": 68
       },
       "id": 46,
       "panels": [],
-      "title": "Go runtime",
+      "title": "Resource usage",
       "type": "row"
     },
     {
@@ -1441,7 +1708,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 53
+        "y": 69
       },
       "id": 44,
       "legend": {
@@ -1538,21 +1805,20 @@
       "bars": false,
       "dashLength": 10,
       "dashes": false,
-      "datasource": "${DS_PROMETHEUS}",
       "fill": 1,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 53
+        "y": 69
       },
-      "id": 42,
+      "id": 57,
       "legend": {
         "avg": false,
         "current": false,
         "max": false,
         "min": false,
-        "show": false,
+        "show": true,
         "total": false,
         "values": false
       },
@@ -1571,10 +1837,10 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(rate(go_gc_duration_seconds_sum{job=\"$job\"}[2m]))",
+          "expr": "rate(process_cpu_seconds_total{job=\"$job\"}[5m])",
           "format": "time_series",
-          "intervalFactor": 2,
-          "legendFormat": "gc duration",
+          "intervalFactor": 1,
+          "legendFormat": "Rate 5m",
           "refId": "A"
         }
       ],
@@ -1582,7 +1848,7 @@
       "timeFrom": null,
       "timeRegions": [],
       "timeShift": null,
-      "title": "GC duration",
+      "title": "CPU",
       "tooltip": {
         "shared": true,
         "sort": 0,
@@ -1598,7 +1864,7 @@
       },
       "yaxes": [
         {
-          "format": "s",
+          "format": "short",
           "label": null,
           "logBase": 1,
           "max": null,
@@ -1630,7 +1896,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 61
+        "y": 77
       },
       "id": 47,
       "legend": {
@@ -1717,7 +1983,93 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 61
+        "y": 77
+      },
+      "id": 42,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": false,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {},
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(rate(go_gc_duration_seconds_sum{job=\"$job\"}[2m]))",
+          "format": "time_series",
+          "intervalFactor": 2,
+          "legendFormat": "gc duration",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "GC duration",
+      "tooltip": {
+        "shared": true,
+        "sort": 0,
+        "value_type": "individual"
+      },
+      "type": "graph",
+      "xaxis": {
+        "buckets": null,
+        "mode": "time",
+        "name": null,
+        "show": true,
+        "values": []
+      },
+      "yaxes": [
+        {
+          "format": "s",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        },
+        {
+          "format": "short",
+          "label": null,
+          "logBase": 1,
+          "max": null,
+          "min": null,
+          "show": true
+        }
+      ],
+      "yaxis": {
+        "align": false,
+        "alignLevel": null
+      }
+    },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "fill": 1,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 0,
+        "y": 85
       },
       "id": 48,
       "legend": {
@@ -1744,7 +2096,7 @@
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(go_threads{job=\"$job\"})",
+          "expr": "sum(process_num_threads{job=\"$job\"})",
           "format": "time_series",
           "intervalFactor": 2,
           "legendFormat": "threads",
@@ -1878,5 +2230,5 @@
   "timezone": "",
   "title": "VictoriaMetrics",
   "uid": "wNf0q_kZk",
-  "version": 1
+  "version": 2
 }

deployment/docker/Makefile

@@ -1,5 +1,5 @@
 DOCKER_NAMESPACE := victoriametrics
-BUILDER_IMAGE := local/builder:go1.12.6
+BUILDER_IMAGE := local/builder:go1.12.9
 CERTS_IMAGE := local/certs:1.0.2
 
 package-certs:
@@ -19,8 +19,9 @@ app-via-docker: package-certs package-builder
 		--mount type=bind,src="$(shell pwd)/gocache-for-docker",dst=/gocache \
 		--env GOCACHE=/gocache \
 		--env GO111MODULE=on \
+		$(DOCKER_OPTS) \
 		$(BUILDER_IMAGE) \
-		go build $(RACE) -mod=vendor -ldflags "-s -w -extldflags '-static' $(GO_BUILDINFO)" -tags 'netgo osusergo' -o bin/$(APP_NAME)-prod $(PKG_PREFIX)/app/$(APP_NAME)
+		go build $(RACE) -mod=vendor -ldflags "-s -w -extldflags '-static' $(GO_BUILDINFO)" -tags 'netgo osusergo' -o bin/$(APP_NAME)$(APP_SUFFIX)-prod $(PKG_PREFIX)/app/$(APP_NAME)
 
 package-via-docker:
 	(docker image ls --format '{{.Repository}}:{{.Tag}}' | grep -q '$(DOCKER_NAMESPACE)/$(APP_NAME):$(PKG_TAG)$(RACE)') || (\

deployment/docker/builder/Dockerfile

@@ -1,2 +1,2 @@
-FROM golang:1.12.6
+FROM golang:1.12.9
 STOPSIGNAL SIGINT

go.mod

@@ -2,15 +2,17 @@ module github.com/VictoriaMetrics/VictoriaMetrics
 
 require (
 	github.com/VictoriaMetrics/fastcache v1.5.1
-	github.com/VictoriaMetrics/metrics v1.5.0
+	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 v1.6.0
 	github.com/valyala/histogram v1.0.1
 	github.com/valyala/quicktemplate v1.1.1
-	golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f
+	golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa
 )
 
 go 1.12

go.sum

@@ -3,8 +3,8 @@ github.com/OneOfOne/xxhash v1.2.5 h1:zl/OfRA6nftbBK9qTohYBJ5xvw6C/oNKizR7cZGl3cI
 github.com/OneOfOne/xxhash v1.2.5/go.mod h1:eZbhyaAYD41SGSSsnmcpxVoRiQ/MPUTjUdIIOT9Um7Q=
 github.com/VictoriaMetrics/fastcache v1.5.1 h1:qHgHjyoNFV7jgucU8QZUuU4gcdhfs8QW1kw68OD2Lag=
 github.com/VictoriaMetrics/fastcache v1.5.1/go.mod h1:+jv9Ckb+za/P1ZRg/sulP5Ni1v49daAVERr0H3CuscE=
-github.com/VictoriaMetrics/metrics v1.5.0 h1:WvQqPn+z9pR1U7J58CgaGiWrN8phNGSpr2xUSxJnfpE=
-github.com/VictoriaMetrics/metrics v1.5.0/go.mod h1:QZAL5yLaXvhSPeib0ahluGo9VK0HXDZHovKaKlpuWvs=
+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 v1.6.0 h1:34qKK75C6Dx9zof2JqUiunfJQ87Up6vTHXABWDyCH+g=
+github.com/valyala/gozstd v1.6.0/go.mod h1:y5Ew47GLlP37EkTB+B4s7r6A5rdaeB7ftbl9zoYiIPQ=
 github.com/valyala/histogram v1.0.1 h1:FzA7n2Tz/wKRMejgu3PV1vw3htAklTjjuoI6z3d4KDg=
 github.com/valyala/histogram v1.0.1/go.mod h1:lQy0xA4wUz2+IUnf97SivorsJIp8FxsnRd6x25q7Mto=
 github.com/valyala/quicktemplate v1.1.1 h1:C58y/wN0FMTi2PR0n3onltemfFabany53j7M6SDDB8k=
 github.com/valyala/quicktemplate v1.1.1/go.mod h1:EH+4AkTd43SvgIbQHYu59/cJyxDoOVRUAfrukLPuGJ4=
 github.com/valyala/tcplisten v0.0.0-20161114210144-ceec8f93295a/go.mod h1:v3UYOV9WzVtRmSR+PDvWpU/qWl4Wa5LApYYX4ZtKbio=
 golang.org/x/net v0.0.0-20180911220305-26e67e76b6c3/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
-golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f h1:25KHgbfyiSm6vwQLbM3zZIe1v9p/3ea4Rz+nnM5K/i4=
-golang.org/x/sys v0.0.0-20190616124812-15dcb6c0061f/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa h1:KIDDMLT1O0Nr7TSxp8xM5tJcdn8tgyAONntO829og1M=
+golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

lib/decimal/decimal_test.go

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

lib/encoding/compress.go

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

lib/encoding/encoding.go

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

lib/encoding/encoding_cgo_test.go

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

lib/encoding/encoding_pure_test.go

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

lib/encoding/encoding_test.go

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

lib/encoding/zstd/zstd_cgo.go

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

lib/encoding/zstd/zstd_pure.go

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

lib/encoding/zstd/zstd_test.go

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

lib/fs/fs.go

@@ -5,12 +5,15 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"regexp"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/metrics"
+	"golang.org/x/sys/unix"
 )
 
 // ReadAtCloser is rand-access read interface.
@@ -87,26 +90,42 @@ func MustSyncPath(path string) {
 	}
 }
 
-// WriteFile writes data to the given file path.
+var tmpFileNum uint64
+
+// WriteFileAtomically atomically writes data to the given file path.
 //
-// WriteFile returns only after the file is fully written
+// WriteFile returns only after the file is fully written and synced
 // to the underlying storage.
-func WriteFile(path string, data []byte) error {
+func WriteFileAtomically(path string, data []byte) error {
+	// Check for the existing file. It is expected that
+	// the WriteFileAtomically function cannot be called concurrently
+	// with the same `path`.
 	if IsPathExist(path) {
 		return fmt.Errorf("cannot create file %q, since it already exists", path)
 	}
-	f, err := filestream.Create(path, false)
+
+	n := atomic.AddUint64(&tmpFileNum, 1)
+	tmpPath := fmt.Sprintf("%s.tmp.%d", path, n)
+	f, err := filestream.Create(tmpPath, false)
 	if err != nil {
-		return fmt.Errorf("cannot create file %q: %s", path, err)
+		return fmt.Errorf("cannot create file %q: %s", tmpPath, err)
 	}
 	if _, err := f.Write(data); err != nil {
 		f.MustClose()
-		return fmt.Errorf("cannot write %d bytes to file %q: %s", len(data), path, err)
+		MustRemoveAll(tmpPath)
+		return fmt.Errorf("cannot write %d bytes to file %q: %s", len(data), tmpPath, err)
 	}
 
 	// Sync and close the file.
 	f.MustClose()
 
+	// Atomically move the file from tmpPath to path.
+	if err := os.Rename(tmpPath, path); err != nil {
+		// do not call MustRemoveAll(tmpPath) here, so the user could inspect
+		// the file contents during investigating the issue.
+		return fmt.Errorf("cannot move %q to %q: %s", tmpPath, path, err)
+	}
+
 	// Sync the containing directory, so the file is guaranteed to appear in the directory.
 	// See https://www.quora.com/When-should-you-fsync-the-containing-directory-in-addition-to-the-file-itself
 	absPath, err := filepath.Abs(path)
@@ -119,6 +138,15 @@ func WriteFile(path string, data []byte) error {
 	return nil
 }
 
+// IsTemporaryFileName returns true if fn matches temporary file name pattern
+// from WriteFileAtomically.
+func IsTemporaryFileName(fn string) bool {
+	return tmpFileNameRe.MatchString(fn)
+}
+
+// tmpFileNameRe is regexp for temporary file name - see WriteFileAtomically for details.
+var tmpFileNameRe = regexp.MustCompile(`\.tmp\.\d+$`)
+
 // MkdirAllIfNotExist creates the given path dir if it isn't exist.
 func MkdirAllIfNotExist(path string) error {
 	if IsPathExist(path) {
@@ -185,6 +213,18 @@ 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 {
@@ -346,3 +386,17 @@ func MustWriteData(w io.Writer, data []byte) {
 		logger.Panicf("BUG: writer wrote %d bytes instead of %d bytes", n, len(data))
 	}
 }
+
+// CreateFlockFile creates flock.lock file in the directory dir
+// and returns the handler to the file.
+func CreateFlockFile(dir string) (*os.File, error) {
+	flockFile := dir + "/flock.lock"
+	flockF, err := os.Create(flockFile)
+	if err != nil {
+		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
+	}
+	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
+		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+	}
+	return flockF, nil
+}

lib/fs/fs_test.go

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

lib/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/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/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_header.go

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

lib/mergeset/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

@@ -134,13 +134,9 @@ func OpenTable(path string) (*Table, error) {
 	}
 
 	// Protect from concurrent opens.
-	flockFile := path + "/flock.lock"
-	flockF, err := os.Create(flockFile)
+	flockF, err := fs.CreateFlockFile(path)
 	if err != nil {
-		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
-	}
-	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
-		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+		return nil, err
 	}
 
 	// Open table parts.
@@ -161,8 +157,8 @@ 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
 }
@@ -242,6 +238,7 @@ type TableMetrics struct {
 
 	BlocksCount uint64
 	ItemsCount  uint64
+	SizeBytes   uint64
 
 	DataBlocksCacheSize     uint64
 	DataBlocksCacheRequests uint64
@@ -274,6 +271,7 @@ 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()
@@ -713,7 +711,7 @@ func (tb *Table) mergeParts(pws []*partWrapper, stopCh <-chan struct{}, isOuterP
 	dstPartPath := ph.Path(tb.path, mergeIdx)
 	fmt.Fprintf(&bb, "%s -> %s\n", tmpPartPath, dstPartPath)
 	txnPath := fmt.Sprintf("%s/txn/%016X", tb.path, mergeIdx)
-	if err := fs.WriteFile(txnPath, bb.B); err != nil {
+	if err := fs.WriteFileAtomically(txnPath, bb.B); err != nil {
 		return fmt.Errorf("cannot create transaction file %q: %s", txnPath, err)
 	}
 
@@ -727,6 +725,7 @@ 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 +760,7 @@ 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
@@ -991,7 +990,12 @@ func runTransactions(txnLock *sync.RWMutex, path string) error {
 	})
 
 	for _, fi := range fis {
-		txnPath := txnDir + "/" + fi.Name()
+		fn := fi.Name()
+		if fs.IsTemporaryFileName(fn) {
+			// Skip temporary files, which could be left after unclean shutdown.
+			continue
+		}
+		txnPath := txnDir + "/" + fn
 		if err := runTransaction(txnLock, path, txnPath); err != nil {
 			return fmt.Errorf("cannot run transaction from %q: %s", txnPath, err)
 		}

lib/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/storage/block.go

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

lib/storage/index_db.go

@@ -18,6 +18,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/mergeset"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	xxhash "github.com/cespare/xxhash/v2"
 )
@@ -52,17 +53,17 @@ type indexDB struct {
 	extDBLock sync.Mutex
 
 	// Cache for fast TagFilters -> TSIDs lookup.
-	tagCache *fastcache.Cache
+	tagCache *workingsetcache.Cache
 
 	// Cache for fast MetricID -> TSID lookup.
-	metricIDCache *fastcache.Cache
+	metricIDCache *workingsetcache.Cache
 
 	// Cache for fast MetricID -> MetricName lookup.
-	metricNameCache *fastcache.Cache
+	metricNameCache *workingsetcache.Cache
 
 	// Cache holding useless TagFilters entries, which have no tag filters
 	// matching low number of metrics.
-	uselessTagFiltersCache *fastcache.Cache
+	uselessTagFiltersCache *workingsetcache.Cache
 
 	indexSearchPool sync.Pool
 
@@ -101,7 +102,20 @@ type indexDB struct {
 }
 
 // openIndexDB opens index db from the given path with the given caches.
-func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (*indexDB, error) {
+func openIndexDB(path string, metricIDCache, metricNameCache *workingsetcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (*indexDB, error) {
+	if metricIDCache == nil {
+		logger.Panicf("BUG: metricIDCache must be non-nil")
+	}
+	if metricNameCache == nil {
+		logger.Panicf("BUG: metricNameCache must be non-nil")
+	}
+	if currHourMetricIDs == nil {
+		logger.Panicf("BUG: currHourMetricIDs must be non-nil")
+	}
+	if prevHourMetricIDs == nil {
+		logger.Panicf("BUG: prevHourMetricIDs must be non-nil")
+	}
+
 	tb, err := mergeset.OpenTable(path)
 	if err != nil {
 		return nil, fmt.Errorf("cannot open indexDB %q: %s", path, err)
@@ -111,18 +125,16 @@ func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, c
 
 	// Do not persist tagCache in files, since it is very volatile.
 	mem := memory.Allowed()
-	tagCache := fastcache.New(mem / 32)
 
 	db := &indexDB{
 		refCount: 1,
 		tb:       tb,
 		name:     name,
 
-		tagCache:        tagCache,
-		metricIDCache:   metricIDCache,
-		metricNameCache: metricNameCache,
-
-		uselessTagFiltersCache: fastcache.New(mem / 128),
+		tagCache:               workingsetcache.New(mem/32, time.Hour),
+		metricIDCache:          metricIDCache,
+		metricNameCache:        metricNameCache,
+		uselessTagFiltersCache: workingsetcache.New(mem/128, time.Hour),
 
 		currHourMetricIDs: currHourMetricIDs,
 		prevHourMetricIDs: prevHourMetricIDs,
@@ -142,12 +154,12 @@ func openIndexDB(path string, metricIDCache, metricNameCache *fastcache.Cache, c
 // IndexDBMetrics contains essential metrics for indexDB.
 type IndexDBMetrics struct {
 	TagCacheSize      uint64
-	TagCacheBytesSize uint64
+	TagCacheSizeBytes uint64
 	TagCacheRequests  uint64
 	TagCacheMisses    uint64
 
 	UselessTagFiltersCacheSize      uint64
-	UselessTagFiltersCacheBytesSize uint64
+	UselessTagFiltersCacheSizeBytes uint64
 	UselessTagFiltersCacheRequests  uint64
 	UselessTagFiltersCacheMisses    uint64
 
@@ -176,14 +188,14 @@ func (db *indexDB) UpdateMetrics(m *IndexDBMetrics) {
 	cs.Reset()
 	db.tagCache.UpdateStats(&cs)
 	m.TagCacheSize += cs.EntriesCount
-	m.TagCacheBytesSize += cs.BytesSize
+	m.TagCacheSizeBytes += cs.BytesSize
 	m.TagCacheRequests += cs.GetBigCalls
 	m.TagCacheMisses += cs.Misses
 
 	cs.Reset()
 	db.uselessTagFiltersCache.UpdateStats(&cs)
 	m.UselessTagFiltersCacheSize += cs.EntriesCount
-	m.UselessTagFiltersCacheBytesSize += cs.BytesSize
+	m.UselessTagFiltersCacheSizeBytes += cs.BytesSize
 	m.UselessTagFiltersCacheRequests += cs.GetBigCalls
 	m.UselessTagFiltersCacheMisses += cs.Misses
 
@@ -261,6 +273,15 @@ func (db *indexDB) decRef() {
 	db.tb.MustClose()
 	db.SetExtDB(nil)
 
+	// Free space occupied by caches owned by db.
+	db.tagCache.Stop()
+	db.uselessTagFiltersCache.Stop()
+
+	db.tagCache = nil
+	db.metricIDCache = nil
+	db.metricNameCache = nil
+	db.uselessTagFiltersCache = nil
+
 	if atomic.LoadUint64(&db.mustDrop) == 0 {
 		return
 	}
@@ -271,20 +292,36 @@ func (db *indexDB) decRef() {
 }
 
 func (db *indexDB) getFromTagCache(key []byte) ([]TSID, bool) {
-	value := db.tagCache.GetBig(nil, key)
-	if len(value) == 0 {
+	compressedBuf := tagBufPool.Get()
+	defer tagBufPool.Put(compressedBuf)
+	compressedBuf.B = db.tagCache.GetBig(compressedBuf.B[:0], key)
+	if len(compressedBuf.B) == 0 {
 		return nil, false
 	}
-	tsids, err := unmarshalTSIDs(nil, value)
+	buf := tagBufPool.Get()
+	defer tagBufPool.Put(buf)
+	var err error
+	buf.B, err = encoding.DecompressZSTD(buf.B[:0], compressedBuf.B)
+	if err != nil {
+		logger.Panicf("FATAL: cannot decompress tsids from tagCache: %s", err)
+	}
+	tsids, err := unmarshalTSIDs(nil, buf.B)
 	if err != nil {
 		logger.Panicf("FATAL: cannot unmarshal tsids from tagCache: %s", err)
 	}
 	return tsids, true
 }
 
+var tagBufPool bytesutil.ByteBufferPool
+
 func (db *indexDB) putToTagCache(tsids []TSID, key []byte) {
-	value := marshalTSIDs(nil, tsids)
-	db.tagCache.SetBig(key, value)
+	buf := tagBufPool.Get()
+	buf.B = marshalTSIDs(buf.B[:0], tsids)
+	compressedBuf := tagBufPool.Get()
+	compressedBuf.B = encoding.CompressZSTDLevel(compressedBuf.B[:0], buf.B, 1)
+	tagBufPool.Put(buf)
+	db.tagCache.SetBig(key, compressedBuf.B)
+	tagBufPool.Put(compressedBuf)
 }
 
 func (db *indexDB) getFromMetricIDCache(dst *TSID, metricID uint64) error {
@@ -321,8 +358,11 @@ func (db *indexDB) putMetricNameToCache(metricID uint64, metricName []byte) {
 	db.metricNameCache.Set(key[:], metricName)
 }
 
-func marshalTagFiltersKeyVersioned(dst []byte, tfss []*TagFilters) []byte {
-	prefix := atomic.LoadUint64(&tagFiltersKeyGen)
+func marshalTagFiltersKey(dst []byte, tfss []*TagFilters, versioned bool) []byte {
+	prefix := ^uint64(0)
+	if versioned {
+		prefix = atomic.LoadUint64(&tagFiltersKeyGen)
+	}
 	dst = encoding.MarshalUint64(dst, prefix)
 	for _, tfs := range tfss {
 		dst = append(dst, 0) // separator between tfs groups.
@@ -920,7 +960,7 @@ func (db *indexDB) searchTSIDs(tfss []*TagFilters, tr TimeRange, maxMetrics int)
 	tfKeyBuf := tagFiltersKeyBufPool.Get()
 	defer tagFiltersKeyBufPool.Put(tfKeyBuf)
 
-	tfKeyBuf.B = marshalTagFiltersKeyVersioned(tfKeyBuf.B[:0], tfss)
+	tfKeyBuf.B = marshalTagFiltersKey(tfKeyBuf.B[:0], tfss, true)
 	tsids, ok := db.getFromTagCache(tfKeyBuf.B)
 	if ok {
 		// Fast path - tsids found in the cache.
@@ -937,7 +977,12 @@ func (db *indexDB) searchTSIDs(tfss []*TagFilters, tr TimeRange, maxMetrics int)
 
 	var extTSIDs []TSID
 	if db.doExtDB(func(extDB *indexDB) {
-		tsids, ok := extDB.getFromTagCache(tfKeyBuf.B)
+		tfKeyExtBuf := tagFiltersKeyBufPool.Get()
+		defer tagFiltersKeyBufPool.Put(tfKeyExtBuf)
+
+		// Data in extDB cannot be changed, so use unversioned keys for tag cache.
+		tfKeyExtBuf.B = marshalTagFiltersKey(tfKeyExtBuf.B[:0], tfss, false)
+		tsids, ok := extDB.getFromTagCache(tfKeyExtBuf.B)
 		if ok {
 			extTSIDs = tsids
 			return
@@ -946,8 +991,8 @@ func (db *indexDB) searchTSIDs(tfss []*TagFilters, tr TimeRange, maxMetrics int)
 		extTSIDs, err = is.searchTSIDs(tfss, tr, maxMetrics)
 		extDB.putIndexSearch(is)
 
-		// Do not store found tsids into extDB.tagCache,
-		// since they will be stored into outer cache instead.
+		sort.Slice(extTSIDs, func(i, j int) bool { return extTSIDs[i].Less(&extTSIDs[j]) })
+		extDB.putToTagCache(extTSIDs, tfKeyExtBuf.B)
 	}) {
 		if err != nil {
 			return nil, err
@@ -1057,6 +1102,15 @@ func mergeTSIDs(a, b []TSID) []TSID {
 }
 
 func (is *indexSearch) searchTSIDs(tfss []*TagFilters, tr TimeRange, maxMetrics int) ([]TSID, error) {
+	// Verify whether `is` contains data for the given tr.
+	ok, err := is.containsTimeRange(tr)
+	if err != nil {
+		return nil, fmt.Errorf("error in containsTimeRange(%s): %s", &tr, err)
+	}
+	if !ok {
+		// Fast path: nothing to search.
+		return nil, nil
+	}
 	metricIDs, err := is.searchMetricIDs(tfss, tr, maxMetrics)
 	if err != nil {
 		return nil, err
@@ -1184,8 +1238,9 @@ func (is *indexSearch) updateMetricIDsByMetricNameMatch(metricIDs, srcMetricIDs
 
 func (is *indexSearch) getTagFilterWithMinMetricIDsCountAdaptive(tfs *TagFilters, maxMetrics int) (*tagFilter, map[uint64]struct{}, error) {
 	kb := &is.kb
-	kb.B = tfs.marshal(kb.B[:0])
+	kb.B = append(kb.B[:0], uselessMultiTagFiltersKeyPrefix)
 	kb.B = encoding.MarshalUint64(kb.B, uint64(maxMetrics))
+	kb.B = tfs.marshal(kb.B)
 	if len(is.db.uselessTagFiltersCache.Get(nil, kb.B)) > 0 {
 		// Skip useless work below, since the tfs doesn't contain tag filters matching less than maxMetrics metrics.
 		return nil, nil, errTooManyMetrics
@@ -1199,20 +1254,23 @@ func (is *indexSearch) getTagFilterWithMinMetricIDsCountAdaptive(tfs *TagFilters
 	}
 	for {
 		minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCount(tfs, maxAllowedMetrics)
-		if err != nil {
-			return nil, nil, err
-		}
-		if len(minMetricIDs) < maxAllowedMetrics {
-			// Found the tag filter with the minimum number of metrics.
-			return minTf, minMetricIDs, nil
+		if err != errTooManyMetrics {
+			if err != nil {
+				return nil, nil, err
+			}
+			if len(minMetricIDs) < maxAllowedMetrics {
+				// Found the tag filter with the minimum number of metrics.
+				return minTf, minMetricIDs, nil
+			}
 		}
 
 		// Too many metrics matched.
 		if maxAllowedMetrics >= maxMetrics {
 			// The tag filter with minimum matching metrics matches at least maxMetrics metrics.
-			kb.B = tfs.marshal(kb.B[:0])
+			kb.B = append(kb.B[:0], uselessMultiTagFiltersKeyPrefix)
 			kb.B = encoding.MarshalUint64(kb.B, uint64(maxMetrics))
-			is.db.uselessTagFiltersCache.Set(kb.B, []byte("1"))
+			kb.B = tfs.marshal(kb.B)
+			is.db.uselessTagFiltersCache.Set(kb.B, uselessTagFilterCacheValue)
 			return nil, nil, errTooManyMetrics
 		}
 
@@ -1226,15 +1284,50 @@ func (is *indexSearch) getTagFilterWithMinMetricIDsCountAdaptive(tfs *TagFilters
 
 var errTooManyMetrics = errors.New("all the tag filters match too many metrics")
 
+const maxDaysForDateMetricIDs = 40
+
+func (is *indexSearch) adjustMaxMetricsAdaptive(tr TimeRange, maxMetrics int) int {
+	minDate := uint64(tr.MinTimestamp) / msecPerDay
+	maxDate := uint64(tr.MaxTimestamp) / msecPerDay
+	if maxDate-minDate > maxDaysForDateMetricIDs {
+		// Cannot reduce maxMetrics for the given time range,
+		// since the it is expensive extracting metricIDs for the given tr.
+		return maxMetrics
+	}
+	hmPrev := is.db.prevHourMetricIDs.Load().(*hourMetricIDs)
+	if !hmPrev.isFull {
+		return maxMetrics
+	}
+	hourMetrics := len(hmPrev.m)
+	if hourMetrics >= 256 && maxMetrics > hourMetrics/4 {
+		// It is cheaper to filter on the hour or day metrics if the minimum
+		// number of matching metrics across tfs exceeds hourMetrics / 4.
+		return hourMetrics / 4
+	}
+	return maxMetrics
+}
+
 func (is *indexSearch) getTagFilterWithMinMetricIDsCount(tfs *TagFilters, maxMetrics int) (*tagFilter, map[uint64]struct{}, error) {
 	var minMetricIDs map[uint64]struct{}
 	var minTf *tagFilter
+	kb := &is.kb
+	uselessTagFilters := 0
 	for i := range tfs.tfs {
 		tf := &tfs.tfs[i]
 		if tf.isNegative {
 			// Skip negative filters.
 			continue
 		}
+
+		kb.B = append(kb.B[:0], uselessSingleTagFilterKeyPrefix)
+		kb.B = encoding.MarshalUint64(kb.B[:0], uint64(maxMetrics))
+		kb.B = tf.Marshal(kb.B)
+		if len(is.db.uselessTagFiltersCache.Get(nil, kb.B)) > 0 {
+			// Skip useless work below, since the tf matches at least maxMetrics metrics.
+			uselessTagFilters++
+			continue
+		}
+
 		metricIDs, err := is.getMetricIDsForTagFilter(tf, maxMetrics)
 		if err != nil {
 			if err == errFallbackToMetricNameMatch {
@@ -1243,20 +1336,32 @@ func (is *indexSearch) getTagFilterWithMinMetricIDsCount(tfs *TagFilters, maxMet
 			}
 			return nil, nil, fmt.Errorf("cannot find MetricIDs for tagFilter %s: %s", tf, err)
 		}
-		if minTf == nil || len(metricIDs) < len(minMetricIDs) {
-			minMetricIDs = metricIDs
-			minTf = tf
-			maxMetrics = len(minMetricIDs)
-			if maxMetrics <= 1 {
-				// There is no need in inspecting other filters, since minTf
-				// already matches 0 or 1 metric.
-				break
-			}
+		if len(metricIDs) >= maxMetrics {
+			// The tf matches at least maxMetrics. Skip it
+			kb.B = append(kb.B[:0], uselessSingleTagFilterKeyPrefix)
+			kb.B = encoding.MarshalUint64(kb.B[:0], uint64(maxMetrics))
+			kb.B = tf.Marshal(kb.B)
+			is.db.uselessTagFiltersCache.Set(kb.B, uselessTagFilterCacheValue)
+			uselessTagFilters++
+			continue
+		}
+
+		minMetricIDs = metricIDs
+		minTf = tf
+		maxMetrics = len(minMetricIDs)
+		if maxMetrics <= 1 {
+			// There is no need in inspecting other filters, since minTf
+			// already matches 0 or 1 metric.
+			break
 		}
 	}
 	if minTf != nil {
 		return minTf, minMetricIDs, nil
 	}
+	if uselessTagFilters == len(tfs.tfs) {
+		// All the tag filters return at least maxMetrics entries.
+		return nil, nil, errTooManyMetrics
+	}
 
 	// There is no positive filter with small number of matching metrics.
 	// Create it, so it matches all the MetricIDs for tfs.commonPrefix.
@@ -1294,7 +1399,7 @@ func matchTagFilters(mn *MetricName, tfs []*tagFilter, kb *bytesutil.ByteBuffer)
 				continue
 			}
 
-			// Found the matching tag name. Match for the value.
+			// Found the matching tag name. Match the value.
 			b := tag.Marshal(kb.B)
 			kb.B = b[:len(kb.B)]
 			ok, err := matchTagFilter(b, tf)
@@ -1307,7 +1412,7 @@ func matchTagFilters(mn *MetricName, tfs []*tagFilter, kb *bytesutil.ByteBuffer)
 			tagMatched = true
 			break
 		}
-		if !tagMatched {
+		if !tagMatched && !tf.isNegative {
 			// Matching tag name wasn't found.
 			return false, nil
 		}
@@ -1376,7 +1481,8 @@ func (is *indexSearch) updateMetricIDsForTagFilters(metricIDs map[uint64]struct{
 	// Sort tag filters for faster ts.Seek below.
 	sort.Slice(tfs.tfs, func(i, j int) bool { return bytes.Compare(tfs.tfs[i].prefix, tfs.tfs[j].prefix) < 0 })
 
-	minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetrics)
+	maxMetricsAdjusted := is.adjustMaxMetricsAdaptive(tr, maxMetrics)
+	minTf, minMetricIDs, err := is.getTagFilterWithMinMetricIDsCountAdaptive(tfs, maxMetricsAdjusted)
 	if err != nil {
 		if err != errTooManyMetrics {
 			return err
@@ -1394,7 +1500,7 @@ func (is *indexSearch) updateMetricIDsForTagFilters(metricIDs map[uint64]struct{
 		maxTimeRangeMetrics := 20 * maxMetrics
 		metricIDsForTimeRange, err := is.getMetricIDsForTimeRange(tr, maxTimeRangeMetrics+1)
 		if err == errMissingMetricIDsForDate {
-			return fmt.Errorf("cannot find tag filter matching less up to %d time series; either increase -search.maxUniqueTimeseries or use more specific tag filters",
+			return fmt.Errorf("cannot find tag filter matching less than %d time series; either increase -search.maxUniqueTimeseries or use more specific tag filters",
 				maxMetrics)
 		}
 		if err != nil {
@@ -1443,6 +1549,13 @@ func (is *indexSearch) updateMetricIDsForTagFilters(metricIDs map[uint64]struct{
 	return nil
 }
 
+const (
+	uselessSingleTagFilterKeyPrefix = 0
+	uselessMultiTagFiltersKeyPrefix = 1
+)
+
+var uselessTagFilterCacheValue = []byte("1")
+
 func (is *indexSearch) getMetricIDsForTagFilter(tf *tagFilter, maxMetrics int) (map[uint64]struct{}, error) {
 	if tf.isNegative {
 		logger.Panicf("BUG: isNegative must be false")
@@ -1456,17 +1569,35 @@ func (is *indexSearch) getMetricIDsForTagFilter(tf *tagFilter, maxMetrics int) (
 		return metricIDs, nil
 	}
 
-	// Slow path - scan all the rows with tf.prefix
+	// Slow path - scan for all the rows with the given prefix.
 	maxLoops := maxMetrics * maxIndexScanLoopsPerMetric
+	err := is.getMetricIDsForTagFilterSlow(tf, maxLoops, func(metricID uint64) bool {
+		metricIDs[metricID] = struct{}{}
+		return len(metricIDs) < maxMetrics
+	})
+	if err != nil {
+		return nil, err
+	}
+	return metricIDs, nil
+}
+
+func (is *indexSearch) getMetricIDsForTagFilterSlow(tf *tagFilter, maxLoops int, f func(metricID uint64) bool) error {
+	if len(tf.orSuffixes) > 0 {
+		logger.Panicf("BUG: the getMetricIDsForTagFilterSlow must be called only for empty tf.orSuffixes; got %s", tf.orSuffixes)
+	}
+
+	// Scan all the rows with tf.prefix and call f on every tf match.
 	loops := 0
 	ts := &is.ts
+	kb := &is.kb
+	var prevMatchingK []byte
+	var prevMatch bool
 	ts.Seek(tf.prefix)
-	for len(metricIDs) < maxMetrics && ts.NextItem() {
+	for ts.NextItem() {
 		loops++
 		if loops > maxLoops {
-			return nil, errFallbackToMetricNameMatch
+			return errFallbackToMetricNameMatch
 		}
-
 		k := ts.Item
 		if !bytes.HasPrefix(k, tf.prefix) {
 			break
@@ -1475,25 +1606,49 @@ func (is *indexSearch) getMetricIDsForTagFilter(tf *tagFilter, maxMetrics int) (
 		// Get MetricID from k (the last 8 bytes).
 		k = k[len(tf.prefix):]
 		if len(k) < 8 {
-			return nil, fmt.Errorf("invald key suffix size; want at least %d bytes; got %d bytes", 8, len(k))
+			return fmt.Errorf("invald key suffix size; want at least %d bytes; got %d bytes", 8, len(k))
 		}
 		v := k[len(k)-8:]
 		k = k[:len(k)-8]
+		metricID := encoding.UnmarshalUint64(v)
+
+		if prevMatch && string(k) == string(prevMatchingK) {
+			// Fast path: the same tag value found.
+			// There is no need in checking it again with potentially
+			// slow tf.matchSuffix, which may call regexp.
+			if !f(metricID) {
+				break
+			}
+			continue
+		}
 
 		ok, err := tf.matchSuffix(k)
 		if err != nil {
-			return nil, fmt.Errorf("error when matching %s: %s", tf, err)
+			return fmt.Errorf("error when matching %s: %s", tf, err)
 		}
 		if !ok {
+			prevMatch = false
+			// Optimization: skip all the metricIDs for the given tag value
+			kb.B = append(kb.B[:0], ts.Item[:len(ts.Item)-8]...)
+			// The last char in kb.B must be tagSeparatorChar. Just increment it
+			// in order to jump to the next tag value.
+			if len(kb.B) == 0 || kb.B[len(kb.B)-1] != tagSeparatorChar || tagSeparatorChar >= 0xff {
+				return fmt.Errorf("data corruption: the last char in k=%X must be %X", kb.B, tagSeparatorChar)
+			}
+			kb.B[len(kb.B)-1]++
+			ts.Seek(kb.B)
 			continue
 		}
-		metricID := encoding.UnmarshalUint64(v)
-		metricIDs[metricID] = struct{}{}
+		prevMatch = true
+		prevMatchingK = append(prevMatchingK[:0], k...)
+		if !f(metricID) {
+			break
+		}
 	}
 	if err := ts.Error(); err != nil {
-		return nil, fmt.Errorf("error when searching for tag filter prefix %q: %s", tf.prefix, err)
+		return fmt.Errorf("error when searching for tag filter prefix %q: %s", tf.prefix, err)
 	}
-	return metricIDs, nil
+	return nil
 }
 
 func (is *indexSearch) updateMetricIDsForOrSuffixesNoFilter(tf *tagFilter, maxMetrics int, metricIDs map[uint64]struct{}) error {
@@ -1585,6 +1740,15 @@ func (is *indexSearch) updateMetricIDsForOrSuffixWithFilter(prefix []byte, metri
 		}
 		metricID := encoding.UnmarshalUint64(v)
 		if metricID != nextMetricID {
+			// Skip metricIDs smaller than the found metricID, since they don't
+			// match anything.
+			if len(sortedFilter) > 0 && metricID > sortedFilter[0] {
+				sortedFilter = sortedFilter[1:]
+				n := sort.Search(len(sortedFilter), func(i int) bool {
+					return metricID <= sortedFilter[i]
+				})
+				sortedFilter = sortedFilter[n:]
+			}
 			continue
 		}
 		if isNegative {
@@ -1617,15 +1781,15 @@ func (is *indexSearch) getMetricIDsForTimeRange(tr TimeRange, maxMetrics int) (m
 
 	// Slow path: collect the metric ids for all the days covering the given tr.
 	atomic.AddUint64(&is.db.dateMetricIDsSearchCalls, 1)
-	minDate := tr.MinTimestamp / msecPerDay
-	maxDate := tr.MaxTimestamp / msecPerDay
-	if maxDate-minDate > 40 {
+	minDate := uint64(tr.MinTimestamp) / msecPerDay
+	maxDate := uint64(tr.MaxTimestamp) / msecPerDay
+	if maxDate-minDate > maxDaysForDateMetricIDs {
 		// Too much dates must be covered. Give up.
 		return nil, errMissingMetricIDsForDate
 	}
 	metricIDs := make(map[uint64]struct{}, maxMetrics)
 	for minDate <= maxDate {
-		if err := is.getMetricIDsForDate(uint64(minDate), metricIDs, maxMetrics); err != nil {
+		if err := is.getMetricIDsForDate(minDate, metricIDs, maxMetrics); err != nil {
 			return nil, err
 		}
 		minDate++
@@ -1637,9 +1801,6 @@ func (is *indexSearch) getMetricIDsForTimeRange(tr TimeRange, maxMetrics int) (m
 func (is *indexSearch) getMetricIDsForRecentHours(tr TimeRange, maxMetrics int) (map[uint64]struct{}, bool) {
 	minHour := uint64(tr.MinTimestamp) / msecPerHour
 	maxHour := uint64(tr.MaxTimestamp) / msecPerHour
-	if is.db.currHourMetricIDs == nil {
-		return nil, false
-	}
 	hmCurr := is.db.currHourMetricIDs.Load().(*hourMetricIDs)
 	if maxHour == hmCurr.hour && minHour == maxHour && hmCurr.isFull {
 		// The tr fits the current hour.
@@ -1650,9 +1811,6 @@ func (is *indexSearch) getMetricIDsForRecentHours(tr TimeRange, maxMetrics int)
 		}
 		return getMetricIDsCopy(hmCurr.m), true
 	}
-	if is.db.prevHourMetricIDs == nil {
-		return nil, false
-	}
 	hmPrev := is.db.prevHourMetricIDs.Load().(*hourMetricIDs)
 	if maxHour == hmPrev.hour && minHour == maxHour && hmPrev.isFull {
 		// The tr fits the previous hour.
@@ -1760,6 +1918,28 @@ func (is *indexSearch) getMetricIDsForDate(date uint64, metricIDs map[uint64]str
 	return nil
 }
 
+func (is *indexSearch) containsTimeRange(tr TimeRange) (bool, error) {
+	ts := &is.ts
+	kb := &is.kb
+
+	// Verify whether the maximum date in `ts` covers tr.MinTimestamp.
+	minDate := uint64(tr.MinTimestamp) / msecPerDay
+	kb.B = marshalCommonPrefix(kb.B[:0], nsPrefixDateToMetricID)
+	kb.B = encoding.MarshalUint64(kb.B, minDate)
+	ts.Seek(kb.B)
+	if !ts.NextItem() {
+		if err := ts.Error(); err != nil {
+			return false, fmt.Errorf("error when searching for minDate=%d, prefix %q: %s", minDate, kb.B, err)
+		}
+		return false, nil
+	}
+	if !bytes.HasPrefix(ts.Item, kb.B[:1]) {
+		// minDate exceeds max date from ts.
+		return false, nil
+	}
+	return true, nil
+}
+
 func (is *indexSearch) updateMetricIDsForCommonPrefix(metricIDs map[uint64]struct{}, commonPrefix []byte, maxMetrics int) error {
 	ts := &is.ts
 	ts.Seek(commonPrefix)
@@ -1807,47 +1987,19 @@ func (is *indexSearch) intersectMetricIDsWithTagFilter(tf *tagFilter, filter map
 
 	// Slow path - scan for all the rows with the given prefix.
 	maxLoops := len(filter) * maxIndexScanLoopsPerMetric
-	loops := 0
-	ts := &is.ts
-	ts.Seek(tf.prefix)
-	for ts.NextItem() {
-		loops++
-		if loops > maxLoops {
-			return nil, errFallbackToMetricNameMatch
-		}
-
-		k := ts.Item
-		if !bytes.HasPrefix(k, tf.prefix) {
-			break
-		}
-
-		// Extract MetricID from k (the last 8 bytes).
-		k = k[len(tf.prefix):]
-		if len(k) < 8 {
-			return nil, fmt.Errorf("cannot extract metricID from k; want at least %d bytes; got %d bytes", 8, len(k))
-		}
-		v := k[len(k)-8:]
-		k = k[:len(k)-8]
-
-		ok, err := tf.matchSuffix(k)
-		if err != nil {
-			return nil, fmt.Errorf("error when matching %s: %s", tf, err)
-		}
-		if !ok {
-			continue
-		}
-		metricID := encoding.UnmarshalUint64(v)
+	err := is.getMetricIDsForTagFilterSlow(tf, maxLoops, func(metricID uint64) bool {
 		if tf.isNegative {
 			// filter must be equal to metricIDs
 			delete(metricIDs, metricID)
-			continue
+			return true
 		}
 		if _, ok := filter[metricID]; ok {
 			metricIDs[metricID] = struct{}{}
 		}
-	}
-	if err := ts.Error(); err != nil {
-		return nil, fmt.Errorf("error searching %q: %s", tf.prefix, err)
+		return true
+	})
+	if err != nil {
+		return nil, err
 	}
 	return metricIDs, nil
 }

lib/storage/index_db_test.go

@@ -7,11 +7,12 @@ import (
 	"math/rand"
 	"os"
 	"reflect"
+	"sync/atomic"
 	"testing"
 	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
-	"github.com/VictoriaMetrics/fastcache"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 )
 
 func TestMarshalUnmarshalTSIDs(t *testing.T) {
@@ -56,12 +57,18 @@ func TestMarshalUnmarshalTSIDs(t *testing.T) {
 }
 
 func TestIndexDBOpenClose(t *testing.T) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
+
+	var hmCurr atomic.Value
+	hmCurr.Store(&hourMetricIDs{})
+	var hmPrev atomic.Value
+	hmPrev.Store(&hourMetricIDs{})
+
 	for i := 0; i < 5; i++ {
-		db, err := openIndexDB("test-index-db", metricIDCache, metricNameCache, nil, nil)
+		db, err := openIndexDB("test-index-db", metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 		if err != nil {
 			t.Fatalf("cannot open indexDB: %s", err)
 		}
@@ -78,12 +85,18 @@ func TestIndexDB(t *testing.T) {
 	const metricGroups = 10
 
 	t.Run("serial", func(t *testing.T) {
-		metricIDCache := fastcache.New(1234)
-		metricNameCache := fastcache.New(1234)
-		defer metricIDCache.Reset()
-		defer metricNameCache.Reset()
+		metricIDCache := workingsetcache.New(1234, time.Hour)
+		metricNameCache := workingsetcache.New(1234, time.Hour)
+		defer metricIDCache.Stop()
+		defer metricNameCache.Stop()
+
+		var hmCurr atomic.Value
+		hmCurr.Store(&hourMetricIDs{})
+		var hmPrev atomic.Value
+		hmPrev.Store(&hourMetricIDs{})
+
 		dbName := "test-index-db-serial"
-		db, err := openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
+		db, err := openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 		if err != nil {
 			t.Fatalf("cannot open indexDB: %s", err)
 		}
@@ -113,7 +126,7 @@ func TestIndexDB(t *testing.T) {
 
 		// Re-open the db and verify it works as expected.
 		db.MustClose()
-		db, err = openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
+		db, err = openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 		if err != nil {
 			t.Fatalf("cannot open indexDB: %s", err)
 		}
@@ -129,12 +142,18 @@ func TestIndexDB(t *testing.T) {
 	})
 
 	t.Run("concurrent", func(t *testing.T) {
-		metricIDCache := fastcache.New(1234)
-		metricNameCache := fastcache.New(1234)
-		defer metricIDCache.Reset()
-		defer metricNameCache.Reset()
+		metricIDCache := workingsetcache.New(1234, time.Hour)
+		metricNameCache := workingsetcache.New(1234, time.Hour)
+		defer metricIDCache.Stop()
+		defer metricNameCache.Stop()
+
+		var hmCurr atomic.Value
+		hmCurr.Store(&hourMetricIDs{})
+		var hmPrev atomic.Value
+		hmPrev.Store(&hourMetricIDs{})
+
 		dbName := "test-index-db-concurrent"
-		db, err := openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
+		db, err := openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 		if err != nil {
 			t.Fatalf("cannot open indexDB: %s", err)
 		}
@@ -293,6 +312,16 @@ func testIndexDBGetOrCreateTSIDByName(db *indexDB, accountsCount, projectsCount,
 }
 
 func testIndexDBCheckTSIDByName(db *indexDB, mns []MetricName, tsids []TSID, isConcurrent bool) error {
+	// fill Date -> MetricID cache
+	date := uint64(timestampFromTime(time.Now())) / msecPerDay
+	for i := range tsids {
+		tsid := &tsids[i]
+		if err := db.storeDateMetricID(date, tsid.MetricID); err != nil {
+			return fmt.Errorf("error in storeDateMetricID(%d, %d): %s", date, tsid.MetricID, err)
+		}
+	}
+	db.tb.DebugFlush()
+
 	hasValue := func(tvs []string, v []byte) bool {
 		for _, tv := range tvs {
 			if string(v) == tv {
@@ -724,8 +753,8 @@ func TestMatchTagFilters(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %s", err)
 	}
-	if ok {
-		t.Fatalf("Shouldn't match")
+	if !ok {
+		t.Fatalf("Should match")
 	}
 	tfs.Reset()
 	if err := tfs.Add([]byte("non-existing-tag"), []byte("foob.+metric"), true, true); err != nil {
@@ -735,8 +764,19 @@ func TestMatchTagFilters(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %s", err)
 	}
-	if ok {
-		t.Fatalf("Shouldn't match")
+	if !ok {
+		t.Fatalf("Should match")
+	}
+	tfs.Reset()
+	if err := tfs.Add([]byte("non-existing-tag"), []byte(".+"), true, true); err != nil {
+		t.Fatalf("cannot add regexp, negative filter: %s", err)
+	}
+	ok, err = matchTagFilters(&mn, toTFPointers(tfs.tfs), &bb)
+	if err != nil {
+		t.Fatalf("unexpected error: %s", err)
+	}
+	if !ok {
+		t.Fatalf("Should match")
 	}
 
 	// Negative match by existing tag
@@ -830,6 +870,17 @@ func TestMatchTagFilters(t *testing.T) {
 	if !ok {
 		t.Fatalf("Should match")
 	}
+	tfs.Reset()
+	if err := tfs.Add([]byte("key 3"), []byte(""), true, false); err != nil {
+		t.Fatalf("cannot add regexp, negative filter: %s", err)
+	}
+	ok, err = matchTagFilters(&mn, toTFPointers(tfs.tfs), &bb)
+	if err != nil {
+		t.Fatalf("unexpected error: %s", err)
+	}
+	if !ok {
+		t.Fatalf("Should match")
+	}
 
 	// Positive match by multiple tags and MetricGroup
 	tfs.Reset()

lib/storage/index_db_timing_test.go

@@ -4,20 +4,28 @@ import (
 	"fmt"
 	"os"
 	"strconv"
+	"sync/atomic"
 	"testing"
+	"time"
 
-	"github.com/VictoriaMetrics/fastcache"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 )
 
 func BenchmarkIndexDBAddTSIDs(b *testing.B) {
 	const recordsPerLoop = 1e3
 
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
+
+	var hmCurr atomic.Value
+	hmCurr.Store(&hourMetricIDs{})
+	var hmPrev atomic.Value
+	hmPrev.Store(&hourMetricIDs{})
+
 	const dbName = "bench-index-db-add-tsids"
-	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
+	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 	if err != nil {
 		b.Fatalf("cannot open indexDB: %s", err)
 	}
@@ -71,82 +79,19 @@ func benchmarkIndexDBAddTSIDs(db *indexDB, tsid *TSID, mn *MetricName, startOffs
 	}
 }
 
-func BenchmarkIndexDBSearchTSIDs(b *testing.B) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
-	const dbName = "bench-index-db-search-tsids"
-	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
-	if err != nil {
-		b.Fatalf("cannot open indexDB: %s", err)
-	}
-	defer func() {
-		db.MustClose()
-		if err := os.RemoveAll(dbName); err != nil {
-			b.Fatalf("cannot remove indexDB: %s", err)
-		}
-	}()
-
-	const recordsCount = 1e5
-
-	// Fill the db with recordsCount records.
-	var mn MetricName
-	mn.MetricGroup = []byte("rps")
-	for i := 0; i < 2; i++ {
-		key := fmt.Sprintf("key_%d", i)
-		value := fmt.Sprintf("value_%d", i)
-		mn.AddTag(key, value)
-	}
-	var tsid TSID
-	var metricName []byte
-	is := db.getIndexSearch()
-	defer db.putIndexSearch(is)
-	for i := 0; i < recordsCount; i++ {
-		mn.sortTags()
-		metricName = mn.Marshal(metricName[:0])
-		if err := is.GetOrCreateTSIDByName(&tsid, metricName); err != nil {
-			b.Fatalf("cannot insert record: %s", err)
-		}
-	}
-
-	b.SetBytes(1)
-	b.ReportAllocs()
-	b.ResetTimer()
-	b.RunParallel(func(pb *testing.PB) {
-		tags := []Tag{
-			{[]byte("key_0"), []byte("value_0")},
-			{[]byte("key_1"), []byte("value_1")},
-		}
-		var tfs TagFilters
-		tfss := []*TagFilters{&tfs}
-		i := 0
-		for pb.Next() {
-			tfs.Reset()
-			for j := range tags {
-				if err := tfs.Add(tags[j].Key, tags[j].Value, false, false); err != nil {
-					panic(fmt.Errorf("BUG: unexpected error: %s", err))
-				}
-			}
-			tsids, err := db.searchTSIDs(tfss, TimeRange{}, 1e5)
-			if err != nil {
-				panic(fmt.Errorf("unexpected error in search for tfs=%s: %s", &tfs, err))
-			}
-			if len(tsids) == 0 && i < recordsCount {
-				panic(fmt.Errorf("zero tsids found for tfs=%s", &tfs))
-			}
-			i++
-		}
-	})
-}
-
 func BenchmarkIndexDBGetTSIDs(b *testing.B) {
-	metricIDCache := fastcache.New(1234)
-	metricNameCache := fastcache.New(1234)
-	defer metricIDCache.Reset()
-	defer metricNameCache.Reset()
+	metricIDCache := workingsetcache.New(1234, time.Hour)
+	metricNameCache := workingsetcache.New(1234, time.Hour)
+	defer metricIDCache.Stop()
+	defer metricNameCache.Stop()
+
+	var hmCurr atomic.Value
+	hmCurr.Store(&hourMetricIDs{})
+	var hmPrev atomic.Value
+	hmPrev.Store(&hourMetricIDs{})
+
 	const dbName = "bench-index-db-get-tsids"
-	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, nil, nil)
+	db, err := openIndexDB(dbName, metricIDCache, metricNameCache, &hmCurr, &hmPrev)
 	if err != nil {
 		b.Fatalf("cannot open indexDB: %s", err)
 	}

lib/storage/inmemory_part.go

@@ -51,7 +51,8 @@ func (mp *inmemoryPart) InitFromRows(rows []rawRow) {
 // It is unsafe re-using mp while the returned part is in use.
 func (mp *inmemoryPart) NewPart() (*part, error) {
 	ph := mp.ph
-	return newPart(&ph, "", mp.metaindexData.NewReader(), &mp.timestampsData, &mp.valuesData, &mp.indexData)
+	size := uint64(len(mp.timestampsData.B) + len(mp.valuesData.B) + len(mp.indexData.B) + len(mp.metaindexData.B))
+	return newPart(&ph, "", size, mp.metaindexData.NewReader(), &mp.timestampsData, &mp.valuesData, &mp.indexData)
 }
 
 func getInmemoryPart() *inmemoryPart {

lib/storage/merge.go

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

lib/storage/metric_name.go

@@ -50,6 +50,12 @@ func (tag *Tag) copyFrom(src *Tag) {
 	tag.Value = append(tag.Value[:0], src.Value...)
 }
 
+func marshalTagValueNoTrailingTagSeparator(dst, src []byte) []byte {
+	dst = marshalTagValue(dst, src)
+	// Remove trailing tagSeparatorChar
+	return dst[:len(dst)-1]
+}
+
 func marshalTagValue(dst, src []byte) []byte {
 	n1 := bytes.IndexByte(src, escapeChar)
 	n2 := bytes.IndexByte(src, tagSeparatorChar)

lib/storage/part.go

@@ -36,6 +36,9 @@ type part struct {
 	// Empty for in-memory part.
 	path string
 
+	// Total size in bytes of part data.
+	size uint64
+
 	timestampsFile fs.ReadAtCloser
 	valuesFile     fs.ReadAtCloser
 	indexFile      fs.ReadAtCloser
@@ -59,6 +62,7 @@ func openFilePart(path string) (*part, error) {
 	if err != nil {
 		return nil, fmt.Errorf("cannot open timestamps file: %s", err)
 	}
+	timestampsSize := fs.MustFileSize(timestampsPath)
 
 	valuesPath := path + "/values.bin"
 	valuesFile, err := fs.OpenReaderAt(valuesPath)
@@ -66,6 +70,7 @@ func openFilePart(path string) (*part, error) {
 		timestampsFile.MustClose()
 		return nil, fmt.Errorf("cannot open values file: %s", err)
 	}
+	valuesSize := fs.MustFileSize(valuesPath)
 
 	indexPath := path + "/index.bin"
 	indexFile, err := fs.OpenReaderAt(indexPath)
@@ -74,6 +79,7 @@ func openFilePart(path string) (*part, error) {
 		valuesFile.MustClose()
 		return nil, fmt.Errorf("cannot open index file: %s", err)
 	}
+	indexSize := fs.MustFileSize(indexPath)
 
 	metaindexPath := path + "/metaindex.bin"
 	metaindexFile, err := filestream.Open(metaindexPath, true)
@@ -83,15 +89,17 @@ func openFilePart(path string) (*part, error) {
 		indexFile.MustClose()
 		return nil, fmt.Errorf("cannot open metaindex file: %s", err)
 	}
+	metaindexSize := fs.MustFileSize(metaindexPath)
 
-	return newPart(&ph, path, metaindexFile, timestampsFile, valuesFile, indexFile)
+	size := timestampsSize + valuesSize + indexSize + metaindexSize
+	return newPart(&ph, path, size, metaindexFile, timestampsFile, valuesFile, indexFile)
 }
 
 // newPart returns new part initialized with the given arguments.
 //
 // The returned part calls MustClose on all the files passed to newPart
 // when calling part.MustClose.
-func newPart(ph *partHeader, path string, metaindexReader filestream.ReadCloser, timestampsFile, valuesFile, indexFile fs.ReadAtCloser) (*part, error) {
+func newPart(ph *partHeader, path string, size uint64, metaindexReader filestream.ReadCloser, timestampsFile, valuesFile, indexFile fs.ReadAtCloser) (*part, error) {
 	var errors []error
 	metaindex, err := unmarshalMetaindexRows(nil, metaindexReader)
 	if err != nil {
@@ -102,6 +110,7 @@ func newPart(ph *partHeader, path string, metaindexReader filestream.ReadCloser,
 	p := &part{
 		ph:             *ph,
 		path:           path,
+		size:           size,
 		timestampsFile: timestampsFile,
 		valuesFile:     valuesFile,
 		indexFile:      indexFile,

lib/storage/part_search.go

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

lib/storage/part_search_test.go

@@ -1247,7 +1247,7 @@ func testPartSearch(t *testing.T, p *part, tsids []TSID, tr TimeRange, expectedR
 
 func testPartSearchSerial(p *part, tsids []TSID, tr TimeRange, expectedRawBlocks []rawBlock) error {
 	var ps partSearch
-	ps.Init(p, tsids, tr)
+	ps.Init(p, tsids, tr, true)
 	var bs []Block
 	for ps.NextBlock() {
 		var b Block

lib/storage/partition.go

@@ -282,6 +282,9 @@ type partitionMetrics struct {
 	SmallIndexBlocksCacheRequests uint64
 	SmallIndexBlocksCacheMisses   uint64
 
+	BigSizeBytes   uint64
+	SmallSizeBytes uint64
+
 	BigRowsCount   uint64
 	SmallRowsCount uint64
 
@@ -326,6 +329,7 @@ func (pt *partition) UpdateMetrics(m *partitionMetrics) {
 		m.BigIndexBlocksCacheMisses += p.ibCache.Misses()
 		m.BigRowsCount += p.ph.RowsCount
 		m.BigBlocksCount += p.ph.BlocksCount
+		m.BigSizeBytes += p.size
 		m.BigPartsRefCount += atomic.LoadUint64(&pw.refCount)
 	}
 
@@ -337,6 +341,7 @@ func (pt *partition) UpdateMetrics(m *partitionMetrics) {
 		m.SmallIndexBlocksCacheMisses += p.ibCache.Misses()
 		m.SmallRowsCount += p.ph.RowsCount
 		m.SmallBlocksCount += p.ph.BlocksCount
+		m.SmallSizeBytes += p.size
 		m.SmallPartsRefCount += atomic.LoadUint64(&pw.refCount)
 	}
 
@@ -1003,7 +1008,7 @@ func (pt *partition) mergeParts(pws []*partWrapper, stopCh <-chan struct{}) erro
 	}
 	fmt.Fprintf(&bb, "%s -> %s\n", tmpPartPath, dstPartPath)
 	txnPath := fmt.Sprintf("%s/txn/%016X", ptPath, mergeIdx)
-	if err := fs.WriteFile(txnPath, bb.B); err != nil {
+	if err := fs.WriteFileAtomically(txnPath, bb.B); err != nil {
 		return fmt.Errorf("cannot create transaction file %q: %s", txnPath, err)
 	}
 
@@ -1013,12 +1018,14 @@ func (pt *partition) mergeParts(pws []*partWrapper, stopCh <-chan struct{}) erro
 	}
 
 	var newPW *partWrapper
+	var newPSize uint64
 	if len(dstPartPath) > 0 {
 		// Open the merged part if it is non-empty.
 		newP, err := openFilePart(dstPartPath)
 		if err != nil {
 			return fmt.Errorf("cannot open merged part %q: %s", dstPartPath, err)
 		}
+		newPSize = newP.size
 		newPW = &partWrapper{
 			p:        newP,
 			refCount: 1,
@@ -1057,7 +1064,7 @@ func (pt *partition) mergeParts(pws []*partWrapper, stopCh <-chan struct{}) erro
 
 	d := time.Since(startTime)
 	if d > 10*time.Second {
-		logger.Infof("merged %d rows in %s at %d rows/sec to %q", outRowsCount, d, int(float64(outRowsCount)/d.Seconds()), dstPartPath)
+		logger.Infof("merged %d rows in %s at %d rows/sec to %q; sizeBytes: %d", outRowsCount, d, int(float64(outRowsCount)/d.Seconds()), dstPartPath, newPSize)
 	}
 
 	return nil
@@ -1360,7 +1367,12 @@ func runTransactions(txnLock *sync.RWMutex, pathPrefix1, pathPrefix2, path strin
 	})
 
 	for _, fi := range fis {
-		txnPath := txnDir + "/" + fi.Name()
+		fn := fi.Name()
+		if fs.IsTemporaryFileName(fn) {
+			// Skip temporary files, which could be left after unclean shutdown.
+			continue
+		}
+		txnPath := txnDir + "/" + fn
 		if err := runTransaction(txnLock, pathPrefix1, pathPrefix2, txnPath); err != nil {
 			return fmt.Errorf("cannot run transaction from %q: %s", txnPath, err)
 		}

lib/storage/partition_search.go

@@ -56,7 +56,7 @@ func (pts *partitionSearch) reset() {
 // Init initializes the search in the given partition for the given tsid and tr.
 //
 // MustClose must be called when partition search is done.
-func (pts *partitionSearch) Init(pt *partition, tsids []TSID, tr TimeRange) {
+func (pts *partitionSearch) Init(pt *partition, tsids []TSID, tr TimeRange, fetchData bool) {
 	if pts.needClosing {
 		logger.Panicf("BUG: missing partitionSearch.MustClose call before the next call to Init")
 	}
@@ -85,7 +85,7 @@ func (pts *partitionSearch) Init(pt *partition, tsids []TSID, tr TimeRange) {
 	}
 	pts.psPool = pts.psPool[:len(pts.pws)]
 	for i, pw := range pts.pws {
-		pts.psPool[i].Init(pw.p, tsids, tr)
+		pts.psPool[i].Init(pw.p, tsids, tr, fetchData)
 	}
 
 	// Initialize the psHeap.

lib/storage/partition_search_test.go

@@ -238,7 +238,7 @@ func testPartitionSearchSerial(pt *partition, tsids []TSID, tr TimeRange, rbsExp
 
 	bs := []Block{}
 	var pts partitionSearch
-	pts.Init(pt, tsids, tr)
+	pts.Init(pt, tsids, tr, true)
 	for pts.NextBlock() {
 		var b Block
 		b.CopyFrom(pts.Block)
@@ -263,7 +263,7 @@ func testPartitionSearchSerial(pt *partition, tsids []TSID, tr TimeRange, rbsExp
 	}
 
 	// verify that empty tsids returns empty result
-	pts.Init(pt, []TSID{}, tr)
+	pts.Init(pt, []TSID{}, tr, true)
 	if pts.NextBlock() {
 		return fmt.Errorf("unexpected block got for an empty tsids list: %+v", pts.Block)
 	}
@@ -271,6 +271,16 @@ func testPartitionSearchSerial(pt *partition, tsids []TSID, tr TimeRange, rbsExp
 		return fmt.Errorf("unexpected error on empty tsids list: %s", err)
 	}
 	pts.MustClose()
+
+	pts.Init(pt, []TSID{}, tr, false)
+	if pts.NextBlock() {
+		return fmt.Errorf("unexpected block got for an empty tsids list: %+v", pts.Block)
+	}
+	if err := pts.Error(); err != nil {
+		return fmt.Errorf("unexpected error on empty tsids list: %s", err)
+	}
+	pts.MustClose()
+
 	return nil
 }
 

lib/storage/search.go

@@ -110,7 +110,7 @@ func (s *Search) reset() {
 // Init initializes s from the given storage, tfss and tr.
 //
 // MustClose must be called when the search is done.
-func (s *Search) Init(storage *Storage, tfss []*TagFilters, tr TimeRange, maxMetrics int) {
+func (s *Search) Init(storage *Storage, tfss []*TagFilters, tr TimeRange, fetchData bool, maxMetrics int) {
 	if s.needClosing {
 		logger.Panicf("BUG: missing MustClose call before the next call to Init")
 	}
@@ -123,7 +123,7 @@ func (s *Search) Init(storage *Storage, tfss []*TagFilters, tr TimeRange, maxMet
 	// It is ok to call Init on error from storage.searchTSIDs.
 	// Init must be called before returning because it will fail
 	// on Seach.MustClose otherwise.
-	s.ts.Init(storage.tb, tsids, tr)
+	s.ts.Init(storage.tb, tsids, tr, fetchData)
 
 	if err != nil {
 		s.err = err

lib/storage/search_test.go

@@ -193,7 +193,7 @@ func testSearch(st *Storage, tr TimeRange, mrs []MetricRow, accountsCount int) e
 		}
 
 		// Search
-		s.Init(st, []*TagFilters{tfs}, tr, 1e5)
+		s.Init(st, []*TagFilters{tfs}, tr, true, 1e5)
 		var mbs []MetricBlock
 		for s.NextMetricBlock() {
 			var b Block

lib/storage/storage.go

@@ -20,8 +20,8 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
-	"golang.org/x/sys/unix"
 )
 
 const maxRetentionMonths = 12 * 100
@@ -40,16 +40,16 @@ type Storage struct {
 	tb *table
 
 	// tsidCache is MetricName -> TSID cache.
-	tsidCache *fastcache.Cache
+	tsidCache *workingsetcache.Cache
 
 	// metricIDCache is MetricID -> TSID cache.
-	metricIDCache *fastcache.Cache
+	metricIDCache *workingsetcache.Cache
 
 	// metricNameCache is MetricID -> MetricName cache.
-	metricNameCache *fastcache.Cache
+	metricNameCache *workingsetcache.Cache
 
 	// dateMetricIDCache is (Date, MetricID) cache.
-	dateMetricIDCache *fastcache.Cache
+	dateMetricIDCache *workingsetcache.Cache
 
 	// Fast cache for MetricID values occured during the current hour.
 	currHourMetricIDs atomic.Value
@@ -65,6 +65,13 @@ type Storage struct {
 
 	currHourMetricIDsUpdaterWG sync.WaitGroup
 	retentionWatcherWG         sync.WaitGroup
+
+	tooSmallTimestampRows uint64
+	tooBigTimestampRows   uint64
+
+	addRowsConcurrencyLimitReached uint64
+	addRowsConcurrencyLimitTimeout uint64
+	addRowsConcurrencyDroppedRows  uint64
 }
 
 // OpenStorage opens storage on the given path with the given number of retention months.
@@ -96,13 +103,10 @@ func OpenStorage(path string, retentionMonths int) (*Storage, error) {
 		return nil, fmt.Errorf("cannot create %q: %s", snapshotsPath, err)
 	}
 
-	flockFile := path + "/flock.lock"
-	flockF, err := os.Create(flockFile)
+	// Protect from concurrent opens.
+	flockF, err := fs.CreateFlockFile(path)
 	if err != nil {
-		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
-	}
-	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
-		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+		return nil, err
 	}
 	s.flockF = flockF
 
@@ -271,26 +275,35 @@ func (s *Storage) idb() *indexDB {
 
 // Metrics contains essential metrics for the Storage.
 type Metrics struct {
+	TooSmallTimestampRows uint64
+	TooBigTimestampRows   uint64
+
+	AddRowsConcurrencyLimitReached uint64
+	AddRowsConcurrencyLimitTimeout uint64
+	AddRowsConcurrencyDroppedRows  uint64
+	AddRowsConcurrencyCapacity     uint64
+	AddRowsConcurrencyCurrent      uint64
+
 	TSIDCacheSize       uint64
-	TSIDCacheBytesSize  uint64
+	TSIDCacheSizeBytes  uint64
 	TSIDCacheRequests   uint64
 	TSIDCacheMisses     uint64
 	TSIDCacheCollisions uint64
 
 	MetricIDCacheSize       uint64
-	MetricIDCacheBytesSize  uint64
+	MetricIDCacheSizeBytes  uint64
 	MetricIDCacheRequests   uint64
 	MetricIDCacheMisses     uint64
 	MetricIDCacheCollisions uint64
 
 	MetricNameCacheSize       uint64
-	MetricNameCacheBytesSize  uint64
+	MetricNameCacheSizeBytes  uint64
 	MetricNameCacheRequests   uint64
 	MetricNameCacheMisses     uint64
 	MetricNameCacheCollisions uint64
 
 	DateMetricIDCacheSize       uint64
-	DateMetricIDCacheBytesSize  uint64
+	DateMetricIDCacheSizeBytes  uint64
 	DateMetricIDCacheRequests   uint64
 	DateMetricIDCacheMisses     uint64
 	DateMetricIDCacheCollisions uint64
@@ -308,10 +321,19 @@ func (m *Metrics) Reset() {
 
 // UpdateMetrics updates m with metrics from s.
 func (s *Storage) UpdateMetrics(m *Metrics) {
+	m.TooSmallTimestampRows += atomic.LoadUint64(&s.tooSmallTimestampRows)
+	m.TooBigTimestampRows += atomic.LoadUint64(&s.tooBigTimestampRows)
+
+	m.AddRowsConcurrencyLimitReached += atomic.LoadUint64(&s.addRowsConcurrencyLimitReached)
+	m.AddRowsConcurrencyLimitTimeout += atomic.LoadUint64(&s.addRowsConcurrencyLimitTimeout)
+	m.AddRowsConcurrencyDroppedRows += atomic.LoadUint64(&s.addRowsConcurrencyDroppedRows)
+	m.AddRowsConcurrencyCapacity = uint64(cap(addRowsConcurrencyCh))
+	m.AddRowsConcurrencyCurrent = uint64(len(addRowsConcurrencyCh))
+
 	var cs fastcache.Stats
 	s.tsidCache.UpdateStats(&cs)
 	m.TSIDCacheSize += cs.EntriesCount
-	m.TSIDCacheBytesSize += cs.BytesSize
+	m.TSIDCacheSizeBytes += cs.BytesSize
 	m.TSIDCacheRequests += cs.GetCalls
 	m.TSIDCacheMisses += cs.Misses
 	m.TSIDCacheCollisions += cs.Collisions
@@ -319,7 +341,7 @@ func (s *Storage) UpdateMetrics(m *Metrics) {
 	cs.Reset()
 	s.metricIDCache.UpdateStats(&cs)
 	m.MetricIDCacheSize += cs.EntriesCount
-	m.MetricIDCacheBytesSize += cs.BytesSize
+	m.MetricIDCacheSizeBytes += cs.BytesSize
 	m.MetricIDCacheRequests += cs.GetCalls
 	m.MetricIDCacheMisses += cs.Misses
 	m.MetricIDCacheCollisions += cs.Collisions
@@ -327,7 +349,7 @@ func (s *Storage) UpdateMetrics(m *Metrics) {
 	cs.Reset()
 	s.metricNameCache.UpdateStats(&cs)
 	m.MetricNameCacheSize += cs.EntriesCount
-	m.MetricNameCacheBytesSize += cs.BytesSize
+	m.MetricNameCacheSizeBytes += cs.BytesSize
 	m.MetricNameCacheRequests += cs.GetCalls
 	m.MetricNameCacheMisses += cs.Misses
 	m.MetricNameCacheCollisions += cs.Collisions
@@ -335,7 +357,7 @@ func (s *Storage) UpdateMetrics(m *Metrics) {
 	cs.Reset()
 	s.dateMetricIDCache.UpdateStats(&cs)
 	m.DateMetricIDCacheSize += cs.EntriesCount
-	m.DateMetricIDCacheBytesSize += cs.BytesSize
+	m.DateMetricIDCacheSizeBytes += cs.BytesSize
 	m.DateMetricIDCacheRequests += cs.GetCalls
 	m.DateMetricIDCacheMisses += cs.Misses
 	m.DateMetricIDCacheCollisions += cs.Collisions
@@ -439,10 +461,10 @@ func (s *Storage) MustClose() {
 	s.idb().MustClose()
 
 	// Save caches.
-	s.mustSaveCache(s.tsidCache, "MetricName->TSID", "metricName_tsid")
-	s.mustSaveCache(s.metricIDCache, "MetricID->TSID", "metricID_tsid")
-	s.mustSaveCache(s.metricNameCache, "MetricID->MetricName", "metricID_metricName")
-	s.mustSaveCache(s.dateMetricIDCache, "Date->MetricID", "date_metricID")
+	s.mustSaveAndStopCache(s.tsidCache, "MetricName->TSID", "metricName_tsid")
+	s.mustSaveAndStopCache(s.metricIDCache, "MetricID->TSID", "metricID_tsid")
+	s.mustSaveAndStopCache(s.metricNameCache, "MetricID->MetricName", "metricID_metricName")
+	s.mustSaveAndStopCache(s.dateMetricIDCache, "Date->MetricID", "date_metricID")
 
 	hmCurr := s.currHourMetricIDs.Load().(*hourMetricIDs)
 	s.mustSaveHourMetricIDs(hmCurr, "curr_hour_metric_ids")
@@ -492,7 +514,7 @@ func (s *Storage) mustLoadHourMetricIDs(hour uint64, name string) *hourMetricIDs
 		src = src[8:]
 		m[metricID] = struct{}{}
 	}
-	logger.Infof("loaded %s from %q in %s; entriesCount: %d; bytesSize: %d", name, path, time.Since(startTime), hmLen, srcOrigLen)
+	logger.Infof("loaded %s from %q in %s; entriesCount: %d; sizeBytes: %d", name, path, time.Since(startTime), hmLen, srcOrigLen)
 	return &hourMetricIDs{
 		m:      m,
 		hour:   hourLoaded,
@@ -518,33 +540,32 @@ func (s *Storage) mustSaveHourMetricIDs(hm *hourMetricIDs, name string) {
 	if err := ioutil.WriteFile(path, dst, 0644); err != nil {
 		logger.Panicf("FATAL: cannot write %d bytes to %q: %s", len(dst), path, err)
 	}
-	logger.Infof("saved %s to %q in %s; entriesCount: %d; bytesSize: %d", name, path, time.Since(startTime), len(hm.m), len(dst))
+	logger.Infof("saved %s to %q in %s; entriesCount: %d; sizeBytes: %d", name, path, time.Since(startTime), len(hm.m), len(dst))
 }
 
-func (s *Storage) mustLoadCache(info, name string, bytesSize int) *fastcache.Cache {
+func (s *Storage) mustLoadCache(info, name string, sizeBytes int) *workingsetcache.Cache {
 	path := s.cachePath + "/" + name
 	logger.Infof("loading %s cache from %q...", info, path)
 	startTime := time.Now()
-	c := fastcache.LoadFromFileOrNew(path, bytesSize)
+	c := workingsetcache.Load(path, sizeBytes, time.Hour)
 	var cs fastcache.Stats
 	c.UpdateStats(&cs)
-	logger.Infof("loaded %s cache from %q in %s; entriesCount: %d; bytesSize: %d",
+	logger.Infof("loaded %s cache from %q in %s; entriesCount: %d; sizeBytes: %d",
 		info, path, time.Since(startTime), cs.EntriesCount, cs.BytesSize)
 	return c
 }
 
-func (s *Storage) mustSaveCache(c *fastcache.Cache, info, name string) {
-	gomaxprocs := runtime.GOMAXPROCS(-1)
+func (s *Storage) mustSaveAndStopCache(c *workingsetcache.Cache, info, name string) {
 	path := s.cachePath + "/" + name
 	logger.Infof("saving %s cache to %q...", info, path)
 	startTime := time.Now()
-	if err := c.SaveToFileConcurrent(path, gomaxprocs); err != nil {
+	if err := c.Save(path); err != nil {
 		logger.Panicf("FATAL: cannot save %s cache to %q: %s", info, path, err)
 	}
 	var cs fastcache.Stats
 	c.UpdateStats(&cs)
-	c.Reset()
-	logger.Infof("saved %s cache to %q in %s; entriesCount: %d; bytesSize: %d",
+	c.Stop()
+	logger.Infof("saved %s cache to %q in %s; entriesCount: %d; sizeBytes: %d",
 		info, path, time.Since(startTime), cs.EntriesCount, cs.BytesSize)
 }
 
@@ -713,15 +734,24 @@ func (s *Storage) AddRows(mrs []MetricRow, precisionBits uint8) error {
 	// Limit the number of concurrent goroutines that may add rows to the storage.
 	// This should prevent from out of memory errors and CPU trashing when too many
 	// goroutines call AddRows.
-	t := timerpool.Get(addRowsTimeout)
 	select {
 	case addRowsConcurrencyCh <- struct{}{}:
-		timerpool.Put(t)
 		defer func() { <-addRowsConcurrencyCh }()
-	case <-t.C:
-		timerpool.Put(t)
-		return fmt.Errorf("Cannot add %d rows to storage in %s, since it is overloaded with %d concurrent writers. Add more CPUs or reduce load",
-			len(mrs), addRowsTimeout, cap(addRowsConcurrencyCh))
+	default:
+		// Sleep for a while until giving up
+		atomic.AddUint64(&s.addRowsConcurrencyLimitReached, 1)
+		t := timerpool.Get(addRowsTimeout)
+		select {
+		case addRowsConcurrencyCh <- struct{}{}:
+			timerpool.Put(t)
+			defer func() { <-addRowsConcurrencyCh }()
+		case <-t.C:
+			timerpool.Put(t)
+			atomic.AddUint64(&s.addRowsConcurrencyLimitTimeout, 1)
+			atomic.AddUint64(&s.addRowsConcurrencyDroppedRows, uint64(len(mrs)))
+			return fmt.Errorf("Cannot add %d rows to storage in %s, since it is overloaded with %d concurrent writers. Add more CPUs or reduce load",
+				len(mrs), addRowsTimeout, cap(addRowsConcurrencyCh))
+		}
 	}
 
 	// Add rows to the storage.
@@ -752,6 +782,7 @@ func (s *Storage) add(rows []rawRow, mrs []MetricRow, precisionBits uint8) ([]ra
 	}
 	rows = rows[:rowsLen+len(mrs)]
 	j := 0
+	minTimestamp, maxTimestamp := s.tb.getMinMaxTimestamps()
 	for i := range mrs {
 		mr := &mrs[i]
 		if math.IsNaN(mr.Value) {
@@ -759,6 +790,16 @@ func (s *Storage) add(rows []rawRow, mrs []MetricRow, precisionBits uint8) ([]ra
 			// doesn't know how to work with them.
 			continue
 		}
+		if mr.Timestamp < minTimestamp {
+			// Skip rows with too small timestamps outside the retention.
+			atomic.AddUint64(&s.tooSmallTimestampRows, 1)
+			continue
+		}
+		if mr.Timestamp > maxTimestamp {
+			// Skip rows with too big timestamps significantly exceeding the current time.
+			atomic.AddUint64(&s.tooBigTimestampRows, 1)
+			continue
+		}
 		r := &rows[rowsLen+j]
 		j++
 		r.Timestamp = mr.Timestamp
@@ -930,7 +971,7 @@ func (s *Storage) putTSIDToCache(tsid *TSID, metricName []byte) {
 	s.tsidCache.Set(metricName, buf)
 }
 
-func openIndexDBTables(path string, metricIDCache, metricNameCache *fastcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (curr, prev *indexDB, err error) {
+func openIndexDBTables(path string, metricIDCache, metricNameCache *workingsetcache.Cache, currHourMetricIDs, prevHourMetricIDs *atomic.Value) (curr, prev *indexDB, err error) {
 	if err := fs.MkdirAllIfNotExist(path); err != nil {
 		return nil, nil, fmt.Errorf("cannot create directory %q: %s", path, err)
 	}

lib/storage/storage_test.go

@@ -502,12 +502,24 @@ func testStorageDeleteMetrics(s *Storage, workerNum int) error {
 		MaxTimestamp: 2e10,
 	}
 	metricBlocksCount := func(tfs *TagFilters) int {
+		// Verify the number of blocks with fetchData=true
 		n := 0
-		sr.Init(s, []*TagFilters{tfs}, tr, 1e5)
+		sr.Init(s, []*TagFilters{tfs}, tr, true, 1e5)
 		for sr.NextMetricBlock() {
 			n++
 		}
 		sr.MustClose()
+
+		// Make sure the number of blocks with fetchData=false is the same.
+		m := 0
+		sr.Init(s, []*TagFilters{tfs}, tr, false, 1e5)
+		for sr.NextMetricBlock() {
+			m++
+		}
+		sr.MustClose()
+		if n != m {
+			return -1
+		}
 		return n
 	}
 	for i := 0; i < metricsCount; i++ {

lib/storage/table.go

@@ -10,7 +10,6 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
-	"golang.org/x/sys/unix"
 )
 
 // table represents a single table with time series data.
@@ -84,13 +83,10 @@ func openTable(path string, retentionMonths int, getDeletedMetricIDs func() map[
 		return nil, fmt.Errorf("cannot create directory for table %q: %s", path, err)
 	}
 
-	flockFile := path + "/flock.lock"
-	flockF, err := os.Create(flockFile)
+	// Protect from concurrent opens.
+	flockF, err := fs.CreateFlockFile(path)
 	if err != nil {
-		return nil, fmt.Errorf("cannot create lock file %q: %s", flockFile, err)
-	}
-	if err := unix.Flock(int(flockF.Fd()), unix.LOCK_EX|unix.LOCK_NB); err != nil {
-		return nil, fmt.Errorf("cannot acquire lock on file %q: %s", flockFile, err)
+		return nil, err
 	}
 
 	// Create directories for small and big partitions if they don't exist yet.
@@ -310,22 +306,14 @@ func (tb *table) AddRows(rows []rawRow) error {
 	// The slowest path - there are rows that don't fit any existing partition.
 	// Create new partitions for these rows.
 	// Do this under tb.ptwsLock.
-	now := timestampFromTime(time.Now())
-	minTimestamp := now - tb.retentionMilliseconds
-	maxTimestamp := now + 2*24*3600*1000 // allow max +2 days from now due to timezones shit :)
+	minTimestamp, maxTimestamp := tb.getMinMaxTimestamps()
 	tb.ptwsLock.Lock()
 	var errors []error
 	for i := range missingRows {
 		r := &missingRows[i]
 
-		if r.Timestamp < minTimestamp {
-			// Silently skip row with too small timestamp, since it should be deleted anyway.
-			continue
-		}
-		if r.Timestamp > maxTimestamp {
-			err := fmt.Errorf("cannot add row %+v with too big timestamp to table %q; the timestamp cannot be bigger than %d (+2 days from now)",
-				r, tb.path, maxTimestamp)
-			errors = append(errors, err)
+		if r.Timestamp < minTimestamp || r.Timestamp > maxTimestamp {
+			// Silently skip row outside retention, since it should be deleted anyway.
 			continue
 		}
 
@@ -359,6 +347,20 @@ func (tb *table) AddRows(rows []rawRow) error {
 	return nil
 }
 
+func (tb *table) getMinMaxTimestamps() (int64, int64) {
+	now := timestampFromTime(time.Now())
+	minTimestamp := now - tb.retentionMilliseconds
+	maxTimestamp := now + 2*24*3600*1000 // allow max +2 days from now due to timezones shit :)
+	if minTimestamp < 0 {
+		// Negative timestamps aren't supported by the storage.
+		minTimestamp = 0
+	}
+	if maxTimestamp < 0 {
+		maxTimestamp = (1 << 63) - 1
+	}
+	return minTimestamp, maxTimestamp
+}
+
 func (tb *table) startRetentionWatcher() {
 	tb.retentionWatcherWG.Add(1)
 	go func() {

lib/storage/table_search.go

@@ -55,7 +55,7 @@ func (ts *tableSearch) reset() {
 // Init initializes the ts.
 //
 // MustClose must be called then the tableSearch is done.
-func (ts *tableSearch) Init(tb *table, tsids []TSID, tr TimeRange) {
+func (ts *tableSearch) Init(tb *table, tsids []TSID, tr TimeRange, fetchData bool) {
 	if ts.needClosing {
 		logger.Panicf("BUG: missing MustClose call before the next call to Init")
 	}
@@ -86,7 +86,7 @@ func (ts *tableSearch) Init(tb *table, tsids []TSID, tr TimeRange) {
 	}
 	ts.ptsPool = ts.ptsPool[:len(ts.ptws)]
 	for i, ptw := range ts.ptws {
-		ts.ptsPool[i].Init(ptw.pt, tsids, tr)
+		ts.ptsPool[i].Init(ptw.pt, tsids, tr, fetchData)
 	}
 
 	// Initialize the ptsHeap.

lib/storage/table_search_test.go

@@ -251,7 +251,7 @@ func testTableSearchSerial(tb *table, tsids []TSID, tr TimeRange, rbsExpected []
 
 	bs := []Block{}
 	var ts tableSearch
-	ts.Init(tb, tsids, tr)
+	ts.Init(tb, tsids, tr, true)
 	for ts.NextBlock() {
 		var b Block
 		b.CopyFrom(ts.Block)
@@ -276,7 +276,7 @@ func testTableSearchSerial(tb *table, tsids []TSID, tr TimeRange, rbsExpected []
 	}
 
 	// verify that empty tsids returns empty result
-	ts.Init(tb, []TSID{}, tr)
+	ts.Init(tb, []TSID{}, tr, true)
 	if ts.NextBlock() {
 		return fmt.Errorf("unexpected block got for an empty tsids list: %+v", ts.Block)
 	}
@@ -284,5 +284,15 @@ func testTableSearchSerial(tb *table, tsids []TSID, tr TimeRange, rbsExpected []
 		return fmt.Errorf("unexpected error on empty tsids list: %s", err)
 	}
 	ts.MustClose()
+
+	ts.Init(tb, []TSID{}, tr, false)
+	if ts.NextBlock() {
+		return fmt.Errorf("unexpected block got for an empty tsids list with fetchData=false: %+v", ts.Block)
+	}
+	if err := ts.Error(); err != nil {
+		return fmt.Errorf("unexpected error on empty tsids list with fetchData=false: %s", err)
+	}
+	ts.MustClose()
+
 	return nil
 }

lib/storage/table_search_timing_test.go

@@ -26,7 +26,11 @@ func BenchmarkTableSearch(b *testing.B) {
 				b.Run(fmt.Sprintf("tsidsCount_%d", tsidsCount), func(b *testing.B) {
 					for _, tsidsSearch := range []int{1, 1e1, 1e2, 1e3, 1e4} {
 						b.Run(fmt.Sprintf("tsidsSearch_%d", tsidsSearch), func(b *testing.B) {
-							benchmarkTableSearch(b, rowsCount, tsidsCount, tsidsSearch)
+							for _, fetchData := range []bool{true, false} {
+								b.Run(fmt.Sprintf("fetchData_%v", fetchData), func(b *testing.B) {
+									benchmarkTableSearch(b, rowsCount, tsidsCount, tsidsSearch, fetchData)
+								})
+							}
 						})
 					}
 				})
@@ -103,9 +107,9 @@ func createBenchTable(b *testing.B, path string, startTimestamp int64, rowsPerIn
 	tb.MustClose()
 }
 
-func benchmarkTableSearch(b *testing.B, rowsCount, tsidsCount, tsidsSearch int) {
+func benchmarkTableSearch(b *testing.B, rowsCount, tsidsCount, tsidsSearch int, fetchData bool) {
 	startTimestamp := timestampFromTime(time.Now()) - 365*24*3600*1000
-	rowsPerInsert := int(maxRawRowsPerPartition)
+	rowsPerInsert := getMaxRawRowsPerPartition()
 
 	tb := openBenchTable(b, startTimestamp, rowsPerInsert, rowsCount, tsidsCount)
 	tr := TimeRange{
@@ -127,7 +131,7 @@ func benchmarkTableSearch(b *testing.B, rowsCount, tsidsCount, tsidsSearch int)
 			for i := range tsids {
 				tsids[i].MetricID = 1 + uint64(i)
 			}
-			ts.Init(tb, tsids, tr)
+			ts.Init(tb, tsids, tr, fetchData)
 			for ts.NextBlock() {
 			}
 			ts.MustClose()

lib/storage/tag_filters.go

@@ -168,15 +168,14 @@ func (tf *tagFilter) Init(commonPrefix, key, value []byte, isNegative, isRegexp
 			tf.isRegexp = false
 		}
 	}
-	tf.prefix = marshalTagValue(tf.prefix, prefix)
-	if tf.isRegexp {
-		// Remove the trailing tagSeparatorChar from the prefix.
-		tf.prefix = tf.prefix[:len(tf.prefix)-1]
-	}
-	if len(expr) == 0 {
+	tf.prefix = marshalTagValueNoTrailingTagSeparator(tf.prefix, prefix)
+	if !tf.isRegexp {
+		// tf contains plain value without regexp.
+		// Add empty orSuffix in order to trigger fast path for orSuffixes
+		// during the search for matching metricIDs.
+		tf.orSuffixes = append(tf.orSuffixes[:0], "")
 		return nil
 	}
-
 	rcv, err := getRegexpFromCache(expr)
 	if err != nil {
 		return err
@@ -187,16 +186,14 @@ func (tf *tagFilter) Init(commonPrefix, key, value []byte, isNegative, isRegexp
 }
 
 func (tf *tagFilter) matchSuffix(b []byte) (bool, error) {
-	if !tf.isRegexp {
-		return len(b) == 0, nil
-	}
-
-	// Remove the trailing tagSeparatorChar before applying the regexp.
+	// Remove the trailing tagSeparatorChar.
 	if len(b) == 0 || b[len(b)-1] != tagSeparatorChar {
 		return false, fmt.Errorf("unexpected end of b; want %d; b=%q", tagSeparatorChar, b)
 	}
 	b = b[:len(b)-1]
-
+	if !tf.isRegexp {
+		return len(b) == 0, nil
+	}
 	ok := tf.reSuffixMatch(b)
 	return ok, nil
 }

lib/storage/tag_filters_test.go

@@ -54,22 +54,23 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 	key := []byte("key")
 	var tf tagFilter
 
-	tv := func(s string) string {
-		return string(marshalTagValue(nil, []byte(s)))
+	tvNoTrailingTagSeparator := func(s string) string {
+		return string(marshalTagValueNoTrailingTagSeparator(nil, []byte(s)))
 	}
 	init := func(value string, isNegative, isRegexp bool, expectedPrefix string) {
 		t.Helper()
 		if err := tf.Init(commonPrefix, key, []byte(value), isNegative, isRegexp); err != nil {
 			t.Fatalf("unexpected error: %s", err)
 		}
-		prefix := string(commonPrefix) + tv(string(key)) + expectedPrefix
+		prefix := string(commonPrefix) + string(marshalTagValue(nil, []byte(key))) + expectedPrefix
 		if prefix != string(tf.prefix) {
 			t.Fatalf("unexpected tf.prefix; got %q; want %q", tf.prefix, prefix)
 		}
 	}
 	match := func(suffix string) {
 		t.Helper()
-		ok, err := tf.matchSuffix([]byte(suffix))
+		suffixEscaped := marshalTagValue(nil, []byte(suffix))
+		ok, err := tf.matchSuffix(suffixEscaped)
 		if err != nil {
 			t.Fatalf("unexpected error: %s", err)
 		}
@@ -79,7 +80,8 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 	}
 	mismatch := func(suffix string) {
 		t.Helper()
-		ok, err := tf.matchSuffix([]byte(suffix))
+		suffixEscaped := marshalTagValue(nil, []byte(suffix))
+		ok, err := tf.matchSuffix(suffixEscaped)
 		if err != nil {
 			t.Fatalf("unexpected error: %s", err)
 		}
@@ -92,7 +94,7 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 		value := "xx"
 		isNegative := false
 		isRegexp := false
-		expectedPrefix := tv(value)
+		expectedPrefix := tvNoTrailingTagSeparator(value)
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Plain value must match empty suffix only
@@ -104,7 +106,7 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 		value := "xx"
 		isNegative := true
 		isRegexp := false
-		expectedPrefix := tv(value)
+		expectedPrefix := tvNoTrailingTagSeparator(value)
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Negaitve plain value must match all except empty suffix
@@ -119,7 +121,7 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 		value := "http"
 		isNegative := false
 		isRegexp := true
-		expectedPrefix := tv("http")
+		expectedPrefix := tvNoTrailingTagSeparator(value)
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match only empty suffix
@@ -132,7 +134,7 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 		value := "http"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := tv("http")
+		expectedPrefix := tvNoTrailingTagSeparator(value)
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match all except empty suffix
@@ -147,164 +149,220 @@ func TestTagFilterMatchSuffix(t *testing.T) {
 		value := "http.*"
 		isNegative := false
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match any suffix
-		match(tv(""))
-		match(tv("x"))
-		match(tv("http"))
-		match(tv("foobar"))
+		match("")
+		match("x")
+		match("http")
+		match("foobar")
 	})
 	t.Run("negative-regexp-prefix-any-suffix", func(t *testing.T) {
 		value := "http.*"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Mustn't match any suffix
-		mismatch(tv(""))
-		mismatch(tv("x"))
-		mismatch(tv("xhttp"))
-		mismatch(tv("http"))
-		mismatch(tv("httpsdf"))
-		mismatch(tv("foobar"))
+		mismatch("")
+		mismatch("x")
+		mismatch("xhttp")
+		mismatch("http")
+		mismatch("httpsdf")
+		mismatch("foobar")
 	})
 	t.Run("regexp-prefix-contains-suffix", func(t *testing.T) {
 		value := "http.*foo.*"
 		isNegative := false
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match any suffix with `foo`
-		mismatch(tv(""))
-		mismatch(tv("x"))
-		mismatch(tv("http"))
-		match(tv("foo"))
-		match(tv("foobar"))
-		match(tv("xfoobar"))
-		match(tv("xfoo"))
+		mismatch("")
+		mismatch("x")
+		mismatch("http")
+		match("foo")
+		match("foobar")
+		match("xfoobar")
+		match("xfoo")
 	})
 	t.Run("negative-regexp-prefix-contains-suffix", func(t *testing.T) {
 		value := "http.*foo.*"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match any suffix without `foo`
-		match(tv(""))
-		match(tv("x"))
-		match(tv("http"))
-		mismatch(tv("foo"))
-		mismatch(tv("foobar"))
-		mismatch(tv("xfoobar"))
-		mismatch(tv("xfoo"))
-		mismatch(tv("httpfoo"))
-		mismatch(tv("httpfoobar"))
-		mismatch(tv("httpxfoobar"))
-		mismatch(tv("httpxfoo"))
+		match("")
+		match("x")
+		match("http")
+		mismatch("foo")
+		mismatch("foobar")
+		mismatch("xfoobar")
+		mismatch("xfoo")
+		mismatch("httpfoo")
+		mismatch("httpfoobar")
+		mismatch("httpxfoobar")
+		mismatch("httpxfoo")
 	})
 	t.Run("negative-regexp-noprefix-contains-suffix", func(t *testing.T) {
 		value := ".*foo.*"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := ""
+		expectedPrefix := tvNoTrailingTagSeparator("")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match anything not matching `.*foo.*`
-		match(tv(""))
-		match(tv("x"))
-		match(tv("http"))
-		mismatch(tv("foo"))
-		mismatch(tv("foobar"))
-		mismatch(tv("xfoobar"))
-		mismatch(tv("xfoo"))
+		match("")
+		match("x")
+		match("http")
+		mismatch("foo")
+		mismatch("foobar")
+		mismatch("xfoobar")
+		mismatch("xfoo")
 	})
 	t.Run("regexp-prefix-special-suffix", func(t *testing.T) {
 		value := "http.*bar"
 		isNegative := false
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match suffix ending on bar
-		mismatch(tv(""))
-		mismatch(tv("x"))
-		match(tv("bar"))
-		mismatch(tv("barx"))
-		match(tv("foobar"))
-		mismatch(tv("foobarx"))
+		mismatch("")
+		mismatch("x")
+		match("bar")
+		mismatch("barx")
+		match("foobar")
+		mismatch("foobarx")
 	})
 	t.Run("negative-regexp-prefix-special-suffix", func(t *testing.T) {
 		value := "http.*bar"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Mustn't match suffix ending on bar
-		match(tv(""))
-		mismatch(tv("bar"))
-		mismatch(tv("xhttpbar"))
-		mismatch(tv("httpbar"))
-		match(tv("httpbarx"))
-		mismatch(tv("httpxybar"))
-		match(tv("httpxybarx"))
-		mismatch(tv("ahttpxybar"))
+		match("")
+		mismatch("bar")
+		mismatch("xhttpbar")
+		mismatch("httpbar")
+		match("httpbarx")
+		mismatch("httpxybar")
+		match("httpxybarx")
+		mismatch("ahttpxybar")
 	})
 	t.Run("negative-regexp-noprefix-special-suffix", func(t *testing.T) {
 		value := ".*bar"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := ""
+		expectedPrefix := tvNoTrailingTagSeparator("")
 		init(value, isNegative, isRegexp, expectedPrefix)
 
 		// Must match all except the regexp from value
-		match(tv(""))
-		mismatch(tv("bar"))
-		mismatch(tv("xhttpbar"))
-		match(tv("barx"))
-		match(tv("pbarx"))
+		match("")
+		mismatch("bar")
+		mismatch("xhttpbar")
+		match("barx")
+		match("pbarx")
 	})
 	t.Run("regexp-or-suffixes", func(t *testing.T) {
 		value := "http(foo|bar)"
 		isNegative := false
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 		if !reflect.DeepEqual(tf.orSuffixes, []string{"bar", "foo"}) {
 			t.Fatalf("unexpected orSuffixes; got %q; want %q", tf.orSuffixes, []string{"bar", "foo"})
 		}
 
 		// Must match foo or bar suffix
-		mismatch(tv(""))
-		mismatch(tv("x"))
-		match(tv("bar"))
-		mismatch(tv("barx"))
-		match(tv("foo"))
-		mismatch(tv("foobar"))
+		mismatch("")
+		mismatch("x")
+		match("bar")
+		mismatch("barx")
+		match("foo")
+		mismatch("foobar")
 	})
 	t.Run("negative-regexp-or-suffixes", func(t *testing.T) {
 		value := "http(foo|bar)"
 		isNegative := true
 		isRegexp := true
-		expectedPrefix := "http"
+		expectedPrefix := tvNoTrailingTagSeparator("http")
 		init(value, isNegative, isRegexp, expectedPrefix)
 		if !reflect.DeepEqual(tf.orSuffixes, []string{"bar", "foo"}) {
 			t.Fatalf("unexpected or suffixes; got %q; want %q", tf.orSuffixes, []string{"bar", "foo"})
 		}
 
 		// Mustn't match foo or bar suffix
-		match(tv(""))
-		match(tv("x"))
-		mismatch(tv("foo"))
-		match(tv("fooa"))
-		match(tv("xfooa"))
-		mismatch(tv("bar"))
-		match(tv("xhttpbar"))
+		match("")
+		match("x")
+		mismatch("foo")
+		match("fooa")
+		match("xfooa")
+		mismatch("bar")
+		match("xhttpbar")
+	})
+	t.Run("non-empty-string-regexp-negative-match", func(t *testing.T) {
+		value := ".+"
+		isNegative := true
+		isRegexp := true
+		expectedPrefix := tvNoTrailingTagSeparator("")
+		init(value, isNegative, isRegexp, expectedPrefix)
+		if len(tf.orSuffixes) != 0 {
+			t.Fatalf("unexpected non-zero number of or suffixes: %d; %q", len(tf.orSuffixes), tf.orSuffixes)
+		}
+
+		match("")
+		mismatch("x")
+		mismatch("foo")
+	})
+	t.Run("non-empty-string-regexp-match", func(t *testing.T) {
+		value := ".+"
+		isNegative := false
+		isRegexp := true
+		expectedPrefix := tvNoTrailingTagSeparator("")
+		init(value, isNegative, isRegexp, expectedPrefix)
+		if len(tf.orSuffixes) != 0 {
+			t.Fatalf("unexpected non-zero number of or suffixes: %d; %q", len(tf.orSuffixes), tf.orSuffixes)
+		}
+
+		mismatch("")
+		match("x")
+		match("foo")
+	})
+	t.Run("match-all-regexp-negative-match", func(t *testing.T) {
+		value := ".*"
+		isNegative := true
+		isRegexp := true
+		expectedPrefix := tvNoTrailingTagSeparator("")
+		init(value, isNegative, isRegexp, expectedPrefix)
+		if len(tf.orSuffixes) != 0 {
+			t.Fatalf("unexpected non-zero number of or suffixes: %d; %q", len(tf.orSuffixes), tf.orSuffixes)
+		}
+
+		mismatch("")
+		mismatch("x")
+		mismatch("foo")
+	})
+	t.Run("match-all-regexp-match", func(t *testing.T) {
+		value := ".*"
+		isNegative := false
+		isRegexp := true
+		expectedPrefix := tvNoTrailingTagSeparator("")
+		init(value, isNegative, isRegexp, expectedPrefix)
+		if len(tf.orSuffixes) != 0 {
+			t.Fatalf("unexpected non-zero number of or suffixes: %d; %q", len(tf.orSuffixes), tf.orSuffixes)
+		}
+
+		match("")
+		match("x")
+		match("foo")
 	})
 }
 

lib/workingsetcache/cache.go

@@ -0,0 +1,255 @@
+package workingsetcache
+
+import (
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/VictoriaMetrics/fastcache"
+)
+
+// Cache is a cache for working set entries.
+//
+// The cache evicts inactive entries after the given expireDuration.
+// Recently accessed entries survive expireDuration.
+//
+// Comparing to fastcache, this cache minimizes the required RAM size
+// to values smaller than maxBytes.
+type Cache struct {
+	curr atomic.Value
+	prev atomic.Value
+
+	// skipPrev indicates whether to use only curr and skip prev.
+	//
+	// This flag is set if curr is filled for more than 50% space.
+	// In this case using prev would result in RAM waste,
+	// it is better to use only curr cache with doubled size.
+	skipPrev uint64
+
+	// mu serializes access to curr, prev and skipPrev
+	// in expirationWorker and cacheSizeWatcher.
+	mu sync.Mutex
+
+	wg     sync.WaitGroup
+	stopCh chan struct{}
+
+	misses uint64
+}
+
+// Load loads the cache from filePath and limits its size to maxBytes
+// and evicts inactive entires after expireDuration.
+//
+// Stop must be called on the returned cache when it is no longer needed.
+func Load(filePath string, maxBytes int, expireDuration time.Duration) *Cache {
+	// Split maxBytes between curr and prev caches.
+	maxBytes /= 2
+	curr := fastcache.LoadFromFileOrNew(filePath, maxBytes)
+	return newWorkingSetCache(curr, maxBytes, expireDuration)
+}
+
+// New creates new cache with the given maxBytes size and the given expireDuration
+// for inactive entries.
+//
+// Stop must be called on the returned cache when it is no longer needed.
+func New(maxBytes int, expireDuration time.Duration) *Cache {
+	// Split maxBytes between curr and prev caches.
+	maxBytes /= 2
+	curr := fastcache.New(maxBytes)
+	return newWorkingSetCache(curr, maxBytes, expireDuration)
+}
+
+func newWorkingSetCache(curr *fastcache.Cache, maxBytes int, expireDuration time.Duration) *Cache {
+	prev := fastcache.New(1024)
+	var c Cache
+	c.curr.Store(curr)
+	c.prev.Store(prev)
+	c.stopCh = make(chan struct{})
+
+	c.wg.Add(1)
+	go func() {
+		defer c.wg.Done()
+		c.expirationWorker(maxBytes, expireDuration)
+	}()
+	c.wg.Add(1)
+	go func() {
+		defer c.wg.Done()
+		c.cacheSizeWatcher(maxBytes)
+	}()
+	return &c
+}
+
+func (c *Cache) expirationWorker(maxBytes int, expireDuration time.Duration) {
+	t := time.NewTicker(expireDuration / 2)
+	for {
+		select {
+		case <-c.stopCh:
+			t.Stop()
+			return
+		case <-t.C:
+		}
+
+		c.mu.Lock()
+		if atomic.LoadUint64(&c.skipPrev) != 0 {
+			// Expire prev cache and create fresh curr cache.
+			// Do not reuse prev cache, since it can have too big capacity.
+			prev := c.prev.Load().(*fastcache.Cache)
+			prev.Reset()
+			curr := c.curr.Load().(*fastcache.Cache)
+			c.prev.Store(curr)
+			curr = fastcache.New(maxBytes)
+			c.curr.Store(curr)
+		}
+		c.mu.Unlock()
+	}
+}
+
+func (c *Cache) cacheSizeWatcher(maxBytes int) {
+	t := time.NewTicker(time.Minute)
+	for {
+		select {
+		case <-c.stopCh:
+			t.Stop()
+			return
+		case <-t.C:
+		}
+		var cs fastcache.Stats
+		curr := c.curr.Load().(*fastcache.Cache)
+		curr.UpdateStats(&cs)
+		if cs.BytesSize < uint64(maxBytes)/2 {
+			continue
+		}
+
+		// curr cache size exceeds 50% of its capacity. It is better
+		// to double the size of curr cache and stop using prev cache,
+		// since this will result in higher summary cache capacity.
+		c.mu.Lock()
+		curr.Reset()
+		prev := c.prev.Load().(*fastcache.Cache)
+		prev.Reset()
+		curr = fastcache.New(maxBytes * 2)
+		c.curr.Store(curr)
+		atomic.StoreUint64(&c.skipPrev, 1)
+		c.mu.Unlock()
+		return
+	}
+}
+
+// Save safes the cache to filePath.
+func (c *Cache) Save(filePath string) error {
+	curr := c.curr.Load().(*fastcache.Cache)
+	concurrency := runtime.GOMAXPROCS(-1)
+	return curr.SaveToFileConcurrent(filePath, concurrency)
+}
+
+// Stop stops the cache.
+//
+// The cache cannot be used after the Stop call.
+func (c *Cache) Stop() {
+	close(c.stopCh)
+	c.wg.Wait()
+
+	c.Reset()
+}
+
+// Reset resets the cache.
+func (c *Cache) Reset() {
+	prev := c.prev.Load().(*fastcache.Cache)
+	prev.Reset()
+	curr := c.curr.Load().(*fastcache.Cache)
+	curr.Reset()
+
+	c.misses = 0
+}
+
+// UpdateStats updates fcs with cache stats.
+func (c *Cache) UpdateStats(fcs *fastcache.Stats) {
+	curr := c.curr.Load().(*fastcache.Cache)
+	fcsOrig := *fcs
+	curr.UpdateStats(fcs)
+	if atomic.LoadUint64(&c.skipPrev) != 0 {
+		return
+	}
+
+	fcs.Misses = fcsOrig.Misses + atomic.LoadUint64(&c.misses)
+	fcsOrig.Reset()
+	prev := c.prev.Load().(*fastcache.Cache)
+	prev.UpdateStats(&fcsOrig)
+	fcs.EntriesCount += fcsOrig.EntriesCount
+	fcs.BytesSize += fcsOrig.BytesSize
+}
+
+// Get appends the found value for the given key to dst and returns the result.
+func (c *Cache) Get(dst, key []byte) []byte {
+	curr := c.curr.Load().(*fastcache.Cache)
+	result := curr.Get(dst, key)
+	if len(result) > len(dst) {
+		// Fast path - the entry is found in the current cache.
+		return result
+	}
+	if atomic.LoadUint64(&c.skipPrev) != 0 {
+		return result
+	}
+
+	// Search for the entry in the previous cache.
+	prev := c.prev.Load().(*fastcache.Cache)
+	result = prev.Get(dst, key)
+	if len(result) <= len(dst) {
+		// Nothing found.
+		atomic.AddUint64(&c.misses, 1)
+		return result
+	}
+	// Cache the found entry in the current cache.
+	curr.Set(key, result[len(dst):])
+	return result
+}
+
+// Has verifies whether the cahce contains the given key.
+func (c *Cache) Has(key []byte) bool {
+	curr := c.curr.Load().(*fastcache.Cache)
+	if curr.Has(key) {
+		return true
+	}
+	if atomic.LoadUint64(&c.skipPrev) != 0 {
+		return false
+	}
+	prev := c.prev.Load().(*fastcache.Cache)
+	return prev.Has(key)
+}
+
+// Set sets the given value for the given key.
+func (c *Cache) Set(key, value []byte) {
+	curr := c.curr.Load().(*fastcache.Cache)
+	curr.Set(key, value)
+}
+
+// GetBig appends the found value for the given key to dst and returns the result.
+func (c *Cache) GetBig(dst, key []byte) []byte {
+	curr := c.curr.Load().(*fastcache.Cache)
+	result := curr.GetBig(dst, key)
+	if len(result) > len(dst) {
+		// Fast path - the entry is found in the current cache.
+		return result
+	}
+	if atomic.LoadUint64(&c.skipPrev) != 0 {
+		return result
+	}
+
+	// Search for the entry in the previous cache.
+	prev := c.prev.Load().(*fastcache.Cache)
+	result = prev.GetBig(dst, key)
+	if len(result) <= len(dst) {
+		// Nothing found.
+		atomic.AddUint64(&c.misses, 1)
+		return result
+	}
+	// Cache the found entry in the current cache.
+	curr.SetBig(key, result[len(dst):])
+	return result
+}
+
+// SetBig sets the given value for the given key.
+func (c *Cache) SetBig(key, value []byte) {
+	curr := c.curr.Load().(*fastcache.Cache)
+	curr.SetBig(key, value)
+}

package/VAR_BUILD

@@ -0,0 +1 @@
+1

package/VAR_VERSION

@@ -0,0 +1 @@
+1.7.0