app/vmstorage: add -bigMergeConcurrency and -smallMergeConcurrency flags for tuning the maximum number of CPU cores used during merges

The previous information became outdated after v1.28.0, since now each row in the inverted index
can refer to multiple time series.

.github/workflows/main.yml

@@ -0,0 +1,44 @@
+name: main
+on:
+  - push
+  - pull_request
+jobs:
+  build:
+    name: Build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Setup Go
+        uses: actions/setup-go@v1
+        with:
+          go-version: 1.13
+        id: go
+      - name: Code checkout
+        uses: actions/checkout@v1
+      - name: Dependencies
+        env:
+          GO111MODULE: off
+        run: |
+          go get -v golang.org/x/lint/golint
+          go get -u github.com/kisielk/errcheck
+      - name: Build
+        env:
+          GO111MODULE: on
+        run: |
+            export PATH=$PATH:$(go env GOPATH)/bin # temporary fix. See https://github.com/actions/setup-go/issues/14
+            make check-all
+            git diff --exit-code
+            make test-full
+            make test-pure
+            make test-full-386
+            make victoria-metrics
+            make victoria-metrics-pure
+            make victoria-metrics-arm
+            make victoria-metrics-arm64
+            GOOS=freebsd go build -mod=vendor ./app/victoria-metrics
+            GOOS=darwin go build -mod=vendor ./app/victoria-metrics
+      - name: Publish coverage
+        uses: codecov/codecov-action@v1.0.0
+        with:
+          token: ${{secrets.CODECOV_TOKEN}}
+          file: ./coverage.txt
+

.travis.yml

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

Makefile

@@ -1,7 +1,7 @@
 PKG_PREFIX := github.com/VictoriaMetrics/VictoriaMetrics
 
 BUILDINFO_TAG ?= $(shell echo $$(git describe --long --all | tr '/' '-')$$( \
-	      git diff-index --quiet HEAD -- || echo '-dirty-'$$(git diff-index -u HEAD | sha1sum | grep -oP '^.{8}')))
+	      git diff-index --quiet HEAD -- || echo '-dirty-'$$(git diff-index -u HEAD | openssl sha1 | cut -c 10-17)))
 
 PKG_TAG ?= $(shell git tag -l --points-at HEAD)
 ifeq ($(PKG_TAG),)
@@ -61,6 +61,9 @@ test-pure:
 test-full:
 	GO111MODULE=on go test -tags=integration -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
 
+test-full-386:
+	GO111MODULE=on GOARCH=386 go test -tags=integration -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
+
 benchmark:
 	GO111MODULE=on go test -mod=vendor -bench=. ./lib/...
 	GO111MODULE=on go test -mod=vendor -bench=. ./app/...
@@ -75,6 +78,12 @@ vendor-update:
 	GO111MODULE=on go mod tidy
 	GO111MODULE=on go mod vendor
 
+app-local:
+	CGO_ENABLED=1 GO111MODULE=on go build $(RACE) -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/$(APP_NAME)$(RACE) $(PKG_PREFIX)/app/$(APP_NAME)
+
+app-local-pure:
+	CGO_ENABLED=0 GO111MODULE=on go build $(RACE) -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/$(APP_NAME)-pure$(RACE) $(PKG_PREFIX)/app/$(APP_NAME)
+
 quicktemplate-gen: install-qtc
 	qtc
 
@@ -83,7 +92,7 @@ install-qtc:
 
 
 golangci-lint: install-golangci-lint
-	golangci-lint run --exclude '(SA4003|SA1019):' -D errcheck
+	golangci-lint run --exclude '(SA4003|SA1019):' -D errcheck -D structcheck
 
 install-golangci-lint:
 	which golangci-lint || GO111MODULE=off go get -u github.com/golangci/golangci-lint/cmd/golangci-lint

README.md

@@ -2,7 +2,7 @@
 [![Slack](https://img.shields.io/badge/join%20slack-%23victoriametrics-brightgreen.svg)](http://slack.victoriametrics.com/)
 [![GitHub license](https://img.shields.io/github/license/VictoriaMetrics/VictoriaMetrics.svg)](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/LICENSE)
 [![Go Report](https://goreportcard.com/badge/github.com/VictoriaMetrics/VictoriaMetrics)](https://goreportcard.com/report/github.com/VictoriaMetrics/VictoriaMetrics)
-[![Build Status](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics.svg?branch=master)](https://travis-ci.org/VictoriaMetrics/VictoriaMetrics)
+[![Build Status](https://github.com/VictoriaMetrics/VictoriaMetrics/workflows/main/badge.svg)](https://github.com/VictoriaMetrics/VictoriaMetrics/actions)
 [![codecov](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics/branch/master/graph/badge.svg)](https://codecov.io/gh/VictoriaMetrics/VictoriaMetrics)
 
 <img alt="Victoria Metrics" src="logo.png">
@@ -21,7 +21,7 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
 
 * Supports [Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/), so it can be used as Prometheus drop-in replacement in Grafana.
   Additionally, VictoriaMetrics extends PromQL with opt-in [useful features](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/ExtendedPromQL).
-* Global query view. Multiple Prometheus instances may write data into VictoriaMetrics. Later this data may be used in a single query.
+* Supports global query view. Multiple Prometheus instances may write data into VictoriaMetrics. Later this data may be used in a single query.
 * High performance and good scalability for both [inserts](https://medium.com/@valyala/high-cardinality-tsdb-benchmarks-victoriametrics-vs-timescaledb-vs-influxdb-13e6ee64dd6b)
   and [selects](https://medium.com/@valyala/when-size-matters-benchmarking-victoriametrics-vs-timescale-and-influxdb-6035811952d4).
   [Outperforms InfluxDB and TimescaleDB by up to 20x](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
@@ -38,13 +38,14 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   * All the data is stored in a single directory pointed by `-storageDataPath` flag.
   * Easy backups from [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
 * Storage is protected from corruption on unclean shutdown (i.e. hardware reset or `kill -9`) thanks to [the storage architecture](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
-* Supports metrics' ingestion and backfilling via the following protocols:
+* Supports metrics' ingestion and [backfilling](#backfilling) via the following protocols:
   * [Prometheus remote write API](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write)
   * [InfluxDB line protocol](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/)
   * [Graphite plaintext protocol](https://graphite.readthedocs.io/en/latest/feeding-carbon.html) with [tags](https://graphite.readthedocs.io/en/latest/tags.html#carbon)
     if `-graphiteListenAddr` is set.
   * [OpenTSDB put message](http://opentsdb.net/docs/build/html/api_telnet/put.html) if `-opentsdbListenAddr` is set.
-* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars and industrial telemetry.
+  * [HTTP OpenTSDB /api/put requests](http://opentsdb.net/docs/build/html/api_http/put.html) if `-opentsdbHTTPListenAddr` is set.
+* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars, industrial telemetry and various Enterprise workloads.
 * Has open source [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
 
 
@@ -86,6 +87,9 @@ Cluster version is available [here](https://github.com/VictoriaMetrics/VictoriaM
   - [Tuning](#tuning)
   - [Monitoring](#monitoring)
   - [Troubleshooting](#troubleshooting)
+  - [Backfilling](#backfilling)
+  - [Profiling](#profiling)
+- [Integrations](#integrations)
 - [Roadmap](#roadmap)
 - [Contacts](#contacts)
 - [Community and contributions](#community-and-contributions)
@@ -104,11 +108,12 @@ or [docker image](https://hub.docker.com/r/victoriametrics/victoria-metrics/) wi
 
 The following command-line flags are used the most:
 
-* `-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory.
-* `-retentionPeriod` - retention period in months for the data. Older data is automatically deleted.
+* `-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory. Default path is `victoria-metrics-data` in current working directory.
+* `-retentionPeriod` - retention period in months for the data. Older data is automatically deleted. Default period is 1 month.
 * `-httpListenAddr` - TCP address to listen to for http requests. By default, it listens port `8428` on all the network interfaces.
 * `-graphiteListenAddr` - TCP and UDP address to listen to for Graphite data. By default, it is disabled.
-* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data. By default, it is disabled.
+* `-opentsdbListenAddr` - TCP and UDP address to listen to for OpenTSDB data over telnet protocol. By default, it is disabled.
+* `-opentsdbHTTPListenAddr` - TCP address to listen to for HTTP OpenTSDB data over `/api/put`. By default, it is disabled.
 
 Pass `-help` to see all the available flags with description and default values.
 
@@ -152,7 +157,7 @@ The label name may be arbitrary - `datacenter` is just an example. The label val
 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,
+It is recommended upgrading Prometheus to [v2.12.0](https://github.com/prometheus/prometheus/releases) or newer,
 since the previous versions may have issues with `remote_write`.
 
 
@@ -167,7 +172,7 @@ http://<victoriametrics-addr>:8428
 Substitute `<victoriametrics-addr>` with the hostname or IP address of VictoriaMetrics.
 
 Then build graphs with the created datasource using [Prometheus query language](https://prometheus.io/docs/prometheus/latest/querying/basics/).
-VictoriaMetrics supports native PromQL and [extends it with useful features](ExtendedPromQL).
+VictoriaMetrics supports native PromQL and [extends it with useful features](https://github.com/VictoriaMetrics/VictoriaMetrics/wiki/ExtendedPromQL).
 
 
 ### How to upgrade VictoriaMetrics?
@@ -182,6 +187,9 @@ Follow the following steps during the upgrade:
 2) Wait until the process stops. This can take a few seconds.
 3) Start the upgraded VictoriaMetrics.
 
+Prometheus doesn't drop data during VictoriaMetrics restart.
+See [this article](https://grafana.com/blog/2019/03/25/whats-new-in-prometheus-2.8-wal-based-remote-write/) for details.
+
 
 ### How to apply new config to VictoriaMetrics?
 
@@ -191,6 +199,9 @@ VictoriaMetrics must be restarted for applying new config:
 2) Wait until the process stops. This can take a few seconds.
 3) Start VictoriaMetrics with the new config.
 
+Prometheus doesn't drop data during VictoriaMetrics restart.
+See [this article](https://grafana.com/blog/2019/03/25/whats-new-in-prometheus-2.8-wal-based-remote-write/) for details.
+
 
 ### How to send data from InfluxDB-compatible agents such as [Telegraf](https://www.influxdata.com/time-series-platform/telegraf/)?
 
@@ -205,7 +216,8 @@ For instance, put the following lines into `Telegraf` config, so it sends data t
 Do not forget substituting `<victoriametrics-addr>` with the real address where VictoriaMetrics runs.
 
 VictoriaMetrics maps Influx data using the following rules:
-* [`db` query arg](https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint) is mapped into `db` label value.
+* [`db` query arg](https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint) is mapped into `db` label value
+  unless `db` tag exists in the Influx line.
 * Field names are mapped to time series names prefixed with `{measurement}{separator}` value,
   where `{separator}` equals to `_` by default. It can be changed with `-influxMeasurementFieldSeparator` command-line flag.
   See also `-influxSkipSingleField` command-line flag.
@@ -236,14 +248,14 @@ An arbitrary number of lines delimited by '\n' may be sent in a single request.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match={__name__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
 
 ```
-{"metric":{"__name__":"measurement.field1","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560272508147]}
-{"metric":{"__name__":"measurement.field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
+{"metric":{"__name__":"measurement_field1","tag1":"value1","tag2":"value2"},"values":[123],"timestamps":[1560272508147]}
+{"metric":{"__name__":"measurement_field2","tag1":"value1","tag2":"value2"},"values":[1.23],"timestamps":[1560272508147]}
 ```
 
 Note that Influx line protocol expects [timestamps in *nanoseconds* by default](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/#timestamp),
@@ -274,7 +286,7 @@ An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match={__name__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -294,8 +306,13 @@ or via [go-graphite/carbonapi](https://github.com/go-graphite/carbonapi/blob/mas
 
 ### How to send data from OpenTSDB-compatible agents?
 
+VictoriaMetrics supports [telnet put protocol](http://opentsdb.net/docs/build/html/api_telnet/put.html)
+and [HTTP /api/put requests](http://opentsdb.net/docs/build/html/api_http/put.html) for ingesting OpenTSDB data.
+
+#### Sending data via `telnet put` protocol
+
 1) Enable OpenTSDB receiver in VictoriaMetrics by setting `-opentsdbListenAddr` command line flag. For instance,
-the following command will enable OpenTSDB receiver in VictoriaMetrics on TCP and UDP port `4242`:
+the following command enables OpenTSDB receiver in VictoriaMetrics on TCP and UDP port `4242`:
 
 ```
 /path/to/victoria-metrics-prod -opentsdbListenAddr=:4242
@@ -314,7 +331,7 @@ An arbitrary number of lines delimited by `\n` may be sent in one go.
 After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
 
 ```
-curl -G 'http://localhost:8428/api/v1/export' --data-urlencode 'match={__name__!=""}'
+curl -G 'http://localhost:8428/api/v1/export' -d 'match={__name__!=""}'
 ```
 
 The `/api/v1/export` endpoint should return the following response:
@@ -324,6 +341,44 @@ The `/api/v1/export` endpoint should return the following response:
 ```
 
 
+#### Sending OpenTSDB data via HTTP `/api/put` requests
+
+1) Enable HTTP server for OpenTSDB `/api/put` requests by setting `-opentsdbHTTPListenAddr` command line flag. For instance,
+the following command enables OpenTSDB HTTP server on port `4242`:
+
+```
+/path/to/victoria-metrics-prod -opentsdbHTTPListenAddr=:4242
+```
+
+2) Send data to the given address from OpenTSDB-compatible agents.
+
+Example for writing a single data point:
+
+```
+curl -H 'Content-Type: application/json' -d '{"metric":"x.y.z","value":45.34,"tags":{"t1":"v1","t2":"v2"}}' http://localhost:4242/api/put
+```
+
+Example for writing multiple data points in a single request:
+
+```
+curl -H 'Content-Type: application/json' -d '[{"metric":"foo","value":45.34},{"metric":"bar","value":43}]' http://localhost:4242/api/put
+```
+
+After that the data may be read via [/api/v1/export](#how-to-export-time-series) endpoint:
+
+```
+curl -G 'http://localhost:8428/api/v1/export' -d 'match[]=x.y.z' -d 'match[]=foo' -d 'match[]=bar'
+```
+
+The `/api/v1/export` endpoint should return the following response:
+
+```
+{"metric":{"__name__":"foo"},"values":[45.34],"timestamps":[1566464846000]}
+{"metric":{"__name__":"bar"},"values":[43],"timestamps":[1566464846000]}
+{"metric":{"__name__":"x.y.z","t1":"v1","t2":"v2"},"values":[45.34],"timestamps":[1566464763000]}
+```
+
+
 ### How to build from sources
 
 We recommend using either [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) or
@@ -430,6 +485,9 @@ where `<timeseries_selector_for_delete>` may contain any [time series selector](
 for metrics to delete. After that all the time series matching the given selector are deleted. Storage space for
 the deleted time series isn't freed instantly - it is freed during subsequent merges of data files.
 
+It is recommended verifying which metrics will be deleted with the call to `http://<victoria-metrics-addr>:8428/api/v1/series?match[]=<timeseries_selector_for_delete>`
+before actually deleting the metrics.
+
 
 ### How to export time series?
 
@@ -454,7 +512,7 @@ at `http://<victoriametrics-addr>:8428/federate?match[]=<timeseries_selector_for
 
 Optional `start` and `end` args may be added to the request in order to scrape the last point for each selected time series on the `[start ... end]` interval.
 `start` and `end` may contain either unix timestamp in seconds or [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) values. By default, the last point
-on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but can be overridden.
+on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but it can be overridden.
 For instance, `/federate?match[]=up&max_lookback=1h` would return last points on the `[now - 1h ... now]` interval. This may be useful for time series federation
 with scrape intervals exceeding `5m`.
 
@@ -470,7 +528,7 @@ A rough estimation of the required resources for ingestion path:
   VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited by `-memory.allowedPercent` flag.
 
 * CPU cores: a CPU core per 300K inserted data points per second. So, ~4 CPU cores are required for processing
-  the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data.
+  the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data or for time series with high number of labels.
   See [this article](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) for details.
   If you see lower numbers per CPU core, then it is likely active time series info doesn't fit caches,
   so you need more RAM for lowering CPU usage.
@@ -595,6 +653,14 @@ For example, substitute `-graphiteListenAddr=:2003` with `-graphiteListenAddr=<i
 * 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.
+* The recommended filesystem is `ext4`, the recommended persistent storage is [persistent HDD-based disk on GCP](https://cloud.google.com/compute/docs/disks/#pdspecs),
+  since it is protected from hardware failures via internal replication and it can be [resized on the fly](https://cloud.google.com/compute/docs/disks/add-persistent-disk#resize_pd).
+  If you plan storing more than 1TB of data on `ext4` partition or plan extending it to more than 16TB,
+  then the following options are recommended to pass to `mkfs.ext4`:
+
+```
+mkfs.ext4 ... -O 64bit,huge_file,extent -T huge
+```
 
 
 ### Monitoring
@@ -607,10 +673,7 @@ 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.
+* `vm_rows{type="indexdb"}` - the number of rows in inverted index. High value for this number usually mean high churn rate for time series.
 * Sum of `vm_rows{type="storage/big"}` and `vm_rows{type="storage/small"}` - total number of `(timestamp, value)` data points
   in the database.
 * Sum of all the `vm_cache_size_bytes` metrics - the total size of all the caches in the database.
@@ -621,6 +684,9 @@ The most interesting metrics are:
 
 ### Troubleshooting
 
+* It is recommended to use default command-line flag values (i.e. don't set them explicitly) until the need
+  in tweaking these flag values arises.
+
 * If VictoriaMetrics works slowly and eats more than a CPU core per 100K ingested data points per second,
   then it is likely you have too many active time series for the current amount of RAM.
   It is recommended increasing the amount of RAM on the node with VictoriaMetrics in order to improve
@@ -637,6 +703,43 @@ The most interesting metrics are:
   of data loss stored in the broken parts. In the future, `vmrecover` tool will be created
   for automatic recovering from such errors.
 
+
+### Backfilling
+
+Make sure that configured `-retentionPeriod` covers timestamps for the backfilled data.
+
+It is recommended disabling query cache with `-search.disableCache` command-line flag when writing
+historical data with timestamps from the past, since the cache assumes that the data is written with
+the current timestamps. Query cache can be enabled after the backfilling is complete.
+
+
+### Profiling
+
+VictoriaMetrics provides handlers for collecting the following [Go profiles](https://blog.golang.org/profiling-go-programs):
+
+- Memory profile. It can be collected with the following command:
+```
+curl -s http://<victoria-metrics-host>:8428/debug/pprof/heap > mem.pprof
+```
+
+- CPU profile. It can be collected with the following command:
+```
+curl -s http://<victoria-metrics-host>:8428/debug/pprof/profile > cpu.pprof
+```
+
+The command for collecting CPU profile waits for 30 seconds before returning.
+
+The collected profiles may be analyzed with [go tool pprof](https://github.com/google/pprof).
+
+
+## Integrations
+
+* [netdata](https://github.com/netdata/netdata) can push data into VictoriaMetrics via `Prometheus remote_write API`.
+  See [these docs](https://github.com/netdata/netdata#integrations).
+* [go-graphite/carbonapi](https://github.com/go-graphite/carbonapi) can use VictoriaMetrics as time series backend.
+  See [this example](/blob/master/cmd/carbonapi/carbonapi.example.prometheus.yaml).
+
+
 ## Roadmap
 
 - [ ] Replication [#118](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/118)
@@ -659,8 +762,8 @@ Contact us with any questions regarding VictoriaMetrics at [info@victoriametrics
 Feel free asking any questions regarding VictoriaMetrics:
 
 - [slack](http://slack.victoriametrics.com/)
-- [telergam-en](https://t.me/VictoriaMetrics_en)
-- [telergam-ru](https://t.me/VictoriaMetrics_ru1)
+- [telegram-en](https://t.me/VictoriaMetrics_en)
+- [telegram-ru](https://t.me/VictoriaMetrics_ru1)
 - [google groups](https://groups.google.com/forum/#!forum/victorametrics-users)
 
 

app/victoria-metrics/Makefile

@@ -1,7 +1,7 @@
 # All these commands must run from repository root.
 
 victoria-metrics:
-	GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics ./app/victoria-metrics
+	APP_NAME=victoria-metrics $(MAKE) app-local
 
 victoria-metrics-prod:
 	APP_NAME=victoria-metrics $(MAKE) app-via-docker
@@ -32,8 +32,14 @@ victoria-metrics-arm64:
 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-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-386 ./app/victoria-metrics
+
+victoria-metrics-386-prod:
+	APP_NAME=victoria-metrics APP_SUFFIX='-386' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=386' $(MAKE) app-via-docker
+
 victoria-metrics-pure:
-	GO111MODULE=on CGO_ENABLED=0 go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-pure ./app/victoria-metrics
+	APP_NAME=victoria-metrics $(MAKE) app-local-pure
 
 victoria-metrics-pure-prod:
 	APP_NAME=victoria-metrics APP_SUFFIX='-pure' DOCKER_OPTS='--env CGO_ENABLED=0' $(MAKE) app-via-docker

app/victoria-metrics/main.go

@@ -9,6 +9,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/procutil"
@@ -43,6 +44,8 @@ func main() {
 	vmstorage.Stop()
 	vmselect.Stop()
 
+	fs.MustStopDirRemover()
+
 	logger.Infof("the VictoriaMetrics has been stopped in %s", time.Since(startTime))
 }
 

app/victoria-metrics/main_test.go

@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"flag"
 	"fmt"
+	"io"
 	"io/ioutil"
 	"log"
 	"net"
@@ -18,26 +19,31 @@ import (
 	"testing"
 	"time"
 
+	testutil "github.com/VictoriaMetrics/VictoriaMetrics/app/victoria-metrics/test"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
 const (
-	testFixturesDir        = "testdata"
-	testStorageSuffix      = "vm-test-storage"
-	testHTTPListenAddr     = ":7654"
-	testStatsDListenAddr   = ":2003"
-	testOpenTSDBListenAddr = ":4242"
-	testLogLevel           = "INFO"
+	testFixturesDir            = "testdata"
+	testStorageSuffix          = "vm-test-storage"
+	testHTTPListenAddr         = ":7654"
+	testStatsDListenAddr       = ":2003"
+	testOpenTSDBListenAddr     = ":4242"
+	testOpenTSDBHTTPListenAddr = ":4243"
+	testLogLevel               = "INFO"
 )
 
 const (
-	testReadHTTPPath   = "http://127.0.0.1" + testHTTPListenAddr
-	testWriteHTTPPath  = "http://127.0.0.1" + testHTTPListenAddr + "/write"
-	testHealthHTTPPath = "http://127.0.0.1" + testHTTPListenAddr + "/health"
+	testReadHTTPPath          = "http://127.0.0.1" + testHTTPListenAddr
+	testWriteHTTPPath         = "http://127.0.0.1" + testHTTPListenAddr + "/write"
+	testOpenTSDBWriteHTTPPath = "http://127.0.0.1" + testOpenTSDBHTTPListenAddr + "/api/put"
+	testPromWriteHTTPPath     = "http://127.0.0.1" + testHTTPListenAddr + "/api/v1/write"
+	testHealthHTTPPath        = "http://127.0.0.1" + testHTTPListenAddr + "/health"
 )
 
 const (
@@ -50,18 +56,69 @@ var (
 )
 
 type test struct {
-	Name   string `json:"name"`
-	Data   string `json:"data"`
-	Query  string `json:"query"`
-	Result []Row  `json:"result"`
+	Name             string     `json:"name"`
+	Data             []string   `json:"data"`
+	Query            []string   `json:"query"`
+	ResultMetrics    []Metric   `json:"result_metrics"`
+	ResultSeries     Series     `json:"result_series"`
+	ResultQuery      Query      `json:"result_query"`
+	ResultQueryRange QueryRange `json:"result_query_range"`
+	Issue            string     `json:"issue"`
 }
 
-type Row struct {
+type Metric struct {
 	Metric     map[string]string `json:"metric"`
 	Values     []float64         `json:"values"`
 	Timestamps []int64           `json:"timestamps"`
 }
 
+func (r *Metric) UnmarshalJSON(b []byte) error {
+	type plain Metric
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
+type Series struct {
+	Status string              `json:"status"`
+	Data   []map[string]string `json:"data"`
+}
+type Query struct {
+	Status string    `json:"status"`
+	Data   QueryData `json:"data"`
+}
+type QueryData struct {
+	ResultType string            `json:"resultType"`
+	Result     []QueryDataResult `json:"result"`
+}
+
+type QueryDataResult struct {
+	Metric map[string]string `json:"metric"`
+	Value  []interface{}     `json:"value"`
+}
+
+func (r *QueryDataResult) UnmarshalJSON(b []byte) error {
+	type plain QueryDataResult
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
+type QueryRange struct {
+	Status string         `json:"status"`
+	Data   QueryRangeData `json:"data"`
+}
+type QueryRangeData struct {
+	ResultType string                 `json:"resultType"`
+	Result     []QueryRangeDataResult `json:"result"`
+}
+
+type QueryRangeDataResult struct {
+	Metric map[string]string `json:"metric"`
+	Values [][]interface{}   `json:"values"`
+}
+
+func (r *QueryRangeDataResult) UnmarshalJSON(b []byte) error {
+	type plain QueryRangeDataResult
+	return json.Unmarshal(testutil.PopulateTimeTpl(b, insertionTime), (*plain)(r))
+}
+
 func TestMain(m *testing.M) {
 	setUp()
 	code := m.Run()
@@ -92,7 +149,7 @@ func setUp() {
 
 func processFlags() {
 	flag.Parse()
-	for _, fs := range []struct {
+	for _, fv := range []struct {
 		flag  string
 		value string
 	}{
@@ -101,10 +158,11 @@ func processFlags() {
 		{flag: "graphiteListenAddr", value: testStatsDListenAddr},
 		{flag: "opentsdbListenAddr", value: testOpenTSDBListenAddr},
 		{flag: "loggerLevel", value: testLogLevel},
+		{flag: "opentsdbHTTPListenAddr", value: testOpenTSDBHTTPListenAddr},
 	} {
 		// panics if flag doesn't exist
-		if err := flag.Lookup(fs.flag).Value.Set(fs.value); err != nil {
-			log.Fatalf("unable to set %q with value %q, err: %v", fs.flag, fs.value, err)
+		if err := flag.Lookup(fv.flag).Value.Set(fv.value); err != nil {
+			log.Fatalf("unable to set %q with value %q, err: %v", fv.flag, fv.value, err)
 		}
 	}
 }
@@ -121,67 +179,126 @@ func waitFor(timeout time.Duration, f func() bool) error {
 }
 
 func tearDown() {
+	if err := httpserver.Stop(*httpListenAddr); err != nil {
+		log.Printf("cannot stop the webservice: %s", err)
+	}
 	vminsert.Stop()
 	vmstorage.Stop()
 	vmselect.Stop()
-	if err := httpserver.Stop(*httpListenAddr); err != nil {
-		log.Fatalf("cannot stop the webservice: %s", err)
-	}
-	os.RemoveAll(storagePath)
+	fs.MustRemoveAll(storagePath)
+	fs.MustStopDirRemover()
 }
 
 func TestWriteRead(t *testing.T) {
 	t.Run("write", testWrite)
 	time.Sleep(1 * time.Second)
 	vmstorage.Stop()
-
 	// open storage after stop in write
 	vmstorage.InitWithoutMetrics()
 	t.Run("read", testRead)
 }
 
 func testWrite(t *testing.T) {
+	t.Run("prometheus", func(t *testing.T) {
+		for _, test := range readIn("prometheus", t, insertionTime) {
+			s := newSuite(t)
+			r := testutil.WriteRequest{}
+			s.noError(json.Unmarshal([]byte(strings.Join(test.Data, "\n")), &r.Timeseries))
+			data, err := testutil.Compress(r)
+			s.greaterThan(len(r.Timeseries), 0)
+			if err != nil {
+				t.Errorf("error compressing %v %s", r, err)
+				t.Fail()
+			}
+			httpWrite(t, testPromWriteHTTPPath, bytes.NewBuffer(data))
+		}
+	})
+
 	t.Run("influxdb", func(t *testing.T) {
-		for _, test := range readIn("influxdb", t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
+		for _, x := range readIn("influxdb", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				httpWrite(t, testWriteHTTPPath, test.Data)
+				httpWrite(t, testWriteHTTPPath, bytes.NewBufferString(strings.Join(test.Data, "\n")))
 			})
 		}
 	})
 	t.Run("graphite", func(t *testing.T) {
-		for _, test := range readIn("graphite", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+		for _, x := range readIn("graphite", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				tcpWrite(t, "127.0.0.1"+testStatsDListenAddr, test.Data)
+				tcpWrite(t, "127.0.0.1"+testStatsDListenAddr, strings.Join(test.Data, "\n"))
 			})
 		}
 	})
 	t.Run("opentsdb", func(t *testing.T) {
-		for _, test := range readIn("opentsdb", t, fmt.Sprintf("%d", insertionTime.Unix())) {
+		for _, x := range readIn("opentsdb", t, insertionTime) {
+			test := x
 			t.Run(test.Name, func(t *testing.T) {
 				t.Parallel()
-				tcpWrite(t, "127.0.0.1"+testOpenTSDBListenAddr, test.Data)
+				tcpWrite(t, "127.0.0.1"+testOpenTSDBListenAddr, strings.Join(test.Data, "\n"))
+			})
+		}
+	})
+	t.Run("opentsdbhttp", func(t *testing.T) {
+		for _, x := range readIn("opentsdbhttp", t, insertionTime) {
+			test := x
+			t.Run(test.Name, func(t *testing.T) {
+				t.Parallel()
+				logger.Infof("writing %s", test.Data)
+				httpWrite(t, testOpenTSDBWriteHTTPPath, bytes.NewBufferString(strings.Join(test.Data, "\n")))
 			})
 		}
 	})
 }
 
 func testRead(t *testing.T) {
-	for _, engine := range []string{"graphite", "opentsdb", "influxdb"} {
+	for _, engine := range []string{"prometheus", "graphite", "opentsdb", "influxdb", "opentsdbhttp"} {
 		t.Run(engine, func(t *testing.T) {
-			for _, test := range readIn(engine, t, fmt.Sprintf("%d", insertionTime.UnixNano())) {
-				test := test
+			for _, x := range readIn(engine, t, insertionTime) {
+				test := x
 				t.Run(test.Name, func(t *testing.T) {
 					t.Parallel()
-					rowContains(t, httpRead(t, testReadHTTPPath, test.Query), test.Result)
+					for _, q := range test.Query {
+						q = testutil.PopulateTimeTplString(q, insertionTime)
+						if test.Issue != "" {
+							test.Issue = "Regression in " + test.Issue
+						}
+						switch true {
+						case strings.HasPrefix(q, "/api/v1/export"):
+							if err := checkMetricsResult(httpReadMetrics(t, testReadHTTPPath, q), test.ResultMetrics); err != nil {
+								t.Fatalf("Export. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/series"):
+							s := Series{}
+							httpReadStruct(t, testReadHTTPPath, q, &s)
+							if err := checkSeriesResult(s, test.ResultSeries); err != nil {
+								t.Fatalf("Series. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/query_range"):
+							queryResult := QueryRange{}
+							httpReadStruct(t, testReadHTTPPath, q, &queryResult)
+							if err := checkQueryRangeResult(queryResult, test.ResultQueryRange); err != nil {
+								t.Fatalf("Query Range. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						case strings.HasPrefix(q, "/api/v1/query"):
+							queryResult := Query{}
+							httpReadStruct(t, testReadHTTPPath, q, &queryResult)
+							if err := checkQueryResult(queryResult, test.ResultQuery); err != nil {
+								t.Fatalf("Query. %s fails with error %s.%s", q, err, test.Issue)
+							}
+						default:
+							t.Fatalf("unsupported read query %s", q)
+						}
+					}
 				})
 			}
 		})
 	}
 }
 
-func readIn(readFor string, t *testing.T, timeStr string) []test {
+func readIn(readFor string, t *testing.T, insertTime time.Time) []test {
 	t.Helper()
 	s := newSuite(t)
 	var tt []test
@@ -193,7 +310,9 @@ func readIn(readFor string, t *testing.T, timeStr string) []test {
 		s.noError(err)
 		item := test{}
 		s.noError(json.Unmarshal(b, &item))
-		item.Data = strings.Replace(item.Data, "{TIME}", timeStr, 1)
+		for i := range item.Data {
+			item.Data[i] = testutil.PopulateTimeTplString(item.Data[i], insertTime)
+		}
 		tt = append(tt, item)
 		return nil
 	}))
@@ -203,10 +322,10 @@ func readIn(readFor string, t *testing.T, timeStr string) []test {
 	return tt
 }
 
-func httpWrite(t *testing.T, address string, data string) {
+func httpWrite(t *testing.T, address string, r io.Reader) {
 	t.Helper()
 	s := newSuite(t)
-	resp, err := http.Post(address, "", bytes.NewBufferString(data))
+	resp, err := http.Post(address, "", r)
 	s.noError(err)
 	s.noError(resp.Body.Close())
 	s.equalInt(resp.StatusCode, 204)
@@ -223,35 +342,122 @@ func tcpWrite(t *testing.T, address string, data string) {
 	s.equalInt(n, len(data))
 }
 
-func httpRead(t *testing.T, address, query string) []Row {
+func httpReadMetrics(t *testing.T, address, query string) []Metric {
 	t.Helper()
 	s := newSuite(t)
 	resp, err := http.Get(address + query)
 	s.noError(err)
 	defer resp.Body.Close()
 	s.equalInt(resp.StatusCode, 200)
-	var rows []Row
+	var rows []Metric
 	for dec := json.NewDecoder(resp.Body); dec.More(); {
-		var row Row
+		var row Metric
 		s.noError(dec.Decode(&row))
 		rows = append(rows, row)
 	}
 	return rows
 }
-
-func rowContains(t *testing.T, rows, contains []Row) {
+func httpReadStruct(t *testing.T, address, query string, dst interface{}) {
 	t.Helper()
-	for _, r := range rows {
-		contains = removeIfFound(r, contains)
-	}
-	if len(contains) > 0 {
-		t.Fatalf("result rows %+v not found in %+v", contains, rows)
-	}
+	s := newSuite(t)
+	resp, err := http.Get(address + query)
+	s.noError(err)
+	defer resp.Body.Close()
+	s.equalInt(resp.StatusCode, 200)
+	s.noError(json.NewDecoder(resp.Body).Decode(dst))
 }
 
-func removeIfFound(r Row, contains []Row) []Row {
+func checkMetricsResult(got, want []Metric) error {
+	for _, r := range append([]Metric(nil), got...) {
+		want = removeIfFoundMetrics(r, want)
+	}
+	if len(want) > 0 {
+		return fmt.Errorf("exptected metrics %+v not found in %+v", want, got)
+	}
+	return nil
+}
+
+func removeIfFoundMetrics(r Metric, contains []Metric) []Metric {
+	for i, item := range contains {
+		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Values, item.Values) &&
+			reflect.DeepEqual(r.Timestamps, item.Timestamps) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkSeriesResult(got, want Series) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	wantData := append([]map[string]string(nil), want.Data...)
+	for _, r := range got.Data {
+		wantData = removeIfFoundSeries(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected seria(s) %+v not found in %+v", wantData, got.Data)
+	}
+	return nil
+}
+
+func removeIfFoundSeries(r map[string]string, contains []map[string]string) []map[string]string {
+	for i, item := range contains {
+		if reflect.DeepEqual(r, item) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkQueryResult(got, want Query) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	if got.Data.ResultType != want.Data.ResultType {
+		return fmt.Errorf("result type mismatch %q - %q", want.Data.ResultType, got.Data.ResultType)
+	}
+	wantData := append([]QueryDataResult(nil), want.Data.Result...)
+	for _, r := range got.Data.Result {
+		wantData = removeIfFoundQueryData(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected query result %+v not found in %+v", wantData, got.Data.Result)
+	}
+	return nil
+}
+
+func removeIfFoundQueryData(r QueryDataResult, contains []QueryDataResult) []QueryDataResult {
+	for i, item := range contains {
+		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Value[0], item.Value[0]) && reflect.DeepEqual(r.Value[1], item.Value[1]) {
+			contains[i] = contains[len(contains)-1]
+			return contains[:len(contains)-1]
+		}
+	}
+	return contains
+}
+
+func checkQueryRangeResult(got, want QueryRange) error {
+	if got.Status != want.Status {
+		return fmt.Errorf("status mismatch %q - %q", want.Status, got.Status)
+	}
+	if got.Data.ResultType != want.Data.ResultType {
+		return fmt.Errorf("result type mismatch %q - %q", want.Data.ResultType, got.Data.ResultType)
+	}
+	wantData := append([]QueryRangeDataResult(nil), want.Data.Result...)
+	for _, r := range got.Data.Result {
+		wantData = removeIfFoundQueryRangeData(r, wantData)
+	}
+	if len(wantData) > 0 {
+		return fmt.Errorf("expected query range result %+v not found in %+v", wantData, got.Data.Result)
+	}
+	return nil
+}
+
+func removeIfFoundQueryRangeData(r QueryRangeDataResult, contains []QueryRangeDataResult) []QueryRangeDataResult {
 	for i, item := range contains {
-		// todo check time
 		if reflect.DeepEqual(r.Metric, item.Metric) && reflect.DeepEqual(r.Values, item.Values) {
 			contains[i] = contains[len(contains)-1]
 			return contains[:len(contains)-1]
@@ -279,3 +485,11 @@ func (s *suite) equalInt(a, b int) {
 		s.t.FailNow()
 	}
 }
+
+func (s *suite) greaterThan(a, b int) {
+	s.t.Helper()
+	if a <= b {
+		s.t.Errorf("%d less or equal then %d", a, b)
+		s.t.FailNow()
+	}
+}

app/victoria-metrics/test/parser.go

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

app/victoria-metrics/test/parser_test.go

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

app/victoria-metrics/test/prom_types.go

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

app/victoria-metrics/test/prom_writter.go

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

app/vminsert/common/gzip_reader.go

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

app/vminsert/common/insert_ctx.go

@@ -2,9 +2,11 @@ package common
 
 import (
 	"fmt"
+	"net/http"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
@@ -99,7 +101,10 @@ func (ctx *InsertCtx) AddLabel(name, value string) {
 // FlushBufs flushes buffered rows to the underlying storage.
 func (ctx *InsertCtx) FlushBufs() error {
 	if err := vmstorage.AddRows(ctx.mrs); err != nil {
-		return fmt.Errorf("cannot store metrics: %s", err)
+		return &httpserver.ErrorWithStatusCode{
+			Err:        fmt.Errorf("cannot store metrics: %s", err),
+			StatusCode: http.StatusServiceUnavailable,
+		}
 	}
 	return nil
 }

app/vminsert/concurrencylimiter/concurrencylimiter.go

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

app/vminsert/graphite/parser.go

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

app/vminsert/graphite/parser_test.go

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

app/vminsert/graphite/parser_timing_test.go

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

app/vminsert/graphite/request_handler.go

@@ -85,11 +85,7 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			return false
 		}
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		graphiteUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal graphite plaintext protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
-	}
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
 
 	// Fill missing timestamps with the current timestamp rounded to seconds.
 	currentTimestamp := time.Now().Unix()
@@ -136,9 +132,8 @@ func (ctx *pushCtx) reset() {
 }
 
 var (
-	graphiteReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
-	graphiteReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
-	graphiteUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="graphite"}`)
+	graphiteReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
+	graphiteReadErrors = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/influx/parser.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
 	"github.com/valyala/fastjson/fastfloat"
 )
 
@@ -41,13 +43,8 @@ func (rs *Rows) Reset() {
 // See https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/
 //
 // s must be unchanged until rs is in use.
-func (rs *Rows) Unmarshal(s string) error {
-	var err error
-	rs.Rows, rs.tagsPool, rs.fieldsPool, err = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0], rs.fieldsPool[:0])
-	if err != nil {
-		return err
-	}
-	return err
+func (rs *Rows) Unmarshal(s string) {
+	rs.Rows, rs.tagsPool, rs.fieldsPool = unmarshalRows(rs.Rows[:0], s, rs.tagsPool[:0], rs.fieldsPool[:0])
 }
 
 // Row is a single influx row.
@@ -65,9 +62,8 @@ func (r *Row) reset() {
 	r.Timestamp = 0
 }
 
-func (r *Row) unmarshal(s string, tagsPool []Tag, fieldsPool []Field) ([]Tag, []Field, error) {
+func (r *Row) unmarshal(s string, tagsPool []Tag, fieldsPool []Field, noEscapeChars bool) ([]Tag, []Field, error) {
 	r.reset()
-	noEscapeChars := strings.IndexByte(s, '\\') < 0
 	n := nextUnescapedChar(s, ' ', noEscapeChars)
 	if n < 0 {
 		return tagsPool, fieldsPool, fmt.Errorf("cannot find Whitespace I in %q", s)
@@ -141,9 +137,6 @@ func (tag *Tag) unmarshal(s string, noEscapeChars bool) error {
 		return fmt.Errorf("missing tag value for %q", s)
 	}
 	tag.Key = unescapeTagValue(s[:n], noEscapeChars)
-	if len(tag.Key) == 0 {
-		return fmt.Errorf("tag key cannot be empty")
-	}
 	tag.Value = unescapeTagValue(s[n+1:], noEscapeChars)
 	return nil
 }
@@ -177,41 +170,51 @@ func (f *Field) unmarshal(s string, noEscapeChars, hasQuotedFields bool) error {
 	return nil
 }
 
-func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]Row, []Tag, []Field, error) {
+func unmarshalRows(dst []Row, s string, tagsPool []Tag, fieldsPool []Field) ([]Row, []Tag, []Field) {
+	noEscapeChars := strings.IndexByte(s, '\\') < 0
 	for len(s) > 0 {
 		n := strings.IndexByte(s, '\n')
-		if n == 0 {
-			// Skip empty line
-			s = s[1:]
-			continue
-		}
-		if cap(dst) > len(dst) {
-			dst = dst[:len(dst)+1]
-		} else {
-			dst = append(dst, Row{})
-		}
-		r := &dst[len(dst)-1]
 		if n < 0 {
 			// The last line.
-			var err error
-			tagsPool, fieldsPool, err = r.unmarshal(s, tagsPool, fieldsPool)
-			if err != nil {
-				err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s, err)
-				return dst, tagsPool, fieldsPool, err
-			}
-			return dst, tagsPool, fieldsPool, nil
-		}
-		var err error
-		tagsPool, fieldsPool, err = r.unmarshal(s[:n], tagsPool, fieldsPool)
-		if err != nil {
-			err = fmt.Errorf("cannot unmarshal Influx line %q: %s", s[:n], err)
-			return dst, tagsPool, fieldsPool, err
+			return unmarshalRow(dst, s, tagsPool, fieldsPool, noEscapeChars)
 		}
+		dst, tagsPool, fieldsPool = unmarshalRow(dst, s[:n], tagsPool, fieldsPool, noEscapeChars)
 		s = s[n+1:]
 	}
-	return dst, tagsPool, fieldsPool, nil
+	return dst, tagsPool, fieldsPool
 }
 
+func unmarshalRow(dst []Row, s string, tagsPool []Tag, fieldsPool []Field, noEscapeChars bool) ([]Row, []Tag, []Field) {
+	if len(s) > 0 && s[len(s)-1] == '\r' {
+		s = s[:len(s)-1]
+	}
+	if len(s) == 0 {
+		// Skip empty line
+		return dst, tagsPool, fieldsPool
+	}
+	if s[0] == '#' {
+		// Skip comment
+		return dst, tagsPool, fieldsPool
+	}
+
+	if cap(dst) > len(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, Row{})
+	}
+	r := &dst[len(dst)-1]
+	var err error
+	tagsPool, fieldsPool, err = r.unmarshal(s, tagsPool, fieldsPool, noEscapeChars)
+	if err != nil {
+		dst = dst[:len(dst)-1]
+		logger.Errorf("cannot unmarshal Influx line %q: %s; skipping it", s, err)
+		invalidLines.Inc()
+	}
+	return dst, tagsPool, fieldsPool
+}
+
+var invalidLines = metrics.NewCounter(`vm_rows_invalid_total{type="influx"}`)
+
 func unmarshalTags(dst []Tag, s string, noEscapeChars bool) ([]Tag, error) {
 	for {
 		if cap(dst) > len(dst) {
@@ -223,14 +226,22 @@ func unmarshalTags(dst []Tag, s string, noEscapeChars bool) ([]Tag, error) {
 		n := nextUnescapedChar(s, ',', noEscapeChars)
 		if n < 0 {
 			if err := tag.unmarshal(s, noEscapeChars); err != nil {
-				return dst, err
+				return dst[:len(dst)-1], err
+			}
+			if len(tag.Key) == 0 || len(tag.Value) == 0 {
+				// Skip empty tag
+				dst = dst[:len(dst)-1]
 			}
 			return dst, nil
 		}
 		if err := tag.unmarshal(s[:n], noEscapeChars); err != nil {
-			return dst, err
+			return dst[:len(dst)-1], err
 		}
 		s = s[n+1:]
+		if len(tag.Key) == 0 || len(tag.Value) == 0 {
+			// Skip empty tag
+			dst = dst[:len(dst)-1]
+		}
 	}
 }
 

app/vminsert/influx/parser_test.go

@@ -74,13 +74,15 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	f := func(s string) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
 		}
 
 		// Try again
-		if err := rows.Unmarshal(s); err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
+		rows.Unmarshal(s)
+		if len(rows.Rows) != 0 {
+			t.Fatalf("expecting zero rows; got %d rows", len(rows.Rows))
 		}
 	}
 
@@ -94,12 +96,8 @@ func TestRowsUnmarshalFailure(t *testing.T) {
 	// Missing tag value
 	f("foo,bar")
 	f("foo,bar baz")
-	f("foo,bar= baz")
 	f("foo,bar=123, 123")
 
-	// Missing tag name
-	f("foo,=bar baz=234")
-
 	// Missing field value
 	f("foo bar")
 	f("foo bar=")
@@ -122,17 +120,13 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	f := func(s string, rowsExpected *Rows) {
 		t.Helper()
 		var rows Rows
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
 
 		// Try unmarshaling again
-		if err := rows.Unmarshal(s); err != nil {
-			t.Fatalf("cannot unmarshal %q: %s", s, err)
-		}
+		rows.Unmarshal(s)
 		if !reflect.DeepEqual(rows.Rows, rowsExpected.Rows) {
 			t.Fatalf("unexpected rows;\ngot\n%+v;\nwant\n%+v", rows.Rows, rowsExpected.Rows)
 		}
@@ -146,6 +140,12 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	// Empty line
 	f("", &Rows{})
 	f("\n\n", &Rows{})
+	f("\n\r\n", &Rows{})
+
+	// Comment
+	f("\n# foobar\n", &Rows{})
+	f("#foobar baz", &Rows{})
+	f("#foobar baz\n#sss", &Rows{})
 
 	// Minimal line without tags and timestamp
 	f("foo bar=123", &Rows{
@@ -157,6 +157,15 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			}},
 		}},
 	})
+	f("# comment\nfoo bar=123\r\n#comment2 sdsf dsf", &Rows{
+		Rows: []Row{{
+			Measurement: "foo",
+			Fields: []Field{{
+				Key:   "bar",
+				Value: 123,
+			}},
+		}},
+	})
 	f("foo bar=123\n", &Rows{
 		Rows: []Row{{
 			Measurement: "foo",
@@ -216,7 +225,7 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Line with empty tag values
-	f("foo,tag1=xyz,tagN=,tag2=43as bar=123", &Rows{
+	f("foo,tag1=xyz,tagN=,tag2=43as,=xxx bar=123", &Rows{
 		Rows: []Row{{
 			Measurement: "foo",
 			Tags: []Tag{
@@ -224,10 +233,6 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 					Key:   "tag1",
 					Value: "xyz",
 				},
-				{
-					Key:   "tagN",
-					Value: "",
-				},
 				{
 					Key:   "tag2",
 					Value: "43as",
@@ -309,11 +314,11 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 	})
 
 	// Escape chars
-	f(`fo\,bar\=baz,x\==\\a\,\=\q\  \\\a\=\,=4.34`, &Rows{
+	f(`fo\,bar\=baz,x\=\b=\\a\,\=\q\  \\\a\=\,=4.34`, &Rows{
 		Rows: []Row{{
 			Measurement: `fo,bar=baz`,
 			Tags: []Tag{{
-				Key:   `x=`,
+				Key:   `x=\b`,
 				Value: `\a,=\q `,
 			}},
 			Fields: []Field{{
@@ -348,6 +353,34 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+
+	// Multiple lines with invalid line in the middle.
+	f("foo,tag=xyz field=1.23 48934\n"+
+		"invalid line\n"+
+		"bar x=-1i\n\n", &Rows{
+		Rows: []Row{
+			{
+				Measurement: "foo",
+				Tags: []Tag{{
+					Key:   "tag",
+					Value: "xyz",
+				}},
+				Fields: []Field{{
+					Key:   "field",
+					Value: 1.23,
+				}},
+				Timestamp: 48934,
+			},
+			{
+				Measurement: "bar",
+				Fields: []Field{{
+					Key:   "x",
+					Value: -1,
+				}},
+			},
+		},
+	})
+
 	// No newline after the second line.
 	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/82
 	f("foo,tag=xyz field=1.23 48934\n"+
@@ -374,4 +407,24 @@ func TestRowsUnmarshalSuccess(t *testing.T) {
 			},
 		},
 	})
+
+	f("x,y=z,g=p:\\ \\ 5432\\,\\ gp\\ mon\\ [lol]\\ con10\\ cmd5\\ SELECT f=1", &Rows{
+		Rows: []Row{{
+			Measurement: "x",
+			Tags: []Tag{
+				{
+					Key:   "y",
+					Value: "z",
+				},
+				{
+					Key:   "g",
+					Value: "p:  5432, gp mon [lol] con10 cmd5 SELECT",
+				},
+			},
+			Fields: []Field{{
+				Key:   "f",
+				Value: 1,
+			}},
+		}},
+	})
 }

app/vminsert/influx/parser_timing_test.go

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

app/vminsert/influx/request_handler.go

@@ -1,7 +1,6 @@
 package influx
 
 import (
-	"compress/gzip"
 	"flag"
 	"fmt"
 	"io"
@@ -41,11 +40,11 @@ func insertHandlerInternal(req *http.Request) error {
 
 	r := req.Body
 	if req.Header.Get("Content-Encoding") == "gzip" {
-		zr, err := getGzipReader(r)
+		zr, err := common.GetGzipReader(r)
 		if err != nil {
 			return fmt.Errorf("cannot read gzipped influx line protocol data: %s", err)
 		}
-		defer putGzipReader(zr)
+		defer common.PutGzipReader(zr)
 		r = zr
 	}
 
@@ -91,11 +90,17 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	for i := range rows {
 		r := &rows[i]
 		ic.Labels = ic.Labels[:0]
-		ic.AddLabel("db", db)
+		hasDBLabel := false
 		for j := range r.Tags {
 			tag := &r.Tags[j]
+			if tag.Key == "db" {
+				hasDBLabel = true
+			}
 			ic.AddLabel(tag.Key, tag.Value)
 		}
+		if len(db) > 0 && !hasDBLabel {
+			ic.AddLabel("db", db)
+		}
 		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
 		ctx.metricGroupBuf = append(ctx.metricGroupBuf[:0], r.Measurement...)
 		skipFieldKey := len(r.Fields) == 1 && *skipSingleField
@@ -120,25 +125,6 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	return ic.FlushBufs()
 }
 
-func getGzipReader(r io.Reader) (*gzip.Reader, error) {
-	v := gzipReaderPool.Get()
-	if v == nil {
-		return gzip.NewReader(r)
-	}
-	zr := v.(*gzip.Reader)
-	if err := zr.Reset(r); err != nil {
-		return nil, err
-	}
-	return zr, nil
-}
-
-func putGzipReader(zr *gzip.Reader) {
-	_ = zr.Close()
-	gzipReaderPool.Put(zr)
-}
-
-var gzipReaderPool sync.Pool
-
 func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 	if ctx.err != nil {
 		return false
@@ -151,11 +137,7 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 		}
 		return false
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		influxUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal influx line protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
-	}
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
 
 	// Adjust timestamps according to tsMultiplier
 	currentTs := time.Now().UnixNano() / 1e6
@@ -184,9 +166,8 @@ func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
 }
 
 var (
-	influxReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
-	influxReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
-	influxUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="influx"}`)
+	influxReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
+	influxReadErrors = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
 )
 
 type pushCtx struct {

app/vminsert/main.go

@@ -10,19 +10,25 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/graphite"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/influx"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdbhttp"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/prometheus"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 )
 
 var (
-	graphiteListenAddr   = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
-	opentsdbListenAddr   = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB put messages. Usually :4242 must be set. Doesn't work if empty")
-	maxInsertRequestSize = flag.Int("maxInsertRequestSize", 32*1024*1024, "The maximum size of a single insert request in bytes")
+	graphiteListenAddr     = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
+	opentsdbListenAddr     = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB put messages. Usually :4242 must be set. Doesn't work if empty")
+	opentsdbHTTPListenAddr = flag.String("opentsdbHTTPListenAddr", "", "TCP address to listen for OpentTSDB HTTP put requests. Usually :4242 must be set. Doesn't work if empty")
+	maxInsertRequestSize   = flag.Int("maxInsertRequestSize", 32*1024*1024, "The maximum size of a single insert request in bytes")
+	maxLabelsPerTimeseries = flag.Int("maxLabelsPerTimeseries", 30, "The maximum number of labels accepted per time series. Superflouos labels are dropped")
 )
 
 // Init initializes vminsert.
 func Init() {
+	storage.SetMaxLabelsPerTimeseries(*maxLabelsPerTimeseries)
+
 	concurrencylimiter.Init()
 	if len(*graphiteListenAddr) > 0 {
 		go graphite.Serve(*graphiteListenAddr)
@@ -30,6 +36,9 @@ func Init() {
 	if len(*opentsdbListenAddr) > 0 {
 		go opentsdb.Serve(*opentsdbListenAddr)
 	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		go opentsdbhttp.Serve(*opentsdbHTTPListenAddr, int64(*maxInsertRequestSize))
+	}
 }
 
 // Stop stops vminsert.
@@ -40,6 +49,9 @@ func Stop() {
 	if len(*opentsdbListenAddr) > 0 {
 		opentsdb.Stop()
 	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttp.Stop()
+	}
 }
 
 // RequestHandler is a handler for Prometheus remote storage write API

app/vminsert/opentsdb/parser.go

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

app/vminsert/opentsdb/parser_test.go

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

app/vminsert/opentsdb/parser_timing_test.go

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

app/vminsert/opentsdb/request_handler.go

@@ -85,15 +85,21 @@ func (ctx *pushCtx) Read(r io.Reader) bool {
 			return false
 		}
 	}
-	if err := ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf)); err != nil {
-		opentsdbUnmarshalErrors.Inc()
-		ctx.err = fmt.Errorf("cannot unmarshal OpenTSDB put protocol data with size %d: %s", len(ctx.reqBuf), err)
-		return false
+	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
+
+	// Fill in missing timestamps
+	currentTimestamp := time.Now().Unix()
+	rows := ctx.Rows.Rows
+	for i := range rows {
+		r := &rows[i]
+		if r.Timestamp == 0 {
+			r.Timestamp = currentTimestamp
+		}
 	}
 
 	// Convert timestamps from seconds to milliseconds
-	for i := range ctx.Rows.Rows {
-		ctx.Rows.Rows[i].Timestamp *= 1e3
+	for i := range rows {
+		rows[i].Timestamp *= 1e3
 	}
 	return true
 }
@@ -125,9 +131,8 @@ func (ctx *pushCtx) reset() {
 }
 
 var (
-	opentsdbReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
-	opentsdbReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
-	opentsdbUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="opentsdb"}`)
+	opentsdbReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
+	opentsdbReadErrors = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
 )
 
 func getPushCtx() *pushCtx {

app/vminsert/opentsdbhttp/parser.go

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

app/vminsert/opentsdbhttp/parser_test.go

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

app/vminsert/opentsdbhttp/parser_timing_test.go

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

app/vminsert/opentsdbhttp/request_handler.go

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

app/vminsert/opentsdbhttp/server.go

@@ -0,0 +1,70 @@
+package opentsdbhttp
+
+import (
+	"context"
+	"net/http"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	writeRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/put", protocol="opentsdb-http"}`)
+	writeErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/put", protocol="opentsdb-http"}`)
+)
+
+var (
+	httpServer     *http.Server
+	httpAddr       string
+	maxRequestSize int64
+)
+
+// Serve starts HTTP OpenTSDB server on the given addr.
+func Serve(addr string, maxReqSize int64) {
+	logger.Infof("starting HTTP OpenTSDB server at %q", addr)
+	httpAddr = addr
+	maxRequestSize = maxReqSize
+	httpServer = &http.Server{
+		Addr:         addr,
+		Handler:      http.HandlerFunc(requestHandler),
+		ReadTimeout:  30 * time.Second,
+		WriteTimeout: 10 * time.Second,
+	}
+	go func() {
+		err := httpServer.ListenAndServe()
+		if err == http.ErrServerClosed {
+			return
+		}
+		if err != nil {
+			logger.Fatalf("error serving HTTP OpenTSDB: %s", err)
+		}
+	}()
+}
+
+// requestHandler handles HTTP OpenTSDB insert request.
+func requestHandler(w http.ResponseWriter, r *http.Request) {
+	switch r.URL.Path {
+	case "/api/put":
+		writeRequests.Inc()
+		if err := insertHandler(r, maxRequestSize); err != nil {
+			writeErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return
+		}
+		w.WriteHeader(http.StatusNoContent)
+	default:
+		httpserver.Errorf(w, "unexpected path requested on HTTP OpenTSDB server: %q", r.URL.Path)
+	}
+}
+
+// Stop stops HTTP OpenTSDB server.
+func Stop() {
+	logger.Infof("stopping HTTP OpenTSDB server at %q...", httpAddr)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := httpServer.Shutdown(ctx); err != nil {
+		logger.Fatalf("cannot close HTTP OpenTSDB server: %s", err)
+	}
+}

app/vmselect/main.go

@@ -2,6 +2,7 @@ package vmselect
 
 import (
 	"flag"
+	"fmt"
 	"net/http"
 	"runtime"
 	"strings"
@@ -70,7 +71,11 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		case <-t.C:
 			timerpool.Put(t)
 			concurrencyLimitTimeout.Inc()
-			httpserver.Errorf(w, "cannot handle more than %d concurrent requests", cap(concurrencyCh))
+			err := &httpserver.ErrorWithStatusCode{
+				Err:        fmt.Errorf("cannot handle more than %d concurrent requests", cap(concurrencyCh)),
+				StatusCode: http.StatusServiceUnavailable,
+			}
+			httpserver.Errorf(w, "%s", err)
 			return true
 		}
 	}
@@ -185,7 +190,10 @@ func sendPrometheusError(w http.ResponseWriter, r *http.Request, err error) {
 	logger.Errorf("error in %q: %s", r.URL.Path, err)
 
 	w.Header().Set("Content-Type", "application/json")
-	statusCode := 422
+	statusCode := http.StatusUnprocessableEntity
+	if esc, ok := err.(*httpserver.ErrorWithStatusCode); ok {
+		statusCode = esc.StatusCode
+	}
 	w.WriteHeader(statusCode)
 	prometheus.WriteErrorResponse(w, statusCode, err)
 }

app/vmselect/netstorage/fadvise_darwin.go

@@ -4,6 +4,6 @@ import (
 	"os"
 )
 
-func mustFadviseRandomRead(f *os.File) {
+func mustFadviseSequentialRead(f *os.File) {
 	// Do nothing :)
 }

app/vmselect/netstorage/fadvise_freebsd.go

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

app/vmselect/netstorage/fadvise_linux.go

@@ -7,9 +7,9 @@ import (
 	"golang.org/x/sys/unix"
 )
 
-func mustFadviseRandomRead(f *os.File) {
+func mustFadviseSequentialRead(f *os.File) {
 	fd := int(f.Fd())
-	if err := unix.Fadvise(int(fd), 0, 0, unix.FADV_RANDOM|unix.FADV_WILLNEED); err != nil {
-		logger.Panicf("FATAL: error returned from unix.Fadvise(RANDOM|WILLNEED): %s", err)
+	if err := unix.Fadvise(int(fd), 0, 0, unix.FADV_SEQUENTIAL|unix.FADV_WILLNEED); err != nil {
+		logger.Panicf("FATAL: error returned from unix.Fadvise(SEQUENTIAL|WILLNEED): %s", err)
 	}
 }

app/vmselect/netstorage/netstorage.go

@@ -19,9 +19,9 @@ import (
 )
 
 var (
-	maxTagKeysPerSearch   = flag.Int("search.maxTagKeys", 10e3, "The maximum number of tag keys returned per search")
-	maxTagValuesPerSearch = flag.Int("search.maxTagValues", 10e3, "The maximum number of tag values returned per search")
-	maxMetricsPerSearch   = flag.Int("search.maxUniqueTimeseries", 100e3, "The maximum number of unique time series each search can scan")
+	maxTagKeysPerSearch   = flag.Int("search.maxTagKeys", 100e3, "The maximum number of tag keys returned per search")
+	maxTagValuesPerSearch = flag.Int("search.maxTagValues", 100e3, "The maximum number of tag values returned per search")
+	maxMetricsPerSearch   = flag.Int("search.maxUniqueTimeseries", 300e3, "The maximum number of unique time series each search can scan")
 )
 
 // Result is a single timeseries result.
@@ -484,9 +484,12 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadli
 	tbf := getTmpBlocksFile()
 	m := make(map[string][]tmpBlockAddr)
 	blocksRead := 0
+	bb := tmpBufPool.Get()
+	defer tmpBufPool.Put(bb)
 	for sr.NextMetricBlock() {
 		blocksRead++
-		addr, err := tbf.WriteBlock(sr.MetricBlock.Block)
+		bb.B = storage.MarshalBlock(bb.B[:0], sr.MetricBlock.Block)
+		addr, err := tbf.WriteBlockData(bb.B)
 		if err != nil {
 			putTmpBlocksFile(tbf)
 			return nil, fmt.Errorf("cannot write data block #%d to temporary blocks file: %s", blocksRead, err)
@@ -520,6 +523,15 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadli
 		pts.metricName = metricName
 		pts.addrs = addrs
 	}
+
+	// Sort rss.packedTimeseries by the first addr offset in order
+	// to reduce the number of disk seeks during unpacking in RunParallel.
+	// In this case tmpBlocksFile must be read almost sequentially.
+	sort.Slice(rss.packedTimeseries, func(i, j int) bool {
+		pts := rss.packedTimeseries
+		return pts[i].addrs[0].offset < pts[j].addrs[0].offset
+	})
+
 	return &rss, nil
 }
 

app/vmselect/netstorage/tmp_blocks_file.go

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

app/vmselect/netstorage/tmp_blocks_file_test.go

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

app/vmselect/prometheus/export.qtpl

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

app/vmselect/prometheus/export.qtpl.go

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

app/vmselect/prometheus/federate.qtpl

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

app/vmselect/prometheus/federate.qtpl.go

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

app/vmselect/prometheus/prometheus.go

@@ -21,17 +21,17 @@ import (
 )
 
 var (
+	latencyOffset = flag.Duration("search.latencyOffset", time.Second*60, "The time when data points become visible in query results after the colection. "+
+		"Too small value can result in incomplete last points for query results")
 	maxQueryDuration = flag.Duration("search.maxQueryDuration", time.Second*30, "The maximum time for search query execution")
 	maxQueryLen      = flag.Int("search.maxQueryLen", 16*1024, "The maximum search query length in bytes")
+	maxLookback      = flag.Duration("search.maxLookback", 0, "Synonim to `-search.lookback-delta` from Prometheus. "+
+		"The value is dynamically detected from interval between time series datapoints if not set. It can be overriden on per-query basis via `max_lookback` arg")
 )
 
 // Default step used if not set.
 const defaultStep = 5 * 60 * 1000
 
-// Latency for data processing pipeline, i.e. the time between data is ignested
-// into the system and the time it becomes visible to search.
-const latencyOffset = 60 * 1000
-
 // FederateHandler implements /federate . See https://prometheus.io/docs/prometheus/latest/federation/
 func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	startTime := time.Now()
@@ -43,11 +43,14 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		return fmt.Errorf("missing `match[]` arg")
 	}
-	maxLookback, err := getDuration(r, "max_lookback", defaultStep)
+	lookbackDelta, err := getMaxLookback(r)
 	if err != nil {
 		return err
 	}
-	start, err := getTime(r, "start", ct-maxLookback)
+	if lookbackDelta <= 0 {
+		lookbackDelta = defaultStep
+	}
+	start, err := getTime(r, "start", ct-lookbackDelta)
 	if err != nil {
 		return err
 	}
@@ -463,17 +466,22 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	step, err := getDuration(r, "step", latencyOffset)
+	queryOffset := getLatencyOffsetMilliseconds()
+	step, err := getDuration(r, "step", queryOffset)
 	if err != nil {
 		return err
 	}
 	deadline := getDeadline(r)
+	lookbackDelta, err := getMaxLookback(r)
+	if err != nil {
+		return err
+	}
 
 	if len(query) > *maxQueryLen {
 		return fmt.Errorf(`too long query; got %d bytes; mustn't exceed %d bytes`, len(query), *maxQueryLen)
 	}
-	if ct-start < latencyOffset {
-		start -= latencyOffset
+	if ct-start < queryOffset {
+		start -= queryOffset
 	}
 	if childQuery, windowStr, offsetStr := promql.IsMetricSelectorWithRollup(query); childQuery != "" {
 		var window int64
@@ -503,10 +511,11 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 	}
 
 	ec := promql.EvalConfig{
-		Start:    start,
-		End:      start,
-		Step:     step,
-		Deadline: deadline,
+		Start:         start,
+		End:           start,
+		Step:          step,
+		Deadline:      deadline,
+		LookbackDelta: lookbackDelta,
 	}
 	result, err := promql.Exec(&ec, query, true)
 	if err != nil {
@@ -546,6 +555,10 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	}
 	deadline := getDeadline(r)
 	mayCache := !getBool(r, "nocache")
+	lookbackDelta, err := getMaxLookback(r)
+	if err != nil {
+		return err
+	}
 
 	// Validate input args.
 	if len(query) > *maxQueryLen {
@@ -557,29 +570,68 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	if err := promql.ValidateMaxPointsPerTimeseries(start, end, step); err != nil {
 		return err
 	}
-	start, end = promql.AdjustStartEnd(start, end, step)
+	if mayCache {
+		start, end = promql.AdjustStartEnd(start, end, step)
+	}
 
 	ec := promql.EvalConfig{
-		Start:    start,
-		End:      end,
-		Step:     step,
-		Deadline: deadline,
-		MayCache: mayCache,
+		Start:         start,
+		End:           end,
+		Step:          step,
+		Deadline:      deadline,
+		MayCache:      mayCache,
+		LookbackDelta: lookbackDelta,
 	}
 	result, err := promql.Exec(&ec, query, false)
 	if err != nil {
 		return fmt.Errorf("cannot execute %q: %s", query, err)
 	}
-	if ct-end < latencyOffset {
+	queryOffset := getLatencyOffsetMilliseconds()
+	if ct-end < queryOffset {
 		result = adjustLastPoints(result)
 	}
 
+	// Remove NaN values as Prometheus does.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/153
+	removeNaNValuesInplace(result)
+
 	w.Header().Set("Content-Type", "application/json")
 	WriteQueryRangeResponse(w, result)
 	queryRangeDuration.UpdateDuration(startTime)
 	return nil
 }
 
+func removeNaNValuesInplace(tss []netstorage.Result) {
+	for i := range tss {
+		ts := &tss[i]
+		hasNaNs := false
+		for _, v := range ts.Values {
+			if math.IsNaN(v) {
+				hasNaNs = true
+				break
+			}
+		}
+		if !hasNaNs {
+			// Fast path: nothing to remove.
+			continue
+		}
+
+		// Slow path: remove NaNs.
+		srcTimestamps := ts.Timestamps
+		dstValues := ts.Values[:0]
+		dstTimestamps := ts.Timestamps[:0]
+		for j, v := range ts.Values {
+			if math.IsNaN(v) {
+				continue
+			}
+			dstValues = append(dstValues, v)
+			dstTimestamps = append(dstTimestamps, srcTimestamps[j])
+		}
+		ts.Values = dstValues
+		ts.Timestamps = dstTimestamps
+	}
+}
+
 var queryRangeDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/query_range"}`)
 
 // adjustLastPoints substitutes the last point values with the previous
@@ -689,6 +741,11 @@ func getDuration(r *http.Request, argKey string, defaultValue int64) (int64, err
 
 const maxDurationMsecs = 100 * 365 * 24 * 3600 * 1000
 
+func getMaxLookback(r *http.Request) (int64, error) {
+	d := int64(*maxLookback / time.Millisecond)
+	return getDuration(r, "max_lookback", d)
+}
+
 func getDeadline(r *http.Request) netstorage.Deadline {
 	d, err := getDuration(r, "timeout", 0)
 	if err != nil {
@@ -727,3 +784,11 @@ func getTagFilterssFromMatches(matches []string) ([][]storage.TagFilter, error)
 	}
 	return tagFilterss, nil
 }
+
+func getLatencyOffsetMilliseconds() int64 {
+	d := int64(*latencyOffset / time.Millisecond)
+	if d <= 1000 {
+		d = 1000
+	}
+	return d
+}

app/vmselect/prometheus/prometheus_test.go

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

app/vmselect/prometheus/series_count_response.qtpl

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

app/vmselect/prometheus/series_count_response.qtpl.go

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

app/vmselect/promql/aggr.go

@@ -353,6 +353,25 @@ func aggrFuncCountValues(afa *aggrFuncArg) ([]*timeseries, error) {
 	if err != nil {
 		return nil, err
 	}
+
+	// Remove dstLabel from grouping like Prometheus does.
+	modifier := &afa.ae.Modifier
+	switch strings.ToLower(modifier.Op) {
+	case "without":
+		modifier.Args = append(modifier.Args, dstLabel)
+	case "by":
+		dstArgs := modifier.Args[:0]
+		for _, arg := range modifier.Args {
+			if arg == dstLabel {
+				continue
+			}
+			dstArgs = append(dstArgs, arg)
+		}
+		modifier.Args = dstArgs
+	default:
+		// Do nothing
+	}
+
 	afe := func(tss []*timeseries) []*timeseries {
 		m := make(map[float64]bool)
 		for _, ts := range tss {

app/vmselect/promql/aggr_incremental_test.go

@@ -179,7 +179,8 @@ func compareValues(vs1, vs2 []float64) error {
 			}
 			continue
 		}
-		if v1 != v2 {
+		eps := math.Abs(v1 - v2)
+		if eps > 1e-14 {
 			return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
 		}
 	}

app/vmselect/promql/arch_386.go

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

app/vmselect/promql/binary_op.go

@@ -292,24 +292,14 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	}
 
 	// Slow path: `vector op vector` or `a op {on|ignoring} {group_left|group_right} b`
-	ensureOneX := func(side string, tss []*timeseries) error {
-		if len(tss) == 0 {
-			logger.Panicf("BUG: tss must contain at least one value")
-		}
-		if len(tss) == 1 {
-			return nil
-		}
-		if mergeNonOverlappingTimeseries(tss) {
-			return nil
-		}
-		return fmt.Errorf(`duplicate timeseries on the %s side of %s %s: %s and %s`, side, be.Op, be.GroupModifier.AppendString(nil),
-			stringMetricTags(&tss[0].MetricName), stringMetricTags(&tss[1].MetricName))
-	}
-
 	var rvsLeft, rvsRight []*timeseries
 	mLeft, mRight := createTimeseriesMapByTagSet(be, left, right)
 	joinOp := strings.ToLower(be.JoinModifier.Op)
-	joinTags := be.JoinModifier.Args
+	groupOp := strings.ToLower(be.GroupModifier.Op)
+	if len(groupOp) == 0 {
+		groupOp = "ignoring"
+	}
+	groupTags := be.GroupModifier.Args
 	for k, tssLeft := range mLeft {
 		tssRight := mRight[k]
 		if len(tssRight) == 0 {
@@ -317,37 +307,38 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 		}
 		switch joinOp {
 		case "group_left":
-			if err := ensureOneX("right", tssRight); err != nil {
+			var err error
+			rvsLeft, rvsRight, err = groupJoin("right", be, rvsLeft, rvsRight, tssLeft, tssRight)
+			if err != nil {
 				return nil, nil, nil, err
 			}
-			src := tssRight[0]
-			for _, ts := range tssLeft {
-				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
-				rvsLeft = append(rvsLeft, ts)
-				rvsRight = append(rvsRight, src)
-			}
 		case "group_right":
-			if err := ensureOneX("left", tssLeft); err != nil {
+			var err error
+			rvsRight, rvsLeft, err = groupJoin("left", be, rvsRight, rvsLeft, tssRight, tssLeft)
+			if err != nil {
 				return nil, nil, nil, err
 			}
-			src := tssLeft[0]
-			for _, ts := range tssRight {
-				ts.MetricName.AddMissingTags(joinTags, &src.MetricName)
-				rvsLeft = append(rvsLeft, src)
-				rvsRight = append(rvsRight, ts)
-			}
 		case "":
-			if err := ensureOneX("left", tssLeft); err != nil {
+			if err := ensureSingleTimeseries("left", be, tssLeft); err != nil {
 				return nil, nil, nil, err
 			}
-			if err := ensureOneX("right", tssRight); err != nil {
+			if err := ensureSingleTimeseries("right", be, tssRight); err != nil {
 				return nil, nil, nil, err
 			}
-			resetMetricGroupIfRequired(be, tssLeft[0])
-			rvsLeft = append(rvsLeft, tssLeft[0])
+			tsLeft := tssLeft[0]
+			resetMetricGroupIfRequired(be, tsLeft)
+			switch groupOp {
+			case "on":
+				tsLeft.MetricName.RemoveTagsOn(groupTags)
+			case "ignoring":
+				tsLeft.MetricName.RemoveTagsIgnoring(groupTags)
+			default:
+				logger.Panicf("BUG: unexpected binary op modifier %q", groupOp)
+			}
+			rvsLeft = append(rvsLeft, tsLeft)
 			rvsRight = append(rvsRight, tssRight[0])
 		default:
-			return nil, nil, nil, fmt.Errorf(`unexpected join modifier %q`, joinOp)
+			logger.Panicf("BUG: unexpected join modifier %q", joinOp)
 		}
 	}
 	dst := rvsLeft
@@ -357,6 +348,90 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	return rvsLeft, rvsRight, dst, nil
 }
 
+func ensureSingleTimeseries(side string, be *binaryOpExpr, tss []*timeseries) error {
+	if len(tss) == 0 {
+		logger.Panicf("BUG: tss must contain at least one value")
+	}
+	for len(tss) > 1 {
+		if !mergeNonOverlappingTimeseries(tss[0], tss[len(tss)-1]) {
+			return fmt.Errorf(`duplicate time series on the %s side of %s %s: %s and %s`, side, be.Op, be.GroupModifier.AppendString(nil),
+				stringMetricTags(&tss[0].MetricName), stringMetricTags(&tss[len(tss)-1].MetricName))
+		}
+		tss = tss[:len(tss)-1]
+	}
+	return nil
+}
+
+func groupJoin(singleTimeseriesSide string, be *binaryOpExpr, rvsLeft, rvsRight, tssLeft, tssRight []*timeseries) ([]*timeseries, []*timeseries, error) {
+	joinTags := be.JoinModifier.Args
+	var m map[string]*timeseries
+	for _, tsLeft := range tssLeft {
+		resetMetricGroupIfRequired(be, tsLeft)
+		if len(tssRight) == 1 {
+			// Easy case - right part contains only a single matching time series.
+			tsLeft.MetricName.AddMissingTags(joinTags, &tssRight[0].MetricName)
+			rvsLeft = append(rvsLeft, tsLeft)
+			rvsRight = append(rvsRight, tssRight[0])
+			continue
+		}
+
+		// Hard case - right part contains multiple matching time series.
+		// Verify it doesn't result in duplicate MetricName values after adding missing tags.
+		if m == nil {
+			m = make(map[string]*timeseries, len(tssRight))
+		} else {
+			for k := range m {
+				delete(m, k)
+			}
+		}
+		bb := bbPool.Get()
+		for _, tsRight := range tssRight {
+			var tsCopy timeseries
+			tsCopy.CopyFromShallowTimestamps(tsLeft)
+			tsCopy.MetricName.AddMissingTags(joinTags, &tsRight.MetricName)
+			bb.B = marshalMetricTagsSorted(bb.B[:0], &tsCopy.MetricName)
+			if tsExisting := m[string(bb.B)]; tsExisting != nil {
+				// Try merging tsExisting with tsRight if they don't overlap.
+				if mergeNonOverlappingTimeseries(tsExisting, tsRight) {
+					continue
+				}
+				return nil, nil, fmt.Errorf("duplicate time series on the %s side of `%s %s %s`: %s and %s",
+					singleTimeseriesSide, be.Op, be.GroupModifier.AppendString(nil), be.JoinModifier.AppendString(nil),
+					stringMetricTags(&tsExisting.MetricName), stringMetricTags(&tsRight.MetricName))
+			}
+			m[string(bb.B)] = tsRight
+			rvsLeft = append(rvsLeft, &tsCopy)
+			rvsRight = append(rvsRight, tsRight)
+		}
+		bbPool.Put(bb)
+	}
+	return rvsLeft, rvsRight, nil
+}
+
+func mergeNonOverlappingTimeseries(dst, src *timeseries) bool {
+	// Verify whether the time series can be merged.
+	srcValues := src.Values
+	dstValues := dst.Values
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if math.IsNaN(v) {
+			continue
+		}
+		if !math.IsNaN(dstValues[i]) {
+			return false
+		}
+	}
+
+	// Time series can be merged. Merge them.
+	for i, v := range srcValues {
+		if math.IsNaN(v) {
+			continue
+		}
+		dstValues[i] = v
+	}
+	return true
+}
+
 func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
 	if isBinaryOpCmp(be.Op) && !be.Bool {
 		// Do not reset MetricGroup for non-boolean `compare` binary ops like Prometheus does.
@@ -533,26 +608,3 @@ func isScalar(arg []*timeseries) bool {
 	}
 	return len(mn.Tags) == 0
 }
-
-func mergeNonOverlappingTimeseries(tss []*timeseries) bool {
-	if len(tss) < 2 {
-		logger.Panicf("BUG: expecting at least two timeseries. Got %d", len(tss))
-	}
-
-	// Check whether time series in tss overlap.
-	var dst timeseries
-	dst.CopyFromShallowTimestamps(tss[0])
-	dstValues := dst.Values
-	for _, ts := range tss[1:] {
-		for i, value := range ts.Values {
-			if math.IsNaN(dstValues[i]) {
-				dstValues[i] = value
-			} else if !math.IsNaN(value) {
-				// Time series overlap.
-				return false
-			}
-		}
-	}
-	tss[0].CopyFromShallowTimestamps(&dst)
-	return true
-}

app/vmselect/promql/eval.go

@@ -70,6 +70,9 @@ type EvalConfig struct {
 
 	MayCache bool
 
+	// LookbackDelta is analog to `-query.lookback-delta` from Prometheus.
+	LookbackDelta int64
+
 	timestamps     []int64
 	timestampsOnce sync.Once
 }
@@ -82,6 +85,7 @@ func newEvalConfig(src *EvalConfig) *EvalConfig {
 	ec.Step = src.Step
 	ec.Deadline = src.Deadline
 	ec.MayCache = src.MayCache
+	ec.LookbackDelta = src.LookbackDelta
 
 	// do not copy src.timestamps - they must be generated again.
 	return &ec
@@ -290,10 +294,10 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 		return fe, nrf
 	}
 	if re, ok := e.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
-		// e = rollupExpr(metricExpr)
+		// e = metricExpr[d]
 		fe := &funcExpr{
 			Name: "default_rollup",
 			Args: []expr{re},
@@ -315,15 +319,17 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 		if me.IsEmpty() {
 			return nil, nil
 		}
+		// e = rollupFunc(metricExpr)
 		return &funcExpr{
 			Name: fe.Name,
 			Args: []expr{me},
 		}, nrf
 	}
 	if re, ok := arg.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() {
+		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
+		// e = rollupFunc(metricExpr[d])
 		return fe, nrf
 	}
 	return nil, nil
@@ -368,8 +374,8 @@ func getRollupExprArg(arg expr) *rollupExpr {
 			Expr: arg,
 		}
 	}
-	if len(re.Step) == 0 && !re.InheritStep {
-		// Return standard rollup if it doesn't set step.
+	if !re.ForSubquery() {
+		// Return standard rollup if it doesn't contain subquery.
 		return re
 	}
 	me, ok := re.Expr.(*metricExpr)
@@ -463,7 +469,7 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	}
 
 	sharedTimestamps := getTimestamps(ec.Start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, ec.Start, ec.End, ec.Step, window, sharedTimestamps)
+	preFunc, rcs := getRollupConfigs(name, rf, ec.Start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
 	tss := make([]*timeseries, 0, len(tssSQ)*len(rcs))
 	var tssLock sync.Mutex
 	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
@@ -584,7 +590,7 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 		return tss, nil
 	}
 	sharedTimestamps := getTimestamps(start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, start, ec.End, ec.Step, window, sharedTimestamps)
+	preFunc, rcs := getRollupConfigs(name, rf, start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
 
 	// Verify timeseries fit available memory after the rollup.
 	// Take into account points from tssCached.
@@ -687,7 +693,8 @@ func doRollupForTimeseries(rc *rollupConfig, tsDst *timeseries, mnSrc *storage.M
 	tsDst.denyReuse = true
 }
 
-func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64, sharedTimestamps []int64) (func(values []float64, timestamps []int64), []*rollupConfig) {
+func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64, lookbackDelta int64, sharedTimestamps []int64) (
+	func(values []float64, timestamps []int64), []*rollupConfig) {
 	preFunc := func(values []float64, timestamps []int64) {}
 	if rollupFuncsRemoveCounterResets[name] {
 		preFunc = func(values []float64, timestamps []int64) {
@@ -703,6 +710,7 @@ func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64
 			Step:            step,
 			Window:          window,
 			MayAdjustWindow: rollupFuncsMayAdjustWindow[name],
+			LookbackDelta:   lookbackDelta,
 			Timestamps:      sharedTimestamps,
 		}
 	}

app/vmselect/promql/exec.go

@@ -105,14 +105,14 @@ func maySortResults(e expr, tss []*timeseries) bool {
 func timeseriesToResult(tss []*timeseries, maySort bool) ([]netstorage.Result, error) {
 	tss = removeNaNs(tss)
 	result := make([]netstorage.Result, len(tss))
-	m := make(map[string]bool)
+	m := make(map[string]struct{}, len(tss))
 	bb := bbPool.Get()
 	for i, ts := range tss {
 		bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
-		if m[string(bb.B)] {
+		if _, ok := m[string(bb.B)]; ok {
 			return nil, fmt.Errorf(`duplicate output timeseries: %s%s`, ts.MetricName.MetricGroup, stringMetricName(&ts.MetricName))
 		}
-		m[string(bb.B)] = true
+		m[string(bb.B)] = struct{}{}
 
 		rs := &result[i]
 		rs.MetricNameMarshaled = append(rs.MetricNameMarshaled[:0], bb.B...)
@@ -194,11 +194,14 @@ type parseCacheValue struct {
 }
 
 type parseCache struct {
-	m  map[string]*parseCacheValue
-	mu sync.RWMutex
+	// Move atomic counters to the top of struct for 8-byte alignment on 32-bit arch.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
 
 	requests uint64
 	misses   uint64
+
+	m  map[string]*parseCacheValue
+	mu sync.RWMutex
 }
 
 func (pc *parseCache) Requests() uint64 {

app/vmselect/promql/exec_test.go

@@ -369,6 +369,17 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run("timestamp(time()>=1600)", func(t *testing.T) {
+		t.Parallel()
+		q := `timestamp(time()>=1600)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, nan, 1600, 1800, 2000},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run("time()/100", func(t *testing.T) {
 		t.Parallel()
 		q := `time()/100`
@@ -1302,6 +1313,44 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`label_value()`, func(t *testing.T) {
+		t.Parallel()
+		q := `with (
+			x = (
+				label_set(time(), "foo", "123.456", "__name__", "aaa"),
+				label_set(-time(), "foo", "bar", "__name__", "bbb"),
+				label_set(-time(), "__name__", "bxs"),
+				label_set(-time(), "foo", "45", "bar", "xs"),
+			)
+		)
+		sort(x + label_value(x, "foo"))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{-955, -1155, -1355, -1555, -1755, -1955},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("bar"),
+				Value: []byte("xs"),
+			},
+			{
+				Key:   []byte("foo"),
+				Value: []byte("45"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1123.456, 1323.456, 1523.456, 1723.456, 1923.456, 2123.456},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("123.456"),
+		}}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
 	t.Run(`label_transform(mismatch)`, func(t *testing.T) {
 		t.Parallel()
 		q := `label_transform(time(), "__name__", "foobar", "xx")`
@@ -1788,10 +1837,6 @@ func TestExecSuccess(t *testing.T) {
 			Timestamps: timestampsExpected,
 		}
 		r.MetricName.Tags = []storage.Tag{
-			{
-				Key:   []byte("aa"),
-				Value: []byte("bb"),
-			},
 			{
 				Key:   []byte("foo"),
 				Value: []byte("bar"),
@@ -1813,17 +1858,60 @@ func TestExecSuccess(t *testing.T) {
 				Key:   []byte("foo"),
 				Value: []byte("bar"),
 			},
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`vector * on(foo) group_left(additional_tag) duplicate_timeseries_differ_by_additional_tag`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort(label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_left(op) (
+			label_set(time() < 1400, "foo", "bar", "op", "le"),
+			label_set(time() >= 1400, "foo", "bar", "op", "ge"),
+		))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, nan, nan, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("op"),
+				Value: []byte("le"),
+			},
 			{
 				Key:   []byte("xx"),
 				Value: []byte("yy"),
 			},
 		}
-		resultExpected := []netstorage.Result{r}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("op"),
+				Value: []byte("ge"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2}
 		f(q, resultExpected)
 	})
-	t.Run(`vector * on(foo) group_left() duplicate_timeseries`, func(t *testing.T) {
+	t.Run(`vector * on(foo) duplicate_nonoverlapping_timeseries`, func(t *testing.T) {
 		t.Parallel()
-		q := `label_set(time()/10, "foo", "bar") + on(foo) group_left() (
+		q := `label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) (
 			label_set(time() < 1400, "foo", "bar", "op", "le"),
 			label_set(time() >= 1400, "foo", "bar", "op", "ge"),
 		)`
@@ -1832,13 +1920,105 @@ func TestExecSuccess(t *testing.T) {
 			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
 			Timestamps: timestampsExpected,
 		}
-		r1.MetricName.Tags = []storage.Tag{{
-			Key:   []byte("foo"),
-			Value: []byte("bar"),
-		}}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+		}
 		resultExpected := []netstorage.Result{r1}
 		f(q, resultExpected)
 	})
+	t.Run(`vector * on(foo) group_left() duplicate_nonoverlapping_timeseries`, func(t *testing.T) {
+		t.Parallel()
+		q := `label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_left() (
+			label_set(time() < 1400, "foo", "bar", "op", "le"),
+			label_set(time() >= 1400, "foo", "bar", "op", "ge"),
+		)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1}
+		f(q, resultExpected)
+	})
+	t.Run(`vector * on(foo) group_left(__name__)`, func(t *testing.T) {
+		t.Parallel()
+		q := `label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_left(__name__)
+			label_set(time(), "foo", "bar", "__name__", "aaa")`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.MetricGroup = []byte("aaa")
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1}
+		f(q, resultExpected)
+	})
+	t.Run(`vector * on(foo) group_right()`, func(t *testing.T) {
+		t.Parallel()
+		q := `sort(label_set(time()/10, "foo", "bar", "xx", "yy", "__name__", "qwert") + on(foo) group_right(xx) (
+			label_set(time(), "foo", "bar", "__name__", "aaa"),
+			label_set(time()+3, "foo", "bar", "__name__", "yyy","ppp", "123"),
+		))`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1100, 1320, 1540, 1760, 1980, 2200},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1103, 1323, 1543, 1763, 1983, 2203},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("ppp"),
+				Value: []byte("123"),
+			},
+			{
+				Key:   []byte("xx"),
+				Value: []byte("yy"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2}
+		f(q, resultExpected)
+	})
 	t.Run(`vector * on() group_left scalar`, func(t *testing.T) {
 		t.Parallel()
 		q := `sort_desc((label_set(time(), "foo", "bar") or label_set(10, "foo", "qwert")) * on() group_left 2)`
@@ -1933,10 +2113,6 @@ func TestExecSuccess(t *testing.T) {
 			Timestamps: timestampsExpected,
 		}
 		r.MetricName.Tags = []storage.Tag{
-			{
-				Key:   []byte("t1"),
-				Value: []byte("v123"),
-			},
 			{
 				Key:   []byte("t2"),
 				Value: []byte("v3"),
@@ -2042,10 +2218,6 @@ func TestExecSuccess(t *testing.T) {
 			Timestamps: timestampsExpected,
 		}
 		r.MetricName.Tags = []storage.Tag{
-			{
-				Key:   []byte("t1"),
-				Value: []byte("v123"),
-			},
 			{
 				Key:   []byte("t2"),
 				Value: []byte("v3"),
@@ -2160,21 +2332,78 @@ func TestExecSuccess(t *testing.T) {
 	})
 	t.Run(`histogram_quantile(negative-bucket-count)`, func(t *testing.T) {
 		t.Parallel()
-		q := `sort(histogram_quantile(0.6,
+		q := `histogram_quantile(0.6,
 			label_set(90, "foo", "bar", "le", "10")
 			or label_set(-100, "foo", "bar", "le", "30")
 			or label_set(300, "foo", "bar", "le", "+Inf")
-		))`
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{30, 30, 30, 30, 30, 30},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(90, "foo", "bar", "le", "10")
+			or label_set(NaN, "foo", "bar", "le", "30")
+			or label_set(300, "foo", "bar", "le", "+Inf")
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{30, 30, 30, 30, 30, 30},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.2,
+			label_set(0, "foo", "bar", "le", "10")
+			or label_set(100, "foo", "bar", "le", "30")
+			or label_set(300, "foo", "bar", "le", "+Inf")
+		)`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{22, 22, 22, 22, 22, 22},
+			Timestamps: timestampsExpected,
+		}
+		r.MetricName.Tags = []storage.Tag{{
+			Key:   []byte("foo"),
+			Value: []byte("bar"),
+		}}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`histogram_quantile(zero-bucket-count)`, func(t *testing.T) {
+		t.Parallel()
+		q := `histogram_quantile(0.6,
+			label_set(0, "foo", "bar", "le", "10")
+			or label_set(0, "foo", "bar", "le", "30")
+			or label_set(0, "foo", "bar", "le", "+Inf")
+		)`
 		resultExpected := []netstorage.Result{}
 		f(q, resultExpected)
 	})
 	t.Run(`histogram_quantile(nan-bucket-count)`, func(t *testing.T) {
 		t.Parallel()
-		q := `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")
-		))`
+		q := `histogram_quantile(0.6,
+			label_set(nan, "foo", "bar", "le", "10")
+			or label_set(nan, "foo", "bar", "le", "30")
+			or label_set(nan, "foo", "bar", "le", "+Inf")
+		)`
 		resultExpected := []netstorage.Result{}
 		f(q, resultExpected)
 	})
@@ -2453,6 +2682,28 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r}
 		f(q, resultExpected)
 	})
+	t.Run(`increases_over_time`, func(t *testing.T) {
+		t.Parallel()
+		q := `increases_over_time(rand(0)[200s:10s])`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{11, 9, 9, 12, 9, 8},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
+	t.Run(`decreases_over_time`, func(t *testing.T) {
+		t.Parallel()
+		q := `decreases_over_time(rand(0)[200s:10s])`
+		r := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{9, 11, 11, 8, 11, 12},
+			Timestamps: timestampsExpected,
+		}
+		resultExpected := []netstorage.Result{r}
+		f(q, resultExpected)
+	})
 	t.Run(`limitk(-1)`, func(t *testing.T) {
 		t.Parallel()
 		q := `limitk(-1, label_set(10, "foo", "bar") or label_set(time()/150, "baz", "sss"))`
@@ -3305,7 +3556,7 @@ func TestExecSuccess(t *testing.T) {
 		}}
 		r4 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{0.85, 0.94, 0.97, 0.93, 0.98, 0.92},
+			Values:     []float64{0.9, 0.94, 0.97, 0.93, 0.98, 0.92},
 			Timestamps: timestampsExpected,
 		}
 		r4.MetricName.Tags = []storage.Tag{{
@@ -3353,7 +3604,7 @@ func TestExecSuccess(t *testing.T) {
 		q := `sort(rollup(time()[:50s]))`
 		r1 := netstorage.Result{
 			MetricName: metricNameExpected,
-			Values:     []float64{850, 1050, 1250, 1450, 1650, 1850},
+			Values:     []float64{800, 1000, 1200, 1400, 1600, 1800},
 			Timestamps: timestampsExpected,
 		}
 		r1.MetricName.Tags = []storage.Tag{{
@@ -3723,6 +3974,107 @@ func TestExecSuccess(t *testing.T) {
 		resultExpected := []netstorage.Result{r1, r2, r3, r4, r5, r6}
 		f(q, resultExpected)
 	})
+	t.Run(`count_values by (xxx)`, func(t *testing.T) {
+		t.Parallel()
+		q := `count_values("xxx", label_set(10, "foo", "bar", "xxx", "aaa") or label_set(floor(time()/600), "foo", "bar", "baz", "xx")) by (xxx)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, nan, nan, nan, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("1"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, 1, 1, 1, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("2"),
+			},
+		}
+		r3 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, nan, nan, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r3.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("3"),
+			},
+		}
+		r4 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, 1, 1, 1, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r4.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("10"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2, r3, r4}
+		f(q, resultExpected)
+	})
+	t.Run(`count_values without (baz)`, func(t *testing.T) {
+		t.Parallel()
+		q := `count_values("xxx", label_set(floor(time()/600), "foo", "bar")) without (baz)`
+		r1 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{1, nan, nan, nan, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r1.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("1"),
+			},
+		}
+		r2 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, 1, 1, 1, nan, nan},
+			Timestamps: timestampsExpected,
+		}
+		r2.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("2"),
+			},
+		}
+		r3 := netstorage.Result{
+			MetricName: metricNameExpected,
+			Values:     []float64{nan, nan, nan, nan, 1, 1},
+			Timestamps: timestampsExpected,
+		}
+		r3.MetricName.Tags = []storage.Tag{
+			{
+				Key:   []byte("foo"),
+				Value: []byte("bar"),
+			},
+			{
+				Key:   []byte("xxx"),
+				Value: []byte("3"),
+			},
+		}
+		resultExpected := []netstorage.Result{r1, r2, r3}
+		f(q, resultExpected)
+	})
 }
 
 func TestExecError(t *testing.T) {

app/vmselect/promql/parser.go

@@ -1550,6 +1550,10 @@ type rollupExpr struct {
 	InheritStep bool
 }
 
+func (re *rollupExpr) ForSubquery() bool {
+	return len(re.Step) > 0 || re.InheritStep
+}
+
 func (re *rollupExpr) AppendString(dst []byte) []byte {
 	needParens := func() bool {
 		if _, ok := re.Expr.(*rollupExpr); ok {

app/vmselect/promql/regexp_cache.go

@@ -51,11 +51,14 @@ type regexpCacheValue struct {
 }
 
 type regexpCache struct {
-	m  map[string]*regexpCacheValue
-	mu sync.RWMutex
+	// Move atomic counters to the top of struct for 8-byte alignment on 32-bit arch.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
 
 	requests uint64
 	misses   uint64
+
+	m  map[string]*regexpCacheValue
+	mu sync.RWMutex
 }
 
 func (rc *regexpCache) Requests() uint64 {

app/vmselect/promql/rollup.go

@@ -38,21 +38,23 @@ 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),
+	"sum2_over_time":      newRollupFuncOneArg(rollupSum2),
+	"geomean_over_time":   newRollupFuncOneArg(rollupGeomean),
+	"first_over_time":     newRollupFuncOneArg(rollupFirst),
+	"last_over_time":      newRollupFuncOneArg(rollupLast),
+	"distinct_over_time":  newRollupFuncOneArg(rollupDistinct),
+	"increases_over_time": newRollupFuncOneArg(rollupIncreases),
+	"decreases_over_time": newRollupFuncOneArg(rollupDecreases),
+	"integrate":           newRollupFuncOneArg(rollupIntegrate),
+	"ideriv":              newRollupFuncOneArg(rollupIderiv),
+	"lifetime":            newRollupFuncOneArg(rollupLifetime),
+	"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{
@@ -147,6 +149,9 @@ type rollupConfig struct {
 	MayAdjustWindow bool
 
 	Timestamps []int64
+
+	// LoookbackDelta is the analog to `-query.lookback-delta` from Prometheus world.
+	LookbackDelta int64
 }
 
 var (
@@ -184,6 +189,9 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	dstValues = decimal.ExtendFloat64sCapacity(dstValues, len(rc.Timestamps))
 
 	maxPrevInterval := getMaxPrevInterval(timestamps)
+	if rc.LookbackDelta > 0 && maxPrevInterval > rc.LookbackDelta {
+		maxPrevInterval = rc.LookbackDelta
+	}
 	window := rc.Window
 	if window <= 0 {
 		window = rc.Step
@@ -197,23 +205,21 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 
 	i := 0
 	j := 0
+	ni := 0
+	nj := 0
 	for _, tEnd := range rc.Timestamps {
 		tStart := tEnd - window
-		n := sort.Search(len(timestamps)-i, func(n int) bool {
-			return timestamps[i+n] > tStart
-		})
-		i += n
+		ni = seekFirstTimestampIdxAfter(timestamps[i:], tStart, ni)
+		i += ni
 		if j < i {
 			j = i
 		}
-		n = sort.Search(len(timestamps)-j, func(n int) bool {
-			return timestamps[j+n] > tEnd
-		})
-		j += n
+		nj = seekFirstTimestampIdxAfter(timestamps[j:], tEnd, nj)
+		j += nj
 
 		rfa.prevValue = nan
 		rfa.prevTimestamp = tStart - maxPrevInterval
-		if i > 0 && timestamps[i-1] > rfa.prevTimestamp {
+		if i < len(timestamps) && i > 0 && timestamps[i-1] > rfa.prevTimestamp {
 			rfa.prevValue = values[i-1]
 			rfa.prevTimestamp = timestamps[i-1]
 		}
@@ -229,6 +235,46 @@ func (rc *rollupConfig) Do(dstValues []float64, values []float64, timestamps []i
 	return dstValues
 }
 
+func seekFirstTimestampIdxAfter(timestamps []int64, seekTimestamp int64, nHint int) int {
+	if len(timestamps) == 0 || timestamps[0] > seekTimestamp {
+		return 0
+	}
+	startIdx := nHint - 2
+	if startIdx < 0 {
+		startIdx = 0
+	}
+	if startIdx >= len(timestamps) {
+		startIdx = len(timestamps) - 1
+	}
+	endIdx := nHint + 2
+	if endIdx > len(timestamps) {
+		endIdx = len(timestamps)
+	}
+	if startIdx > 0 && timestamps[startIdx] <= seekTimestamp {
+		timestamps = timestamps[startIdx:]
+		endIdx -= startIdx
+	} else {
+		startIdx = 0
+	}
+	if endIdx < len(timestamps) && timestamps[endIdx] > seekTimestamp {
+		timestamps = timestamps[:endIdx]
+	}
+	if len(timestamps) < 16 {
+		// Fast path: the number of timestamps to search is small, so scan them all.
+		for i, timestamp := range timestamps {
+			if timestamp > seekTimestamp {
+				return startIdx + i
+			}
+		}
+		return startIdx + len(timestamps)
+	}
+	// Slow path: too big len(timestamps), so use binary search.
+	i := sort.Search(len(timestamps), func(n int) bool {
+		return n >= 0 && n < len(timestamps) && timestamps[n] > seekTimestamp
+	})
+	return startIdx + i
+}
+
 func getMaxPrevInterval(timestamps []int64) int64 {
 	if len(timestamps) < 2 {
 		return int64(maxSilenceInterval)
@@ -493,11 +539,14 @@ func rollupAvg(rfa *rollupFuncArg) float64 {
 func rollupMin(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
+	minValue := rfa.prevValue
 	values := rfa.values
-	if len(values) == 0 {
-		return rfa.prevValue
+	if math.IsNaN(minValue) {
+		if len(values) == 0 {
+			return nan
+		}
+		minValue = values[0]
 	}
-	minValue := values[0]
 	for _, v := range values {
 		if v < minValue {
 			minValue = v
@@ -509,11 +558,14 @@ func rollupMin(rfa *rollupFuncArg) float64 {
 func rollupMax(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.
+	maxValue := rfa.prevValue
 	values := rfa.values
-	if len(values) == 0 {
-		return rfa.prevValue
+	if math.IsNaN(maxValue) {
+		if len(values) == 0 {
+			return nan
+		}
+		maxValue = values[0]
 	}
-	maxValue := values[0]
 	for _, v := range values {
 		if v > maxValue {
 			maxValue = v
@@ -527,7 +579,10 @@ func rollupSum(rfa *rollupFuncArg) float64 {
 	// before calling rollup funcs.
 	values := rfa.values
 	if len(values) == 0 {
-		return rfa.prevValue
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
 	}
 	var sum float64
 	for _, v := range values {
@@ -782,6 +837,37 @@ func rollupChanges(rfa *rollupFuncArg) float64 {
 	return float64(n)
 }
 
+func rollupIncreases(rfa *rollupFuncArg) float64 {
+	// There is no need in handling NaNs here, since they must be cleaned up
+	// before calling rollup funcs.
+	values := rfa.values
+	if len(values) == 0 {
+		if math.IsNaN(rfa.prevValue) {
+			return nan
+		}
+		return 0
+	}
+	prevValue := rfa.prevValue
+	if math.IsNaN(prevValue) {
+		prevValue = values[0]
+		values = values[1:]
+	}
+	if len(values) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range values {
+		if v > prevValue {
+			n++
+		}
+		prevValue = v
+	}
+	return float64(n)
+}
+
+// `decreases_over_time` logic is the same as `resets` logic.
+var rollupDecreases = rollupResets
+
 func rollupResets(rfa *rollupFuncArg) float64 {
 	// There is no need in handling NaNs here, since they must be cleaned up
 	// before calling rollup funcs.

app/vmselect/promql/rollup_result_cache.go

@@ -53,8 +53,7 @@ func InitRollupResultCache(cachePath string) {
 		c = workingsetcache.New(cacheSize, time.Hour)
 	}
 	if *disableCache {
-		c.Stop()
-		c = nil
+		c.Reset()
 	}
 
 	stats := &fastcache.Stats{}
@@ -99,10 +98,8 @@ func InitRollupResultCache(cachePath string) {
 // StopRollupResultCache closes the rollupResult cache.
 func StopRollupResultCache() {
 	if len(rollupResultCachePath) == 0 {
-		if !*disableCache {
-			rollupResultCacheV.c.Stop()
-			rollupResultCacheV.c = nil
-		}
+		rollupResultCacheV.c.Stop()
+		rollupResultCacheV.c = nil
 		return
 	}
 	logger.Infof("saving rollupResult cache to %q...", rollupResultCachePath)

app/vmselect/promql/rollup_test.go

@@ -182,7 +182,8 @@ func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpecte
 				t.Fatalf("unexpected value; got %v; want %v", v, vExpected)
 			}
 		} else {
-			if v != vExpected {
+			eps := math.Abs(v - vExpected)
+			if eps > 1e-14 {
 				t.Fatalf("unexpected value; got %v; want %v", v, vExpected)
 			}
 		}
@@ -293,6 +294,8 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f("integrate", 61.0275)
 	f("distinct_over_time", 8)
 	f("ideriv", 0)
+	f("decreases_over_time", 5)
+	f("increases_over_time", 5)
 }
 
 func TestRollupNewRollupFuncError(t *testing.T) {
@@ -358,7 +361,7 @@ func TestRollupNoWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{2, 0, 0, 0, 0, 0, 0, 0}
+		valuesExpected := []float64{2, 0, 0, 0, nan, nan, nan, nan}
 		timestampsExpected := []int64{120, 124, 128, 132, 136, 140, 144, 148}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -389,7 +392,7 @@ func TestRollupWindowNoPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{34, 34, 34, nan}
+		valuesExpected := []float64{nan, nan, nan, nan}
 		timestampsExpected := []int64{161, 171, 181, 191}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -420,7 +423,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{12, 44, 34, 34}
+		valuesExpected := []float64{12, 44, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -465,7 +468,7 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{44, 34, 34, 34}
+		valuesExpected := []float64{44, 34, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -479,12 +482,57 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 54, 44, 34}
+		valuesExpected := []float64{nan, 54, 44, nan}
 		timestampsExpected := []int64{0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
 }
 
+func TestRollupFuncsLookbackDelta(t *testing.T) {
+	t.Run("1", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 1,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, 12, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("7", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 7,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, 12, 44, 44, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("0", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{34, 12, 12, 44, 44, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+}
+
 func TestRollupFuncsNoWindow(t *testing.T) {
 	t.Run("first", func(t *testing.T) {
 		rc := rollupConfig{
@@ -524,7 +572,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 21, 12, 32, 34}
+		valuesExpected := []float64{nan, 21, 12, 12, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -584,7 +632,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{10, 50, 90, 130}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("lifetime", func(t *testing.T) {
+	t.Run("lifetime_1", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupLifetime,
 			Start:  0,
@@ -598,7 +646,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("lifetime", func(t *testing.T) {
+	t.Run("lifetime_2", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupLifetime,
 			Start:  0,
@@ -612,7 +660,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("scrape_interval", func(t *testing.T) {
+	t.Run("scrape_interval_1", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupScrapeInterval,
 			Start:  0,
@@ -626,7 +674,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("scrape_interval", func(t *testing.T) {
+	t.Run("scrape_interval_2", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupScrapeInterval,
 			Start:  0,
@@ -752,7 +800,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
-	t.Run("distinct", func(t *testing.T) {
+	t.Run("distinct_over_time_1", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupDistinct,
 			Start:  0,
@@ -766,6 +814,20 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("distinct_over_time_2", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupDistinct,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 80,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 4, 7, 6, 3}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 }
 
 func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExpected []float64, timestampsExpected []int64) {
@@ -796,7 +858,7 @@ func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExp
 			}
 			continue
 		}
-		if v != vExpected {
+		if math.Abs(v-vExpected) > 1e-15 {
 			t.Fatalf("unexpected value at values[%d]; got %f; want %f\nvalues=\n%v\nvaluesExpected=\n%v",
 				i, v, vExpected, values, valuesExpected)
 		}

app/vmselect/promql/transform.go

@@ -63,6 +63,7 @@ var transformFuncs = map[string]transformFunc{
 	"label_copy":         transformLabelCopy,
 	"label_move":         transformLabelMove,
 	"label_transform":    transformLabelTransform,
+	"label_value":        transformLabelValue,
 	"union":              transformUnion,
 	"":                   transformUnion, // empty func is a synonim to union
 	"keep_last_value":    transformKeepLastValue,
@@ -308,8 +309,16 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 	bbPool.Put(bb)
 
 	// Calculate quantile for each group in m
-	lastNonInf := func(xss []x) float64 {
-		for len(xss) > 0 && math.IsInf(xss[len(xss)-1].le, 0) {
+
+	lastNonInf := func(i int, xss []x) float64 {
+		for len(xss) > 0 {
+			xsLast := xss[len(xss)-1]
+			if xsLast.ts.Values[i] == 0 {
+				return nan
+			}
+			if !math.IsInf(xsLast.le, 0) {
+				break
+			}
 			xss = xss[:len(xss)-1]
 		}
 		if len(xss) == 0 {
@@ -318,27 +327,38 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 		return xss[len(xss)-1].le
 	}
 	quantile := func(i int, phis []float64, xss []x) float64 {
-		vPrev := float64(0)
-		lePrev := float64(0)
 		phi := phis[i]
 		if math.IsNaN(phi) {
 			return nan
 		}
-		// Verify for broken buckets with NaN or negative values.
+		// Fix broken buckets.
+		// They are already sorted by le, so their values must be in ascending order,
+		// since the next bucket value includes all the previous buckets.
+		vPrev := float64(0)
 		for _, xs := range xss {
 			v := xs.ts.Values[i]
-			if math.IsNaN(v) || v < 0 {
-				// Broken bucket.
-				return nan
+			if math.IsNaN(v) || v < vPrev {
+				xs.ts.Values[i] = vPrev
+			} else {
+				vPrev = v
 			}
 		}
+		if len(xss) == 0 {
+			return nan
+		}
 		if phi < 0 {
 			return -inf
 		}
 		if phi > 1 {
 			return inf
 		}
-		vReq := xss[len(xss)-1].ts.Values[i] * phi
+		vLast := xss[len(xss)-1].ts.Values[i]
+		if vLast == 0 {
+			return nan
+		}
+		vReq := vLast * phi
+		vPrev = 0
+		lePrev := float64(0)
 		for _, xs := range xss {
 			v := xs.ts.Values[i]
 			le := xs.le
@@ -348,16 +368,16 @@ func transformHistogramQuantile(tfa *transformFuncArg) ([]*timeseries, error) {
 				continue
 			}
 			if math.IsInf(le, 0) {
-				return lastNonInf(xss)
+				return lastNonInf(i, xss)
 			}
 			if v == vPrev {
 				return lePrev
 			}
 			return lePrev + (le-lePrev)*(vReq-vPrev)/(v-vPrev)
 		}
-		return lastNonInf(xss)
+		return lastNonInf(i, xss)
 	}
-	var rvs []*timeseries
+	rvs := make([]*timeseries, 0, len(m))
 	for _, xss := range m {
 		sort.Slice(xss, func(i, j int) bool {
 			return xss[i].le < xss[j].le
@@ -881,6 +901,33 @@ func labelReplace(tss []*timeseries, srcLabel string, r *regexp.Regexp, dstLabel
 	return tss, nil
 }
 
+func transformLabelValue(tfa *transformFuncArg) ([]*timeseries, error) {
+	args := tfa.args
+	if err := expectTransformArgsNum(args, 2); err != nil {
+		return nil, err
+	}
+	labelName, err := getString(args[1], 1)
+	if err != nil {
+		return nil, fmt.Errorf("cannot get label name: %s", err)
+	}
+	rvs := args[0]
+	for _, ts := range rvs {
+		ts.MetricName.ResetMetricGroup()
+		labelValue := ts.MetricName.GetTagValue(labelName)
+		v, err := strconv.ParseFloat(string(labelValue), 64)
+		if err != nil {
+			v = nan
+		}
+		values := ts.Values
+		for i := range values {
+			values[i] = v
+		}
+	}
+	// Do not remove timeseries with only NaN values, so `default` could be applied to them:
+	// label_value(q, "label") default 123
+	return rvs, nil
+}
+
 func transformLn(v float64) float64 {
 	return math.Log(v)
 }
@@ -1074,7 +1121,10 @@ func transformTimestamp(tfa *transformFuncArg) ([]*timeseries, error) {
 		ts.MetricName.ResetMetricGroup()
 		values := ts.Values
 		for i, t := range ts.Timestamps {
-			values[i] = float64(t) / 1e3
+			v := values[i]
+			if !math.IsNaN(v) {
+				values[i] = float64(t) / 1e3
+			}
 		}
 	}
 	return rvs, nil

app/vmstorage/main.go

@@ -24,6 +24,9 @@ var (
 
 	// DataPath is a path to storage data.
 	DataPath = flag.String("storageDataPath", "victoria-metrics-data", "Path to storage data")
+
+	bigMergeConcurrency   = flag.Int("bigMergeConcurrency", 0, "The maximum number of CPU cores to use for big merges. Default value is used if set to 0")
+	smallMergeConcurrency = flag.Int("smallMergeConcurrency", 0, "The maximum number of CPU cores to use for small merges. Default value is used if set to 0")
 )
 
 // Init initializes vmstorage.
@@ -39,6 +42,10 @@ func InitWithoutMetrics() {
 	if err := encoding.CheckPrecisionBits(uint8(*precisionBits)); err != nil {
 		logger.Fatalf("invalid `-precisionBits`: %s", err)
 	}
+
+	storage.SetBigMergeWorkersCount(*bigMergeConcurrency)
+	storage.SetSmallMergeWorkersCount(*smallMergeConcurrency)
+
 	logger.Infof("opening storage at %q with retention period %d months", *DataPath, *retentionPeriod)
 	startTime := time.Now()
 	WG = syncwg.WaitGroup{}

dashboards/victoriametrics.json

@@ -14,7 +14,7 @@
       "type": "grafana",
       "id": "grafana",
       "name": "Grafana",
-      "version": "6.2.1"
+      "version": "6.3.5"
     },
     {
       "type": "panel",
@@ -60,12 +60,12 @@
       }
     ]
   },
-  "description": "Overview for single node VictoriaMetrics v1.22.2 or higher",
+  "description": "Overview for single node VictoriaMetrics v1.28.0 or higher",
   "editable": true,
   "gnetId": 10229,
   "graphTooltip": 0,
   "id": null,
-  "iteration": 1563651131627,
+  "iteration": 1572208904768,
   "links": [
     {
       "icon": "doc",
@@ -121,7 +121,6 @@
         {
           "targetBlank": true,
           "title": "VictoriaMetrics releases",
-          "type": "absolute",
           "url": "https://github.com/VictoriaMetrics/VictoriaMetrics/releases"
         }
       ],
@@ -490,6 +489,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "* `*` - unsupported query path\n* `/write` - insert into VM\n* `/metrics` - query VM system metrics\n* `/query` - query instant values\n* `/query_range` - query over a range of time\n* `/series` - match a certain label set\n* `/label/{}/values` - query a list of label values (variables mostly)",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -513,7 +513,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null as zero",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -580,6 +582,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "The less time it takes is better.\n* `*` - 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,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -603,7 +606,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null as zero",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -670,6 +675,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "Shows the number of active time series with new data points inserted during the last hour. High value may result in ingestion slowdown. \n\nSee following link for details:",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -692,12 +698,13 @@
         {
           "targetBlank": true,
           "title": "troubleshooting",
-          "type": "absolute",
           "url": "https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#troubleshooting"
         }
       ],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -764,6 +771,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited with -`memory.allowedPercent` flag. Line `max allowed` shows max allowed memory size for cache.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -784,7 +792,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -865,6 +875,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "* `*` - unsupported query path\n* `/write` - insert into VM\n* `/metrics` - query VM system metrics\n* `/query` - query instant values\n* `/query_range` - query over a range of time\n* `/series` - match a certain label set\n* `/label/{}/values` - query a list of label values (variables mostly)",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -888,7 +899,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null as zero",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -953,55 +966,74 @@
       "dashLength": 10,
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
-      "description": "",
+      "description": "Shows how many ongoing insertions are taking place.\n* `max` - equal to number of CPU * 2\n* `current` - current number of goroutines busy with inserting rows into storage\n\nWhen `current` hits `max` constantly, it means storage is overloaded and require more CPU.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 12,
         "y": 27
       },
-      "id": 49,
+      "id": 59,
       "legend": {
-        "avg": false,
-        "current": false,
+        "alignAsTable": true,
+        "avg": true,
+        "current": true,
+        "hideEmpty": false,
+        "hideZero": false,
         "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",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
       "renderer": "flot",
-      "seriesOverrides": [],
+      "seriesOverrides": [
+        {
+          "alias": "max",
+          "color": "#C4162A"
+        }
+      ],
       "spaceLength": 10,
       "stack": false,
       "steppedLine": false,
       "targets": [
         {
-          "expr": "sum(rate(vm_tcplistener_accepts_total{job=\"$job\"}[$__interval]))",
+          "expr": "sum(vm_concurrent_addrows_capacity{job=\"$job\"})",
           "format": "time_series",
-          "hide": false,
           "intervalFactor": 1,
-          "legendFormat": "connections",
+          "legendFormat": "max",
           "refId": "A"
+        },
+        {
+          "expr": "sum(vm_concurrent_addrows_current{job=\"$job\"})",
+          "format": "time_series",
+          "intervalFactor": 1,
+          "legendFormat": "current",
+          "refId": "B"
         }
       ],
       "thresholds": [],
       "timeFrom": null,
       "timeRegions": [],
       "timeShift": null,
-      "title": "TCP connections rate",
+      "title": "Concurrent inserts",
       "tooltip": {
         "shared": true,
-        "sort": 0,
+        "sort": 2,
         "value_type": "individual"
       },
       "type": "graph",
@@ -1014,7 +1046,7 @@
       },
       "yaxes": [
         {
-          "decimals": null,
+          "decimals": 0,
           "format": "short",
           "label": null,
           "logBase": 1,
@@ -1023,6 +1055,7 @@
           "show": true
         },
         {
+          "decimals": 0,
           "format": "short",
           "label": null,
           "logBase": 1,
@@ -1044,6 +1077,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1064,7 +1098,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1125,6 +1161,98 @@
         "alignLevel": null
       }
     },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "description": "",
+      "fill": 1,
+      "fillGradient": 0,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 12,
+        "y": 35
+      },
+      "id": 49,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": false,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {
+        "dataLinks": []
+      },
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(rate(vm_tcplistener_accepts_total{job=\"$job\"}[$__interval]))",
+          "format": "time_series",
+          "hide": false,
+          "intervalFactor": 1,
+          "legendFormat": "connections",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "TCP connections rate",
+      "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": {
@@ -1146,6 +1274,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "How many datapoints are inserted into storage per second",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1168,7 +1297,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null as zero",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1242,6 +1373,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "How many datapoints are in RAM queue waiting to be written into storage. The less is better.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1262,7 +1394,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1344,6 +1478,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "How many datapoints are in the storage.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1364,7 +1499,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1431,6 +1568,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "Data parts of LSM tree.\nHigh number of parts could be an evidence of slow merge performance - check the resource utilization.\n* `indexdb` - inverted index\n* `storage/small` - recently added parts of data ingested into storage(hot data)\n* `storage/big` -  small parts gradually merged into big parts (cold data)",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1451,7 +1589,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1518,6 +1658,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "Shows amount of on-disk space occupied by data points.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1538,7 +1679,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1605,6 +1748,7 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "Shows amount of on-disk space occupied by inverted index.",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
@@ -1625,7 +1769,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1683,13 +1829,105 @@
         "alignLevel": null
       }
     },
+    {
+      "aliasColors": {},
+      "bars": false,
+      "dashLength": 10,
+      "dashes": false,
+      "datasource": "${DS_PROMETHEUS}",
+      "description": "Shows how many rows were ignored on insertion due to corrupted or out of retention timestamps.",
+      "fill": 1,
+      "fillGradient": 0,
+      "gridPos": {
+        "h": 8,
+        "w": 12,
+        "x": 0,
+        "y": 68
+      },
+      "id": 58,
+      "legend": {
+        "avg": false,
+        "current": false,
+        "max": false,
+        "min": false,
+        "show": false,
+        "total": false,
+        "values": false
+      },
+      "lines": true,
+      "linewidth": 1,
+      "links": [],
+      "nullPointMode": "null",
+      "options": {
+        "dataLinks": []
+      },
+      "percentage": false,
+      "pointradius": 2,
+      "points": false,
+      "renderer": "flot",
+      "seriesOverrides": [],
+      "spaceLength": 10,
+      "stack": false,
+      "steppedLine": false,
+      "targets": [
+        {
+          "expr": "sum(vm_rows_ignored_total{job=\"$job\"}) by (reason) > 0",
+          "format": "time_series",
+          "hide": false,
+          "intervalFactor": 1,
+          "legendFormat": "{{reason}}",
+          "refId": "A"
+        }
+      ],
+      "thresholds": [],
+      "timeFrom": null,
+      "timeRegions": [],
+      "timeShift": null,
+      "title": "Rows ignored",
+      "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": 68
+        "y": 76
       },
       "id": 46,
       "panels": [],
@@ -1704,11 +1942,12 @@
       "datasource": "${DS_PROMETHEUS}",
       "description": "",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 69
+        "y": 77
       },
       "id": 44,
       "legend": {
@@ -1724,7 +1963,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1806,11 +2047,12 @@
       "dashLength": 10,
       "dashes": false,
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 69
+        "y": 77
       },
       "id": 57,
       "legend": {
@@ -1826,7 +2068,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1892,11 +2136,12 @@
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 77
+        "y": 85
       },
       "id": 47,
       "legend": {
@@ -1912,7 +2157,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -1979,11 +2226,12 @@
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 77
+        "y": 85
       },
       "id": 42,
       "legend": {
@@ -1999,7 +2247,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -2065,11 +2315,12 @@
       "dashes": false,
       "datasource": "${DS_PROMETHEUS}",
       "fill": 1,
+      "fillGradient": 0,
       "gridPos": {
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 85
+        "y": 93
       },
       "id": 48,
       "legend": {
@@ -2085,7 +2336,9 @@
       "linewidth": 1,
       "links": [],
       "nullPointMode": "null",
-      "options": {},
+      "options": {
+        "dataLinks": []
+      },
       "percentage": false,
       "pointradius": 2,
       "points": false,
@@ -2147,7 +2400,7 @@
     }
   ],
   "refresh": "30s",
-  "schemaVersion": 18,
+  "schemaVersion": 19,
   "style": "dark",
   "tags": [],
   "templating": {
@@ -2230,5 +2483,5 @@
   "timezone": "",
   "title": "VictoriaMetrics",
   "uid": "wNf0q_kZk",
-  "version": 2
-}
+  "version": 3
+}

deployment/docker/Makefile

@@ -1,5 +1,5 @@
 DOCKER_NAMESPACE := victoriametrics
-BUILDER_IMAGE := local/builder:go1.12.9
+BUILDER_IMAGE := local/builder:go1.13.3
 CERTS_IMAGE := local/certs:1.0.2
 
 package-certs:
@@ -21,7 +21,8 @@ app-via-docker: package-certs package-builder
 		--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)$(APP_SUFFIX)-prod $(PKG_PREFIX)/app/$(APP_NAME)
+		go build $(RACE) -mod=vendor -trimpath -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:1.12.9
+FROM golang:1.13.3
 STOPSIGNAL SIGINT

deployment/docker/docker-compose.yml

@@ -2,7 +2,7 @@ version: '3.5'
 services:
   prometheus:
     container_name: prometheus
-    image: prom/prometheus:v2.10.0
+    image: prom/prometheus:v2.12.0
     depends_on:
       - "victoriametrics"
     ports:
@@ -35,7 +35,7 @@ services:
     restart: always
   grafana:
     container_name: grafana
-    image: grafana/grafana:6.2.1
+    image: grafana/grafana:6.3.5
     entrypoint: >
       /bin/sh -c "
       cd /var/lib/grafana &&

go.mod

@@ -1,18 +1,18 @@
 module github.com/VictoriaMetrics/VictoriaMetrics
 
 require (
-	github.com/VictoriaMetrics/fastcache v1.5.1
-	github.com/VictoriaMetrics/metrics v1.7.1
-	github.com/cespare/xxhash/v2 v2.0.1-0.20190104013014-3767db7a7e18
+	github.com/VictoriaMetrics/fastcache v1.5.2
+	github.com/VictoriaMetrics/metrics v1.7.2
+	github.com/cespare/xxhash/v2 v2.1.0
 	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/klauspost/compress v1.9.1
 	github.com/valyala/fastjson v1.4.1
-	github.com/valyala/gozstd v1.6.0
+	github.com/valyala/fastrand v1.0.0
+	github.com/valyala/gozstd v1.6.2
 	github.com/valyala/histogram v1.0.1
-	github.com/valyala/quicktemplate v1.1.1
-	golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa
+	github.com/valyala/quicktemplate v1.3.1
+	golang.org/x/sys v0.0.0-20191027211539-f8518d3b3627
 )
 
 go 1.12

go.sum

@@ -1,16 +1,18 @@
 github.com/OneOfOne/xxhash v1.2.2/go.mod h1:HSdplMjZKSmBqAxg5vPj2TmRDmfkzw+cTzAElWljhcU=
 github.com/OneOfOne/xxhash v1.2.5 h1:zl/OfRA6nftbBK9qTohYBJ5xvw6C/oNKizR7cZGl3cI=
 github.com/OneOfOne/xxhash v1.2.5/go.mod h1:eZbhyaAYD41SGSSsnmcpxVoRiQ/MPUTjUdIIOT9Um7Q=
-github.com/VictoriaMetrics/fastcache v1.5.1 h1:qHgHjyoNFV7jgucU8QZUuU4gcdhfs8QW1kw68OD2Lag=
-github.com/VictoriaMetrics/fastcache v1.5.1/go.mod h1:+jv9Ckb+za/P1ZRg/sulP5Ni1v49daAVERr0H3CuscE=
-github.com/VictoriaMetrics/metrics v1.7.1 h1:g2qrY6Upn8rvlvR40cGHFY0crwi4hpqF0n9vJMNsCSg=
-github.com/VictoriaMetrics/metrics v1.7.1/go.mod h1:LU2j9qq7xqZYXz8tF3/RQnB2z2MbZms5TDiIg9/NHiQ=
+github.com/VictoriaMetrics/fastcache v1.5.2 h1:Erd8iIuBAL9kke8JzM4+WxkKuFkHh3ktwLanJvDgR44=
+github.com/VictoriaMetrics/fastcache v1.5.2/go.mod h1:+jv9Ckb+za/P1ZRg/sulP5Ni1v49daAVERr0H3CuscE=
+github.com/VictoriaMetrics/metrics v1.7.2 h1:PzC0SEo5lbbNK7xaYwclCCdoaIGRmXOfflIMF3LpSW4=
+github.com/VictoriaMetrics/metrics v1.7.2/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=
 github.com/cespare/xxhash v1.1.0/go.mod h1:XrSqR1VqqWfGrhpAt58auRo0WTKS1nRRg3ghfAqPWnc=
 github.com/cespare/xxhash/v2 v2.0.1-0.20190104013014-3767db7a7e18 h1:pl4eWIqvFe/Kg3zkn7NxevNzILnZYWDCG7qbA1CJik0=
 github.com/cespare/xxhash/v2 v2.0.1-0.20190104013014-3767db7a7e18/go.mod h1:HD5P3vAIAh+Y2GAxg0PrPN1P8WkepXGpjbUPDHJqqKM=
+github.com/cespare/xxhash/v2 v2.1.0 h1:yTUvW7Vhb89inJ+8irsUqiWjh8iT6sQPZiQzI6ReGkA=
+github.com/cespare/xxhash/v2 v2.1.0/go.mod h1:dgIUBU3pDso/gPgZ1osOZ0iQf77oPR28Tjxl5dIMyVM=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 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=
@@ -20,8 +22,8 @@ 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/compress v1.9.1 h1:TWy0o9J9c6LK9C8t7Msh6IAJNXbsU/nvKLTQUU5HdaY=
+github.com/klauspost/compress v1.9.1/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=
@@ -29,8 +31,6 @@ github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZb
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/spaolacci/murmur3 v0.0.0-20180118202830-f09979ecbc72/go.mod h1:JwIasOWyU6f++ZhiEuf87xNszmSA2myDM2Kzu9HwQUA=
 github.com/spaolacci/murmur3 v1.0.1-0.20190317074736-539464a789e9/go.mod h1:JwIasOWyU6f++ZhiEuf87xNszmSA2myDM2Kzu9HwQUA=
-github.com/spaolacci/murmur3 v1.1.0 h1:7c1g84S4BPRrfL5Xrdp6fOJ206sU9y293DDHaoy0bLI=
-github.com/spaolacci/murmur3 v1.1.0/go.mod h1:JwIasOWyU6f++ZhiEuf87xNszmSA2myDM2Kzu9HwQUA=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/testify v1.3.0 h1:TivCn/peBQ7UY8ooIcPgZFpTNSz0Q2U6UrFlUfqbe0Q=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
@@ -41,13 +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 v1.6.0 h1:34qKK75C6Dx9zof2JqUiunfJQ87Up6vTHXABWDyCH+g=
-github.com/valyala/gozstd v1.6.0/go.mod h1:y5Ew47GLlP37EkTB+B4s7r6A5rdaeB7ftbl9zoYiIPQ=
+github.com/valyala/gozstd v1.6.2 h1:MgBfNm0I8IKm51LUTTKfO9vi4BtmoH7kBXeUvgaiZVU=
+github.com/valyala/gozstd v1.6.2/go.mod h1:y5Ew47GLlP37EkTB+B4s7r6A5rdaeB7ftbl9zoYiIPQ=
 github.com/valyala/histogram v1.0.1 h1:FzA7n2Tz/wKRMejgu3PV1vw3htAklTjjuoI6z3d4KDg=
 github.com/valyala/histogram v1.0.1/go.mod h1:lQy0xA4wUz2+IUnf97SivorsJIp8FxsnRd6x25q7Mto=
-github.com/valyala/quicktemplate v1.1.1 h1:C58y/wN0FMTi2PR0n3onltemfFabany53j7M6SDDB8k=
-github.com/valyala/quicktemplate v1.1.1/go.mod h1:EH+4AkTd43SvgIbQHYu59/cJyxDoOVRUAfrukLPuGJ4=
+github.com/valyala/quicktemplate v1.3.1 h1:V9Ixd/ONuoT6C1ipx8XR2dNGSDgIVnvT4ezZ38ZWllU=
+github.com/valyala/quicktemplate v1.3.1/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 v0.0.0-20190804053845-51ab0e2deafa h1:KIDDMLT1O0Nr7TSxp8xM5tJcdn8tgyAONntO829og1M=
-golang.org/x/sys v0.0.0-20190804053845-51ab0e2deafa/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20191027211539-f8518d3b3627 h1:/FZUR3d/QsXe4AcJyJFCc40TOj3y6Hs23Y3YJlvVkWo=
+golang.org/x/sys v0.0.0-20191027211539-f8518d3b3627/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

lib/decimal/decimal.go

@@ -265,61 +265,83 @@ var (
 // For instance, for f = -1.234 it returns v = -1234, e = -3.
 //
 // FromFloat doesn't work properly with NaN values, so don't pass them here.
-func FromFloat(f float64) (v int64, e int16) {
-	if math.IsInf(f, 0) {
-		// Special case for Inf
-		if math.IsInf(f, 1) {
-			return vInfPos, 0
-		}
-		return vInfNeg, 0
-	}
-
-	minus := false
-	if f < 0 {
-		f = -f
-		minus = true
-	}
+func FromFloat(f float64) (int64, int16) {
 	if f == 0 {
-		// Special case for 0.0 and -0.0
 		return 0, 0
 	}
-	v, e = positiveFloatToDecimal(f)
-	if minus {
-		v = -v
+	if math.IsInf(f, 0) {
+		return fromFloatInf(f)
 	}
-	if v == 0 {
-		e = 0
-	} else if v > vMax {
-		v = vMax
-	} else if v < vMin {
+	if f > 0 {
+		v, e := positiveFloatToDecimal(f)
+		if v > vMax {
+			v = vMax
+		}
+		return v, e
+	}
+	v, e := positiveFloatToDecimal(-f)
+	v = -v
+	if v < vMin {
 		v = vMin
 	}
 	return v, e
 }
 
+func fromFloatInf(f float64) (int64, int16) {
+	// Special case for Inf
+	if math.IsInf(f, 1) {
+		return vInfPos, 0
+	}
+	return vInfNeg, 0
+}
+
 func positiveFloatToDecimal(f float64) (int64, int16) {
+	// There is no need in checking for f == 0, since it should be already checked by the caller.
+	u := uint64(f)
+	if float64(u) != f {
+		return positiveFloatToDecimalSlow(f)
+	}
+	// Fast path for integers.
+	if u < 1<<55 && u%10 != 0 {
+		return int64(u), 0
+	}
+	return getDecimalAndScale(u)
+}
+
+func getDecimalAndScale(u uint64) (int64, int16) {
 	var scale int16
-	v := int64(f)
-	if f == float64(v) {
-		// Fast path for integers.
-		u := uint64(v)
-		if u%10 != 0 {
-			return v, 0
-		}
-		// Minimize v by converting trailing zeros to scale.
+	for u >= 1<<55 {
+		// Remove trailing garbage bits left after float64->uint64 conversion,
+		// since float64 contains only 53 significant bits.
+		// See https://en.wikipedia.org/wiki/Double-precision_floating-point_format
 		u /= 10
 		scale++
-		for u != 0 && u%10 == 0 {
-			u /= 10
-			scale++
-		}
+	}
+	if u%10 != 0 {
 		return int64(u), scale
 	}
+	// Minimize v by converting trailing zeros to scale.
+	u /= 10
+	scale++
+	for u != 0 && u%10 == 0 {
+		u /= 10
+		scale++
+	}
+	return int64(u), scale
+}
 
+func positiveFloatToDecimalSlow(f float64) (int64, int16) {
 	// Slow path for floating point numbers.
+	var scale int16
+	prec := conversionPrecision
 	if f > 1e6 || f < 1e-6 {
 		// Normalize f, so it is in the small range suitable
 		// for the next loop.
+		if f > 1e6 {
+			// Increase conversion precision for big numbers.
+			// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/213
+			prec = 1e15
+		}
 		_, exp := math.Frexp(f)
 		scale = int16(float64(exp) * math.Ln2 / math.Ln10)
 		f *= math.Pow10(-int(scale))
@@ -327,13 +349,13 @@ func positiveFloatToDecimal(f float64) (int64, int16) {
 
 	// Multiply f by 100 until the fractional part becomes
 	// too small comparing to integer part.
-	for f < conversionPrecision {
+	for f < prec {
 		x, frac := math.Modf(f)
-		if frac*conversionPrecision < x {
+		if frac*prec < x {
 			f = x
 			break
 		}
-		if (1-frac)*conversionPrecision < x {
+		if (1-frac)*prec < x {
 			f = x + 1
 			break
 		}

lib/decimal/decimal_test.go

@@ -7,6 +7,44 @@ import (
 	"testing"
 )
 
+func TestPositiveFloatToDecimal(t *testing.T) {
+	f := func(f float64, decimalExpected int64, exponentExpected int16) {
+		t.Helper()
+		decimal, exponent := positiveFloatToDecimal(f)
+		if decimal != decimalExpected {
+			t.Fatalf("unexpected decimal for positiveFloatToDecimal(%f); got %d; want %d", f, decimal, decimalExpected)
+		}
+		if exponent != exponentExpected {
+			t.Fatalf("unexpected exponent for positiveFloatToDecimal(%f); got %d; want %d", f, exponent, exponentExpected)
+		}
+	}
+	f(0, 0, 1) // The exponent is 1 is OK here. See comment in positiveFloatToDecimal.
+	f(1, 1, 0)
+	f(30, 3, 1)
+	f(12345678900000000, 123456789, 8)
+	f(12345678901234567, 12345678901234568, 0)
+	f(1234567890123456789, 12345678901234567, 2)
+	f(12345678901234567890, 12345678901234567, 3)
+	f(18446744073670737131, 18446744073670737, 3)
+	f(123456789012345678901, 12345678901234568, 4)
+	f(1<<53, 1<<53, 0)
+	f(1<<54, 18014398509481984, 0)
+	f(1<<55, 3602879701896396, 1)
+	f(1<<62, 4611686018427387, 3)
+	f(1<<63, 9223372036854775, 3)
+	f(1<<64, 18446744073709548, 3)
+	f(1<<65, 368934881474191, 5)
+	f(1<<66, 737869762948382, 5)
+	f(1<<67, 1475739525896764, 5)
+
+	f(0.1, 1, -1)
+	f(123456789012345678e-5, 12345678901234568, -4)
+	f(1234567890123456789e-10, 12345678901234568, -8)
+	f(1234567890123456789e-14, 1234567890123, -8)
+	f(1234567890123456789e-17, 12345678901234, -12)
+	f(1234567890123456789e-20, 1234567890123, -14)
+}
+
 func TestAppendDecimalToFloat(t *testing.T) {
 	testAppendDecimalToFloat(t, []int64{}, 0, nil)
 	testAppendDecimalToFloat(t, []int64{0}, 0, []float64{0})
@@ -168,7 +206,7 @@ func TestAppendFloatToDecimal(t *testing.T) {
 	// no-op
 	testAppendFloatToDecimal(t, []float64{}, nil, 0)
 	testAppendFloatToDecimal(t, []float64{0}, []int64{0}, 0)
-	testAppendFloatToDecimal(t, []float64{0, 1, -1, 12345678, -123456789}, []int64{0, 1, -1, 12345678, -123456789}, 0)
+	testAppendFloatToDecimal(t, []float64{0, -0, 1, -1, 12345678, -123456789}, []int64{0, 0, 1, -1, 12345678, -123456789}, 0)
 
 	// upExp
 	testAppendFloatToDecimal(t, []float64{-24, 0, 4.123, 0.3}, []int64{-24000, 0, 4123, 300}, -3)
@@ -248,8 +286,8 @@ func TestFloatToDecimal(t *testing.T) {
 
 	f(math.Inf(1), vInfPos, 0)
 	f(math.Inf(-1), vInfNeg, 0)
-	f(1<<63-1, 922337203685, 7)
-	f(-1<<63, -922337203685, 7)
+	f(1<<63-1, 9223372036854775, 3)
+	f(-1<<63, -9223372036854775, 3)
 
 	// Test precision loss due to conversionPrecision.
 	f(0.1234567890123456, 12345678901234, -14)

lib/encoding/int.go

@@ -1,6 +1,7 @@
 package encoding
 
 import (
+	"encoding/binary"
 	"fmt"
 	"sync"
 )
@@ -12,8 +13,8 @@ func MarshalUint16(dst []byte, u uint16) []byte {
 
 // UnmarshalUint16 returns unmarshaled uint32 from src.
 func UnmarshalUint16(src []byte) uint16 {
-	_ = src[1]
-	return uint16(src[0])<<8 | uint16(src[1])
+	// This is faster than the manual conversion.
+	return binary.BigEndian.Uint16(src)
 }
 
 // MarshalUint32 appends marshaled v to dst and returns the result.
@@ -23,8 +24,8 @@ func MarshalUint32(dst []byte, u uint32) []byte {
 
 // UnmarshalUint32 returns unmarshaled uint32 from src.
 func UnmarshalUint32(src []byte) uint32 {
-	_ = src[3]
-	return uint32(src[0])<<24 | uint32(src[1])<<16 | uint32(src[2])<<8 | uint32(src[3])
+	// This is faster than the manual conversion.
+	return binary.BigEndian.Uint32(src)
 }
 
 // MarshalUint64 appends marshaled v to dst and returns the result.
@@ -34,8 +35,8 @@ func MarshalUint64(dst []byte, u uint64) []byte {
 
 // UnmarshalUint64 returns unmarshaled uint64 from src.
 func UnmarshalUint64(src []byte) uint64 {
-	_ = src[7]
-	return uint64(src[0])<<56 | uint64(src[1])<<48 | uint64(src[2])<<40 | uint64(src[3])<<32 | uint64(src[4])<<24 | uint64(src[5])<<16 | uint64(src[6])<<8 | uint64(src[7])
+	// This is faster than the manual conversion.
+	return binary.BigEndian.Uint64(src)
 }
 
 // MarshalInt16 appends marshaled v to dst and returns the result.
@@ -48,8 +49,8 @@ func MarshalInt16(dst []byte, v int16) []byte {
 
 // UnmarshalInt16 returns unmarshaled int16 from src.
 func UnmarshalInt16(src []byte) int16 {
-	_ = src[1]
-	u := uint16(src[0])<<8 | uint16(src[1])
+	// This is faster than the manual conversion.
+	u := binary.BigEndian.Uint16(src)
 	v := int16(u>>1) ^ (int16(u<<15) >> 15) // zig-zag decoding without branching.
 	return v
 }
@@ -64,8 +65,8 @@ func MarshalInt64(dst []byte, v int64) []byte {
 
 // UnmarshalInt64 returns unmarshaled int64 from src.
 func UnmarshalInt64(src []byte) int64 {
-	_ = src[7]
-	u := uint64(src[0])<<56 | uint64(src[1])<<48 | uint64(src[2])<<40 | uint64(src[3])<<32 | uint64(src[4])<<24 | uint64(src[5])<<16 | uint64(src[6])<<8 | uint64(src[7])
+	// This is faster than the manual conversion.
+	u := binary.BigEndian.Uint64(src)
 	v := int64(u>>1) ^ (int64(u<<63) >> 63) // zig-zag decoding without branching.
 	return v
 }

lib/encoding/int_timing_test.go

@@ -6,6 +6,33 @@ import (
 	"testing"
 )
 
+func BenchmarkMarshalUint64(b *testing.B) {
+	b.ReportAllocs()
+	b.SetBytes(1)
+	b.RunParallel(func(pb *testing.PB) {
+		var dst []byte
+		var sink uint64
+		for pb.Next() {
+			dst = MarshalUint64(dst[:0], sink)
+			sink += uint64(len(dst))
+		}
+		atomic.AddUint64(&Sink, sink)
+	})
+}
+
+func BenchmarkUnmarshalUint64(b *testing.B) {
+	b.ReportAllocs()
+	b.SetBytes(1)
+	b.RunParallel(func(pb *testing.PB) {
+		var sink uint64
+		for pb.Next() {
+			v := UnmarshalUint64(testMarshaledUint64Data)
+			sink += v
+		}
+		atomic.AddUint64(&Sink, sink)
+	})
+}
+
 func BenchmarkMarshalInt64(b *testing.B) {
 	b.ReportAllocs()
 	b.SetBytes(1)
@@ -120,3 +147,4 @@ func benchmarkUnmarshalVarInt64s(b *testing.B, maxValue int64) {
 }
 
 var testMarshaledInt64Data = MarshalInt64(nil, 1234567890)
+var testMarshaledUint64Data = MarshalUint64(nil, 1234567890)

lib/filestream/filestream_freebsd.go

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

lib/fs/dir_remover.go

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

lib/fs/fs.go

@@ -6,9 +6,7 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
-	"strings"
 	"sync/atomic"
-	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
@@ -94,7 +92,7 @@ var tmpFileNum uint64
 
 // WriteFileAtomically atomically writes data to the given file path.
 //
-// WriteFile returns only after the file is fully written and synced
+// WriteFileAtomically returns only after the file is fully written and synced
 // to the underlying storage.
 func WriteFileAtomically(path string, data []byte) error {
 	// Check for the existing file. It is expected that
@@ -248,62 +246,7 @@ func mustSyncParentDirIfExists(path string) {
 //
 // It properly handles NFS issue https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
 func MustRemoveAll(path string) {
-	err := os.RemoveAll(path)
-	if err == nil {
-		// Make sure the parent directory doesn't contain references
-		// to the current directory.
-		mustSyncParentDirIfExists(path)
-		return
-	}
-	if !isTemporaryNFSError(err) {
-		logger.Panicf("FATAL: cannot remove %q: %s", path, err)
-	}
-	// NFS prevents from removing directories with open files.
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
-	// Schedule for later directory removal.
-	select {
-	case removeDirCh <- path:
-	default:
-		logger.Panicf("FATAL: cannot schedule %s for removal, since the removal queue is full (%d entries)", path, cap(removeDirCh))
-	}
-}
-
-var removeDirCh = make(chan string, 1024)
-
-func dirRemover() {
-	for path := range removeDirCh {
-		attempts := 0
-		for {
-			err := os.RemoveAll(path)
-			if err == nil {
-				break
-			}
-			if !isTemporaryNFSError(err) {
-				logger.Panicf("FATAL: cannot remove %q: %s", path, err)
-			}
-			// NFS prevents from removing directories with open files.
-			// Sleep for a while and try again in the hope open files will be closed.
-			// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 .
-			attempts++
-			if attempts > 10 {
-				logger.Panicf("FATAL: cannot remove %q in %d attempts: %s", path, attempts, err)
-			}
-			time.Sleep(100 * time.Millisecond)
-		}
-		// Make sure the parent directory doesn't contain references
-		// to the current directory.
-		mustSyncParentDirIfExists(path)
-	}
-}
-
-func isTemporaryNFSError(err error) bool {
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/61 for details.
-	errStr := err.Error()
-	return strings.Contains(errStr, "directory not empty") || strings.Contains(errStr, "device or resource busy")
-}
-
-func init() {
-	go dirRemover()
+	_ = mustRemoveAll(path)
 }
 
 // HardLinkFiles makes hard links for all the files from srcDir in dstDir.
@@ -400,3 +343,20 @@ func CreateFlockFile(dir string) (*os.File, error) {
 	}
 	return flockF, nil
 }
+
+// MustGetFreeSpace returns free space for the given directory path.
+func MustGetFreeSpace(path string) uint64 {
+	d, err := os.Open(path)
+	if err != nil {
+		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", path, err)
+	}
+	defer MustClose(d)
+
+	fd := d.Fd()
+	var stat unix.Statfs_t
+	if err := unix.Fstatfs(int(fd), &stat); err != nil {
+		logger.Panicf("FATAL: cannot determine free disk space on %q: %s", path, err)
+	}
+	freeSpace := uint64(stat.Bavail) * uint64(stat.Bsize)
+	return freeSpace
+}

lib/httpserver/httpserver.go

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

lib/memory/memory.go

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

lib/memory/memory_bsd.go

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

lib/memory/memory_darwin.go

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

lib/memory/memory_linux.go

@@ -9,12 +9,17 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
+const maxInt = int(^uint(0) >> 1)
+
 func sysTotalMemory() int {
 	var si syscall.Sysinfo_t
 	if err := syscall.Sysinfo(&si); err != nil {
 		logger.Panicf("FATAL: error in syscall.Sysinfo: %s", err)
 	}
-	totalMem := int(si.Totalram) * int(si.Unit)
+	totalMem := maxInt
+	if uint64(maxInt)/uint64(si.Totalram) > uint64(si.Unit) {
+		totalMem = int(uint64(si.Totalram) * uint64(si.Unit))
+	}
 
 	// Try determining the amount of memory inside docker container.
 	// See https://stackoverflow.com/questions/42187085/check-mem-limit-within-a-docker-container .

lib/memory/sysctl.go

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

lib/mergeset/block_header.go

@@ -158,7 +158,7 @@ func unmarshalBlockHeaders(dst []blockHeader, src []byte, blockHeadersCount int)
 	newBHS := dst[dstLen:]
 
 	// Verify that block headers are sorted by firstItem.
-	if !sort.SliceIsSorted(newBHS, func(i, j int) bool { return string(newBHS[i].firstItem) <= string(newBHS[j].firstItem) }) {
+	if !sort.SliceIsSorted(newBHS, func(i, j int) bool { return string(newBHS[i].firstItem) < string(newBHS[j].firstItem) }) {
 		return nil, fmt.Errorf("block headers must be sorted by firstItem; unmarshaled unsorted block headers: %#v", newBHS)
 	}
 

lib/mergeset/encoding.go

@@ -2,6 +2,7 @@ package mergeset
 
 import (
 	"fmt"
+	"os"
 	"sort"
 	"strings"
 	"sync"
@@ -11,10 +12,20 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
+type byteSliceSorter [][]byte
+
+func (s byteSliceSorter) Len() int { return len(s) }
+func (s byteSliceSorter) Less(i, j int) bool {
+	return string(s[i]) < string(s[j])
+}
+func (s byteSliceSorter) Swap(i, j int) {
+	s[i], s[j] = s[j], s[i]
+}
+
 type inmemoryBlock struct {
 	commonPrefix []byte
 	data         []byte
-	items        [][]byte
+	items        byteSliceSorter
 }
 
 func (ib *inmemoryBlock) Reset() {
@@ -77,12 +88,9 @@ func (ib *inmemoryBlock) Add(x []byte) bool {
 // It must fit CPU cache size, i.e. 64KB for the current CPUs.
 const maxInmemoryBlockSize = 64 * 1024
 
-func (ib *inmemoryBlock) itemsLess(i, j int) bool {
-	return string(ib.items[i]) < string(ib.items[j])
-}
-
 func (ib *inmemoryBlock) sort() {
-	sort.Slice(ib.items, ib.itemsLess)
+	// Use sort.Sort instead of sort.Slice in order to eliminate memory allocation.
+	sort.Sort(&ib.items)
 	bb := bbPool.Get()
 	b := bytesutil.Resize(bb.B, len(ib.data))
 	b = b[:0]
@@ -120,7 +128,8 @@ func checkMarshalType(mt marshalType) error {
 }
 
 func (ib *inmemoryBlock) isSorted() bool {
-	return sort.SliceIsSorted(ib.items, ib.itemsLess)
+	// Use sort.IsSorted instead of sort.SliceIsSorted in order to eliminate memory allocation.
+	return sort.IsSorted(&ib.items)
 }
 
 // MarshalUnsortedData marshals unsorted items from ib to sb.
@@ -138,6 +147,10 @@ func (ib *inmemoryBlock) MarshalUnsortedData(sb *storageBlock, firstItemDst, com
 	return ib.marshalData(sb, firstItemDst, commonPrefixDst, compressLevel)
 }
 
+var isInTest = func() bool {
+	return strings.HasSuffix(os.Args[0], ".test")
+}()
+
 // MarshalUnsortedData marshals sorted items from ib to sb.
 //
 // It also:
@@ -146,17 +159,22 @@ func (ib *inmemoryBlock) MarshalUnsortedData(sb *storageBlock, firstItemDst, com
 // - returns the number of items encoded including the first item.
 // - returns the marshal type used for the encoding.
 func (ib *inmemoryBlock) MarshalSortedData(sb *storageBlock, firstItemDst, commonPrefixDst []byte, compressLevel int) ([]byte, []byte, uint32, marshalType) {
-	// if !ib.isSorted() {
-	//	logger.Panicf("BUG: %d items must be sorted; items:\n%s", len(ib.items), ib.debugItemsString())
-	// }
+	if isInTest && !ib.isSorted() {
+		logger.Panicf("BUG: %d items must be sorted; items:\n%s", len(ib.items), ib.debugItemsString())
+	}
 	ib.updateCommonPrefix()
 	return ib.marshalData(sb, firstItemDst, commonPrefixDst, compressLevel)
 }
 
 func (ib *inmemoryBlock) debugItemsString() string {
 	var sb strings.Builder
+	var prevItem []byte
 	for i, item := range ib.items {
+		if string(item) < string(prevItem) {
+			fmt.Fprintf(&sb, "!!! the next item is smaller than the previous item !!!\n")
+		}
 		fmt.Fprintf(&sb, "%05d %X\n", i, item)
+		prevItem = item
 	}
 	return sb.String()
 }
@@ -175,7 +193,7 @@ func (ib *inmemoryBlock) marshalData(sb *storageBlock, firstItemDst, commonPrefi
 	firstItemDst = append(firstItemDst, ib.items[0]...)
 	commonPrefixDst = append(commonPrefixDst, ib.commonPrefix...)
 
-	if len(ib.data)-len(ib.commonPrefix)*len(ib.items) < 64 || len(ib.items) < 10 {
+	if len(ib.data)-len(ib.commonPrefix)*len(ib.items) < 64 || len(ib.items) < 2 {
 		// Use plain encoding form small block, since it is cheaper.
 		ib.marshalDataPlain(sb)
 		return firstItemDst, commonPrefixDst, uint32(len(ib.items)), marshalTypePlain

lib/mergeset/merge.go

@@ -5,18 +5,31 @@ import (
 	"fmt"
 	"sync"
 	"sync/atomic"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
+// PrepareBlockCallback can transform the passed items allocated at the given data.
+//
+// The callback is called during merge before flushing full block of the given items
+// to persistent storage.
+//
+// The callback must return sorted items. The first and the last item must be unchanged.
+// The callback can re-use data and items for storing the result.
+type PrepareBlockCallback func(data []byte, items [][]byte) ([]byte, [][]byte)
+
 // mergeBlockStreams merges bsrs and writes result to bsw.
 //
 // It also fills ph.
 //
+// prepareBlock is optional.
+//
 // The function immediately returns when stopCh is closed.
 //
 // It also atomically adds the number of items merged to itemsMerged.
-func mergeBlockStreams(ph *partHeader, bsw *blockStreamWriter, bsrs []*blockStreamReader, stopCh <-chan struct{}, itemsMerged *uint64) error {
+func mergeBlockStreams(ph *partHeader, bsw *blockStreamWriter, bsrs []*blockStreamReader, prepareBlock PrepareBlockCallback, stopCh <-chan struct{}, itemsMerged *uint64) error {
 	bsm := bsmPool.Get().(*blockStreamMerger)
-	if err := bsm.Init(bsrs); err != nil {
+	if err := bsm.Init(bsrs, prepareBlock); err != nil {
 		return fmt.Errorf("cannot initialize blockStreamMerger: %s", err)
 	}
 	err := bsm.Merge(bsw, ph, stopCh, itemsMerged)
@@ -39,15 +52,24 @@ var bsmPool = &sync.Pool{
 }
 
 type blockStreamMerger struct {
+	prepareBlock PrepareBlockCallback
+
 	bsrHeap bsrHeap
 
 	// ib is a scratch block with pending items.
 	ib inmemoryBlock
 
 	phFirstItemCaught bool
+
+	// This are auxiliary buffers used in flushIB
+	// for consistency checks after prepareBlock call.
+	firstItem []byte
+	lastItem  []byte
 }
 
 func (bsm *blockStreamMerger) reset() {
+	bsm.prepareBlock = nil
+
 	for i := range bsm.bsrHeap {
 		bsm.bsrHeap[i] = nil
 	}
@@ -57,8 +79,9 @@ func (bsm *blockStreamMerger) reset() {
 	bsm.phFirstItemCaught = false
 }
 
-func (bsm *blockStreamMerger) Init(bsrs []*blockStreamReader) error {
+func (bsm *blockStreamMerger) Init(bsrs []*blockStreamReader, prepareBlock PrepareBlockCallback) error {
 	bsm.reset()
+	bsm.prepareBlock = prepareBlock
 	for _, bsr := range bsrs {
 		if bsr.Next() {
 			bsm.bsrHeap = append(bsm.bsrHeap, bsr)
@@ -95,25 +118,23 @@ again:
 
 	bsr := heap.Pop(&bsm.bsrHeap).(*blockStreamReader)
 
-	if !bsm.phFirstItemCaught {
-		ph.firstItem = append(ph.firstItem[:0], bsr.Block.items[0]...)
-		bsm.phFirstItemCaught = true
-	}
-
 	var nextItem []byte
 	hasNextItem := false
 	if len(bsm.bsrHeap) > 0 {
 		nextItem = bsm.bsrHeap[0].bh.firstItem
 		hasNextItem = true
 	}
-	for bsr.blockItemIdx < len(bsr.Block.items) && (!hasNextItem || string(bsr.Block.items[bsr.blockItemIdx]) <= string(nextItem)) {
-		if bsm.ib.Add(bsr.Block.items[bsr.blockItemIdx]) {
-			bsr.blockItemIdx++
+	for bsr.blockItemIdx < len(bsr.Block.items) {
+		item := bsr.Block.items[bsr.blockItemIdx]
+		if hasNextItem && string(item) > string(nextItem) {
+			break
+		}
+		if !bsm.ib.Add(item) {
+			// The bsm.ib is full. Flush it to bsw and continue.
+			bsm.flushIB(bsw, ph, itemsMerged)
 			continue
 		}
-
-		// The bsm.ib is full. Flush it to bsw and continue.
-		bsm.flushIB(bsw, ph, itemsMerged)
+		bsr.blockItemIdx++
 	}
 	if bsr.blockItemIdx == len(bsr.Block.items) {
 		// bsr.Block is fully read. Proceed to the next block.
@@ -139,9 +160,35 @@ func (bsm *blockStreamMerger) flushIB(bsw *blockStreamWriter, ph *partHeader, it
 		// Nothing to flush.
 		return
 	}
-	itemsCount := uint64(len(bsm.ib.items))
-	ph.itemsCount += itemsCount
-	atomic.AddUint64(itemsMerged, itemsCount)
+	atomic.AddUint64(itemsMerged, uint64(len(bsm.ib.items)))
+	if bsm.prepareBlock != nil {
+		bsm.firstItem = append(bsm.firstItem[:0], bsm.ib.items[0]...)
+		bsm.lastItem = append(bsm.lastItem[:0], bsm.ib.items[len(bsm.ib.items)-1]...)
+		bsm.ib.data, bsm.ib.items = bsm.prepareBlock(bsm.ib.data, bsm.ib.items)
+		if len(bsm.ib.items) == 0 {
+			// Nothing to flush
+			return
+		}
+		// Consistency checks after prepareBlock call.
+		firstItem := bsm.ib.items[0]
+		if string(firstItem) != string(bsm.firstItem) {
+			logger.Panicf("BUG: prepareBlock must return first item equal to the original first item;\ngot\n%X\nwant\n%X", firstItem, bsm.firstItem)
+		}
+		lastItem := bsm.ib.items[len(bsm.ib.items)-1]
+		if string(lastItem) != string(bsm.lastItem) {
+			logger.Panicf("BUG: prepareBlock must return last item equal to the original last item;\ngot\n%X\nwant\n%X", lastItem, bsm.lastItem)
+		}
+		// Verify whether the bsm.ib.items are sorted only in tests, since this
+		// can be expensive check in prod for items with long common prefix.
+		if isInTest && !bsm.ib.isSorted() {
+			logger.Panicf("BUG: prepareBlock must return sorted items;\ngot\n%s", bsm.ib.debugItemsString())
+		}
+	}
+	ph.itemsCount += uint64(len(bsm.ib.items))
+	if !bsm.phFirstItemCaught {
+		ph.firstItem = append(ph.firstItem[:0], bsm.ib.items[0]...)
+		bsm.phFirstItemCaught = true
+	}
 	ph.lastItem = append(ph.lastItem[:0], bsm.ib.items[len(bsm.ib.items)-1]...)
 	bsw.WriteBlock(&bsm.ib)
 	bsm.ib.Reset()

lib/mergeset/merge_test.go

@@ -30,14 +30,14 @@ func TestMultilevelMerge(t *testing.T) {
 	var dstIP1 inmemoryPart
 	var bsw1 blockStreamWriter
 	bsw1.InitFromInmemoryPart(&dstIP1, 0)
-	if err := mergeBlockStreams(&dstIP1.ph, &bsw1, bsrs[:5], nil, &itemsMerged); err != nil {
+	if err := mergeBlockStreams(&dstIP1.ph, &bsw1, bsrs[:5], nil, nil, &itemsMerged); err != nil {
 		t.Fatalf("cannot merge first level part 1: %s", err)
 	}
 
 	var dstIP2 inmemoryPart
 	var bsw2 blockStreamWriter
 	bsw2.InitFromInmemoryPart(&dstIP2, 0)
-	if err := mergeBlockStreams(&dstIP2.ph, &bsw2, bsrs[5:], nil, &itemsMerged); err != nil {
+	if err := mergeBlockStreams(&dstIP2.ph, &bsw2, bsrs[5:], nil, nil, &itemsMerged); err != nil {
 		t.Fatalf("cannot merge first level part 2: %s", err)
 	}
 
@@ -54,7 +54,7 @@ func TestMultilevelMerge(t *testing.T) {
 		newTestBlockStreamReader(&dstIP2),
 	}
 	bsw.InitFromInmemoryPart(&dstIP, 0)
-	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrsTop, nil, &itemsMerged); err != nil {
+	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrsTop, nil, nil, &itemsMerged); err != nil {
 		t.Fatalf("cannot merge second level: %s", err)
 	}
 	if itemsMerged != uint64(len(items)) {
@@ -76,7 +76,7 @@ func TestMergeForciblyStop(t *testing.T) {
 	ch := make(chan struct{})
 	var itemsMerged uint64
 	close(ch)
-	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrs, ch, &itemsMerged); err != errForciblyStopped {
+	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrs, nil, ch, &itemsMerged); err != errForciblyStopped {
 		t.Fatalf("unexpected error during merge: got %v; want %v", err, errForciblyStopped)
 	}
 	if itemsMerged != 0 {
@@ -120,7 +120,7 @@ func testMergeBlockStreamsSerial(blocksToMerge, maxItemsPerBlock int) error {
 	var dstIP inmemoryPart
 	var bsw blockStreamWriter
 	bsw.InitFromInmemoryPart(&dstIP, 0)
-	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrs, nil, &itemsMerged); err != nil {
+	if err := mergeBlockStreams(&dstIP.ph, &bsw, bsrs, nil, nil, &itemsMerged); err != nil {
 		return fmt.Errorf("cannot merge block streams: %s", err)
 	}
 	if itemsMerged != uint64(len(items)) {

lib/mergeset/part.go

@@ -5,6 +5,7 @@ import (
 	"path/filepath"
 	"sync"
 	"sync/atomic"
+	"unsafe"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
@@ -13,7 +14,7 @@ import (
 
 func getMaxCachedIndexBlocksPerPart() int {
 	maxCachedIndexBlocksPerPartOnce.Do(func() {
-		n := memory.Allowed() / 1024 / 1024 / 2
+		n := memory.Allowed() / 1024 / 1024 / 4
 		if n == 0 {
 			n = 10
 		}
@@ -29,7 +30,7 @@ var (
 
 func getMaxCachedInmemoryBlocksPerPart() int {
 	maxCachedInmemoryBlocksPerPartOnce.Do(func() {
-		n := memory.Allowed() / 1024 / 1024 / 2
+		n := memory.Allowed() / 1024 / 1024 / 4
 		if n == 0 {
 			n = 10
 		}
@@ -43,7 +44,7 @@ var (
 	maxCachedInmemoryBlocksPerPartOnce sync.Once
 )
 
-type part struct {
+type partInternals struct {
 	ph partHeader
 
 	path string
@@ -55,7 +56,14 @@ type part struct {
 	indexFile fs.ReadAtCloser
 	itemsFile fs.ReadAtCloser
 	lensFile  fs.ReadAtCloser
+}
 
+type part struct {
+	partInternals
+
+	// Align atomic counters inside caches by 8 bytes on 32-bit architectures.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212 .
+	_         [(8 - (unsafe.Sizeof(partInternals{}) % 8)) % 8]byte
 	idxbCache indexBlockCache
 	ibCache   inmemoryBlockCache
 }
@@ -114,15 +122,15 @@ func newPart(ph *partHeader, path string, size uint64, metaindexReader filestrea
 	}
 	metaindexReader.MustClose()
 
-	p := &part{
-		path: path,
-		size: size,
-		mrs:  mrs,
+	var p part
+	p.path = path
+	p.size = size
+	p.mrs = mrs
+
+	p.indexFile = indexFile
+	p.itemsFile = itemsFile
+	p.lensFile = lensFile
 
-		indexFile: indexFile,
-		itemsFile: itemsFile,
-		lensFile:  lensFile,
-	}
 	p.ph.CopyFrom(ph)
 	p.idxbCache.Init()
 	p.ibCache.Init()
@@ -133,7 +141,7 @@ func newPart(ph *partHeader, path string, size uint64, metaindexReader filestrea
 		p.MustClose()
 		return nil, err
 	}
-	return p, nil
+	return &p, nil
 }
 
 func (p *part) MustClose() {
@@ -165,12 +173,15 @@ func putIndexBlock(idxb *indexBlock) {
 var indexBlockPool sync.Pool
 
 type indexBlockCache struct {
+	// Atomically updated counters must go first in the struct, so they are properly
+	// aligned to 8 bytes on 32-bit architectures.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
+	requests uint64
+	misses   uint64
+
 	m         map[uint64]*indexBlock
 	missesMap map[uint64]uint64
 	mu        sync.RWMutex
-
-	requests uint64
-	misses   uint64
 }
 
 func (idxbc *indexBlockCache) Init() {
@@ -274,12 +285,15 @@ func (idxbc *indexBlockCache) Misses() uint64 {
 }
 
 type inmemoryBlockCache struct {
+	// Atomically updated counters must go first in the struct, so they are properly
+	// aligned to 8 bytes on 32-bit architectures.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
+	requests uint64
+	misses   uint64
+
 	m         map[inmemoryBlockCacheKey]*inmemoryBlock
 	missesMap map[inmemoryBlockCacheKey]uint64
 	mu        sync.RWMutex
-
-	requests uint64
-	misses   uint64
 }
 
 type inmemoryBlockCacheKey struct {

lib/mergeset/part_search.go

@@ -31,6 +31,8 @@ type partSearch struct {
 	// Pointer to inmemory block, which may be reused.
 	inmemoryBlockReuse *inmemoryBlock
 
+	shouldCacheBlock func(item []byte) bool
+
 	idxbCache *indexBlockCache
 	ibCache   *inmemoryBlockCache
 
@@ -59,6 +61,7 @@ func (ps *partSearch) reset() {
 		putInmemoryBlock(ps.inmemoryBlockReuse)
 		ps.inmemoryBlockReuse = nil
 	}
+	ps.shouldCacheBlock = nil
 	ps.idxbCache = nil
 	ps.ibCache = nil
 	ps.err = nil
@@ -75,7 +78,7 @@ func (ps *partSearch) reset() {
 // Init initializes ps for search in the p.
 //
 // Use Seek for search in p.
-func (ps *partSearch) Init(p *part) {
+func (ps *partSearch) Init(p *part, shouldCacheBlock func(item []byte) bool) {
 	ps.reset()
 
 	ps.p = p
@@ -324,6 +327,16 @@ func (ps *partSearch) readIndexBlock(mr *metaindexRow) (*indexBlock, error) {
 }
 
 func (ps *partSearch) getInmemoryBlock(bh *blockHeader) (*inmemoryBlock, bool, error) {
+	if ps.shouldCacheBlock != nil {
+		if !ps.shouldCacheBlock(bh.firstItem) {
+			ib, err := ps.readInmemoryBlock(bh)
+			if err != nil {
+				return nil, false, err
+			}
+			return ib, true, nil
+		}
+	}
+
 	var ibKey inmemoryBlockCacheKey
 	ibKey.Init(bh)
 	ib := ps.ibCache.Get(ibKey)
@@ -371,7 +384,7 @@ func binarySearchKey(items [][]byte, key []byte) int {
 	i, j := uint(0), n
 	for i < j {
 		h := uint(i+j) >> 1
-		if string(key) > string(items[h]) {
+		if h >= 0 && h < uint(len(items)) && string(key) > string(items[h]) {
 			i = h + 1
 		} else {
 			j = h