lib/promscrape: consistently update /targets page after SIGHUP

This arg allows limiting the number of data points that may be exported on a single line.

.github/ISSUE_TEMPLATE/bug_report.md

@@ -26,5 +26,16 @@ $ ./victoria-metrics-prod --version
 victoria-metrics-20190730-121249-heads-single-node-0-g671d9e55
 ```
 
+**Used command-line flags**
+Command-line flags are listed as `flag{name="httpListenAddr", value=":443"} 1` lines at `/metrics` page.
+See the following docs for details:
+
+* [monitoring for single-node VictoriaMetrics](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#monitoring)
+* [montioring for VictoriaMetrics cluster](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/cluster/README.md#monitoring)
+
 **Additional context**
-Add any other context about the problem here such as error logs, `/metrics` output, screenshots from [the official Grafana dashboard for VictoriaMetrics](https://grafana.com/dashboards/10229).
+Add any other context about the problem here such as error logs from VictoriaMetrics and Prometheus,
+`/metrics` output, screenshots from the official Grafana dashboards for VictoriaMetrics:
+
+* [Grafana dashboard for single-node VictoriaMetrics](https://grafana.com/dashboards/10229)
+* [Grafana dashboard for VictoriaMetrics cluster](https://grafana.com/grafana/dashboards/11176)

.github/workflows/github-pages.yml

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

.github/workflows/main.yml

@@ -1,7 +1,13 @@
 name: main
 on:
-  - push
-  - pull_request
+  push:
+    paths-ignore:
+      - 'docs/**'
+      - '**.md'
+  pull_request:
+    paths-ignore:
+      - 'docs/**'
+      - '**.md'
 jobs:
   build:
     name: Build
@@ -24,19 +30,21 @@ jobs:
         env:
           GO111MODULE: on
         run: |
-            export PATH=$PATH:$(go env GOPATH)/bin # temporary fix. See https://github.com/actions/setup-go/issues/14
-            make check-all
-            git diff --exit-code
-            make test-full
-            make test-pure
-            make victoria-metrics
-            make victoria-metrics-pure
-            make victoria-metrics-arm
-            make victoria-metrics-arm64
-            GOOS=freebsd go build -mod=vendor ./app/victoria-metrics
-            GOOS=darwin go build -mod=vendor ./app/victoria-metrics
+          export PATH=$PATH:$(go env GOPATH)/bin # temporary fix. See https://github.com/actions/setup-go/issues/14
+          make check-all
+          git diff --exit-code
+          make test-full
+          make test-pure
+          make test-full-386
+          make victoria-metrics
+          make victoria-metrics-pure
+          make victoria-metrics-arm
+          make victoria-metrics-arm64
+          make vmutils
+          GOOS=freebsd go build -mod=vendor ./app/victoria-metrics
+          GOOS=darwin go build -mod=vendor ./app/victoria-metrics
       - name: Publish coverage
-        uses: codecov/codecov-action@v1.0.0
+        uses: codecov/codecov-action@v1.0.6
         with:
           token: ${{secrets.CODECOV_TOKEN}}
           file: ./coverage.txt

.github/workflows/wiki.yml

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

.gitignore

@@ -1,3 +1,4 @@
+/tmp
 /tags
 /pkg
 *.pprof
@@ -7,6 +8,7 @@
 *.swp
 /gocache-for-docker
 /victoria-metrics-data
+/vmagent-remotewrite-data
 /vmstorage-data
 /vmselect-cache
 /package/temp-deb-*

LICENSE

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

Makefile

@@ -11,7 +11,10 @@ endif
 GO_BUILDINFO = -X '$(PKG_PREFIX)/lib/buildinfo.Version=$(APP_NAME)-$(shell date -u +'%Y%m%d-%H%M%S')-$(BUILDINFO_TAG)'
 
 all: \
-	victoria-metrics-prod
+	victoria-metrics-prod \
+	vmagent-prod \
+	vmbackup-prod \
+	vmrestore-prod
 
 include app/*/Makefile
 include deployment/*/Makefile
@@ -19,12 +22,40 @@ include deployment/*/Makefile
 clean:
 	rm -rf bin/*
 
-publish: publish-victoria-metrics
+publish: \
+	publish-victoria-metrics \
+	publish-vmagent \
+	publish-vmbackup \
+	publish-vmrestore
 
-package: package-victoria-metrics
+package: \
+	package-victoria-metrics \
+	package-vmagent \
+	package-vmbackup \
+	package-vmrestore
 
-release: victoria-metrics-prod
-	cd bin && tar czf victoria-metrics-$(PKG_TAG).tar.gz victoria-metrics-prod
+vmutils: \
+	vmagent \
+	vmbackup \
+	vmrestore
+
+release: \
+	release-victoria-metrics \
+	release-vmutils
+
+release-victoria-metrics: victoria-metrics-prod
+	cd bin && tar czf victoria-metrics-$(PKG_TAG).tar.gz victoria-metrics-prod && \
+		sha256sum victoria-metrics-$(PKG_TAG).tar.gz > victoria-metrics-$(PKG_TAG)_checksums.txt
+
+release-vmutils: \
+	vmagent-prod \
+	vmbackup-prod \
+	vmrestore-prod
+	cd bin && tar czf vmutils-$(PKG_TAG).tar.gz vmagent-prod vmbackup-prod vmrestore-prod && \
+		sha256sum vmutils-$(PKG_TAG).tar.gz > vmutils-$(PKG_TAG)_checksums.txt
+
+pprof-cpu:
+	go tool pprof -trim_path=github.com/VictoriaMetrics/VictoriaMetrics@ $(PPROF_FILE)
 
 fmt:
 	GO111MODULE=on gofmt -l -w -s ./lib
@@ -39,13 +70,17 @@ lint: install-golint
 	golint app/...
 
 install-golint:
-	which golint || GO111MODULE=off go get -u github.com/golang/lint/golint
+	which golint || GO111MODULE=off go get -u golang.org/x/lint/golint
 
 errcheck: install-errcheck
 	errcheck -exclude=errcheck_excludes.txt ./lib/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vminsert/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vmselect/...
 	errcheck -exclude=errcheck_excludes.txt ./app/vmstorage/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmagent/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmbackup/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmrestore/...
+	errcheck -exclude=errcheck_excludes.txt ./app/vmalert/...
 
 install-errcheck:
 	which errcheck || GO111MODULE=off go get -u github.com/kisielk/errcheck
@@ -53,13 +88,19 @@ install-errcheck:
 check-all: fmt vet lint errcheck golangci-lint
 
 test:
-	GO111MODULE=on go test -tags=integration -mod=vendor ./lib/... ./app/...
+	GO111MODULE=on go test -mod=vendor ./lib/... ./app/...
+
+test-race:
+	GO111MODULE=on go test -mod=vendor -race ./lib/... ./app/...
 
 test-pure:
-	GO111MODULE=on CGO_ENABLED=0 go test -tags=integration -mod=vendor ./lib/... ./app/...
+	GO111MODULE=on CGO_ENABLED=0 go test -mod=vendor ./lib/... ./app/...
 
 test-full:
-	GO111MODULE=on go test -tags=integration -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
+	GO111MODULE=on go test -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
+
+test-full-386:
+	GO111MODULE=on GOARCH=386 go test -mod=vendor -coverprofile=coverage.txt -covermode=atomic ./lib/... ./app/...
 
 benchmark:
 	GO111MODULE=on go test -mod=vendor -bench=. ./lib/...
@@ -89,7 +130,7 @@ 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

app/victoria-metrics/Makefile

@@ -3,12 +3,50 @@
 victoria-metrics:
 	APP_NAME=victoria-metrics $(MAKE) app-local
 
+victoria-metrics-race:
+	APP_NAME=victoria-metrics RACE=-race $(MAKE) app-local
+
 victoria-metrics-prod:
 	APP_NAME=victoria-metrics $(MAKE) app-via-docker
 
+victoria-metrics-pure-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-pure
+
+victoria-metrics-amd64-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-amd64
+
+victoria-metrics-arm-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-arm
+
+victoria-metrics-arm64-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-arm64
+
+victoria-metrics-ppc64le-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-ppc64le
+
+victoria-metrics-386-prod:
+	APP_NAME=victoria-metrics $(MAKE) app-via-docker-386
+
 package-victoria-metrics:
-	APP_NAME=victoria-metrics \
-	$(MAKE) package-via-docker
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker
+
+package-victoria-metrics-pure:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-pure
+
+package-victoria-metrics-amd64:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-amd64
+
+package-victoria-metrics-arm:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-arm
+
+package-victoria-metrics-arm64:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-arm64
+
+package-victoria-metrics-ppc64le:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-ppc64le
+
+package-victoria-metrics-386:
+	APP_NAME=victoria-metrics $(MAKE) package-via-docker-386
 
 publish-victoria-metrics:
 	APP_NAME=victoria-metrics $(MAKE) publish-via-docker
@@ -20,24 +58,24 @@ run-victoria-metrics:
 	ARGS='-graphiteListenAddr=:2003 -opentsdbListenAddr=:4242 -retentionPeriod=12 -search.maxUniqueTimeseries=1000000 -search.maxQueryDuration=10m' \
 	$(MAKE) run-via-docker
 
+victoria-metrics-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-amd64 ./app/victoria-metrics
+
 victoria-metrics-arm:
 	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm ./app/victoria-metrics
 
-victoria-metrics-arm-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-arm' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm' $(MAKE) app-via-docker
-
 victoria-metrics-arm64:
 	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-arm64 ./app/victoria-metrics
 
-victoria-metrics-arm64-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-arm64' DOCKER_OPTS='--env CGO_ENABLED=0 --env GOARCH=arm64' $(MAKE) app-via-docker
+victoria-metrics-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-ppc64le ./app/victoria-metrics
+
+victoria-metrics-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/victoria-metrics-386 ./app/victoria-metrics
 
 victoria-metrics-pure:
 	APP_NAME=victoria-metrics $(MAKE) app-local-pure
 
-victoria-metrics-pure-prod:
-	APP_NAME=victoria-metrics APP_SUFFIX='-pure' DOCKER_OPTS='--env CGO_ENABLED=0' $(MAKE) app-via-docker
-
 ### Packaging as DEB - amd64
 victoria-metrics-package-deb: victoria-metrics-prod
 	./package/package_deb.sh amd64

app/victoria-metrics/deployment/Dockerfile

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

app/victoria-metrics/main.go

@@ -9,44 +9,55 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/procutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
 
-var httpListenAddr = flag.String("httpListenAddr", ":8428", "TCP address to listen for http connections")
+var (
+	httpListenAddr    = flag.String("httpListenAddr", ":8428", "TCP address to listen for http connections")
+	minScrapeInterval = flag.Duration("dedup.minScrapeInterval", 0, "Remove superflouos samples from time series if they are located closer to each other than this duration. "+
+		"This may be useful for reducing overhead when multiple identically configured Prometheus instances write data to the same VictoriaMetrics. "+
+		"Deduplication is disabled if the -dedup.minScrapeInterval is 0")
+)
 
 func main() {
-	flag.Parse()
+	envflag.Parse()
 	buildinfo.Init()
 	logger.Init()
-	logger.Infof("starting VictoraMetrics at %q...", *httpListenAddr)
+	logger.Infof("starting VictoriaMetrics at %q...", *httpListenAddr)
 	startTime := time.Now()
+	storage.SetMinScrapeIntervalForDeduplication(*minScrapeInterval)
 	vmstorage.Init()
 	vmselect.Init()
 	vminsert.Init()
+	startSelfScraper()
 
 	go httpserver.Serve(*httpListenAddr, requestHandler)
-	logger.Infof("started VictoriaMetrics in %s", time.Since(startTime))
+	logger.Infof("started VictoriaMetrics in %.3f seconds", time.Since(startTime).Seconds())
 
 	sig := procutil.WaitForSigterm()
 	logger.Infof("received signal %s", sig)
 
+	stopSelfScraper()
+
 	logger.Infof("gracefully shutting down webservice at %q", *httpListenAddr)
 	startTime = time.Now()
 	if err := httpserver.Stop(*httpListenAddr); err != nil {
 		logger.Fatalf("cannot stop the webservice: %s", err)
 	}
 	vminsert.Stop()
-	logger.Infof("successfully shut down the webservice in %s", time.Since(startTime))
+	logger.Infof("successfully shut down the webservice in %.3f seconds", time.Since(startTime).Seconds())
 
 	vmstorage.Stop()
 	vmselect.Stop()
 
 	fs.MustStopDirRemover()
 
-	logger.Infof("the VictoriaMetrics has been stopped in %s", time.Since(startTime))
+	logger.Infof("the VictoriaMetrics has been stopped in %.3f seconds", time.Since(startTime).Seconds())
 }
 
 func requestHandler(w http.ResponseWriter, r *http.Request) bool {

app/victoria-metrics/main_test.go

@@ -1,5 +1,3 @@
-// +build integration
-
 package main
 
 import (
@@ -23,6 +21,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
@@ -148,7 +147,7 @@ func setUp() {
 }
 
 func processFlags() {
-	flag.Parse()
+	envflag.Parse()
 	for _, fv := range []struct {
 		flag  string
 		value string
@@ -186,7 +185,6 @@ func tearDown() {
 	vmstorage.Stop()
 	vmselect.Stop()
 	fs.MustRemoveAll(storagePath)
-	fs.MustStopDirRemover()
 }
 
 func TestWriteRead(t *testing.T) {
@@ -303,6 +301,9 @@ func readIn(readFor string, t *testing.T, insertTime time.Time) []test {
 	s := newSuite(t)
 	var tt []test
 	s.noError(filepath.Walk(filepath.Join(testFixturesDir, readFor), func(path string, info os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
 		if filepath.Ext(path) != ".json" {
 			return nil
 		}

app/victoria-metrics/self_scraper.go

@@ -0,0 +1,99 @@
+package main
+
+import (
+	"flag"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/prometheus"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+)
+
+var selfScrapeInterval = flag.Duration("selfScrapeInterval", 0, "Interval for self-scraping own metrics at /metrics page")
+
+var selfScraperStopCh chan struct{}
+var selfScraperWG sync.WaitGroup
+
+func startSelfScraper() {
+	selfScraperStopCh = make(chan struct{})
+	selfScraperWG.Add(1)
+	go func() {
+		defer selfScraperWG.Done()
+		selfScraper(*selfScrapeInterval)
+	}()
+}
+
+func stopSelfScraper() {
+	close(selfScraperStopCh)
+	selfScraperWG.Wait()
+}
+
+func selfScraper(scrapeInterval time.Duration) {
+	if scrapeInterval <= 0 {
+		// Self-scrape is disabled.
+		return
+	}
+	logger.Infof("started self-scraping `/metrics` page with interval %.3f seconds", scrapeInterval.Seconds())
+
+	var bb bytesutil.ByteBuffer
+	var rows prometheus.Rows
+	var mrs []storage.MetricRow
+	var labels []prompb.Label
+	t := time.NewTicker(scrapeInterval)
+	var currentTimestamp int64
+	for {
+		select {
+		case <-selfScraperStopCh:
+			t.Stop()
+			logger.Infof("stopped self-scraping `/metrics` page")
+			return
+		case currentTime := <-t.C:
+			currentTimestamp = currentTime.UnixNano() / 1e6
+		}
+		bb.Reset()
+		httpserver.WritePrometheusMetrics(&bb)
+		s := bytesutil.ToUnsafeString(bb.B)
+		rows.Reset()
+		rows.Unmarshal(s)
+		mrs = mrs[:0]
+		for i := range rows.Rows {
+			r := &rows.Rows[i]
+			labels = labels[:0]
+			labels = addLabel(labels, "", r.Metric)
+			labels = addLabel(labels, "job", "victoria-metrics")
+			labels = addLabel(labels, "instance", "self")
+			for j := range r.Tags {
+				t := &r.Tags[j]
+				labels = addLabel(labels, t.Key, t.Value)
+			}
+			if len(mrs) < cap(mrs) {
+				mrs = mrs[:len(mrs)+1]
+			} else {
+				mrs = append(mrs, storage.MetricRow{})
+			}
+			mr := &mrs[len(mrs)-1]
+			mr.MetricNameRaw = storage.MarshalMetricNameRaw(mr.MetricNameRaw[:0], labels)
+			mr.Timestamp = currentTimestamp
+			mr.Value = r.Value
+		}
+		logger.Infof("writing %d rows at timestamp %d", len(mrs), currentTimestamp)
+		vmstorage.AddRows(mrs)
+	}
+}
+
+func addLabel(dst []prompb.Label, key, value string) []prompb.Label {
+	if len(dst) < cap(dst) {
+		dst = dst[:len(dst)+1]
+	} else {
+		dst = append(dst, prompb.Label{})
+	}
+	lb := &dst[len(dst)-1]
+	lb.Name = bytesutil.ToUnsafeBytes(key)
+	lb.Value = bytesutil.ToUnsafeBytes(value)
+	return dst
+}

app/victoria-metrics/test/parser.go

@@ -1,5 +1,3 @@
-// +build integration
-
 package test
 
 import (
@@ -15,12 +13,15 @@ var (
 	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))
 }
 
-func PopulateTimeTpl(b []byte, t time.Time) []byte {
+// 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 {
@@ -40,7 +41,7 @@ func PopulateTimeTpl(b []byte, t time.Time) []byte {
 		case `TIME_NS`:
 			return []byte(fmt.Sprintf("%d", t.UnixNano()))
 		default:
-			log.Fatalf("unkown time pattern %s in %s", parts[0], repl)
+			log.Fatalf("unknown time pattern %s in %s", parts[0], repl)
 		}
 		return repl
 	})

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

@@ -1,18 +1,18 @@
-// +build integration
-
-// Source https://github.com/prometheus/prometheus/blob/master/prompb/remote.pb.go . Code is copy pasted and cleaned up
 package test
 
+// Source https://github.com/prometheus/prometheus/blob/master/prompb/remote.pb.go . Code is copy pasted and cleaned up
 import (
 	"encoding/binary"
 	"math"
 	"math/bits"
 )
 
+// WriteRequest is write request
 type WriteRequest struct {
 	Timeseries []TimeSeries `protobuf:"bytes,1,rep,name=timeseries,proto3" json:"timeseries"`
 }
 
+// Size returns m size in bytes after marshaling.
 func (m *WriteRequest) Size() (n int) {
 	if m == nil {
 		return 0
@@ -31,6 +31,7 @@ func sovRemote(x uint64) (n int) {
 	return (bits.Len64(x|1) + 6) / 7
 }
 
+// Marshal marshals m.
 func (m *WriteRequest) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -41,11 +42,13 @@ func (m *WriteRequest) Marshal() (dAtA []byte, err error) {
 	return dAtA[:n], nil
 }
 
+// MarshalTo marshals m to dAtA
 func (m *WriteRequest) MarshalTo(dAtA []byte) (int, error) {
 	size := m.Size()
 	return m.MarshalToSizedBuffer(dAtA[:size])
 }
 
+// MarshalToSizedBuffer marshals m to dAtA.
 func (m *WriteRequest) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	i := len(dAtA)
 	if len(m.Timeseries) > 0 {
@@ -77,11 +80,13 @@ func encodeVarintRemote(dAtA []byte, offset int, v uint64) int {
 	return base
 }
 
+// Sample is time series sample.
 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"`
 }
 
+// Reset resets m.
 func (m *Sample) Reset() { *m = Sample{} }
 
 // TimeSeries represents samples and labels for a single time series.
@@ -90,21 +95,27 @@ type TimeSeries struct {
 	Samples []Sample `protobuf:"bytes,2,rep,name=samples,proto3" json:"samples"`
 }
 
+// Reset resets m.
 func (m *TimeSeries) Reset() { *m = TimeSeries{} }
 
+// Label is time series label.
 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"`
 }
 
+// Reset resets m.
 func (m *Label) Reset() { *m = Label{} }
 
+// Labels is a set of labels.
 type Labels struct {
 	Labels []Label `protobuf:"bytes,1,rep,name=labels,proto3" json:"labels"`
 }
 
+// Reset resets m.
 func (m *Labels) Reset() { *m = Labels{} }
 
+// Marshal marshals m.
 func (m *Sample) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -115,11 +126,13 @@ func (m *Sample) Marshal() (dAtA []byte, err error) {
 	return dAtA[:n], nil
 }
 
+// MarshalTo marshals m to dAtA.
 func (m *Sample) MarshalTo(dAtA []byte) (int, error) {
 	size := m.Size()
 	return m.MarshalToSizedBuffer(dAtA[:size])
 }
 
+// MarshalToSizedBuffer marshals m to dAtA.
 func (m *Sample) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	i := len(dAtA)
 	if m.Timestamp != 0 {
@@ -136,6 +149,7 @@ func (m *Sample) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	return len(dAtA) - i, nil
 }
 
+// Marshal marshals m.
 func (m *TimeSeries) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -146,11 +160,13 @@ func (m *TimeSeries) Marshal() (dAtA []byte, err error) {
 	return dAtA[:n], nil
 }
 
+// MarshalTo marshals m to dAtA.
 func (m *TimeSeries) MarshalTo(dAtA []byte) (int, error) {
 	size := m.Size()
 	return m.MarshalToSizedBuffer(dAtA[:size])
 }
 
+// MarshalToSizedBuffer marshals m to dAtA.
 func (m *TimeSeries) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	i := len(dAtA)
 	if len(m.Samples) > 0 {
@@ -184,6 +200,7 @@ func (m *TimeSeries) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	return len(dAtA) - i, nil
 }
 
+// Marshal marshals m.
 func (m *Label) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -194,11 +211,13 @@ func (m *Label) Marshal() (dAtA []byte, err error) {
 	return dAtA[:n], nil
 }
 
+// MarshalTo marshals m to dAtA.
 func (m *Label) MarshalTo(dAtA []byte) (int, error) {
 	size := m.Size()
 	return m.MarshalToSizedBuffer(dAtA[:size])
 }
 
+// MarshalToSizedBuffer marshals m to dAtA.
 func (m *Label) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	i := len(dAtA)
 	_ = i
@@ -221,6 +240,7 @@ func (m *Label) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	return len(dAtA) - i, nil
 }
 
+// Marshal marshals m.
 func (m *Labels) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -231,11 +251,13 @@ func (m *Labels) Marshal() (dAtA []byte, err error) {
 	return dAtA[:n], nil
 }
 
+// MarshalTo marshals m to dAtA.
 func (m *Labels) MarshalTo(dAtA []byte) (int, error) {
 	size := m.Size()
 	return m.MarshalToSizedBuffer(dAtA[:size])
 }
 
+// MarshalToSizedBuffer marshals m to dAtA.
 func (m *Labels) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	i := len(dAtA)
 	if len(m.Labels) > 0 {
@@ -267,6 +289,7 @@ func encodeVarintTypes(dAtA []byte, offset int, v uint64) int {
 	return base
 }
 
+// Size returns the size of marshaled m.
 func (m *Sample) Size() (n int) {
 	if m == nil {
 		return 0
@@ -280,6 +303,7 @@ func (m *Sample) Size() (n int) {
 	return n
 }
 
+// Size returns the size of marshaled m.
 func (m *TimeSeries) Size() (n int) {
 	if m == nil {
 		return 0
@@ -301,6 +325,7 @@ func (m *TimeSeries) Size() (n int) {
 	return n
 }
 
+// Size returns the size of marshaled m.
 func (m *Label) Size() (n int) {
 	if m == nil {
 		return 0
@@ -318,6 +343,7 @@ func (m *Label) Size() (n int) {
 	return n
 }
 
+// Size returns the size of marshaled m.
 func (m *Labels) Size() (n int) {
 	if m == nil {
 		return 0

app/victoria-metrics/test/prom_writter.go

@@ -1,9 +1,8 @@
-// +build integration
-
 package test
 
 import "github.com/golang/snappy"
 
+// Compress marshals and compresses wr.
 func Compress(wr WriteRequest) ([]byte, error) {
 	data, err := wr.Marshal()
 	if err != nil {

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

@@ -3,12 +3,14 @@
   "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 2 {TIME_S-2m}",
-    "not_nan_not_inf;item=y 1 {TIME_S-1m}",
-    "not_nan_not_inf;item=y 2 {TIME_S-2m}"],
-  "query": ["/api/v1/query_range?query=1/(not_nan_not_inf-1)!=inf!=nan&start={TIME_S-2m}&end={TIME_S}&step=60"],
+    "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":"x"},"values":[["{TIME_S-2m}","1"]]},{"metric":{"item":"y"},"values":[["{TIME_S-2m}","1"]]}]}}
+      "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 +1,21 @@
+{
+  "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-120s}","3"],
+	      ["{TIME_S-60s}","2"],
+	      ["{TIME_S-30s}","1"],
+	      ["{TIME_S-20s}","1"]
+      ]}]}}
+}

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

@@ -0,0 +1,30 @@
+{
+  "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-60s}","2"],
+	      ["{TIME_S-50s}","2"],
+	      ["{TIME_S-40s}","2"],
+	      ["{TIME_S-30s}","1"],
+	      ["{TIME_S-20s}","1"],
+	      ["{TIME_S-10s}","1"],
+	      ["{TIME_S-0s}","1"]
+      ]}]}}
+}

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

@@ -2,13 +2,17 @@
   "name": "not-nan-as-missing-data",
   "issue": "https://github.com/VictoriaMetrics/VictoriaMetrics/issues/153",
   "data": [
-    "not_nan_as_missing_data;item=x 1 {TIME_S-1m}",
     "not_nan_as_missing_data;item=x 2 {TIME_S-2m}",
-    "not_nan_as_missing_data;item=y 3 {TIME_S-1m}",
-    "not_nan_as_missing_data;item=y 4 {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"]]}]}}
+      "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/vmagent/Makefile

@@ -0,0 +1,76 @@
+# All these commands must run from repository root.
+
+vmagent:
+	APP_NAME=vmagent $(MAKE) app-local
+
+vmagent-race:
+	APP_NAME=vmagent RACE=-race $(MAKE) app-local
+
+vmagent-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker
+
+vmagent-pure-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-pure
+
+vmagent-amd64-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-amd64
+
+vmagent-arm-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-arm
+
+vmagent-arm64-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-arm64
+
+vmagent-ppc64le-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-ppc64le
+
+vmagent-386-prod:
+	APP_NAME=vmagent $(MAKE) app-via-docker-386
+
+package-vmagent:
+	APP_NAME=vmagent $(MAKE) package-via-docker
+
+package-vmagent-pure:
+	APP_NAME=vmagent $(MAKE) package-via-docker-pure
+
+package-vmagent-amd64:
+	APP_NAME=vmagent $(MAKE) package-via-docker-amd64
+
+package-vmagent-arm:
+	APP_NAME=vmagent $(MAKE) package-via-docker-arm
+
+package-vmagent-arm64:
+	APP_NAME=vmagent $(MAKE) package-via-docker-arm64
+
+package-vmagent-ppc64le:
+	APP_NAME=vmagent $(MAKE) package-via-docker-ppc64le
+
+package-vmagent-386:
+	APP_NAME=vmagent $(MAKE) package-via-docker-386
+
+publish-vmagent:
+	APP_NAME=vmagent $(MAKE) publish-via-docker
+
+run-vmagent:
+	mkdir -p vmagent-data
+	DOCKER_OPTS='-v $(shell pwd)/vmagent-data:/vmagent-data' \
+	APP_NAME=vmagent \
+	$(MAKE) run-via-docker
+
+vmagent-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmagent-amd64 ./app/vmagent
+
+vmagent-arm:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmagent-arm ./app/vmagent
+
+vmagent-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmagent-arm64 ./app/vmagent
+
+vmagent-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmagent-ppc64le ./app/vmagent
+
+vmagent-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmagent-386 ./app/vmagent
+
+vmagent-pure:
+	APP_NAME=vmagent $(MAKE) app-local-pure

app/vmagent/README.md

@@ -0,0 +1,213 @@
+## vmagent
+
+`vmagent` is a tiny but brave agent, which helps you collecting metrics from various sources
+and storing them to [VictoriaMetrics](https://github.com/VictoriaMetrics/VictoriaMetrics).
+
+<img alt="vmagent" src="vmagent.png">
+
+
+### Motivation
+
+While VictoriaMetrics provides an efficient solution to store and observe metrics, our users needed something fast
+and RAM friendly to scrape metrics from Prometheus-compatible exporters to VictoriaMetrics.
+Also, we found that users’ infrastructure is like snowflakes - never alike, and we decided to add more flexibility
+to `vmagent` (like the ability to push metrics instead of pulling them). We did our best and plan to do even more.
+
+
+### Features
+
+* Can be used as drop-in replacement for Prometheus for scraping targets such as [node_exporter](https://github.com/prometheus/node_exporter).
+  See [Quick Start](#quick-start) for details.
+* Can add, remove and modify labels via Prometheus relabeling. See [these docs](#relabeling) for details.
+* Accepts data via all the ingestion protocols supported by VictoriaMetrics:
+  * Influx line protocol via `http://<vmagent>:8429/write`. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-send-data-from-influxdb-compatible-agents-such-as-telegraf).
+  * Graphite plaintext protocol if `-graphiteListenAddr` command-line flag is set. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-send-data-from-graphite-compatible-agents-such-as-statsd).
+  * OpenTSDB telnet and http protocols if `-opentsdbListenAddr` command-line flag is set. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-send-data-from-opentsdb-compatible-agents).
+  * Prometheus remote write protocol via `http://<vmagent>:8429/api/v1/write`.
+  * JSON lines import protocol via `http://<vmagent>:8429/api/v1/import`. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-import-time-series-data).
+  * Arbitrary CSV data via `http://<vmagent>:8429/api/v1/import/csv`. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-import-csv-data).
+* Can replicate collected metrics simultaneously to multiple remote storage systems.
+* Works in environments with unstable connections to remote storage. If the remote storage is unavailable, the collected metrics
+  are buffered at `-remoteWrite.tmpDataPath`. The buffered metrics are sent to remote storage as soon as connection
+  to remote storage is recovered. The maximum disk usage for the buffer can be limited with `-remoteWrite.maxDiskUsagePerURL`.
+* Uses lower amounts of RAM, CPU, disk IO and network bandwidth comparing to Prometheus.
+
+
+### Quick Start
+
+Just download `vmutils-*` archive from [releases page](https://github.com/VictoriaMetrics/VictoriaMetrics/releases), unpack it
+and pass the following flags to `vmagent` binary in order to start scraping Prometheus targets:
+
+* `-promscrape.config` with the path to Prometheus config file (it is usually located at `/etc/prometheus/prometheus.yml`)
+* `-remoteWrite.url` with the remote storage endpoint such as VictoriaMetrics. Multiple `-remoteWrite.url` args can be set in parallel
+  in order to replicate data concurrently to multiple remote storage systems.
+
+Example command line:
+
+```
+/path/to/vmagent -promscrape.config=/path/to/prometheus.yml -remoteWrite.url=https://victoria-metrics-host:8428/api/v1/write
+```
+
+If you need collecting only Influx data, then the following command line would be enough:
+
+```
+/path/to/vmagent -remoteWrite.url=https://victoria-metrics-host:8428/api/v1/write
+```
+
+Then send Influx data to `http://vmagent-host:8429`. See [these docs](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-send-data-from-influxdb-compatible-agents-such-as-telegraf) for more details.
+
+`vmagent` is also available in [docker images](https://hub.docker.com/r/victoriametrics/vmagent/).
+
+Pass `-help` to `vmagent` in order to see the full list of supported command-line flags with their descriptions.
+
+
+### Use cases
+
+
+#### IoT and Edge monitoring
+
+`vmagent` can run and collect metrics in IoT and industrial networks with unreliable or scheduled connections to the remote storage.
+It buffers the collected data in local files until the connection to remote storage becomes available and then sends the buffered
+data to the remote storage. It re-tries sending the data to remote storage on any errors.
+The maximum buffer size can be limited with `-remoteWrite.maxDiskUsagePerURL`.
+
+`vmagent` works on various architectures from IoT world - 32-bit arm, 64-bit arm, ppc64, 386, amd64.
+See [the corresponding Makefile rules](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/Makefile) for details.
+
+
+#### Drop-in replacement for Prometheus
+
+If you use Prometheus only for scraping metrics from various targets and forwarding these metrics to remote storage,
+then `vmagent` can replace such Prometheus setup. Usually `vmagent` requires lower amounts of RAM, CPU and network bandwidth comparing to Prometheus for such setup.
+See [these docs](#how-to-collect-metrics-in-prometheus-format) for details.
+
+
+#### Replication and high availability
+
+`vmagent` replicates the collected metrics among multiple remote storage instances configured via `-remoteWrite.url` args.
+If a single remote storage instance temporarily goes out of service, then the collected data remains available in another remote storage instances.
+`vmagent` buffers the collected data in files at `-remoteWrite.tmpDataPath` until the remote storage becomes available again.
+Then it sends the buffered data to the remote storage in order to prevent data gaps in the remote storage.
+
+
+#### Relabeling and filtering
+
+`vmagent` can add, remove or update labels on the collected data before sending it to remote storage. Additionally,
+it can remove unneeded samples via Prometheus-like relabeling before sending the collected data to remote storage.
+See [these docs](#relabeling) for details.
+
+
+#### Splitting data streams among multiple systems
+
+`vmagent` supports splitting of the collected data among muliple destinations with the help of `-remoteWrite.urlRelabelConfig`,
+which is applied independently for each configured `-remoteWrite.url` destination. For instance, it is possible to replicate or split
+data among long-term remote storage, short-term remote storage and real-time analytical system [built on top of Kafka](https://github.com/Telefonica/prometheus-kafka-adapter).
+Note that each destination can receive its own subset of the collected data thanks to per-destination relabeling via `-remoteWrite.urlRelabelConfig`.
+
+
+
+### How to collect metrics in Prometheus format
+
+Pass the path to `prometheus.yml` to `-promscrape.config` command-line flag. `vmagent` takes into account the following
+sections from [Prometheus config file](https://prometheus.io/docs/prometheus/latest/configuration/configuration/):
+
+* `global`
+* `scrape_configs`
+
+All the other sections are ignored, including [remote_write](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write) section.
+Use `-remoteWrite.*` command-line flags instead for configuring remote write settings.
+
+The following scrape types in [scrape_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config) section are supported:
+
+* `static_configs` - for scraping statically defined targets. See [these docs](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#static_config) for details.
+* `file_sd_configs` - for scraping targets defined in external files aka file-based service discover.
+  See [these docs](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#file_sd_config) for details.
+
+The following service discovery mechanisms will be added to `vmagent` soon:
+
+* [kubernetes_sd_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#kubernetes_sd_config)
+* [ec2_sd_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#ec2_sd_config)
+* [gce_sd_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#gce_sd_config)
+* [consul_sd_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#consul_sd_config)
+* [dns_sd_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#dns_sd_config)
+
+
+File feature requests at [our issue tracker](https://github.com/VictoriaMetrics/VictoriaMetrics/issues) if you need other service discovery mechanisms to be supported by `vmagent`.
+
+
+### Adding labels to metrics
+
+Labels can be added to metrics via the following mechanisms:
+
+* Via `global -> external_labels` section in `-promscrape.config` file. These labels are added only to metrics scraped from targets configured in `-promscrape.config` file.
+* Via `-remoteWrite.label` command-line flag. These labels are added to all the collected metrics before sending them to `-remoteWrite.url`.
+
+
+### Relabeling
+
+`vmagent` supports [Prometheus relabeling](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#relabel_config).
+Additionally it provides the following extra actions:
+
+* `replace_all`: replaces all the occurences of `regex` in the values of `source_labels` with the `replacement` and stores the result in the `target_label`.
+* `labelmap_all`: replaces all the occurences of `regex` in all the label names with the `replacement`.
+
+The relabeling can be defined in the following places:
+
+* At `scrape_config -> relabel_configs` section in `-promscrape.config` file. This relabeling is applied to targets when parsing the file during `vmagent` startup
+  or during config reload after sending `SIGHUP` signal to `vmagent`  via `kill -HUP`.
+* At `scrape_config -> metric_relabel_configs` section in `-promscrape.config` file. This relabeling is applied to metrics after each scrape for the configured targets.
+* At `-remoteWrite.relabelConfig` file. This relabeling is aplied to all the collected metrics before sending them to remote storage.
+* At `-remoteWrite.urlRelabelConfig` files. This relabeling is applied to metrics before sending them to the corresponding `-remoteWrite.url`.
+
+Read more about relabeling in the following articles:
+
+* [Life of a label](https://www.robustperception.io/life-of-a-label)
+* [Discarding targets and timeseries with relabeling](https://www.robustperception.io/relabelling-can-discard-targets-timeseries-and-alerts)
+* [Dropping labels at scrape time](https://www.robustperception.io/dropping-metrics-at-scrape-time-with-prometheus)
+* [Extracting labels from legacy metric names](https://www.robustperception.io/extracting-labels-from-legacy-metric-names)
+* [relabel_configs vs metric_relabel_configs](https://www.robustperception.io/relabel_configs-vs-metric_relabel_configs)
+
+
+### Monitoring
+
+`vmagent` exports various metrics in Prometheus exposition format at `http://vmagent-host:8429/metrics` page. It is recommended setting up regular scraping of this page
+either via `vmagent` itself or via Prometheus, so the exported metrics could be analyzed later.
+
+`vmagent` also exports target statuses at `http://vmagent-host:8429/targets` page in plaintext format.
+
+
+### Troubleshooting
+
+* It is recommended increasing the maximum number of open files in the system (`ulimit -n`) when scraping big number of targets,
+  since `vmagent` establishes at least a single TCP connection per each target.
+
+* It is recommended increasing `-remoteWrite.queues` if `vmagent` collects more than 100K samples per second
+  and `vmagent_remotewrite_pending_data_bytes` metric exported by `vmagent` at `/metrics` page constantly grows.
+
+* `vmagent` buffers scraped data at `-remoteWrite.tmpDataPath` directory until it is sent to `-remoteWrite.url`.
+  The directory can grow big when remote storage is unvailable during extended periods of time and if `-remoteWrite.maxDiskUsagePerURL` isn't set.
+  If you don't want sending all the data from the directory to remote storage, just stop `vmagent` and delete the directory.
+
+
+### How to build from sources
+
+It is recommended using [binary releases](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) - `vmagent` is located in `vmutils-*` archives there.
+
+
+#### Development build
+
+1. [Install Go](https://golang.org/doc/install). The minimum supported version is Go 1.12.
+2. Run `make vmagent` from the root folder of the repository.
+   It builds `vmagent` binary and puts it into the `bin` folder.
+
+#### Production build
+
+1. [Install docker](https://docs.docker.com/install/).
+2. Run `make vmagent-prod` from the root folder of the repository.
+   It builds `vmagent-prod` binary and puts it into the `bin` folder.
+
+#### Building docker images
+
+Run `make package-vmagent`. It builds `victoriametrics/vmagent:<PKG_TAG>` docker image locally.
+`<PKG_TAG>` is auto-generated image tag, which depends on source code in the repository.
+The `<PKG_TAG>` may be manually set via `PKG_TAG=foobar make package-vmagent`.

app/vmagent/common/push_ctx.go

@@ -0,0 +1,70 @@
+package common
+
+import (
+	"runtime"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+)
+
+// PushCtx is a context used for populating WriteRequest.
+type PushCtx struct {
+	WriteRequest prompbmarshal.WriteRequest
+
+	// Labels contains flat list of all the labels used in WriteRequest.
+	Labels []prompbmarshal.Label
+
+	// Samples contains flat list of all the samples used in WriteRequest.
+	Samples []prompbmarshal.Sample
+}
+
+// Reset resets ctx.
+func (ctx *PushCtx) Reset() {
+	tss := ctx.WriteRequest.Timeseries
+	for i := range tss {
+		ts := &tss[i]
+		ts.Labels = nil
+		ts.Samples = nil
+	}
+	ctx.WriteRequest.Timeseries = ctx.WriteRequest.Timeseries[:0]
+
+	labels := ctx.Labels
+	for i := range labels {
+		label := &labels[i]
+		label.Name = ""
+		label.Value = ""
+	}
+	ctx.Labels = ctx.Labels[:0]
+
+	ctx.Samples = ctx.Samples[:0]
+}
+
+// GetPushCtx returns PushCtx from pool.
+//
+// Call PutPushCtx when the ctx is no longer needed.
+func GetPushCtx() *PushCtx {
+	select {
+	case ctx := <-pushCtxPoolCh:
+		return ctx
+	default:
+		if v := pushCtxPool.Get(); v != nil {
+			return v.(*PushCtx)
+		}
+		return &PushCtx{}
+	}
+}
+
+// PutPushCtx returns ctx to the pool.
+//
+// ctx mustn't be used after returning to the pool.
+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/vmagent/csvimport/request_handler.go

@@ -0,0 +1,63 @@
+package csvimport
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/csvimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="csvimport"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="csvimport"}`)
+)
+
+// InsertHandler processes csv data from req.
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range rows {
+		r := &rows[i]
+		labelsLen := len(labels)
+		labels = append(labels, prompbmarshal.Label{
+			Name:  "__name__",
+			Value: r.Metric,
+		})
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  tag.Key,
+				Value: tag.Value,
+			})
+		}
+		samples = append(samples, prompbmarshal.Sample{
+			Value:     r.Value,
+			Timestamp: r.Timestamp,
+		})
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[len(samples)-1:],
+		})
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return nil
+}

app/vmagent/deployment/Dockerfile

@@ -0,0 +1,8 @@
+ARG certs_image
+FROM $certs_image AS certs
+FROM scratch
+COPY --from=certs /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/ca-certificates.crt
+ARG src_binary
+COPY $src_binary ./vmagent-prod
+EXPOSE 8429
+ENTRYPOINT ["/vmagent-prod"]

app/vmagent/graphite/request_handler.go

@@ -0,0 +1,65 @@
+package graphite
+
+import (
+	"io"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/graphite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="graphite"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="graphite"}`)
+)
+
+// InsertHandler processes remote write for graphite plaintext protocol.
+//
+// See https://graphite.readthedocs.io/en/latest/feeding-carbon.html#the-plaintext-protocol
+func InsertHandler(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range rows {
+		r := &rows[i]
+		labelsLen := len(labels)
+		labels = append(labels, prompbmarshal.Label{
+			Name:  "__name__",
+			Value: r.Metric,
+		})
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  tag.Key,
+				Value: tag.Value,
+			})
+		}
+		samples = append(samples, prompbmarshal.Sample{
+			Value:     r.Value,
+			Timestamp: r.Timestamp,
+		})
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[len(samples)-1:],
+		})
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return nil
+}

app/vmagent/influx/request_handler.go

@@ -0,0 +1,167 @@
+package influx
+
+import (
+	"flag"
+	"io"
+	"net/http"
+	"runtime"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/influx"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	measurementFieldSeparator = flag.String("influxMeasurementFieldSeparator", "_", "Separator for '{measurement}{separator}{field_name}' metric name when inserted via Influx line protocol")
+	skipSingleField           = flag.Bool("influxSkipSingleField", false, "Uses '{measurement}' instead of '{measurement}{separator}{field_name}' for metic name if Influx line contains only a single field")
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="influx"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="influx"}`)
+)
+
+// InsertHandlerForReader processes remote write for influx line protocol.
+//
+// See https://github.com/influxdata/telegraf/tree/master/plugins/inputs/socket_listener/
+func InsertHandlerForReader(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, false, "", "", insertRows)
+	})
+}
+
+// InsertHandlerForHTTP processes remote write for influx line protocol.
+//
+// See https://github.com/influxdata/influxdb/blob/4cbdc197b8117fee648d62e2e5be75c6575352f0/tsdb/README.md
+func InsertHandlerForHTTP(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		isGzipped := req.Header.Get("Content-Encoding") == "gzip"
+		q := req.URL.Query()
+		precision := q.Get("precision")
+		// Read db tag from https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint
+		db := q.Get("db")
+		return parser.ParseStream(req.Body, isGzipped, precision, db, insertRows)
+	})
+}
+
+func insertRows(db string, rows []parser.Row) error {
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+
+	rowsTotal := 0
+	tssDst := ctx.ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.ctx.Labels[:0]
+	samples := ctx.ctx.Samples[:0]
+	commonLabels := ctx.commonLabels[:0]
+	buf := ctx.buf[:0]
+	for i := range rows {
+		r := &rows[i]
+		commonLabels = commonLabels[:0]
+		hasDBLabel := false
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			if tag.Key == "db" {
+				hasDBLabel = true
+			}
+			commonLabels = append(commonLabels, prompbmarshal.Label{
+				Name:  tag.Key,
+				Value: tag.Value,
+			})
+		}
+		if len(db) > 0 && !hasDBLabel {
+			commonLabels = append(commonLabels, prompbmarshal.Label{
+				Name:  "db",
+				Value: db,
+			})
+		}
+		ctx.metricGroupBuf = append(ctx.metricGroupBuf[:0], r.Measurement...)
+		skipFieldKey := len(r.Fields) == 1 && *skipSingleField
+		if len(ctx.metricGroupBuf) > 0 && !skipFieldKey {
+			ctx.metricGroupBuf = append(ctx.metricGroupBuf, *measurementFieldSeparator...)
+		}
+		for j := range r.Fields {
+			f := &r.Fields[j]
+			bufLen := len(buf)
+			buf = append(buf, ctx.metricGroupBuf...)
+			if !skipFieldKey {
+				buf = append(buf, f.Key...)
+			}
+			metricGroup := bytesutil.ToUnsafeString(buf[bufLen:])
+			labelsLen := len(labels)
+			labels = append(labels, prompbmarshal.Label{
+				Name:  "__name__",
+				Value: metricGroup,
+			})
+			labels = append(labels, commonLabels...)
+			samples = append(samples, prompbmarshal.Sample{
+				Timestamp: r.Timestamp,
+				Value:     f.Value,
+			})
+			tssDst = append(tssDst, prompbmarshal.TimeSeries{
+				Labels:  labels[labelsLen:],
+				Samples: samples[len(samples)-1:],
+			})
+		}
+		rowsTotal += len(r.Fields)
+	}
+	ctx.buf = buf
+	ctx.ctx.WriteRequest.Timeseries = tssDst
+	ctx.ctx.Labels = labels
+	ctx.ctx.Samples = samples
+	ctx.commonLabels = commonLabels
+	remotewrite.Push(&ctx.ctx.WriteRequest)
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+
+	return nil
+}
+
+type pushCtx struct {
+	ctx            common.PushCtx
+	commonLabels   []prompbmarshal.Label
+	metricGroupBuf []byte
+	buf            []byte
+}
+
+func (ctx *pushCtx) reset() {
+	ctx.ctx.Reset()
+
+	commonLabels := ctx.commonLabels
+	for i := range commonLabels {
+		label := &commonLabels[i]
+		label.Name = ""
+		label.Value = ""
+	}
+
+	ctx.metricGroupBuf = ctx.metricGroupBuf[:0]
+	ctx.buf = ctx.buf[:0]
+}
+
+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/vmagent/main.go

@@ -0,0 +1,180 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/csvimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/graphite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/influx"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/opentsdb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/opentsdbhttp"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/promremotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/vmimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	graphiteserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/graphite"
+	influxserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/influx"
+	opentsdbserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/opentsdb"
+	opentsdbhttpserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/opentsdbhttp"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/procutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promscrape"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	httpListenAddr = flag.String("httpListenAddr", ":8429", "TCP address to listen for http connections. "+
+		"Set this flag to empty value in order to disable listening on any port. This mode may be useful for running multiple vmagent instances on the same server. "+
+		"Note that /targets and /metrics pages aren't available if -httpListenAddr=''")
+	influxListenAddr   = flag.String("influxListenAddr", "", "TCP and UDP address to listen for Influx line protocol data. Usually :8189 must be set. Doesn't work if empty")
+	graphiteListenAddr = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
+	opentsdbListenAddr = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB metrics. "+
+		"Telnet put messages and HTTP /api/put messages are simultaneously served on TCP port. "+
+		"Usually :4242 must be set. Doesn't work if empty")
+	opentsdbHTTPListenAddr = flag.String("opentsdbHTTPListenAddr", "", "TCP address to listen for OpentTSDB HTTP put requests. Usually :4242 must be set. Doesn't work if empty")
+)
+
+var (
+	influxServer       *influxserver.Server
+	graphiteServer     *graphiteserver.Server
+	opentsdbServer     *opentsdbserver.Server
+	opentsdbhttpServer *opentsdbhttpserver.Server
+)
+
+func main() {
+	envflag.Parse()
+	buildinfo.Init()
+	logger.Init()
+	logger.Infof("starting vmagent at %q...", *httpListenAddr)
+	startTime := time.Now()
+	remotewrite.Init()
+	writeconcurrencylimiter.Init()
+	if len(*influxListenAddr) > 0 {
+		influxServer = influxserver.MustStart(*influxListenAddr, influx.InsertHandlerForReader)
+	}
+	if len(*graphiteListenAddr) > 0 {
+		graphiteServer = graphiteserver.MustStart(*graphiteListenAddr, graphite.InsertHandler)
+	}
+	if len(*opentsdbListenAddr) > 0 {
+		opentsdbServer = opentsdbserver.MustStart(*opentsdbListenAddr, opentsdb.InsertHandler, opentsdbhttp.InsertHandler)
+	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttpServer = opentsdbhttpserver.MustStart(*opentsdbHTTPListenAddr, opentsdbhttp.InsertHandler)
+	}
+
+	promscrape.Init(remotewrite.Push)
+
+	if len(*httpListenAddr) > 0 {
+		go httpserver.Serve(*httpListenAddr, requestHandler)
+	}
+	logger.Infof("started vmagent in %.3f seconds", time.Since(startTime).Seconds())
+
+	sig := procutil.WaitForSigterm()
+	logger.Infof("received signal %s", sig)
+
+	startTime = time.Now()
+	if len(*httpListenAddr) > 0 {
+		logger.Infof("gracefully shutting down webservice at %q", *httpListenAddr)
+		if err := httpserver.Stop(*httpListenAddr); err != nil {
+			logger.Fatalf("cannot stop the webservice: %s", err)
+		}
+		logger.Infof("successfully shut down the webservice in %.3f seconds", time.Since(startTime).Seconds())
+	}
+
+	promscrape.Stop()
+
+	if len(*influxListenAddr) > 0 {
+		influxServer.MustStop()
+	}
+	if len(*graphiteListenAddr) > 0 {
+		graphiteServer.MustStop()
+	}
+	if len(*opentsdbListenAddr) > 0 {
+		opentsdbServer.MustStop()
+	}
+	if len(*opentsdbHTTPListenAddr) > 0 {
+		opentsdbhttpServer.MustStop()
+	}
+	remotewrite.Stop()
+
+	logger.Infof("successfully stopped vmagent in %.3f seconds", time.Since(startTime).Seconds())
+}
+
+func requestHandler(w http.ResponseWriter, r *http.Request) bool {
+	path := strings.Replace(r.URL.Path, "//", "/", -1)
+	switch path {
+	case "/api/v1/write":
+		prometheusWriteRequests.Inc()
+		if err := promremotewrite.InsertHandler(r); err != nil {
+			prometheusWriteErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
+	case "/api/v1/import":
+		vmimportRequests.Inc()
+		if err := vmimport.InsertHandler(r); err != nil {
+			vmimportErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
+	case "/api/v1/import/csv":
+		csvimportRequests.Inc()
+		if err := csvimport.InsertHandler(r); err != nil {
+			csvimportErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
+	case "/write", "/api/v2/write":
+		influxWriteRequests.Inc()
+		if err := influx.InsertHandlerForHTTP(r); err != nil {
+			influxWriteErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
+	case "/query":
+		// Emulate fake response for influx query.
+		// This is required for TSBS benchmark.
+		influxQueryRequests.Inc()
+		fmt.Fprintf(w, `{"results":[{"series":[{"values":[]}]}]}`)
+		return true
+	case "/targets":
+		promscrapeTargetsRequests.Inc()
+		w.Header().Set("Content-Type", "text/plain")
+		promscrape.WriteHumanReadableTargetsStatus(w)
+		return true
+	}
+	return false
+}
+
+var (
+	prometheusWriteRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/api/v1/write", protocol="promremotewrite"}`)
+	prometheusWriteErrors   = metrics.NewCounter(`vmagent_http_request_errors_total{path="/api/v1/write", protocol="promremotewrite"}`)
+
+	vmimportRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/api/v1/import", protocol="vmimport"}`)
+	vmimportErrors   = metrics.NewCounter(`vmagent_http_request_errors_total{path="/api/v1/import", protocol="vmimport"}`)
+
+	csvimportRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/api/v1/import/csv", protocol="csvimport"}`)
+	csvimportErrors   = metrics.NewCounter(`vmagent_http_request_errors_total{path="/api/v1/import/csv", protocol="csvimport"}`)
+
+	influxWriteRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/write", protocol="influx"}`)
+	influxWriteErrors   = metrics.NewCounter(`vmagent_http_request_errors_total{path="/write", protocol="influx"}`)
+
+	influxQueryRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/query", protocol="influx"}`)
+
+	promscrapeTargetsRequests = metrics.NewCounter(`vmagent_http_requests_total{path="/targets"}`)
+)

app/vmagent/opentsdb/request_handler.go

@@ -0,0 +1,65 @@
+package opentsdb
+
+import (
+	"io"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/opentsdb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="opentsdb"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="opentsdb"}`)
+)
+
+// InsertHandler processes remote write for OpenTSDB put protocol.
+//
+// See http://opentsdb.net/docs/build/html/api_telnet/put.html
+func InsertHandler(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range rows {
+		r := &rows[i]
+		labelsLen := len(labels)
+		labels = append(labels, prompbmarshal.Label{
+			Name:  "__name__",
+			Value: r.Metric,
+		})
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  tag.Key,
+				Value: tag.Value,
+			})
+		}
+		samples = append(samples, prompbmarshal.Sample{
+			Value:     r.Value,
+			Timestamp: r.Timestamp,
+		})
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[len(samples)-1:],
+		})
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return nil
+}

app/vmagent/opentsdbhttp/request_handler.go

@@ -0,0 +1,64 @@
+package opentsdbhttp
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/opentsdbhttp"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="opentsdbhttp"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="opentsdbhttp"}`)
+)
+
+// InsertHandler processes HTTP OpenTSDB put requests.
+// See http://opentsdb.net/docs/build/html/api_http/put.html
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range rows {
+		r := &rows[i]
+		labelsLen := len(labels)
+		labels = append(labels, prompbmarshal.Label{
+			Name:  "__name__",
+			Value: r.Metric,
+		})
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  tag.Key,
+				Value: tag.Value,
+			})
+		}
+		samples = append(samples, prompbmarshal.Sample{
+			Value:     r.Value,
+			Timestamp: r.Timestamp,
+		})
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[len(samples)-1:],
+		})
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return nil
+}

app/vmagent/promremotewrite/request_handler.go

@@ -0,0 +1,67 @@
+package promremotewrite
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/promremotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="promremotewrite"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="promremotewrite"}`)
+)
+
+// InsertHandler processes remote write for prometheus.
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(timeseries []prompb.TimeSeries) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	rowsTotal := 0
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range timeseries {
+		ts := &timeseries[i]
+		labelsLen := len(labels)
+		for i := range ts.Labels {
+			label := &ts.Labels[i]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  bytesutil.ToUnsafeString(label.Name),
+				Value: bytesutil.ToUnsafeString(label.Value),
+			})
+		}
+		samplesLen := len(samples)
+		for i := range ts.Samples {
+			sample := &ts.Samples[i]
+			samples = append(samples, prompbmarshal.Sample{
+				Value:     sample.Value,
+				Timestamp: sample.Timestamp,
+			})
+		}
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[samplesLen:],
+		})
+		rowsTotal += len(ts.Samples)
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	return nil
+}

app/vmagent/remotewrite/client.go

@@ -0,0 +1,268 @@
+package remotewrite
+
+import (
+	"crypto/tls"
+	"crypto/x509"
+	"encoding/base64"
+	"flag"
+	"fmt"
+	"io/ioutil"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/persistentqueue"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fasthttp"
+)
+
+var (
+	sendTimeout = flag.Duration("remoteWrite.sendTimeout", time.Minute, "Timeout for sending a single block of data to -remoteWrite.url")
+
+	tlsInsecureSkipVerify = flag.Bool("remoteWrite.tlsInsecureSkipVerify", false, "Whether to skip tls verification when connecting to -remoteWrite.url")
+	tlsCertFile           = flag.String("remoteWrite.tlsCertFile", "", "Optional path to client-side TLS certificate file to use when connecting to -remoteWrite.url")
+	tlsKeyFile            = flag.String("remoteWrite.tlsKeyFile", "", "Optional path to client-side TLS certificate key to use when connecting to -remoteWrite.url")
+	tlsCAFile             = flag.String("remoteWrite.tlsCAFile", "", "Optional path to TLS CA file to use for verifying connections to -remoteWrite.url. "+
+		"By default system CA is used")
+
+	basicAuthUsername = flag.String("remoteWrite.basicAuth.username", "", "Optional basic auth username to use for -remoteWrite.url")
+	basicAuthPassword = flag.String("remoteWrite.basicAuth.password", "", "Optional basic auth password to use for -remoteWrite.url")
+	bearerToken       = flag.String("remoteWrite.bearerToken", "", "Optional bearer auth token to use for -remoteWrite.url")
+)
+
+type client struct {
+	urlLabelValue  string
+	remoteWriteURL string
+	host           string
+	requestURI     string
+	authHeader     string
+	fq             *persistentqueue.FastQueue
+	hc             *fasthttp.HostClient
+
+	requestDuration *metrics.Histogram
+	requestsOKCount *metrics.Counter
+	errorsCount     *metrics.Counter
+	retriesCount    *metrics.Counter
+
+	wg     sync.WaitGroup
+	stopCh chan struct{}
+}
+
+func newClient(remoteWriteURL, urlLabelValue string, fq *persistentqueue.FastQueue, concurrency int) *client {
+	authHeader := ""
+	if len(*basicAuthUsername) > 0 || len(*basicAuthPassword) > 0 {
+		// See https://en.wikipedia.org/wiki/Basic_access_authentication
+		token := *basicAuthUsername + ":" + *basicAuthPassword
+		token64 := base64.StdEncoding.EncodeToString([]byte(token))
+		authHeader = "Basic " + token64
+	}
+	if len(*bearerToken) > 0 {
+		if authHeader != "" {
+			logger.Panicf("FATAL: `-remoteWrite.bearerToken`=%q cannot be set when `-remoteWrite.basicAuth.*` flags are set", *bearerToken)
+		}
+		authHeader = "Bearer " + *bearerToken
+	}
+
+	readTimeout := *sendTimeout
+	if readTimeout <= 0 {
+		readTimeout = time.Minute
+	}
+	writeTimeout := readTimeout
+	var u fasthttp.URI
+	u.Update(remoteWriteURL)
+	scheme := string(u.Scheme())
+	switch scheme {
+	case "http", "https":
+	default:
+		logger.Panicf("FATAL: unsupported scheme in -remoteWrite.url=%q: %q. It must be http or https", remoteWriteURL, scheme)
+	}
+	host := string(u.Host())
+	if len(host) == 0 {
+		logger.Panicf("FATAL: invalid -remoteWrite.url=%q: host cannot be empty. Make sure the url looks like `http://host:port/path`", remoteWriteURL)
+	}
+	requestURI := string(u.RequestURI())
+	isTLS := scheme == "https"
+	var tlsCfg *tls.Config
+	if isTLS {
+		var err error
+		tlsCfg, err = getTLSConfig()
+		if err != nil {
+			logger.Panicf("FATAL: cannot initialize TLS config: %s", err)
+		}
+	}
+	if !strings.Contains(host, ":") {
+		if isTLS {
+			host += ":443"
+		} else {
+			host += ":80"
+		}
+	}
+	maxConns := 2 * concurrency
+	hc := &fasthttp.HostClient{
+		Addr:                host,
+		Name:                "vmagent",
+		Dial:                statDial,
+		DialDualStack:       netutil.TCP6Enabled(),
+		IsTLS:               isTLS,
+		TLSConfig:           tlsCfg,
+		MaxConns:            maxConns,
+		MaxIdleConnDuration: 10 * readTimeout,
+		ReadTimeout:         readTimeout,
+		WriteTimeout:        writeTimeout,
+		MaxResponseBodySize: 1024 * 1024,
+	}
+	c := &client{
+		urlLabelValue:  urlLabelValue,
+		remoteWriteURL: remoteWriteURL,
+		host:           host,
+		requestURI:     requestURI,
+		authHeader:     authHeader,
+		fq:             fq,
+		hc:             hc,
+		stopCh:         make(chan struct{}),
+	}
+	c.requestDuration = metrics.GetOrCreateHistogram(fmt.Sprintf(`vmagent_remotewrite_duration_seconds{url=%q}`, c.urlLabelValue))
+	c.requestsOKCount = metrics.GetOrCreateCounter(fmt.Sprintf(`vmagent_remotewrite_requests_total{url=%q, status_code="2XX"}`, c.urlLabelValue))
+	c.errorsCount = metrics.GetOrCreateCounter(fmt.Sprintf(`vmagent_remotewrite_errors_total{url=%q}`, c.urlLabelValue))
+	c.retriesCount = metrics.GetOrCreateCounter(fmt.Sprintf(`vmagent_remotewrite_retries_count_total{url=%q}`, c.urlLabelValue))
+	for i := 0; i < concurrency; i++ {
+		c.wg.Add(1)
+		go func() {
+			defer c.wg.Done()
+			c.runWorker()
+		}()
+	}
+	logger.Infof("initialized client for -remoteWrite.url=%q", c.remoteWriteURL)
+	return c
+}
+
+func (c *client) MustStop() {
+	close(c.stopCh)
+	c.wg.Wait()
+	logger.Infof("stopped client for -remoteWrite.url=%q", c.remoteWriteURL)
+}
+
+func getTLSConfig() (*tls.Config, error) {
+	var tlsRootCA *x509.CertPool
+	var tlsCertificate *tls.Certificate
+	if *tlsCertFile != "" || *tlsKeyFile != "" {
+		cert, err := tls.LoadX509KeyPair(*tlsCertFile, *tlsKeyFile)
+		if err != nil {
+			return nil, fmt.Errorf("cannot load TLS certificate for -remoteWrite.tlsCertFile=%q and -remoteWrite.tlsKeyFile=%q: %s", *tlsCertFile, *tlsKeyFile, err)
+		}
+		tlsCertificate = &cert
+	}
+	if *tlsCAFile != "" {
+		data, err := ioutil.ReadFile(*tlsCAFile)
+		if err != nil {
+			return nil, fmt.Errorf("cannot read -remoteWrite.tlsCAFile=%q: %s", *tlsCAFile, err)
+		}
+		tlsRootCA = x509.NewCertPool()
+		if !tlsRootCA.AppendCertsFromPEM(data) {
+			return nil, fmt.Errorf("cannot parse data -remoteWrite.tlsCAFile=%q", *tlsCAFile)
+		}
+	}
+	tlsCfg := &tls.Config{
+		RootCAs:            tlsRootCA,
+		ClientSessionCache: tls.NewLRUClientSessionCache(0),
+	}
+	if tlsCertificate != nil {
+		tlsCfg.Certificates = []tls.Certificate{*tlsCertificate}
+	}
+	tlsCfg.InsecureSkipVerify = *tlsInsecureSkipVerify
+	return tlsCfg, nil
+}
+
+func (c *client) runWorker() {
+	var ok bool
+	var block []byte
+	ch := make(chan struct{})
+	for {
+		block, ok = c.fq.MustReadBlock(block[:0])
+		if !ok {
+			return
+		}
+		go func() {
+			c.sendBlock(block)
+			ch <- struct{}{}
+		}()
+		select {
+		case <-ch:
+			// The block has been sent successfully
+			continue
+		case <-c.stopCh:
+			// c must be stopped. Wait for a while in the hope the block will be sent.
+			graceDuration := 5 * time.Second
+			select {
+			case <-ch:
+				// The block has been sent successfully.
+			case <-time.After(graceDuration):
+				logger.Errorf("couldn't sent block with size %d bytes to %q in %.3f seconds during shutdown; dropping it",
+					len(block), c.remoteWriteURL, graceDuration.Seconds())
+			}
+			return
+		}
+	}
+}
+
+func (c *client) sendBlock(block []byte) {
+	req := fasthttp.AcquireRequest()
+	req.SetRequestURI(c.requestURI)
+	req.SetHost(c.host)
+	req.Header.SetMethod("POST")
+	req.Header.Add("Content-Type", "application/x-protobuf")
+	req.Header.Add("Content-Encoding", "snappy")
+	if c.authHeader != "" {
+		req.Header.Set("Authorization", c.authHeader)
+	}
+	req.SetBody(block)
+
+	retryDuration := time.Second
+	resp := fasthttp.AcquireResponse()
+
+again:
+	select {
+	case <-c.stopCh:
+		fasthttp.ReleaseRequest(req)
+		fasthttp.ReleaseResponse(resp)
+		return
+	default:
+	}
+
+	startTime := time.Now()
+	// There is no need in calling DoTimeout, since the timeout is set in c.hc.ReadTimeout.
+	err := c.hc.Do(req, resp)
+	c.requestDuration.UpdateDuration(startTime)
+	if err != nil {
+		c.errorsCount.Inc()
+		retryDuration *= 2
+		if retryDuration > time.Minute {
+			retryDuration = time.Minute
+		}
+		logger.Errorf("couldn't send a block with size %d bytes to %q: %s; re-sending the block in %.3f seconds",
+			len(block), c.remoteWriteURL, err, retryDuration.Seconds())
+		time.Sleep(retryDuration)
+		c.retriesCount.Inc()
+		goto again
+	}
+	statusCode := resp.StatusCode()
+	if statusCode/100 != 2 {
+		metrics.GetOrCreateCounter(fmt.Sprintf(`vmagent_remotewrite_requests_total{url=%q, status_code="%d"}`, c.urlLabelValue, statusCode)).Inc()
+		retryDuration *= 2
+		if retryDuration > time.Minute {
+			retryDuration = time.Minute
+		}
+		logger.Errorf("unexpected status code received after sending a block with size %d bytes to %q: %d; response body=%q; re-sending the block in %.3f seconds",
+			len(block), c.remoteWriteURL, statusCode, resp.Body(), retryDuration.Seconds())
+		time.Sleep(retryDuration)
+		c.retriesCount.Inc()
+		goto again
+	}
+	c.requestsOKCount.Inc()
+
+	// The block has been successfully sent to the remote storage.
+	fasthttp.ReleaseResponse(resp)
+	fasthttp.ReleaseRequest(req)
+}

app/vmagent/remotewrite/pendingseries.go

@@ -0,0 +1,199 @@
+package remotewrite
+
+import (
+	"flag"
+	"sync"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/persistentqueue"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/golang/snappy"
+)
+
+var (
+	flushInterval = flag.Duration("remoteWrite.flushInterval", time.Second, "Interval for flushing the data to remote storage. "+
+		"Higher value reduces network bandwidth usage at the cost of delayed push of scraped data to remote storage")
+	maxUnpackedBlockSize = flag.Int("remoteWrite.maxBlockSize", 32*1024*1024, "The maximum size in bytes of unpacked request to send to remote storage. "+
+		"It shouldn't exceed -maxInsertRequestSize from VictoriaMetrics")
+)
+
+// the maximum number of rows to send per each block.
+const maxRowsPerBlock = 10000
+
+type pendingSeries struct {
+	mu sync.Mutex
+	wr writeRequest
+
+	stopCh            chan struct{}
+	periodicFlusherWG sync.WaitGroup
+}
+
+func newPendingSeries(pushBlock func(block []byte)) *pendingSeries {
+	var ps pendingSeries
+	ps.wr.pushBlock = pushBlock
+	ps.stopCh = make(chan struct{})
+	ps.periodicFlusherWG.Add(1)
+	go func() {
+		defer ps.periodicFlusherWG.Done()
+		ps.periodicFlusher()
+	}()
+	return &ps
+}
+
+func (ps *pendingSeries) MustStop() {
+	close(ps.stopCh)
+	ps.periodicFlusherWG.Wait()
+}
+
+func (ps *pendingSeries) Push(tss []prompbmarshal.TimeSeries) {
+	ps.mu.Lock()
+	ps.wr.push(tss)
+	ps.mu.Unlock()
+}
+
+func (ps *pendingSeries) periodicFlusher() {
+	ticker := time.NewTicker(*flushInterval)
+	defer ticker.Stop()
+	mustStop := false
+	for !mustStop {
+		select {
+		case <-ps.stopCh:
+			mustStop = true
+		case <-ticker.C:
+			if time.Since(ps.wr.lastFlushTime) < *flushInterval/2 {
+				continue
+			}
+		}
+		ps.mu.Lock()
+		ps.wr.flush()
+		ps.mu.Unlock()
+	}
+}
+
+type writeRequest struct {
+	wr            prompbmarshal.WriteRequest
+	pushBlock     func(block []byte)
+	lastFlushTime time.Time
+
+	tss []prompbmarshal.TimeSeries
+
+	labels  []prompbmarshal.Label
+	samples []prompbmarshal.Sample
+	buf     []byte
+}
+
+func (wr *writeRequest) reset() {
+	wr.wr.Timeseries = nil
+
+	for i := range wr.tss {
+		ts := &wr.tss[i]
+		ts.Labels = nil
+		ts.Samples = nil
+	}
+	wr.tss = wr.tss[:0]
+
+	for i := range wr.labels {
+		label := &wr.labels[i]
+		label.Name = ""
+		label.Value = ""
+	}
+	wr.labels = wr.labels[:0]
+
+	wr.samples = wr.samples[:0]
+	wr.buf = wr.buf[:0]
+}
+
+func (wr *writeRequest) flush() {
+	wr.wr.Timeseries = wr.tss
+	wr.lastFlushTime = time.Now()
+	pushWriteRequest(&wr.wr, wr.pushBlock)
+	wr.reset()
+}
+
+func (wr *writeRequest) push(src []prompbmarshal.TimeSeries) {
+	tssDst := wr.tss
+	for i := range src {
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{})
+		dst := &tssDst[len(tssDst)-1]
+		wr.copyTimeSeries(dst, &src[i])
+		if len(wr.tss) >= maxRowsPerBlock {
+			wr.flush()
+			tssDst = wr.tss
+		}
+	}
+	wr.tss = tssDst
+}
+
+func (wr *writeRequest) copyTimeSeries(dst, src *prompbmarshal.TimeSeries) {
+	labelsDst := wr.labels
+	labelsLen := len(wr.labels)
+	samplesDst := wr.samples
+	buf := wr.buf
+	for i := range src.Labels {
+		labelsDst = append(labelsDst, prompbmarshal.Label{})
+		dstLabel := &labelsDst[len(labelsDst)-1]
+		srcLabel := &src.Labels[i]
+
+		buf = append(buf, srcLabel.Name...)
+		dstLabel.Name = bytesutil.ToUnsafeString(buf[len(buf)-len(srcLabel.Name):])
+		buf = append(buf, srcLabel.Value...)
+		dstLabel.Value = bytesutil.ToUnsafeString(buf[len(buf)-len(srcLabel.Value):])
+	}
+	dst.Labels = labelsDst[labelsLen:]
+
+	samplesDst = append(samplesDst, prompbmarshal.Sample{})
+	dstSample := &samplesDst[len(samplesDst)-1]
+	if len(src.Samples) != 1 {
+		logger.Panicf("BUG: unexpected number of samples in time series; got %d; want 1", len(src.Samples))
+	}
+	*dstSample = src.Samples[0]
+	dst.Samples = samplesDst[len(samplesDst)-1:]
+
+	wr.samples = samplesDst
+	wr.labels = labelsDst
+	wr.buf = buf
+}
+
+func pushWriteRequest(wr *prompbmarshal.WriteRequest, pushBlock func(block []byte)) {
+	if len(wr.Timeseries) == 0 {
+		// Nothing to push
+		return
+	}
+	bb := writeRequestBufPool.Get()
+	bb.B = prompbmarshal.MarshalWriteRequest(bb.B[:0], wr)
+	if len(bb.B) <= *maxUnpackedBlockSize {
+		zb := snappyBufPool.Get()
+		zb.B = snappy.Encode(zb.B[:cap(zb.B)], bb.B)
+		writeRequestBufPool.Put(bb)
+		if len(zb.B) <= persistentqueue.MaxBlockSize {
+			pushBlock(zb.B)
+			blockSizeRows.Update(float64(len(wr.Timeseries)))
+			blockSizeBytes.Update(float64(len(zb.B)))
+			snappyBufPool.Put(zb)
+			return
+		}
+		snappyBufPool.Put(zb)
+	} else {
+		writeRequestBufPool.Put(bb)
+	}
+
+	// Too big block. Recursively split it into smaller parts.
+	timeseries := wr.Timeseries
+	n := len(timeseries) / 2
+	wr.Timeseries = timeseries[:n]
+	pushWriteRequest(wr, pushBlock)
+	wr.Timeseries = timeseries[n:]
+	pushWriteRequest(wr, pushBlock)
+	wr.Timeseries = timeseries
+}
+
+var (
+	blockSizeBytes = metrics.NewHistogram(`vmagent_remotewrite_block_size_bytes`)
+	blockSizeRows  = metrics.NewHistogram(`vmagent_remotewrite_block_size_rows`)
+)
+
+var writeRequestBufPool bytesutil.ByteBufferPool
+var snappyBufPool bytesutil.ByteBufferPool

app/vmagent/remotewrite/relabel.go

@@ -0,0 +1,113 @@
+package remotewrite
+
+import (
+	"flag"
+	"strings"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promrelabel"
+)
+
+var (
+	unparsedLabelsGlobal = flagutil.NewArray("remoteWrite.label", "Optional label in the form 'name=value' to add to all the metrics before sending them to -remoteWrite.url. "+
+		"Pass multiple -remoteWrite.label flags in order to add multiple flags to metrics before sending them to remote storage")
+	relabelConfigPathGlobal = flag.String("remoteWrite.relabelConfig", "", "Optional path to file with relabel_config entries. These entries are applied to all the metrics "+
+		"before sending them to -remoteWrite.url. See https://prometheus.io/docs/prometheus/latest/configuration/configuration/#relabel_config for details")
+)
+
+var labelsGlobal []prompbmarshal.Label
+var prcsGlobal []promrelabel.ParsedRelabelConfig
+
+// initRelabelGlobal must be called after parsing command-line flags.
+func initRelabelGlobal() {
+	// Init labelsGlobal
+	labelsGlobal = nil
+	for _, s := range *unparsedLabelsGlobal {
+		n := strings.IndexByte(s, '=')
+		if n < 0 {
+			logger.Panicf("FATAL: missing '=' in `-remoteWrite.label`. It must contain label in the form `name=value`; got %q", s)
+		}
+		labelsGlobal = append(labelsGlobal, prompbmarshal.Label{
+			Name:  s[:n],
+			Value: s[n+1:],
+		})
+	}
+
+	// Init prcsGlobal
+	prcsGlobal = nil
+	if len(*relabelConfigPathGlobal) > 0 {
+		var err error
+		prcsGlobal, err = promrelabel.LoadRelabelConfigs(*relabelConfigPathGlobal)
+		if err != nil {
+			logger.Panicf("FATAL: cannot load relabel configs from -remoteWrite.relabelConfig=%q: %s", *relabelConfigPathGlobal, err)
+		}
+	}
+}
+
+func (rctx *relabelCtx) applyRelabeling(tss []prompbmarshal.TimeSeries, extraLabels []prompbmarshal.Label, prcs []promrelabel.ParsedRelabelConfig) []prompbmarshal.TimeSeries {
+	if len(extraLabels) == 0 && len(prcs) == 0 {
+		// Nothing to change.
+		return tss
+	}
+	tssDst := tss[:0]
+	labels := rctx.labels[:0]
+	for i := range tss {
+		ts := &tss[i]
+		labelsLen := len(labels)
+		labels = append(labels, ts.Labels...)
+		// extraLabels must be added before applying relabeling according to https://prometheus.io/docs/prometheus/latest/configuration/configuration/#remote_write
+		for j := range extraLabels {
+			extraLabel := &extraLabels[j]
+			tmp := promrelabel.GetLabelByName(labels[labelsLen:], extraLabel.Name)
+			if tmp != nil {
+				tmp.Value = extraLabel.Value
+			} else {
+				labels = append(labels, *extraLabel)
+			}
+		}
+		labels = promrelabel.ApplyRelabelConfigs(labels, labelsLen, prcs, true)
+		if len(labels) == labelsLen {
+			// Drop the current time series, since relabeling removed all the labels.
+			continue
+		}
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: ts.Samples,
+		})
+	}
+	rctx.labels = labels
+	return tssDst
+}
+
+type relabelCtx struct {
+	// pool for labels, which are used during the relabeling.
+	labels []prompbmarshal.Label
+}
+
+func (rctx *relabelCtx) reset() {
+	labels := rctx.labels
+	for i := range labels {
+		label := &labels[i]
+		label.Name = ""
+		label.Value = ""
+	}
+	rctx.labels = rctx.labels[:0]
+}
+
+var relabelCtxPool = &sync.Pool{
+	New: func() interface{} {
+		return &relabelCtx{}
+	},
+}
+
+func getRelabelCtx() *relabelCtx {
+	return relabelCtxPool.Get().(*relabelCtx)
+}
+
+func putRelabelCtx(rctx *relabelCtx) {
+	rctx.labels = rctx.labels[:0]
+	relabelCtxPool.Put(rctx)
+}

app/vmagent/remotewrite/remotewrite.go

@@ -0,0 +1,195 @@
+package remotewrite
+
+import (
+	"flag"
+	"fmt"
+	"sync/atomic"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/persistentqueue"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promrelabel"
+	"github.com/VictoriaMetrics/metrics"
+	xxhash "github.com/cespare/xxhash/v2"
+)
+
+var (
+	remoteWriteURLs = flagutil.NewArray("remoteWrite.url", "Remote storage URL to write data to. It must support Prometheus remote_write API. "+
+		"It is recommended using VictoriaMetrics as remote storage. Example url: http://<victoriametrics-host>:8428/api/v1/write . "+
+		"Pass multiple -remoteWrite.url flags in order to write data concurrently to multiple remote storage systems")
+	relabelConfigPaths = flagutil.NewArray("remoteWrite.urlRelabelConfig", "Optional path to relabel config for the corresponding -remoteWrite.url")
+	tmpDataPath        = flag.String("remoteWrite.tmpDataPath", "vmagent-remotewrite-data", "Path to directory where temporary data for remote write component is stored")
+	queues             = flag.Int("remoteWrite.queues", 1, "The number of concurrent queues to each -remoteWrite.url. Set more queues if a single queue "+
+		"isn't enough for sending high volume of collected data to remote storage")
+	showRemoteWriteURL = flag.Bool("remoteWrite.showURL", false, "Whether to show -remoteWrite.url in the exported metrics. "+
+		"It is hidden by default, since it can contain sensistive auth info")
+	maxPendingBytesPerURL = flag.Int("remoteWrite.maxDiskUsagePerURL", 0, "The maximum file-based buffer size in bytes at -remoteWrite.tmpDataPath "+
+		"for each -remoteWrite.url. When buffer size reaches the configured maximum, then old data is dropped when adding new data to the buffer. "+
+		"Buffered data is stored in ~500MB chunks, so the minimum practical value for this flag is 500000000. "+
+		"Disk usage is unlimited if the value is set to 0")
+)
+
+var rwctxs []*remoteWriteCtx
+
+// Init initializes remotewrite.
+//
+// It must be called after flag.Parse().
+//
+// Stop must be called for graceful shutdown.
+func Init() {
+	if len(*remoteWriteURLs) == 0 {
+		logger.Panicf("FATAL: at least one `-remoteWrite.url` must be set")
+	}
+
+	if !*showRemoteWriteURL {
+		// remoteWrite.url can contain authentication codes, so hide it at `/metrics` output.
+		httpserver.RegisterSecretFlag("remoteWrite.url")
+	}
+	initRelabelGlobal()
+
+	maxInmemoryBlocks := memory.Allowed() / len(*remoteWriteURLs) / maxRowsPerBlock / 100
+	if maxInmemoryBlocks > 200 {
+		// There is no much sense in keeping higher number of blocks in memory,
+		// since this means that the producer outperforms consumer and the queue
+		// will continue growing. It is better storing the queue to file.
+		maxInmemoryBlocks = 200
+	}
+	if maxInmemoryBlocks < 2 {
+		maxInmemoryBlocks = 2
+	}
+	for i, remoteWriteURL := range *remoteWriteURLs {
+		relabelConfigPath := ""
+		if i < len(*relabelConfigPaths) {
+			relabelConfigPath = (*relabelConfigPaths)[i]
+		}
+		urlLabelValue := fmt.Sprintf("secret-url-%d", i+1)
+		if *showRemoteWriteURL {
+			urlLabelValue = remoteWriteURL
+		}
+		rwctx := newRemoteWriteCtx(remoteWriteURL, relabelConfigPath, maxInmemoryBlocks, urlLabelValue)
+		rwctxs = append(rwctxs, rwctx)
+	}
+}
+
+// Stop stops remotewrite.
+//
+// It is expected that nobody calls Push during and after the call to this func.
+func Stop() {
+	for _, rwctx := range rwctxs {
+		rwctx.MustStop()
+	}
+	rwctxs = nil
+}
+
+// Push sends wr to remote storage systems set via `-remoteWrite.url`.
+//
+// Each timeseries in wr.Timeseries must contain one sample.
+func Push(wr *prompbmarshal.WriteRequest) {
+	var rctx *relabelCtx
+	if len(prcsGlobal) > 0 || len(labelsGlobal) > 0 {
+		rctx = getRelabelCtx()
+	}
+	tss := wr.Timeseries
+	for len(tss) > 0 {
+		// Process big tss in smaller blocks in order to reduce maxmimum memory usage
+		tssBlock := tss
+		if len(tssBlock) > maxRowsPerBlock {
+			tssBlock = tss[:maxRowsPerBlock]
+			tss = tss[maxRowsPerBlock:]
+		} else {
+			tss = nil
+		}
+		if rctx != nil {
+			tssBlockLen := len(tssBlock)
+			tssBlock = rctx.applyRelabeling(tssBlock, labelsGlobal, prcsGlobal)
+			globalRelabelMetricsDropped.Add(tssBlockLen - len(tssBlock))
+		}
+		for _, rwctx := range rwctxs {
+			rwctx.Push(tssBlock)
+		}
+		if rctx != nil {
+			rctx.reset()
+		}
+	}
+	if rctx != nil {
+		putRelabelCtx(rctx)
+	}
+}
+
+var globalRelabelMetricsDropped = metrics.NewCounter("vmagent_remotewrite_global_relabel_metrics_dropped_total")
+
+type remoteWriteCtx struct {
+	fq         *persistentqueue.FastQueue
+	c          *client
+	prcs       []promrelabel.ParsedRelabelConfig
+	pss        []*pendingSeries
+	pssNextIdx uint64
+
+	relabelMetricsDropped *metrics.Counter
+}
+
+func newRemoteWriteCtx(remoteWriteURL, relabelConfigPath string, maxInmemoryBlocks int, urlLabelValue string) *remoteWriteCtx {
+	h := xxhash.Sum64([]byte(remoteWriteURL))
+	path := fmt.Sprintf("%s/persistent-queue/%016X", *tmpDataPath, h)
+	fq := persistentqueue.MustOpenFastQueue(path, remoteWriteURL, maxInmemoryBlocks, *maxPendingBytesPerURL)
+	_ = metrics.GetOrCreateGauge(fmt.Sprintf(`vmagent_remotewrite_pending_data_bytes{path=%q, url=%q}`, path, urlLabelValue), func() float64 {
+		return float64(fq.GetPendingBytes())
+	})
+	_ = metrics.GetOrCreateGauge(fmt.Sprintf(`vmagent_remotewrite_pending_inmemory_blocks{path=%q, url=%q}`, path, urlLabelValue), func() float64 {
+		return float64(fq.GetInmemoryQueueLen())
+	})
+	c := newClient(remoteWriteURL, urlLabelValue, fq, *queues)
+	var prcs []promrelabel.ParsedRelabelConfig
+	if len(relabelConfigPath) > 0 {
+		var err error
+		prcs, err = promrelabel.LoadRelabelConfigs(relabelConfigPath)
+		if err != nil {
+			logger.Panicf("FATAL: cannot load relabel configs from -remoteWrite.urlRelabelConfig=%q: %s", relabelConfigPath, err)
+		}
+	}
+	pss := make([]*pendingSeries, *queues)
+	for i := range pss {
+		pss[i] = newPendingSeries(fq.MustWriteBlock)
+	}
+	return &remoteWriteCtx{
+		fq:   fq,
+		c:    c,
+		prcs: prcs,
+		pss:  pss,
+
+		relabelMetricsDropped: metrics.GetOrCreateCounter(fmt.Sprintf(`vmagent_remotewrite_relabel_metrics_dropped_total{path=%q, url=%q}`, path, urlLabelValue)),
+	}
+}
+
+func (rwctx *remoteWriteCtx) MustStop() {
+	for _, ps := range rwctx.pss {
+		ps.MustStop()
+	}
+	rwctx.pss = nil
+	rwctx.fq.MustClose()
+	rwctx.fq = nil
+	rwctx.prcs = nil
+	rwctx.c.MustStop()
+	rwctx.c = nil
+
+	rwctx.relabelMetricsDropped = nil
+}
+
+func (rwctx *remoteWriteCtx) Push(tss []prompbmarshal.TimeSeries) {
+	var rctx *relabelCtx
+	if len(rwctx.prcs) > 0 {
+		rctx = getRelabelCtx()
+		tssLen := len(tss)
+		tss = rctx.applyRelabeling(tss, nil, rwctx.prcs)
+		rwctx.relabelMetricsDropped.Add(tssLen - len(tss))
+	}
+	pss := rwctx.pss
+	idx := atomic.AddUint64(&rwctx.pssNextIdx, 1) % uint64(len(pss))
+	pss[idx].Push(tss)
+	if rctx != nil {
+		putRelabelCtx(rctx)
+	}
+}

app/vmagent/remotewrite/statconn.go

@@ -0,0 +1,71 @@
+package remotewrite
+
+import (
+	"net"
+	"sync/atomic"
+
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fasthttp"
+)
+
+func statDial(addr string) (net.Conn, error) {
+	conn, err := fasthttp.Dial(addr)
+	dialsTotal.Inc()
+	if err != nil {
+		dialErrors.Inc()
+		return nil, err
+	}
+	conns.Inc()
+	sc := &statConn{
+		Conn: conn,
+	}
+	return sc, nil
+}
+
+var (
+	dialsTotal = metrics.NewCounter(`vmagent_remotewrite_dials_total`)
+	dialErrors = metrics.NewCounter(`vmagent_remotewrite_dial_errors_total`)
+	conns      = metrics.NewCounter(`vmagent_remotewrite_conns`)
+)
+
+type statConn struct {
+	closed uint64
+	net.Conn
+}
+
+func (sc *statConn) Read(p []byte) (int, error) {
+	n, err := sc.Conn.Read(p)
+	connReadsTotal.Inc()
+	if err != nil {
+		connReadErrors.Inc()
+	}
+	connBytesRead.Add(n)
+	return n, err
+}
+
+func (sc *statConn) Write(p []byte) (int, error) {
+	n, err := sc.Conn.Write(p)
+	connWritesTotal.Inc()
+	if err != nil {
+		connWriteErrors.Inc()
+	}
+	connBytesWritten.Add(n)
+	return n, err
+}
+
+func (sc *statConn) Close() error {
+	err := sc.Conn.Close()
+	if atomic.AddUint64(&sc.closed, 1) == 1 {
+		conns.Dec()
+	}
+	return err
+}
+
+var (
+	connReadsTotal   = metrics.NewCounter(`vmagent_remotewrite_conn_reads_total`)
+	connWritesTotal  = metrics.NewCounter(`vmagent_remotewrite_conn_writes_total`)
+	connReadErrors   = metrics.NewCounter(`vmagent_remotewrite_conn_read_errors_total`)
+	connWriteErrors  = metrics.NewCounter(`vmagent_remotewrite_conn_write_errors_total`)
+	connBytesRead    = metrics.NewCounter(`vmagent_remotewrite_conn_bytes_read_total`)
+	connBytesWritten = metrics.NewCounter(`vmagent_remotewrite_conn_bytes_written_total`)
+)

app/vmagent/vmimport/request_handler.go

@@ -0,0 +1,70 @@
+package vmimport
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmagent/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/vmimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vmagent_rows_inserted_total{type="vmimport"}`)
+	rowsPerInsert = metrics.NewHistogram(`vmagent_rows_per_insert{type="vmimport"}`)
+)
+
+// InsertHandler processes `/api/v1/import` request.
+//
+// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/6
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetPushCtx()
+	defer common.PutPushCtx(ctx)
+
+	rowsTotal := 0
+	tssDst := ctx.WriteRequest.Timeseries[:0]
+	labels := ctx.Labels[:0]
+	samples := ctx.Samples[:0]
+	for i := range rows {
+		r := &rows[i]
+		labelsLen := len(labels)
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			labels = append(labels, prompbmarshal.Label{
+				Name:  bytesutil.ToUnsafeString(tag.Key),
+				Value: bytesutil.ToUnsafeString(tag.Value),
+			})
+		}
+		values := r.Values
+		timestamps := r.Timestamps
+		_ = timestamps[len(values)-1]
+		samplesLen := len(samples)
+		for j, value := range values {
+			samples = append(samples, prompbmarshal.Sample{
+				Value:     value,
+				Timestamp: timestamps[j],
+			})
+		}
+		tssDst = append(tssDst, prompbmarshal.TimeSeries{
+			Labels:  labels[labelsLen:],
+			Samples: samples[samplesLen:],
+		})
+		rowsTotal += len(values)
+	}
+	ctx.WriteRequest.Timeseries = tssDst
+	ctx.Labels = labels
+	ctx.Samples = samples
+	remotewrite.Push(&ctx.WriteRequest)
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	return nil
+}

app/vmalert/README.md

@@ -0,0 +1,41 @@
+## VM Alert
+
+#### Abstract
+The application which accepts the alert rules, executes them on given source, sends(fires) an alert to(in) alert management system
+
+### Components
+
+#### Alert Config Reader
+It accepts yaml config as input parameter in Prometheus format, parses it into Go struct.
+
+#### Source Caller
+Create own watchdog for every alert group (goroutines), which executes alert query on given source and issues an alert if source returns non-empty result. 
+Source can be any service which supports PromQL (MetricsQL). 
+
+#### Alert Management System Provider
+Send positive alert to alert management system, provides interface for every concrete implementation.
+Should be ingratiated with Prometheus alertmanager. 
+
+open questions:
+- do we really need alert group or can just run every alert in own goroutine?
+
+#### Web Server
+Expose metrics
+
+open questions:
+- should the tool provide API or UI for managing alerting rules? Where to store config updated via the API or UI?
+- should the tool provide “alerting rules validation mode” for validating and debugging alerting rules? This mode is useful when creating and debugging alerting rules.
+
+#### Requirements:
+- Stateless
+- Avoid external dependencies if possible
+- Reuse existing code from VictoriaMetrics repo
+- Makefile rules for common tasks – see Makefiles for other apps in the app/ dir 
+- Every package should be covered by tests 
+- Dockerfile
+- Graceful shutdown 
+- Helm template
+- Application uses command line flags for configuration
+
+
+<img alt="VM Alert" src="vmalert.png">

app/vmalert/config/parser.go

@@ -0,0 +1,32 @@
+package config
+
+import "time"
+
+// Annotations basic annotation for alert rule
+type Annotations struct {
+	Summary     string
+	Description string
+}
+
+// Alert basic alert entity rule
+type Alert struct {
+	Name        string
+	Expr        string
+	For         time.Duration
+	Labels      map[string]string
+	Annotations Annotations
+
+	Start time.Time
+	End   time.Time
+}
+
+// Group grouping array of alert
+type Group struct {
+	Name  string
+	Rules []Alert
+}
+
+// Parse parses config from given file
+func Parse(filepath string) ([]Group, error) {
+	return []Group{}, nil
+}

app/vmalert/datasource/datasource.go

@@ -0,0 +1,14 @@
+package datasource
+
+import "context"
+
+// Metrics the data returns from storage
+type Metrics struct{}
+
+// VMStorage represents vmstorage entity with ability to read and write metrics
+type VMStorage struct{}
+
+//Query basic query to the datasource
+func (s *VMStorage) Query(ctx context.Context, query string) ([]Metrics, error) {
+	return nil, nil
+}

app/vmalert/main.go

@@ -0,0 +1,60 @@
+package main
+
+import (
+	"flag"
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/datasource"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/procutil"
+)
+
+var (
+	configPath     = flag.String("config", "config.yaml", "Path to alert configuration file")
+	httpListenAddr = flag.String("httpListenAddr", ":8880", "Address to listen for http connections")
+)
+
+func main() {
+	envflag.Parse()
+	buildinfo.Init()
+	logger.Init()
+
+	logger.Infof("reading alert rules configuration file from %s", *configPath)
+	alertGroups, err := config.Parse(*configPath)
+	if err != nil {
+		logger.Fatalf("Cannot parse configuration file %s", err)
+	}
+	w := &watchdog{storage: &datasource.VMStorage{}}
+	for id := range alertGroups {
+		go func(group config.Group) {
+			w.run(group)
+		}(alertGroups[id])
+	}
+	go func() {
+		httpserver.Serve(*httpListenAddr, func(w http.ResponseWriter, r *http.Request) bool {
+			panic("not implemented")
+		})
+	}()
+	sig := procutil.WaitForSigterm()
+	logger.Infof("service received signal %s", sig)
+	if err := httpserver.Stop(*httpListenAddr); err != nil {
+		logger.Fatalf("cannot stop the webservice: %s", err)
+	}
+	w.stop()
+}
+
+type watchdog struct {
+	storage *datasource.VMStorage
+}
+
+func (w *watchdog) run(a config.Group) {
+
+}
+
+func (w *watchdog) stop() {
+	panic("not implemented")
+}

app/vmalert/provider/alert_manager_request.qtpl

@@ -0,0 +1,26 @@
+{% import (
+    "time"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+) %}
+{% stripspace %}
+
+{% func amRequest(alert *config.Alert, generatorURL string) %}
+{
+	"startsAt":{%q= alert.Start.Format(time.RFC3339Nano) %},
+	"generatorURL": {%q= generatorURL %},
+    {% if !alert.End.IsZero() %}
+    "endsAt":{%q= alert.End.Format(time.RFC3339Nano) %},
+    {% endif %}
+    "labels": {
+        "alertname":{%q= alert.Name %}
+        {% for k,v := range alert.Labels %}
+            ,{%q= k %}:{%q= v %}
+        {% endfor %}
+    },
+    "annotations": {
+        "summary": {%q= alert.Annotations.Summary %},
+        "description": {%q= alert.Annotations.Description %}
+    }
+}
+{% endfunc %}
+{% endstripspace %}

app/vmalert/provider/alert_manager_request.qtpl.go

@@ -0,0 +1,101 @@
+// Code generated by qtc from "alert_manager_request.qtpl". DO NOT EDIT.
+// See https://github.com/valyala/quicktemplate for details.
+
+//line app/vmalert/provider/alert_manager_request.qtpl:1
+package provider
+
+//line app/vmalert/provider/alert_manager_request.qtpl:1
+import (
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+	"time"
+)
+
+//line app/vmalert/provider/alert_manager_request.qtpl:7
+import (
+	qtio422016 "io"
+
+	qt422016 "github.com/valyala/quicktemplate"
+)
+
+//line app/vmalert/provider/alert_manager_request.qtpl:7
+var (
+	_ = qtio422016.Copy
+	_ = qt422016.AcquireByteBuffer
+)
+
+//line app/vmalert/provider/alert_manager_request.qtpl:7
+func streamamRequest(qw422016 *qt422016.Writer, alert *config.Alert, generatorURL string) {
+//line app/vmalert/provider/alert_manager_request.qtpl:7
+	qw422016.N().S(`{"startsAt":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:9
+	qw422016.N().Q(alert.Start.Format(time.RFC3339Nano))
+//line app/vmalert/provider/alert_manager_request.qtpl:9
+	qw422016.N().S(`,"generatorURL":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:10
+	qw422016.N().Q(generatorURL)
+//line app/vmalert/provider/alert_manager_request.qtpl:10
+	qw422016.N().S(`,`)
+//line app/vmalert/provider/alert_manager_request.qtpl:11
+	if !alert.End.IsZero() {
+//line app/vmalert/provider/alert_manager_request.qtpl:11
+		qw422016.N().S(`"endsAt":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:12
+		qw422016.N().Q(alert.End.Format(time.RFC3339Nano))
+//line app/vmalert/provider/alert_manager_request.qtpl:12
+		qw422016.N().S(`,`)
+//line app/vmalert/provider/alert_manager_request.qtpl:13
+	}
+//line app/vmalert/provider/alert_manager_request.qtpl:13
+	qw422016.N().S(`"labels": {"alertname":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:15
+	qw422016.N().Q(alert.Name)
+//line app/vmalert/provider/alert_manager_request.qtpl:16
+	for k, v := range alert.Labels {
+//line app/vmalert/provider/alert_manager_request.qtpl:16
+		qw422016.N().S(`,`)
+//line app/vmalert/provider/alert_manager_request.qtpl:17
+		qw422016.N().Q(k)
+//line app/vmalert/provider/alert_manager_request.qtpl:17
+		qw422016.N().S(`:`)
+//line app/vmalert/provider/alert_manager_request.qtpl:17
+		qw422016.N().Q(v)
+//line app/vmalert/provider/alert_manager_request.qtpl:18
+	}
+//line app/vmalert/provider/alert_manager_request.qtpl:18
+	qw422016.N().S(`},"annotations": {"summary":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:21
+	qw422016.N().Q(alert.Annotations.Summary)
+//line app/vmalert/provider/alert_manager_request.qtpl:21
+	qw422016.N().S(`,"description":`)
+//line app/vmalert/provider/alert_manager_request.qtpl:22
+	qw422016.N().Q(alert.Annotations.Description)
+//line app/vmalert/provider/alert_manager_request.qtpl:22
+	qw422016.N().S(`}}`)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+}
+
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+func writeamRequest(qq422016 qtio422016.Writer, alert *config.Alert, generatorURL string) {
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	qw422016 := qt422016.AcquireWriter(qq422016)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	streamamRequest(qw422016, alert, generatorURL)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	qt422016.ReleaseWriter(qw422016)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+}
+
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+func amRequest(alert *config.Alert, generatorURL string) string {
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	qb422016 := qt422016.AcquireByteBuffer()
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	writeamRequest(qb422016, alert, generatorURL)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	qs422016 := string(qb422016.B)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	qt422016.ReleaseByteBuffer(qb422016)
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+	return qs422016
+//line app/vmalert/provider/alert_manager_request.qtpl:25
+}

app/vmalert/provider/alertmanager.go

@@ -0,0 +1,66 @@
+package provider
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+const alertsPath = "/api/v2/alerts"
+
+var pool = sync.Pool{New: func() interface{} {
+	return &bytes.Buffer{}
+}}
+
+// AlertManager represents integration provider with Prometheus alert manager
+type AlertManager struct {
+	alertURL string
+	argFunc  AlertURLGenerator
+	client   *http.Client
+}
+
+// AlertURLGenerator returns URL to single alert by given name
+type AlertURLGenerator func(name string) string
+
+// NewAlertManager is a constructor for AlertManager
+func NewAlertManager(alertManagerURL string, fn AlertURLGenerator, c *http.Client) *AlertManager {
+	return &AlertManager{
+		alertURL: strings.TrimSuffix(alertManagerURL, "/") + alertsPath,
+		argFunc:  fn,
+		client:   c,
+	}
+}
+
+const (
+	jsonArrayOpen  byte = 91 // [
+	jsonArrayClose byte = 93 // ]
+)
+
+// Send an alert or resolve message
+func (am *AlertManager) Send(alert *config.Alert) error {
+	b := pool.Get().(*bytes.Buffer)
+	b.Reset()
+	defer pool.Put(b)
+	b.WriteByte(jsonArrayOpen)
+	writeamRequest(b, alert, am.argFunc(alert.Name))
+	b.WriteByte(jsonArrayClose)
+	resp, err := am.client.Post(am.alertURL, "application/json", b)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusOK {
+		b.Reset()
+		if _, err := io.Copy(b, resp.Body); err != nil {
+			logger.Errorf("unable to copy error response body to buffer %s", err)
+		}
+		return fmt.Errorf("invalid response from alertmanager %s", b)
+	}
+	return nil
+}

app/vmalert/provider/alertmanager_test.go

@@ -0,0 +1,80 @@
+package provider
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+)
+
+func TestAlertManager_Send(t *testing.T) {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/", func(_ http.ResponseWriter, _ *http.Request) {
+		t.Errorf("should not be called")
+	})
+	c := -1
+	mux.HandleFunc(alertsPath, func(w http.ResponseWriter, r *http.Request) {
+		c++
+		if r.Method != http.MethodPost {
+			t.Errorf("expected POST method got %s", r.Method)
+		}
+		switch c {
+		case 0:
+			conn, _, _ := w.(http.Hijacker).Hijack()
+			_ = conn.Close()
+		case 1:
+			w.WriteHeader(500)
+		case 2:
+			var a []struct {
+				Labels       map[string]string `json:"labels"`
+				StartsAt     time.Time         `json:"startsAt"`
+				EndAt        time.Time         `json:"endsAt"`
+				Annotations  map[string]string `json:"annotations"`
+				GeneratorURL string            `json:"generatorURL"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&a); err != nil {
+				t.Errorf("can not unmarshal data into alert %s", err)
+				t.FailNow()
+			}
+			if len(a) != 1 {
+				t.Errorf("expected 1 alert in array got %d", len(a))
+			}
+			if a[0].GeneratorURL != "alert0" {
+				t.Errorf("exptected alert0 as generatorURL got %s", a[0].GeneratorURL)
+			}
+			if a[0].Labels["alertname"] != "alert0" {
+				t.Errorf("exptected alert0 as alert name got %s", a[0].Labels["alertname"])
+			}
+			if a[0].StartsAt.IsZero() {
+				t.Errorf("exptected non-zero start time")
+			}
+			if a[0].EndAt.IsZero() {
+				t.Errorf("exptected non-zero end time")
+			}
+		}
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+	am := NewAlertManager(srv.URL, func(name string) string {
+		return name
+	}, srv.Client())
+	if err := am.Send(&config.Alert{}); err == nil {
+		t.Error("expected connection error got nil")
+	}
+	if err := am.Send(&config.Alert{}); err == nil {
+		t.Error("expected wrong http code error got nil")
+	}
+	if err := am.Send(&config.Alert{
+		Name:  "alert0",
+		Start: time.Now().UTC(),
+		End:   time.Now().UTC(),
+	}); err != nil {
+		t.Errorf("unexpected error %s", err)
+	}
+	if c != 2 {
+		t.Errorf("expected 2 calls(count from zero) to server got %d", c)
+	}
+}

app/vmalert/provider/common.go

@@ -0,0 +1,8 @@
+package provider
+
+import "github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
+
+// AlertProvider is common interface for alert manager provider
+type AlertProvider interface {
+	Send(rule config.Alert) error
+}

app/vmbackup/Makefile

@@ -0,0 +1,70 @@
+# All these commands must run from repository root.
+
+vmbackup:
+	APP_NAME=vmbackup $(MAKE) app-local
+
+vmbackup-race:
+	APP_NAME=vmbackup RACE=-race $(MAKE) app-local
+
+vmbackup-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker
+
+vmbackup-pure-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-pure
+
+vmbackup-amd64-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-amd64
+
+vmbackup-arm-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-arm
+
+vmbackup-arm64-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-arm64
+
+vmbackup-ppc64le-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-ppc64le
+
+vmbackup-386-prod:
+	APP_NAME=vmbackup $(MAKE) app-via-docker-386
+
+package-vmbackup:
+	APP_NAME=vmbackup $(MAKE) package-via-docker
+
+package-vmbackup-pure:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-pure
+
+package-vmbackup-amd64:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-amd64
+
+package-vmbackup-arm:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-arm
+
+package-vmbackup-arm64:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-arm64
+
+package-vmbackup-ppc64le:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-ppc64le
+
+package-vmbackup-386:
+	APP_NAME=vmbackup $(MAKE) package-via-docker-386
+
+publish-vmbackup:
+	APP_NAME=vmbackup $(MAKE) publish-via-docker
+
+vmbackup-pure:
+	APP_NAME=vmbackup $(MAKE) app-local-pure
+
+vmbackup-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-amd64 ./app/vmbackup
+
+vmbackup-arm:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-arm ./app/vmbackup
+
+vmbackup-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-arm64 ./app/vmbackup
+
+vmbackup-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-ppc64le ./app/vmbackup
+
+vmbackup-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmbackup-386 ./app/vmbackup

app/vmbackup/README.md

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

app/vmbackup/deployment/Dockerfile

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

app/vmbackup/main.go

@@ -0,0 +1,115 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"os"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/actions"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/fslocal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+var (
+	storageDataPath = flag.String("storageDataPath", "victoria-metrics-data", "Path to VictoriaMetrics data. Must match -storageDataPath from VictoriaMetrics or vmstorage")
+	snapshotName    = flag.String("snapshotName", "", "Name for the snapshot to backup. See https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/README.md#how-to-work-with-snapshots")
+	dst             = flag.String("dst", "", "Where to put the backup on the remote storage. "+
+		"Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir\n"+
+		"-dst can point to the previous backup. In this case incremental backup is performed, i.e. only changed data is uploaded")
+	origin            = flag.String("origin", "", "Optional origin directory on the remote storage with old backup for server-side copying when performing full backup. This speeds up full backups")
+	concurrency       = flag.Int("concurrency", 10, "The number of concurrent workers. Higher concurrency may reduce backup duration")
+	maxBytesPerSecond = flag.Int("maxBytesPerSecond", 0, "The maximum upload speed. There is no limit if it is set to 0")
+)
+
+func main() {
+	flag.Usage = usage
+	envflag.Parse()
+	buildinfo.Init()
+
+	srcFS, err := newSrcFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	dstFS, err := newDstFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	originFS, err := newOriginFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	a := &actions.Backup{
+		Concurrency: *concurrency,
+		Src:         srcFS,
+		Dst:         dstFS,
+		Origin:      originFS,
+	}
+	if err := a.Run(); err != nil {
+		logger.Fatalf("cannot create backup: %s", err)
+	}
+}
+
+func usage() {
+	const s = `
+vmbackup performs backups for VictoriaMetrics data from instant snapshots to gcs, s3
+or local filesystem. Backed up data can be restored with vmrestore.
+
+See the docs at https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md .
+`
+
+	f := flag.CommandLine.Output()
+	fmt.Fprintf(f, "%s\n", s)
+	flag.PrintDefaults()
+}
+
+func newSrcFS() (*fslocal.FS, error) {
+	if len(*snapshotName) == 0 {
+		return nil, fmt.Errorf("`-snapshotName` cannot be empty")
+	}
+	snapshotPath := *storageDataPath + "/snapshots/" + *snapshotName
+
+	// Verify the snapshot exists.
+	f, err := os.Open(snapshotPath)
+	if err != nil {
+		return nil, fmt.Errorf("cannot open snapshot at %q: %s", snapshotPath, err)
+	}
+	fi, err := f.Stat()
+	_ = f.Close()
+	if err != nil {
+		return nil, fmt.Errorf("cannot stat %q: %s", snapshotPath, err)
+	}
+	if !fi.IsDir() {
+		return nil, fmt.Errorf("snapshot %q must be a directory", snapshotPath)
+	}
+
+	fs := &fslocal.FS{
+		Dir:               snapshotPath,
+		MaxBytesPerSecond: *maxBytesPerSecond,
+	}
+	if err := fs.Init(); err != nil {
+		return nil, fmt.Errorf("cannot initialize fs: %s", err)
+	}
+	return fs, nil
+}
+
+func newDstFS() (common.RemoteFS, error) {
+	fs, err := actions.NewRemoteFS(*dst)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-dst`=%q: %s", *dst, err)
+	}
+	return fs, nil
+}
+
+func newOriginFS() (common.RemoteFS, error) {
+	if len(*origin) == 0 {
+		return nil, nil
+	}
+	fs, err := actions.NewRemoteFS(*origin)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-origin`=%q: %s", *origin, err)
+	}
+	return fs, nil
+}

app/vminsert/common/insert_ctx.go

@@ -47,7 +47,7 @@ func (ctx *InsertCtx) marshalMetricNameRaw(prefix []byte, labels []prompb.Label)
 	return metricNameRaw[:len(metricNameRaw):len(metricNameRaw)]
 }
 
-// WriteDataPoint writes (timestamp, value) with the given prefix and lables into ctx buffer.
+// WriteDataPoint writes (timestamp, value) with the given prefix and labels into ctx buffer.
 func (ctx *InsertCtx) WriteDataPoint(prefix []byte, labels []prompb.Label, timestamp int64, value float64) {
 	metricNameRaw := ctx.marshalMetricNameRaw(prefix, labels)
 	ctx.addRow(metricNameRaw, timestamp, value)
@@ -78,6 +78,26 @@ func (ctx *InsertCtx) addRow(metricNameRaw []byte, timestamp int64, value float6
 	mr.Value = value
 }
 
+// AddLabelBytes adds (name, value) label to ctx.Labels.
+//
+// name and value must exist until ctx.Labels is used.
+func (ctx *InsertCtx) AddLabelBytes(name, value []byte) {
+	labels := ctx.Labels
+	if cap(labels) > len(labels) {
+		labels = labels[:len(labels)+1]
+	} else {
+		labels = append(labels, prompb.Label{})
+	}
+	label := &labels[len(labels)-1]
+
+	// Do not copy name and value contents for performance reasons.
+	// This reduces GC overhead on the number of objects and allocations.
+	label.Name = name
+	label.Value = value
+
+	ctx.Labels = labels
+}
+
 // AddLabel adds (name, value) label to ctx.Labels.
 //
 // name and value must exist until ctx.Labels is used.

app/vminsert/common/insert_ctx_pool.go

@@ -0,0 +1,36 @@
+package common
+
+import (
+	"runtime"
+	"sync"
+)
+
+// GetInsertCtx returns InsertCtx from the pool.
+//
+// Call PutInsertCtx for returning it to the pool.
+func GetInsertCtx() *InsertCtx {
+	select {
+	case ctx := <-insertCtxPoolCh:
+		return ctx
+	default:
+		if v := insertCtxPool.Get(); v != nil {
+			return v.(*InsertCtx)
+		}
+		return &InsertCtx{}
+	}
+}
+
+// PutInsertCtx returns ctx to the pool.
+//
+// ctx cannot be used after the call.
+func PutInsertCtx(ctx *InsertCtx) {
+	ctx.Reset(0)
+	select {
+	case insertCtxPoolCh <- ctx:
+	default:
+		insertCtxPool.Put(ctx)
+	}
+}
+
+var insertCtxPool sync.Pool
+var insertCtxPoolCh = make(chan *InsertCtx, runtime.GOMAXPROCS(-1))

app/vminsert/csvimport/request_handler.go

@@ -0,0 +1,44 @@
+package csvimport
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/csvimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="csvimport"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="csvimport"}`)
+)
+
+// InsertHandler processes /api/v1/import/csv requests.
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, func(rows []parser.Row) error {
+			return insertRows(rows)
+		})
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetInsertCtx()
+	defer common.PutInsertCtx(ctx)
+
+	ctx.Reset(len(rows))
+	for i := range rows {
+		r := &rows[i]
+		ctx.Labels = ctx.Labels[:0]
+		ctx.AddLabel("", r.Metric)
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			ctx.AddLabel(tag.Key, tag.Value)
+		}
+		ctx.WriteDataPoint(nil, ctx.Labels, r.Timestamp, r.Value)
+	}
+	rowsInserted.Add(len(rows))
+	rowsPerInsert.Update(float64(len(rows)))
+	return ctx.FlushBufs()
+}

app/vminsert/graphite/request_handler.go

@@ -1,161 +1,44 @@
 package graphite
 
 import (
-	"fmt"
 	"io"
-	"net"
-	"runtime"
-	"sync"
-	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/graphite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
 	"github.com/VictoriaMetrics/metrics"
 )
 
 var (
 	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="graphite"}`)
-	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="graphite"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="graphite"}`)
 )
 
-// insertHandler processes remote write for graphite plaintext protocol.
+// InsertHandler processes remote write for graphite plaintext protocol.
 //
 // See https://graphite.readthedocs.io/en/latest/feeding-carbon.html#the-plaintext-protocol
-func insertHandler(r io.Reader) error {
-	return concurrencylimiter.Do(func() error {
-		return insertHandlerInternal(r)
+func InsertHandler(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, insertRows)
 	})
 }
 
-func insertHandlerInternal(r io.Reader) error {
-	ctx := getPushCtx()
-	defer putPushCtx(ctx)
-	for ctx.Read(r) {
-		if err := ctx.InsertRows(); err != nil {
-			return err
-		}
-	}
-	return ctx.Error()
-}
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetInsertCtx()
+	defer common.PutInsertCtx(ctx)
 
-func (ctx *pushCtx) InsertRows() error {
-	rows := ctx.Rows.Rows
-	ic := &ctx.Common
-	ic.Reset(len(rows))
+	ctx.Reset(len(rows))
 	for i := range rows {
 		r := &rows[i]
-		ic.Labels = ic.Labels[:0]
-		ic.AddLabel("", r.Metric)
+		ctx.Labels = ctx.Labels[:0]
+		ctx.AddLabel("", r.Metric)
 		for j := range r.Tags {
 			tag := &r.Tags[j]
-			ic.AddLabel(tag.Key, tag.Value)
+			ctx.AddLabel(tag.Key, tag.Value)
 		}
-		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
+		ctx.WriteDataPoint(nil, ctx.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
 	rowsPerInsert.Update(float64(len(rows)))
-	return ic.FlushBufs()
+	return ctx.FlushBufs()
 }
-
-const flushTimeout = 3 * time.Second
-
-func (ctx *pushCtx) Read(r io.Reader) bool {
-	graphiteReadCalls.Inc()
-	if ctx.err != nil {
-		return false
-	}
-	if c, ok := r.(net.Conn); ok {
-		if err := c.SetReadDeadline(time.Now().Add(flushTimeout)); err != nil {
-			graphiteReadErrors.Inc()
-			ctx.err = fmt.Errorf("cannot set read deadline: %s", err)
-			return false
-		}
-	}
-	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
-	if ctx.err != nil {
-		if ne, ok := ctx.err.(net.Error); ok && ne.Timeout() {
-			// Flush the read data on timeout and try reading again.
-			ctx.err = nil
-		} else {
-			if ctx.err != io.EOF {
-				graphiteReadErrors.Inc()
-				ctx.err = fmt.Errorf("cannot read graphite plaintext protocol data: %s", ctx.err)
-			}
-			return false
-		}
-	}
-	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
-
-	// Fill missing timestamps with the current timestamp rounded to seconds.
-	currentTimestamp := time.Now().Unix()
-	rows := ctx.Rows.Rows
-	for i := range rows {
-		r := &rows[i]
-		if r.Timestamp == 0 {
-			r.Timestamp = currentTimestamp
-		}
-	}
-
-	// Convert timestamps from seconds to milliseconds.
-	for i := range rows {
-		rows[i].Timestamp *= 1e3
-	}
-
-	return true
-}
-
-type pushCtx struct {
-	Rows   Rows
-	Common common.InsertCtx
-
-	reqBuf  []byte
-	tailBuf []byte
-
-	err error
-}
-
-func (ctx *pushCtx) Error() error {
-	if ctx.err == io.EOF {
-		return nil
-	}
-	return ctx.err
-}
-
-func (ctx *pushCtx) reset() {
-	ctx.Rows.Reset()
-	ctx.Common.Reset(0)
-	ctx.reqBuf = ctx.reqBuf[:0]
-	ctx.tailBuf = ctx.tailBuf[:0]
-
-	ctx.err = nil
-}
-
-var (
-	graphiteReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="graphite"}`)
-	graphiteReadErrors = metrics.NewCounter(`vm_read_errors_total{name="graphite"}`)
-)
-
-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/influx/request_handler.go

@@ -2,87 +2,59 @@ package influx
 
 import (
 	"flag"
-	"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"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/influx"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
 	"github.com/VictoriaMetrics/metrics"
 )
 
 var (
-	measurementFieldSeparator = flag.String("influxMeasurementFieldSeparator", "_", "Separator for `{measurement}{separator}{field_name}` metric name when inserted via Influx line protocol")
-	skipSingleField           = flag.Bool("influxSkipSingleField", false, "Uses `{measurement}` instead of `{measurement}{separator}{field_name}` for metic name if Influx line contains only a single field")
+	measurementFieldSeparator = flag.String("influxMeasurementFieldSeparator", "_", "Separator for '{measurement}{separator}{field_name}' metric name when inserted via Influx line protocol")
+	skipSingleField           = flag.Bool("influxSkipSingleField", false, "Uses '{measurement}' instead of '{measurement}{separator}{field_name}' for metic name if Influx line contains only a single field")
 )
 
 var (
 	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="influx"}`)
-	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="influx"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="influx"}`)
 )
 
-// InsertHandler processes remote write for influx line protocol.
+// InsertHandlerForReader processes remote write for influx line protocol.
 //
-// See https://github.com/influxdata/influxdb/blob/4cbdc197b8117fee648d62e2e5be75c6575352f0/tsdb/README.md
-func InsertHandler(req *http.Request) error {
-	return concurrencylimiter.Do(func() error {
-		return insertHandlerInternal(req)
+// See https://github.com/influxdata/telegraf/tree/master/plugins/inputs/socket_listener/
+func InsertHandlerForReader(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, false, "", "", insertRows)
 	})
 }
 
-func insertHandlerInternal(req *http.Request) error {
-	influxReadCalls.Inc()
-
-	r := req.Body
-	if req.Header.Get("Content-Encoding") == "gzip" {
-		zr, err := common.GetGzipReader(r)
-		if err != nil {
-			return fmt.Errorf("cannot read gzipped influx line protocol data: %s", err)
-		}
-		defer common.PutGzipReader(zr)
-		r = zr
-	}
-
-	q := req.URL.Query()
-	tsMultiplier := int64(1e6)
-	switch q.Get("precision") {
-	case "ns":
-		tsMultiplier = 1e6
-	case "u":
-		tsMultiplier = 1e3
-	case "ms":
-		tsMultiplier = 1
-	case "s":
-		tsMultiplier = -1e3
-	case "m":
-		tsMultiplier = -1e3 * 60
-	case "h":
-		tsMultiplier = -1e3 * 3600
-	}
-
-	// Read db tag from https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint
-	db := q.Get("db")
-
-	ctx := getPushCtx()
-	defer putPushCtx(ctx)
-	for ctx.Read(r, tsMultiplier) {
-		if err := ctx.InsertRows(db); err != nil {
-			return err
-		}
-	}
-	return ctx.Error()
+// InsertHandlerForHTTP processes remote write for influx line protocol.
+//
+// See https://github.com/influxdata/influxdb/blob/4cbdc197b8117fee648d62e2e5be75c6575352f0/tsdb/README.md
+func InsertHandlerForHTTP(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		isGzipped := req.Header.Get("Content-Encoding") == "gzip"
+		q := req.URL.Query()
+		precision := q.Get("precision")
+		// Read db tag from https://docs.influxdata.com/influxdb/v1.7/tools/api/#write-http-endpoint
+		db := q.Get("db")
+		return parser.ParseStream(req.Body, isGzipped, precision, db, insertRows)
+	})
 }
 
-func (ctx *pushCtx) InsertRows(db string) error {
-	rows := ctx.Rows.Rows
+func insertRows(db string, rows []parser.Row) error {
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+
 	rowsLen := 0
 	for i := range rows {
-		rowsLen += len(rows[i].Tags)
+		rowsLen += len(rows[i].Fields)
 	}
 	ic := &ctx.Common
 	ic.Reset(rowsLen)
@@ -104,7 +76,7 @@ func (ctx *pushCtx) InsertRows(db string) error {
 		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
 		ctx.metricGroupBuf = append(ctx.metricGroupBuf[:0], r.Measurement...)
 		skipFieldKey := len(r.Fields) == 1 && *skipSingleField
-		if !skipFieldKey {
+		if len(ctx.metricGroupBuf) > 0 && !skipFieldKey {
 			ctx.metricGroupBuf = append(ctx.metricGroupBuf, *measurementFieldSeparator...)
 		}
 		metricGroupPrefixLen := len(ctx.metricGroupBuf)
@@ -125,80 +97,16 @@ func (ctx *pushCtx) InsertRows(db string) error {
 	return ic.FlushBufs()
 }
 
-func (ctx *pushCtx) Read(r io.Reader, tsMultiplier int64) bool {
-	if ctx.err != nil {
-		return false
-	}
-	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
-	if ctx.err != nil {
-		if ctx.err != io.EOF {
-			influxReadErrors.Inc()
-			ctx.err = fmt.Errorf("cannot read influx line protocol data: %s", ctx.err)
-		}
-		return false
-	}
-	ctx.Rows.Unmarshal(bytesutil.ToUnsafeString(ctx.reqBuf))
-
-	// Adjust timestamps according to tsMultiplier
-	currentTs := time.Now().UnixNano() / 1e6
-	if tsMultiplier >= 1 {
-		for i := range ctx.Rows.Rows {
-			row := &ctx.Rows.Rows[i]
-			if row.Timestamp == 0 {
-				row.Timestamp = currentTs
-			} else {
-				row.Timestamp /= tsMultiplier
-			}
-		}
-	} else if tsMultiplier < 0 {
-		tsMultiplier = -tsMultiplier
-		currentTs -= currentTs % tsMultiplier
-		for i := range ctx.Rows.Rows {
-			row := &ctx.Rows.Rows[i]
-			if row.Timestamp == 0 {
-				row.Timestamp = currentTs
-			} else {
-				row.Timestamp *= tsMultiplier
-			}
-		}
-	}
-	return true
-}
-
-var (
-	influxReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="influx"}`)
-	influxReadErrors = metrics.NewCounter(`vm_read_errors_total{name="influx"}`)
-)
-
 type pushCtx struct {
-	Rows   Rows
-	Common common.InsertCtx
-
-	reqBuf         []byte
-	tailBuf        []byte
+	Common         common.InsertCtx
 	metricNameBuf  []byte
 	metricGroupBuf []byte
-
-	err error
-}
-
-func (ctx *pushCtx) Error() error {
-	if ctx.err == io.EOF {
-		return nil
-	}
-	return ctx.err
 }
 
 func (ctx *pushCtx) reset() {
-	ctx.Rows.Reset()
 	ctx.Common.Reset(0)
-
-	ctx.reqBuf = ctx.reqBuf[:0]
-	ctx.tailBuf = ctx.tailBuf[:0]
 	ctx.metricNameBuf = ctx.metricNameBuf[:0]
 	ctx.metricGroupBuf = ctx.metricGroupBuf[:0]
-
-	ctx.err = nil
 }
 
 func getPushCtx() *pushCtx {

app/vminsert/main.go

@@ -6,51 +6,76 @@ import (
 	"net/http"
 	"strings"
 
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/csvimport"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/graphite"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/influx"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/opentsdbhttp"
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/prometheus"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/prompush"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/promremotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/vmimport"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
+	graphiteserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/graphite"
+	influxserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/influx"
+	opentsdbserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/opentsdb"
+	opentsdbhttpserver "github.com/VictoriaMetrics/VictoriaMetrics/lib/ingestserver/opentsdbhttp"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promscrape"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
 	"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")
+	graphiteListenAddr = flag.String("graphiteListenAddr", "", "TCP and UDP address to listen for Graphite plaintext data. Usually :2003 must be set. Doesn't work if empty")
+	influxListenAddr   = flag.String("influxListenAddr", "", "TCP and UDP address to listen for Influx line protocol data. Usually :8189 must be set. Doesn't work if empty")
+	opentsdbListenAddr = flag.String("opentsdbListenAddr", "", "TCP and UDP address to listen for OpentTSDB metrics. "+
+		"Telnet put messages and HTTP /api/put messages are simultaneously served on TCP port. "+
+		"Usually :4242 must be set. Doesn't work if empty")
 	opentsdbHTTPListenAddr = flag.String("opentsdbHTTPListenAddr", "", "TCP address to listen for OpentTSDB HTTP put requests. Usually :4242 must be set. Doesn't work if empty")
-	maxInsertRequestSize   = flag.Int("maxInsertRequestSize", 32*1024*1024, "The maximum size of a single insert request in bytes")
 	maxLabelsPerTimeseries = flag.Int("maxLabelsPerTimeseries", 30, "The maximum number of labels accepted per time series. Superflouos labels are dropped")
 )
 
+var (
+	influxServer       *influxserver.Server
+	graphiteServer     *graphiteserver.Server
+	opentsdbServer     *opentsdbserver.Server
+	opentsdbhttpServer *opentsdbhttpserver.Server
+)
+
 // Init initializes vminsert.
 func Init() {
 	storage.SetMaxLabelsPerTimeseries(*maxLabelsPerTimeseries)
 
-	concurrencylimiter.Init()
+	writeconcurrencylimiter.Init()
+	if len(*influxListenAddr) > 0 {
+		influxServer = influxserver.MustStart(*influxListenAddr, influx.InsertHandlerForReader)
+	}
 	if len(*graphiteListenAddr) > 0 {
-		go graphite.Serve(*graphiteListenAddr)
+		graphiteServer = graphiteserver.MustStart(*graphiteListenAddr, graphite.InsertHandler)
 	}
 	if len(*opentsdbListenAddr) > 0 {
-		go opentsdb.Serve(*opentsdbListenAddr)
+		opentsdbServer = opentsdbserver.MustStart(*opentsdbListenAddr, opentsdb.InsertHandler, opentsdbhttp.InsertHandler)
 	}
 	if len(*opentsdbHTTPListenAddr) > 0 {
-		go opentsdbhttp.Serve(*opentsdbHTTPListenAddr, int64(*maxInsertRequestSize))
+		opentsdbhttpServer = opentsdbhttpserver.MustStart(*opentsdbHTTPListenAddr, opentsdbhttp.InsertHandler)
 	}
+	promscrape.Init(prompush.Push)
 }
 
 // Stop stops vminsert.
 func Stop() {
+	promscrape.Stop()
+	if len(*influxListenAddr) > 0 {
+		influxServer.MustStop()
+	}
 	if len(*graphiteListenAddr) > 0 {
-		graphite.Stop()
+		graphiteServer.MustStop()
 	}
 	if len(*opentsdbListenAddr) > 0 {
-		opentsdb.Stop()
+		opentsdbServer.MustStop()
 	}
 	if len(*opentsdbHTTPListenAddr) > 0 {
-		opentsdbhttp.Stop()
+		opentsdbhttpServer.MustStop()
 	}
 }
 
@@ -60,16 +85,34 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	switch path {
 	case "/api/v1/write":
 		prometheusWriteRequests.Inc()
-		if err := prometheus.InsertHandler(r, int64(*maxInsertRequestSize)); err != nil {
+		if err := promremotewrite.InsertHandler(r); err != nil {
 			prometheusWriteErrors.Inc()
 			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
 			return true
 		}
 		w.WriteHeader(http.StatusNoContent)
 		return true
+	case "/api/v1/import":
+		vmimportRequests.Inc()
+		if err := vmimport.InsertHandler(r); err != nil {
+			vmimportErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
+	case "/api/v1/import/csv":
+		csvimportRequests.Inc()
+		if err := csvimport.InsertHandler(r); err != nil {
+			csvimportErrors.Inc()
+			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
+			return true
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return true
 	case "/write", "/api/v2/write":
 		influxWriteRequests.Inc()
-		if err := influx.InsertHandler(r); err != nil {
+		if err := influx.InsertHandlerForHTTP(r); err != nil {
 			influxWriteErrors.Inc()
 			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
 			return true
@@ -82,6 +125,11 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		influxQueryRequests.Inc()
 		fmt.Fprintf(w, `{"results":[{"series":[{"values":[]}]}]}`)
 		return true
+	case "/targets":
+		promscrapeTargetsRequests.Inc()
+		w.Header().Set("Content-Type", "text/plain")
+		promscrape.WriteHumanReadableTargetsStatus(w)
+		return true
 	default:
 		// This is not our link
 		return false
@@ -89,11 +137,19 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 }
 
 var (
-	prometheusWriteRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/write", protocol="prometheus"}`)
-	prometheusWriteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/write", protocol="prometheus"}`)
+	prometheusWriteRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/write", protocol="promremotewrite"}`)
+	prometheusWriteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/write", protocol="promremotewrite"}`)
+
+	vmimportRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/import", protocol="vmimport"}`)
+	vmimportErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/import", protocol="vmimport"}`)
+
+	csvimportRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/import/csv", protocol="csvimport"}`)
+	csvimportErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/api/v1/import/csv", protocol="csvimport"}`)
 
 	influxWriteRequests = metrics.NewCounter(`vm_http_requests_total{path="/write", protocol="influx"}`)
 	influxWriteErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/write", protocol="influx"}`)
 
 	influxQueryRequests = metrics.NewCounter(`vm_http_requests_total{path="/query", protocol="influx"}`)
+
+	promscrapeTargetsRequests = metrics.NewCounter(`vm_http_requests_total{path="/targets"}`)
 )

app/vminsert/opentsdb/request_handler.go

@@ -1,160 +1,44 @@
 package opentsdb
 
 import (
-	"fmt"
 	"io"
-	"net"
-	"runtime"
-	"sync"
-	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/opentsdb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
 	"github.com/VictoriaMetrics/metrics"
 )
 
 var (
 	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdb"}`)
-	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="opentsdb"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="opentsdb"}`)
 )
 
-// insertHandler processes remote write for OpenTSDB put protocol.
+// InsertHandler processes remote write for OpenTSDB put protocol.
 //
 // See http://opentsdb.net/docs/build/html/api_telnet/put.html
-func insertHandler(r io.Reader) error {
-	return concurrencylimiter.Do(func() error {
-		return insertHandlerInternal(r)
+func InsertHandler(r io.Reader) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(r, insertRows)
 	})
 }
 
-func insertHandlerInternal(r io.Reader) error {
-	ctx := getPushCtx()
-	defer putPushCtx(ctx)
-	for ctx.Read(r) {
-		if err := ctx.InsertRows(); err != nil {
-			return err
-		}
-	}
-	return ctx.Error()
-}
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetInsertCtx()
+	defer common.PutInsertCtx(ctx)
 
-func (ctx *pushCtx) InsertRows() error {
-	rows := ctx.Rows.Rows
-	ic := &ctx.Common
-	ic.Reset(len(rows))
+	ctx.Reset(len(rows))
 	for i := range rows {
 		r := &rows[i]
-		ic.Labels = ic.Labels[:0]
-		ic.AddLabel("", r.Metric)
+		ctx.Labels = ctx.Labels[:0]
+		ctx.AddLabel("", r.Metric)
 		for j := range r.Tags {
 			tag := &r.Tags[j]
-			ic.AddLabel(tag.Key, tag.Value)
+			ctx.AddLabel(tag.Key, tag.Value)
 		}
-		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
+		ctx.WriteDataPoint(nil, ctx.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
 	rowsPerInsert.Update(float64(len(rows)))
-	return ic.FlushBufs()
+	return ctx.FlushBufs()
 }
-
-const flushTimeout = 3 * time.Second
-
-func (ctx *pushCtx) Read(r io.Reader) bool {
-	opentsdbReadCalls.Inc()
-	if ctx.err != nil {
-		return false
-	}
-	if c, ok := r.(net.Conn); ok {
-		if err := c.SetReadDeadline(time.Now().Add(flushTimeout)); err != nil {
-			opentsdbReadErrors.Inc()
-			ctx.err = fmt.Errorf("cannot set read deadline: %s", err)
-			return false
-		}
-	}
-	ctx.reqBuf, ctx.tailBuf, ctx.err = common.ReadLinesBlock(r, ctx.reqBuf, ctx.tailBuf)
-	if ctx.err != nil {
-		if ne, ok := ctx.err.(net.Error); ok && ne.Timeout() {
-			// Flush the read data on timeout and try reading again.
-			ctx.err = nil
-		} else {
-			if ctx.err != io.EOF {
-				opentsdbReadErrors.Inc()
-				ctx.err = fmt.Errorf("cannot read OpenTSDB put protocol data: %s", ctx.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 rows {
-		rows[i].Timestamp *= 1e3
-	}
-	return true
-}
-
-type pushCtx struct {
-	Rows   Rows
-	Common common.InsertCtx
-
-	reqBuf  []byte
-	tailBuf []byte
-
-	err error
-}
-
-func (ctx *pushCtx) Error() error {
-	if ctx.err == io.EOF {
-		return nil
-	}
-	return ctx.err
-}
-
-func (ctx *pushCtx) reset() {
-	ctx.Rows.Reset()
-	ctx.Common.Reset(0)
-	ctx.reqBuf = ctx.reqBuf[:0]
-	ctx.tailBuf = ctx.tailBuf[:0]
-
-	ctx.err = nil
-}
-
-var (
-	opentsdbReadCalls  = metrics.NewCounter(`vm_read_calls_total{name="opentsdb"}`)
-	opentsdbReadErrors = metrics.NewCounter(`vm_read_errors_total{name="opentsdb"}`)
-)
-
-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/opentsdb/server.go

@@ -1,138 +0,0 @@
-package opentsdb
-
-import (
-	"net"
-	"runtime"
-	"strings"
-	"sync"
-	"time"
-
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
-	"github.com/VictoriaMetrics/metrics"
-)
-
-var (
-	writeRequestsTCP = metrics.NewCounter(`vm_opentsdb_requests_total{name="write", net="tcp"}`)
-	writeErrorsTCP   = metrics.NewCounter(`vm_opentsdb_request_errors_total{name="write", net="tcp"}`)
-
-	writeRequestsUDP = metrics.NewCounter(`vm_opentsdb_requests_total{name="write", net="udp"}`)
-	writeErrorsUDP   = metrics.NewCounter(`vm_opentsdb_request_errors_total{name="write", net="udp"}`)
-)
-
-// Serve starts OpenTSDB collector on the given addr.
-func Serve(addr string) {
-	logger.Infof("starting TCP OpenTSDB collector at %q", addr)
-	lnTCP, err := netutil.NewTCPListener("opentsdb", addr)
-	if err != nil {
-		logger.Fatalf("cannot start TCP OpenTSDB collector at %q: %s", addr, err)
-	}
-	listenerTCP = lnTCP
-
-	logger.Infof("starting UDP OpenTSDB collector at %q", addr)
-	lnUDP, err := net.ListenPacket("udp4", addr)
-	if err != nil {
-		logger.Fatalf("cannot start UDP OpenTSDB collector at %q: %s", addr, err)
-	}
-	listenerUDP = lnUDP
-
-	var wg sync.WaitGroup
-	wg.Add(1)
-	go func() {
-		defer wg.Done()
-		serveTCP(listenerTCP)
-		logger.Infof("stopped TCP OpenTSDB collector at %q", addr)
-	}()
-	wg.Add(1)
-	go func() {
-		defer wg.Done()
-		serveUDP(listenerUDP)
-		logger.Infof("stopped UDP OpenTSDB collector at %q", addr)
-	}()
-	wg.Wait()
-}
-
-func serveTCP(ln net.Listener) {
-	for {
-		c, err := ln.Accept()
-		if err != nil {
-			if ne, ok := err.(net.Error); ok {
-				if ne.Temporary() {
-					time.Sleep(time.Second)
-					continue
-				}
-				if strings.Contains(err.Error(), "use of closed network connection") {
-					break
-				}
-				logger.Fatalf("unrecoverable error when accepting TCP OpenTSDB connections: %s", err)
-			}
-			logger.Fatalf("unexpected error when accepting TCP OpenTSDB connections: %s", err)
-		}
-		go func() {
-			writeRequestsTCP.Inc()
-			if err := insertHandler(c); err != nil {
-				writeErrorsTCP.Inc()
-				logger.Errorf("error in TCP OpenTSDB conn %q<->%q: %s", c.LocalAddr(), c.RemoteAddr(), err)
-			}
-			_ = c.Close()
-		}()
-	}
-}
-
-func serveUDP(ln net.PacketConn) {
-	gomaxprocs := runtime.GOMAXPROCS(-1)
-	var wg sync.WaitGroup
-	for i := 0; i < gomaxprocs; i++ {
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
-			var bb bytesutil.ByteBuffer
-			bb.B = bytesutil.Resize(bb.B, 64*1024)
-			for {
-				bb.Reset()
-				bb.B = bb.B[:cap(bb.B)]
-				n, addr, err := ln.ReadFrom(bb.B)
-				if err != nil {
-					writeErrorsUDP.Inc()
-					if ne, ok := err.(net.Error); ok {
-						if ne.Temporary() {
-							time.Sleep(time.Second)
-							continue
-						}
-						if strings.Contains(err.Error(), "use of closed network connection") {
-							break
-						}
-					}
-					logger.Errorf("cannot read OpenTSDB UDP data: %s", err)
-					continue
-				}
-				bb.B = bb.B[:n]
-				writeRequestsUDP.Inc()
-				if err := insertHandler(bb.NewReader()); err != nil {
-					writeErrorsUDP.Inc()
-					logger.Errorf("error in UDP OpenTSDB conn %q<->%q: %s", ln.LocalAddr(), addr, err)
-					continue
-				}
-			}
-		}()
-	}
-	wg.Wait()
-}
-
-var (
-	listenerTCP net.Listener
-	listenerUDP net.PacketConn
-)
-
-// Stop stops the server.
-func Stop() {
-	logger.Infof("stopping TCP OpenTSDB server at %q...", listenerTCP.Addr())
-	if err := listenerTCP.Close(); err != nil {
-		logger.Errorf("cannot close TCP OpenTSDB server: %s", err)
-	}
-	logger.Infof("stopping UDP OpenTSDB server at %q...", listenerUDP.LocalAddr())
-	if err := listenerUDP.Close(); err != nil {
-		logger.Errorf("cannot close UDP OpenTSDB server: %s", err)
-	}
-}

app/vminsert/opentsdbhttp/request_handler.go

@@ -2,149 +2,49 @@ 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"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/opentsdbhttp"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
 	"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"}`)
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="opentsdbhttp"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="opentsdbhttp"}`)
 )
 
-// insertHandler processes HTTP OpenTSDB put requests.
+// 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 InsertHandler(req *http.Request) error {
+	path := req.URL.Path
+	switch path {
+	case "/api/put":
+		return writeconcurrencylimiter.Do(func() error {
+			return parser.ParseStream(req, insertRows)
+		})
+	default:
+		return fmt.Errorf("unexpected path requested on HTTP OpenTSDB server: %q", path)
+	}
 }
 
-func insertHandlerInternal(req *http.Request, maxSize int64) error {
-	opentsdbReadCalls.Inc()
+func insertRows(rows []parser.Row) error {
+	ctx := common.GetInsertCtx()
+	defer common.PutInsertCtx(ctx)
 
-	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
+	ctx.Reset(len(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)
+		ctx.Labels = ctx.Labels[:0]
+		ctx.AddLabel("", r.Metric)
 		for j := range r.Tags {
 			tag := &r.Tags[j]
-			ic.AddLabel(tag.Key, tag.Value)
+			ctx.AddLabel(tag.Key, tag.Value)
 		}
-		ic.WriteDataPoint(nil, ic.Labels, r.Timestamp, r.Value)
+		ctx.WriteDataPoint(nil, ctx.Labels, r.Timestamp, r.Value)
 	}
 	rowsInserted.Add(len(rows))
 	rowsPerInsert.Update(float64(len(rows)))
-	return ic.FlushBufs()
+	return ctx.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

@@ -1,70 +0,0 @@
-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/vminsert/prometheus/request_handler.go

@@ -1,112 +0,0 @@
-package prometheus
-
-import (
-	"fmt"
-	"net/http"
-	"runtime"
-	"sync"
-
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
-	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/concurrencylimiter"
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
-	"github.com/VictoriaMetrics/metrics"
-)
-
-var (
-	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="prometheus"}`)
-	rowsPerInsert = metrics.NewSummary(`vm_rows_per_insert{type="prometheus"}`)
-)
-
-// InsertHandler processes remote write for prometheus.
-func InsertHandler(r *http.Request, maxSize int64) error {
-	return concurrencylimiter.Do(func() error {
-		return insertHandlerInternal(r, maxSize)
-	})
-}
-
-func insertHandlerInternal(r *http.Request, maxSize int64) error {
-	ctx := getPushCtx()
-	defer putPushCtx(ctx)
-	if err := ctx.Read(r, maxSize); err != nil {
-		return err
-	}
-	timeseries := ctx.req.Timeseries
-	rowsLen := 0
-	for i := range timeseries {
-		rowsLen += len(timeseries[i].Samples)
-	}
-	ic := &ctx.Common
-	ic.Reset(rowsLen)
-	rowsTotal := 0
-	for i := range timeseries {
-		ts := &timeseries[i]
-		var metricNameRaw []byte
-		for i := range ts.Samples {
-			r := &ts.Samples[i]
-			metricNameRaw = ic.WriteDataPointExt(metricNameRaw, ts.Labels, r.Timestamp, r.Value)
-		}
-		rowsTotal += len(ts.Samples)
-	}
-	rowsInserted.Add(rowsTotal)
-	rowsPerInsert.Update(float64(rowsTotal))
-	return ic.FlushBufs()
-}
-
-type pushCtx struct {
-	Common common.InsertCtx
-
-	req    prompb.WriteRequest
-	reqBuf []byte
-}
-
-func (ctx *pushCtx) reset() {
-	ctx.Common.Reset(0)
-	ctx.req.Reset()
-	ctx.reqBuf = ctx.reqBuf[:0]
-}
-
-func (ctx *pushCtx) Read(r *http.Request, maxSize int64) error {
-	prometheusReadCalls.Inc()
-
-	var err error
-	ctx.reqBuf, err = prompb.ReadSnappy(ctx.reqBuf[:0], r.Body, maxSize)
-	if err != nil {
-		prometheusReadErrors.Inc()
-		return fmt.Errorf("cannot read prompb.WriteRequest: %s", err)
-	}
-	if err = ctx.req.Unmarshal(ctx.reqBuf); err != nil {
-		prometheusUnmarshalErrors.Inc()
-		return fmt.Errorf("cannot unmarshal prompb.WriteRequest with size %d bytes: %s", len(ctx.reqBuf), err)
-	}
-	return nil
-}
-
-var (
-	prometheusReadCalls       = metrics.NewCounter(`vm_read_calls_total{name="prometheus"}`)
-	prometheusReadErrors      = metrics.NewCounter(`vm_read_errors_total{name="prometheus"}`)
-	prometheusUnmarshalErrors = metrics.NewCounter(`vm_unmarshal_errors_total{name="prometheus"}`)
-)
-
-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/prompush/push.go

@@ -0,0 +1,113 @@
+package prompush
+
+import (
+	"runtime"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="promscrape"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="promscrape"}`)
+)
+
+const maxRowsPerBlock = 10000
+
+// Push pushes wr to storage.
+func Push(wr *prompbmarshal.WriteRequest) {
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+
+	tss := wr.Timeseries
+	for len(tss) > 0 {
+		// Process big tss in smaller blocks in order to reduce maxmimum memory usage
+		tssBlock := tss
+		if len(tssBlock) > maxRowsPerBlock {
+			tssBlock = tss[:maxRowsPerBlock]
+			tss = tss[maxRowsPerBlock:]
+		} else {
+			tss = nil
+		}
+		ctx.push(tssBlock)
+	}
+}
+
+func (ctx *pushCtx) push(tss []prompbmarshal.TimeSeries) {
+	rowsLen := 0
+	for i := range tss {
+		rowsLen += len(tss[i].Samples)
+	}
+	ic := &ctx.Common
+	ic.Reset(rowsLen)
+	rowsTotal := 0
+	labels := ctx.labels[:0]
+	for i := range tss {
+		ts := &tss[i]
+		labels = labels[:0]
+		for j := range ts.Labels {
+			label := &ts.Labels[j]
+			labels = append(labels, prompb.Label{
+				Name:  bytesutil.ToUnsafeBytes(label.Name),
+				Value: bytesutil.ToUnsafeBytes(label.Value),
+			})
+		}
+		var metricNameRaw []byte
+		for i := range ts.Samples {
+			r := &ts.Samples[i]
+			metricNameRaw = ic.WriteDataPointExt(metricNameRaw, labels, r.Timestamp, r.Value)
+		}
+		rowsTotal += len(ts.Samples)
+	}
+	ctx.labels = labels
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	if err := ic.FlushBufs(); err != nil {
+		logger.Errorf("cannot flush promscrape data to storage: %s", err)
+	}
+}
+
+type pushCtx struct {
+	Common common.InsertCtx
+	labels []prompb.Label
+}
+
+func (ctx *pushCtx) reset() {
+	ctx.Common.Reset(0)
+
+	for i := range ctx.labels {
+		label := &ctx.labels[i]
+		label.Name = nil
+		label.Value = nil
+	}
+	ctx.labels = ctx.labels[:0]
+}
+
+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/promremotewrite/request_handler.go

@@ -0,0 +1,47 @@
+package promremotewrite
+
+import (
+	"net/http"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/promremotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="promremotewrite"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="promremotewrite"}`)
+)
+
+// InsertHandler processes remote write for prometheus.
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(timeseries []prompb.TimeSeries) error {
+	ctx := common.GetInsertCtx()
+	defer common.PutInsertCtx(ctx)
+
+	rowsLen := 0
+	for i := range timeseries {
+		rowsLen += len(timeseries[i].Samples)
+	}
+	ctx.Reset(rowsLen)
+	rowsTotal := 0
+	for i := range timeseries {
+		ts := &timeseries[i]
+		var metricNameRaw []byte
+		for i := range ts.Samples {
+			r := &ts.Samples[i]
+			metricNameRaw = ctx.WriteDataPointExt(metricNameRaw, ts.Labels, r.Timestamp, r.Value)
+		}
+		rowsTotal += len(ts.Samples)
+	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	return ctx.FlushBufs()
+}

app/vminsert/vmimport/request_handler.go

@@ -0,0 +1,94 @@
+package vmimport
+
+import (
+	"net/http"
+	"runtime"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vminsert/common"
+	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/vmimport"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+var (
+	rowsInserted  = metrics.NewCounter(`vm_rows_inserted_total{type="vmimport"}`)
+	rowsPerInsert = metrics.NewHistogram(`vm_rows_per_insert{type="vmimport"}`)
+)
+
+// InsertHandler processes `/api/v1/import` request.
+//
+// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/6
+func InsertHandler(req *http.Request) error {
+	return writeconcurrencylimiter.Do(func() error {
+		return parser.ParseStream(req, insertRows)
+	})
+}
+
+func insertRows(rows []parser.Row) error {
+	ctx := getPushCtx()
+	defer putPushCtx(ctx)
+
+	rowsLen := 0
+	for i := range rows {
+		rowsLen += len(rows[i].Values)
+	}
+	ic := &ctx.Common
+	ic.Reset(rowsLen)
+	rowsTotal := 0
+	for i := range rows {
+		r := &rows[i]
+		ic.Labels = ic.Labels[:0]
+		for j := range r.Tags {
+			tag := &r.Tags[j]
+			ic.AddLabelBytes(tag.Key, tag.Value)
+		}
+		ctx.metricNameBuf = storage.MarshalMetricNameRaw(ctx.metricNameBuf[:0], ic.Labels)
+		values := r.Values
+		timestamps := r.Timestamps
+		_ = timestamps[len(values)-1]
+		for j, value := range values {
+			timestamp := timestamps[j]
+			ic.WriteDataPoint(ctx.metricNameBuf, nil, timestamp, value)
+		}
+		rowsTotal += len(values)
+	}
+	rowsInserted.Add(rowsTotal)
+	rowsPerInsert.Update(float64(rowsTotal))
+	return ic.FlushBufs()
+}
+
+type pushCtx struct {
+	Common        common.InsertCtx
+	metricNameBuf []byte
+}
+
+func (ctx *pushCtx) reset() {
+	ctx.Common.Reset(0)
+	ctx.metricNameBuf = ctx.metricNameBuf[:0]
+}
+
+func getPushCtx() *pushCtx {
+	select {
+	case ctx := <-pushCtxPoolCh:
+		return ctx
+	default:
+		if v := pushCtxPool.Get(); v != nil {
+			return v.(*pushCtx)
+		}
+		return &pushCtx{}
+	}
+}
+
+func putPushCtx(ctx *pushCtx) {
+	ctx.reset()
+	select {
+	case pushCtxPoolCh <- ctx:
+	default:
+		pushCtxPool.Put(ctx)
+	}
+}
+
+var pushCtxPool sync.Pool
+var pushCtxPoolCh = make(chan *pushCtx, runtime.GOMAXPROCS(-1))

app/vmrestore/Makefile

@@ -0,0 +1,70 @@
+# All these commands must run from repository root.
+
+vmrestore:
+	APP_NAME=vmrestore $(MAKE) app-local
+
+vmrestore-race:
+	APP_NAME=vmrestore RACE=-race $(MAKE) app-local
+
+vmrestore-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker
+
+vmrestore-pure-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-pure
+
+vmrestore-amd64-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-amd64
+
+vmrestore-arm-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-arm
+
+vmrestore-arm64-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-arm64
+
+vmrestore-ppc64le-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-ppc64le
+
+vmrestore-386-prod:
+	APP_NAME=vmrestore $(MAKE) app-via-docker-386
+
+package-vmrestore:
+	APP_NAME=vmrestore $(MAKE) package-via-docker
+
+package-vmrestore-pure:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-pure
+
+package-vmrestore-amd64:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-amd64
+
+package-vmrestore-arm:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-arm
+
+package-vmrestore-arm64:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-arm64
+
+package-vmrestore-ppc64le:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-ppc64le
+
+package-vmrestore-386:
+	APP_NAME=vmrestore $(MAKE) package-via-docker-386
+
+publish-vmrestore:
+	APP_NAME=vmrestore $(MAKE) publish-via-docker
+
+vmrestore-pure:
+	APP_NAME=vmrestore $(MAKE) app-local-pure
+
+vmrestore-amd64:
+	CGO_ENABLED=1 GOOS=linux GOARCH=amd64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-amd64 ./app/vmrestore
+
+vmrestore-arm:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-arm ./app/vmrestore
+
+vmrestore-arm64:
+	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-arm64 ./app/vmrestore
+
+vmrestore-ppc64le:
+	CGO_ENABLED=0 GOOS=linux GOARCH=ppc64le GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-ppc64le ./app/vmrestore
+
+vmrestore-386:
+	CGO_ENABLED=0 GOOS=linux GOARCH=386 GO111MODULE=on go build -mod=vendor -ldflags "$(GO_BUILDINFO)" -o bin/vmrestore-386 ./app/vmrestore

app/vmrestore/README.md

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

app/vmrestore/deployment/Dockerfile

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

app/vmrestore/main.go

@@ -0,0 +1,81 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/actions"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/common"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/fslocal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/buildinfo"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/envflag"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+)
+
+var (
+	src = flag.String("src", "", "Source path with backup on the remote storage. "+
+		"Example: gcs://bucket/path/to/backup/dir, s3://bucket/path/to/backup/dir or fs:///path/to/local/backup/dir")
+	storageDataPath = flag.String("storageDataPath", "victoria-metrics-data", "Destination path where backup must be restored. "+
+		"VictoriaMetrics must be stopped when restoring from backup. -storageDataPath dir can be non-empty. In this case only missing data is downloaded from backup")
+	concurrency             = flag.Int("concurrency", 10, "The number of concurrent workers. Higher concurrency may reduce restore duration")
+	maxBytesPerSecond       = flag.Int("maxBytesPerSecond", 0, "The maximum download speed. There is no limit if it is set to 0")
+	skipBackupCompleteCheck = flag.Bool("skipBackupCompleteCheck", false, "Whether to skip checking for 'backup complete' file in -src. This may be useful for restoring from old backups, which were created without 'backup complete' file")
+)
+
+func main() {
+	flag.Usage = usage
+	envflag.Parse()
+	buildinfo.Init()
+
+	srcFS, err := newSrcFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	dstFS, err := newDstFS()
+	if err != nil {
+		logger.Fatalf("%s", err)
+	}
+	a := &actions.Restore{
+		Concurrency:             *concurrency,
+		Src:                     srcFS,
+		Dst:                     dstFS,
+		SkipBackupCompleteCheck: *skipBackupCompleteCheck,
+	}
+	if err := a.Run(); err != nil {
+		logger.Fatalf("cannot restore from backup: %s", err)
+	}
+}
+
+func usage() {
+	const s = `
+vmrestore restores VictoriaMetrics data from backups made by vmbackup.
+
+See the docs at https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md .
+`
+
+	f := flag.CommandLine.Output()
+	fmt.Fprintf(f, "%s\n", s)
+	flag.PrintDefaults()
+}
+
+func newDstFS() (*fslocal.FS, error) {
+	if len(*storageDataPath) == 0 {
+		return nil, fmt.Errorf("`-storageDataPath` cannot be empty")
+	}
+	fs := &fslocal.FS{
+		Dir:               *storageDataPath,
+		MaxBytesPerSecond: *maxBytesPerSecond,
+	}
+	if err := fs.Init(); err != nil {
+		return nil, fmt.Errorf("cannot initialize local fs: %s", err)
+	}
+	return fs, nil
+}
+
+func newSrcFS() (common.RemoteFS, error) {
+	fs, err := actions.NewRemoteFS(*src)
+	if err != nil {
+		return nil, fmt.Errorf("cannot parse `-src`=%q: %s", *src, err)
+	}
+	return fs, nil
+}

app/vmselect/main.go

@@ -21,10 +21,26 @@ import (
 
 var (
 	deleteAuthKey         = flag.String("deleteAuthKey", "", "authKey for metrics' deletion via /api/v1/admin/tsdb/delete_series")
-	maxConcurrentRequests = flag.Int("search.maxConcurrentRequests", runtime.GOMAXPROCS(-1)*2, "The maximum number of concurrent search requests. It shouldn't exceed 2*vCPUs for better performance. See also -search.maxQueueDuration")
-	maxQueueDuration      = flag.Duration("search.maxQueueDuration", 10*time.Second, "The maximum time the request waits for execution when -search.maxConcurrentRequests limit is reached")
+	maxConcurrentRequests = flag.Int("search.maxConcurrentRequests", getDefaultMaxConcurrentRequests(), "The maximum number of concurrent search requests. "+
+		"It shouldn't be high, since a single request can saturate all the CPU cores. See also -search.maxQueueDuration")
+	maxQueueDuration  = flag.Duration("search.maxQueueDuration", 10*time.Second, "The maximum time the request waits for execution when -search.maxConcurrentRequests limit is reached")
+	resetCacheAuthKey = flag.String("search.resetCacheAuthKey", "", "Optional authKey for resetting rollup cache via /internal/resetCache call")
 )
 
+func getDefaultMaxConcurrentRequests() int {
+	n := runtime.GOMAXPROCS(-1)
+	if n <= 4 {
+		n *= 2
+	}
+	if n > 16 {
+		// A single request can saturate all the CPU cores, so there is no sense
+		// in allowing higher number of concurrent requests - they will just contend
+		// for unavailable CPU time.
+		n = 16
+	}
+	return n
+}
+
 // Init initializes vmselect
 func Init() {
 	tmpDirPath := *vmstorage.DataPath + "/tmp"
@@ -56,6 +72,7 @@ var (
 
 // RequestHandler handles remote read API requests for Prometheus
 func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
+	startTime := time.Now()
 	// Limit the number of concurrent queries.
 	select {
 	case concurrencyCh <- struct{}{}:
@@ -72,7 +89,9 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 			timerpool.Put(t)
 			concurrencyLimitTimeout.Inc()
 			err := &httpserver.ErrorWithStatusCode{
-				Err:        fmt.Errorf("cannot handle more than %d concurrent requests", cap(concurrencyCh)),
+				Err: fmt.Errorf("cannot handle more than %d concurrent search requests during %s; possible solutions: "+
+					"increase `-search.maxQueueDuration`, increase `-search.maxConcurrentRequests`, increase server capacity",
+					*maxConcurrentRequests, *maxQueueDuration),
 				StatusCode: http.StatusServiceUnavailable,
 			}
 			httpserver.Errorf(w, "%s", err)
@@ -81,13 +100,22 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	}
 
 	path := strings.Replace(r.URL.Path, "//", "/", -1)
+	if path == "/internal/resetRollupResultCache" {
+		if len(*resetCacheAuthKey) > 0 && r.FormValue("authKey") != *resetCacheAuthKey {
+			sendPrometheusError(w, r, fmt.Errorf("invalid authKey=%q for %q", r.FormValue("authKey"), path))
+			return true
+		}
+		promql.ResetRollupResultCache()
+		return true
+	}
+
 	if strings.HasPrefix(path, "/api/v1/label/") {
 		s := r.URL.Path[len("/api/v1/label/"):]
 		if strings.HasSuffix(s, "/values") {
 			labelValuesRequests.Inc()
 			labelName := s[:len(s)-len("/values")]
 			httpserver.EnableCORS(w, r)
-			if err := prometheus.LabelValuesHandler(labelName, w, r); err != nil {
+			if err := prometheus.LabelValuesHandler(startTime, labelName, w, r); err != nil {
 				labelValuesErrors.Inc()
 				sendPrometheusError(w, r, err)
 				return true
@@ -100,7 +128,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/query":
 		queryRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.QueryHandler(w, r); err != nil {
+		if err := prometheus.QueryHandler(startTime, w, r); err != nil {
 			queryErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -109,7 +137,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/query_range":
 		queryRangeRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.QueryRangeHandler(w, r); err != nil {
+		if err := prometheus.QueryRangeHandler(startTime, w, r); err != nil {
 			queryRangeErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -118,7 +146,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/series":
 		seriesRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.SeriesHandler(w, r); err != nil {
+		if err := prometheus.SeriesHandler(startTime, w, r); err != nil {
 			seriesErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -127,7 +155,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/series/count":
 		seriesCountRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.SeriesCountHandler(w, r); err != nil {
+		if err := prometheus.SeriesCountHandler(startTime, w, r); err != nil {
 			seriesCountErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -136,7 +164,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/labels":
 		labelsRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.LabelsHandler(w, r); err != nil {
+		if err := prometheus.LabelsHandler(startTime, w, r); err != nil {
 			labelsErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -145,7 +173,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 	case "/api/v1/labels/count":
 		labelsCountRequests.Inc()
 		httpserver.EnableCORS(w, r)
-		if err := prometheus.LabelsCountHandler(w, r); err != nil {
+		if err := prometheus.LabelsCountHandler(startTime, w, r); err != nil {
 			labelsCountErrors.Inc()
 			sendPrometheusError(w, r, err)
 			return true
@@ -153,7 +181,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		return true
 	case "/api/v1/export":
 		exportRequests.Inc()
-		if err := prometheus.ExportHandler(w, r); err != nil {
+		if err := prometheus.ExportHandler(startTime, w, r); err != nil {
 			exportErrors.Inc()
 			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
 			return true
@@ -161,12 +189,30 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 		return true
 	case "/federate":
 		federateRequests.Inc()
-		if err := prometheus.FederateHandler(w, r); err != nil {
+		if err := prometheus.FederateHandler(startTime, w, r); err != nil {
 			federateErrors.Inc()
 			httpserver.Errorf(w, "error int %q: %s", r.URL.Path, err)
 			return true
 		}
 		return true
+	case "/api/v1/rules":
+		// Return dumb placeholder
+		rulesRequests.Inc()
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprintf(w, "%s", `{"status":"success","data":{"groups":[]}}`)
+		return true
+	case "/api/v1/alerts":
+		// Return dumb placehloder
+		alertsRequests.Inc()
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprintf(w, "%s", `{"status":"success","data":{"alerts":[]}}`)
+		return true
+	case "/api/v1/metadata":
+		// Return dumb placeholder
+		metadataRequests.Inc()
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprintf(w, "%s", `{"status":"success","data":{}}`)
+		return true
 	case "/api/v1/admin/tsdb/delete_series":
 		deleteRequests.Inc()
 		authKey := r.FormValue("authKey")
@@ -174,7 +220,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 			httpserver.Errorf(w, "invalid authKey %q. It must match the value from -deleteAuthKey command line flag", authKey)
 			return true
 		}
-		if err := prometheus.DeleteHandler(r); err != nil {
+		if err := prometheus.DeleteHandler(startTime, r); err != nil {
 			deleteErrors.Inc()
 			httpserver.Errorf(w, "error in %q: %s", r.URL.Path, err)
 			return true
@@ -187,7 +233,7 @@ func RequestHandler(w http.ResponseWriter, r *http.Request) bool {
 }
 
 func sendPrometheusError(w http.ResponseWriter, r *http.Request, err error) {
-	logger.Errorf("error in %q: %s", r.URL.Path, err)
+	logger.Errorf("error in %q: %s", r.RequestURI, err)
 
 	w.Header().Set("Content-Type", "application/json")
 	statusCode := http.StatusUnprocessableEntity
@@ -228,4 +274,8 @@ var (
 
 	federateRequests = metrics.NewCounter(`vm_http_requests_total{path="/federate"}`)
 	federateErrors   = metrics.NewCounter(`vm_http_request_errors_total{path="/federate"}`)
+
+	rulesRequests    = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/rules"}`)
+	alertsRequests   = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/alerts"}`)
+	metadataRequests = metrics.NewCounter(`vm_http_requests_total{path="/api/v1/metadata"}`)
 )

app/vmselect/netstorage/fadvise_darwin.go

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

app/vmselect/netstorage/fadvise_freebsd.go

@@ -1,15 +0,0 @@
-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

@@ -1,15 +0,0 @@
-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/netstorage.go

@@ -92,6 +92,7 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	doneCh := make(chan error)
 
 	// Start workers.
+	rowsProcessedTotal := uint64(0)
 	for i := 0; i < workersCount; i++ {
 		go func(workerID uint) {
 			rs := getResult()
@@ -99,9 +100,10 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 			maxWorkersCount := gomaxprocs / workersCount
 
 			var err error
+			rowsProcessed := 0
 			for pts := range workCh {
 				if time.Until(rss.deadline.Deadline) < 0 {
-					err = fmt.Errorf("timeout exceeded during query execution: %s", rss.deadline.Timeout)
+					err = fmt.Errorf("timeout exceeded during query execution: %s", rss.deadline.String())
 					break
 				}
 				if err = pts.Unpack(rss.tbf, rs, rss.tr, rss.fetchData, maxWorkersCount); err != nil {
@@ -111,8 +113,10 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 					// Skip empty blocks.
 					continue
 				}
+				rowsProcessed += len(rs.Values)
 				f(rs, workerID)
 			}
+			atomic.AddUint64(&rowsProcessedTotal, uint64(rowsProcessed))
 			// Drain the remaining work
 			for range workCh {
 			}
@@ -124,6 +128,7 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	for i := range rss.packedTimeseries {
 		workCh <- &rss.packedTimeseries[i]
 	}
+	seriesProcessedTotal := len(rss.packedTimeseries)
 	rss.packedTimeseries = rss.packedTimeseries[:0]
 	close(workCh)
 
@@ -134,6 +139,8 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 			errors = append(errors, err)
 		}
 	}
+	perQueryRowsProcessed.Update(float64(rowsProcessedTotal))
+	perQuerySeriesProcessed.Update(float64(seriesProcessedTotal))
 	if len(errors) > 0 {
 		// Return just the first error, since other errors
 		// is likely duplicate the first error.
@@ -142,6 +149,9 @@ func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	return nil
 }
 
+var perQueryRowsProcessed = metrics.NewHistogram(`vm_per_query_rows_processed_count`)
+var perQuerySeriesProcessed = metrics.NewHistogram(`vm_per_query_series_processed_count`)
+
 var gomaxprocs = runtime.GOMAXPROCS(-1)
 
 type packedTimeseries struct {
@@ -256,7 +266,7 @@ func mergeSortBlocks(dst *Result, sbh sortBlocksHeap) {
 			dst.Timestamps = append(dst.Timestamps, top.Timestamps[top.NextIdx:]...)
 			dst.Values = append(dst.Values, top.Values[top.NextIdx:]...)
 			putSortBlock(top)
-			return
+			break
 		}
 		sbNext := sbh[0]
 		tsNext := sbNext.Timestamps[sbNext.NextIdx]
@@ -277,8 +287,16 @@ func mergeSortBlocks(dst *Result, sbh sortBlocksHeap) {
 			putSortBlock(top)
 		}
 	}
+
+	timestamps, values := storage.DeduplicateSamples(dst.Timestamps, dst.Values)
+	dedups := len(dst.Timestamps) - len(timestamps)
+	dedupsDuringSelect.Add(dedups)
+	dst.Timestamps = timestamps
+	dst.Values = values
 }
 
+var dedupsDuringSelect = metrics.NewCounter(`vm_deduplicated_samples_total{type="select"}`)
+
 type sortBlock struct {
 	// b is used as a temporary storage for unpacked rows before they
 	// go to Timestamps and Values.
@@ -422,13 +440,10 @@ func GetLabelEntries(deadline Deadline) ([]storage.TagEntry, error) {
 	// Sort labelEntries by the number of label values in each entry.
 	sort.Slice(labelEntries, func(i, j int) bool {
 		a, b := labelEntries[i].Values, labelEntries[j].Values
-		if len(a) < len(b) {
-			return true
+		if len(a) != len(b) {
+			return len(a) > len(b)
 		}
-		if len(a) > len(b) {
-			return false
-		}
-		return labelEntries[i].Key < labelEntries[j].Key
+		return labelEntries[i].Key > labelEntries[j].Key
 	})
 
 	return labelEntries, nil
@@ -452,16 +467,12 @@ func getStorageSearch() *storage.Search {
 }
 
 func putStorageSearch(sr *storage.Search) {
-	n := atomic.LoadUint64(&sr.MissingMetricNamesForMetricID)
-	missingMetricNamesForMetricID.Add(int(n))
 	sr.MustClose()
 	ssPool.Put(sr)
 }
 
 var ssPool sync.Pool
 
-var missingMetricNamesForMetricID = metrics.NewCounter(`vm_missing_metric_names_for_metric_id_total`)
-
 // ProcessSearchQuery performs sq on storage nodes until the given deadline.
 func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadline) (*Results, error) {
 	// Setup search.
@@ -496,7 +507,7 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline Deadli
 		}
 		if time.Until(deadline.Deadline) < 0 {
 			putTmpBlocksFile(tbf)
-			return nil, fmt.Errorf("timeout exceeded while fetching data block #%d from storage: %s", blocksRead, deadline.Timeout)
+			return nil, fmt.Errorf("timeout exceeded while fetching data block #%d from storage: %s", blocksRead, deadline.String())
 		}
 		metricName := sr.MetricBlock.MetricName
 		m[string(metricName)] = append(m[string(metricName)], addr)
@@ -572,13 +583,24 @@ func setupTfss(tagFilterss [][]storage.TagFilter) ([]*storage.TagFilters, error)
 // Deadline contains deadline with the corresponding timeout for pretty error messages.
 type Deadline struct {
 	Deadline time.Time
-	Timeout  time.Duration
+
+	timeout  time.Duration
+	flagHint string
 }
 
 // NewDeadline returns deadline for the given timeout.
-func NewDeadline(timeout time.Duration) Deadline {
+//
+// flagHint must contain a hit for command-line flag, which could be used
+// in order to increase timeout.
+func NewDeadline(timeout time.Duration, flagHint string) Deadline {
 	return Deadline{
 		Deadline: time.Now().Add(timeout),
-		Timeout:  timeout,
+		timeout:  timeout,
+		flagHint: flagHint,
 	}
 }
+
+// String returns human-readable string representation for d.
+func (d *Deadline) String() string {
+	return fmt.Sprintf("%.3f seconds; the timeout can be adjusted with `%s` command-line flag", d.timeout.Seconds(), d.flagHint)
+}

app/vmselect/netstorage/tmp_blocks_file.go

@@ -36,6 +36,9 @@ func maxInmemoryTmpBlocksFile() int {
 	if maxLen < 64*1024 {
 		return 64 * 1024
 	}
+	if maxLen > 4*1024*1024 {
+		return 4 * 1024 * 1024
+	}
 	return maxLen
 }
 
@@ -47,6 +50,7 @@ type tmpBlocksFile struct {
 	buf []byte
 
 	f *os.File
+	r *fs.ReaderAt
 
 	offset uint64
 }
@@ -65,6 +69,7 @@ func putTmpBlocksFile(tbf *tmpBlocksFile) {
 	tbf.MustClose()
 	tbf.buf = tbf.buf[:0]
 	tbf.f = nil
+	tbf.r = nil
 	tbf.offset = 0
 	tmpBlocksFilePool.Put(tbf)
 }
@@ -118,17 +123,20 @@ func (tbf *tmpBlocksFile) Finalize() error {
 	if tbf.f == nil {
 		return nil
 	}
+	fname := tbf.f.Name()
 	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)
+		return fmt.Errorf("cannot write the remaining %d bytes to %q: %s", len(tbf.buf), fname, 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)
+	r, err := fs.OpenReaderAt(fname)
+	if err != nil {
+		logger.Panicf("FATAL: cannot open %q: %s", fname, 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)
+	r.MustFadviseSequentialRead(true)
+	tbf.r = r
 	return nil
 }
 
@@ -140,13 +148,7 @@ func (tbf *tmpBlocksFile) MustReadBlockAt(dst *storage.Block, addr tmpBlockAddr)
 		bb := tmpBufPool.Get()
 		defer tmpBufPool.Put(bb)
 		bb.B = bytesutil.Resize(bb.B, addr.size)
-		n, err := tbf.f.ReadAt(bb.B, int64(addr.offset))
-		if err != nil {
-			logger.Panicf("FATAL: cannot read from %q at %s: %s", tbf.f.Name(), addr, err)
-		}
-		if n != len(bb.B) {
-			logger.Panicf("FATAL: too short number of bytes read at %s; got %d; want %d", addr, n, len(bb.B))
-		}
+		tbf.r.MustReadAt(bb.B, int64(addr.offset))
 		buf = bb.B
 	}
 	tail, err := storage.UnmarshalBlock(dst, buf)
@@ -164,6 +166,10 @@ func (tbf *tmpBlocksFile) MustClose() {
 	if tbf.f == nil {
 		return
 	}
+	if tbf.r != nil {
+		// tbf.r could be nil if Finalize wasn't called.
+		tbf.r.MustClose()
+	}
 	fname := tbf.f.Name()
 
 	// Remove the file at first, then close it.

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

@@ -3,6 +3,7 @@ package prometheus
 import (
 	"flag"
 	"fmt"
+	"io"
 	"math"
 	"net/http"
 	"runtime"
@@ -15,26 +16,28 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/promql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/fastjson/fastfloat"
 	"github.com/valyala/quicktemplate"
 )
 
 var (
-	maxQueryDuration = flag.Duration("search.maxQueryDuration", time.Second*30, "The maximum time for search query execution")
-	maxQueryLen      = flag.Int("search.maxQueryLen", 16*1024, "The maximum search query length in bytes")
+	latencyOffset = flag.Duration("search.latencyOffset", time.Second*30, "The time when data points become visible in query results after the colection. "+
+		"Too small value can result in incomplete last points for query results")
+	maxExportDuration = flag.Duration("search.maxExportDuration", time.Hour*24*30, "The maximum duration for /api/v1/export call")
+	maxQueryDuration  = flag.Duration("search.maxQueryDuration", time.Second*30, "The maximum duration for search query execution")
+	maxQueryLen       = flag.Int("search.maxQueryLen", 16*1024, "The maximum search query length in bytes")
+	maxLookback       = flag.Duration("search.maxLookback", 0, "Synonim to -search.lookback-delta from Prometheus. "+
+		"The value is dynamically detected from interval between time series datapoints if not set. It can be overridden on per-query basis via max_lookback arg")
 )
 
 // Default step used if not set.
 const defaultStep = 5 * 60 * 1000
 
-// Latency for data processing pipeline, i.e. the time between data is ignested
-// into the system and the time it becomes visible to search.
-const latencyOffset = 60 * 1000
-
 // FederateHandler implements /federate . See https://prometheus.io/docs/prometheus/latest/federation/
-func FederateHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
+func FederateHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 	if err := r.ParseForm(); err != nil {
 		return fmt.Errorf("cannot parse request form values: %s", err)
@@ -43,11 +46,14 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	if len(matches) == 0 {
 		return fmt.Errorf("missing `match[]` arg")
 	}
-	maxLookback, err := getDuration(r, "max_lookback", defaultStep)
+	lookbackDelta, err := getMaxLookback(r)
 	if err != nil {
 		return err
 	}
-	start, err := getTime(r, "start", ct-maxLookback)
+	if lookbackDelta <= 0 {
+		lookbackDelta = defaultStep
+	}
+	start, err := getTime(r, "start", ct-lookbackDelta)
 	if err != nil {
 		return err
 	}
@@ -55,7 +61,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 	if start >= end {
 		start = end - defaultStep
 	}
@@ -102,8 +108,7 @@ func FederateHandler(w http.ResponseWriter, r *http.Request) error {
 var federateDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/federate"}`)
 
 // ExportHandler exports data in raw format from /api/v1/export.
-func ExportHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
+func ExportHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 	if err := r.ParseForm(); err != nil {
 		return fmt.Errorf("cannot parse request form values: %s", err)
@@ -126,12 +131,13 @@ func ExportHandler(w http.ResponseWriter, r *http.Request) error {
 		return err
 	}
 	format := r.FormValue("format")
-	deadline := getDeadline(r)
+	maxRowsPerLine := int(fastfloat.ParseInt64BestEffort(r.FormValue("max_rows_per_line")))
+	deadline := getDeadlineForExport(r)
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
-	if err := exportHandler(w, matches, start, end, format, deadline); err != nil {
-		return err
+	if err := exportHandler(w, matches, start, end, format, maxRowsPerLine, deadline); err != nil {
+		return fmt.Errorf("error when exporting data for queries=%q on the time range (start=%d, end=%d): %s", matches, start, end, err)
 	}
 	exportDuration.UpdateDuration(startTime)
 	return nil
@@ -139,10 +145,38 @@ func ExportHandler(w http.ResponseWriter, r *http.Request) error {
 
 var exportDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/export"}`)
 
-func exportHandler(w http.ResponseWriter, matches []string, start, end int64, format string, deadline netstorage.Deadline) error {
+func exportHandler(w http.ResponseWriter, matches []string, start, end int64, format string, maxRowsPerLine int, deadline netstorage.Deadline) error {
 	writeResponseFunc := WriteExportStdResponse
 	writeLineFunc := WriteExportJSONLine
-	contentType := "application/json"
+	if maxRowsPerLine > 0 {
+		writeLineFunc = func(w io.Writer, rs *netstorage.Result) {
+			valuesOrig := rs.Values
+			timestampsOrig := rs.Timestamps
+			values := valuesOrig
+			timestamps := timestampsOrig
+			for len(values) > 0 {
+				var valuesChunk []float64
+				var timestampsChunk []int64
+				if len(values) > maxRowsPerLine {
+					valuesChunk = values[:maxRowsPerLine]
+					timestampsChunk = timestamps[:maxRowsPerLine]
+					values = values[maxRowsPerLine:]
+					timestamps = timestamps[maxRowsPerLine:]
+				} else {
+					valuesChunk = values
+					timestampsChunk = timestamps
+					values = nil
+					timestamps = nil
+				}
+				rs.Values = valuesChunk
+				rs.Timestamps = timestampsChunk
+				WriteExportJSONLine(w, rs)
+			}
+			rs.Values = valuesOrig
+			rs.Timestamps = timestampsOrig
+		}
+	}
+	contentType := "application/stream+json"
 	if format == "prometheus" {
 		contentType = "text/plain"
 		writeLineFunc = WriteExportPrometheusLine
@@ -195,8 +229,7 @@ func exportHandler(w http.ResponseWriter, matches []string, start, end int64, fo
 // DeleteHandler processes /api/v1/admin/tsdb/delete_series prometheus API request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#delete-series
-func DeleteHandler(r *http.Request) error {
-	startTime := time.Now()
+func DeleteHandler(startTime time.Time, r *http.Request) error {
 	if err := r.ParseForm(); err != nil {
 		return fmt.Errorf("cannot parse request form values: %s", err)
 	}
@@ -230,9 +263,8 @@ var deleteDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 // LabelValuesHandler processes /api/v1/label/<labelName>/values request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#querying-label-values
-func LabelValuesHandler(labelName string, w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
-	deadline := getDeadline(r)
+func LabelValuesHandler(startTime time.Time, labelName string, w http.ResponseWriter, r *http.Request) error {
+	deadline := getDeadlineForQuery(r)
 
 	if err := r.ParseForm(); err != nil {
 		return fmt.Errorf("cannot parse form values: %s", err)
@@ -282,8 +314,21 @@ func labelValuesWithMatches(labelName string, matches []string, start, end int64
 	if err != nil {
 		return nil, err
 	}
+
+	// Add `labelName!=''` tag filter in order to filter out series without the labelName.
+	// There is no need in adding `__name__!=''` filter, since all the time series should
+	// already have non-empty name.
+	if labelName != "__name__" {
+		key := []byte(labelName)
+		for i, tfs := range tagFilterss {
+			tagFilterss[i] = append(tfs, storage.TagFilter{
+				Key:        key,
+				IsNegative: true,
+			})
+		}
+	}
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
 	sq := &storage.SearchQuery{
 		MinTimestamp: start,
@@ -321,9 +366,8 @@ func labelValuesWithMatches(labelName string, matches []string, start, end int64
 var labelValuesDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/label/{}/values"}`)
 
 // LabelsCountHandler processes /api/v1/labels/count request.
-func LabelsCountHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
-	deadline := getDeadline(r)
+func LabelsCountHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
+	deadline := getDeadlineForQuery(r)
 	labelEntries, err := netstorage.GetLabelEntries(deadline)
 	if err != nil {
 		return fmt.Errorf(`cannot obtain label entries: %s`, err)
@@ -340,12 +384,39 @@ var labelsCountDuration = metrics.NewSummary(`vm_request_duration_seconds{path="
 // LabelsHandler processes /api/v1/labels request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#getting-label-names
-func LabelsHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
-	deadline := getDeadline(r)
-	labels, err := netstorage.GetLabels(deadline)
-	if err != nil {
-		return fmt.Errorf("cannot obtain labels: %s", err)
+func LabelsHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
+	deadline := getDeadlineForQuery(r)
+
+	if err := r.ParseForm(); err != nil {
+		return fmt.Errorf("cannot parse form values: %s", err)
+	}
+	var labels []string
+	if len(r.Form["match[]"]) == 0 && len(r.Form["start"]) == 0 && len(r.Form["end"]) == 0 {
+		var err error
+		labels, err = netstorage.GetLabels(deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain labels: %s", err)
+		}
+	} else {
+		// Extended functionality that allows filtering by label filters and time range
+		// i.e. /api/v1/labels?match[]=foobar{baz="abc"}&start=...&end=...
+		matches := r.Form["match[]"]
+		if len(matches) == 0 {
+			matches = []string{"{__name__!=''}"}
+		}
+		ct := currentTime()
+		end, err := getTime(r, "end", ct)
+		if err != nil {
+			return err
+		}
+		start, err := getTime(r, "start", end-defaultStep)
+		if err != nil {
+			return err
+		}
+		labels, err = labelsWithMatches(matches, start, end, deadline)
+		if err != nil {
+			return fmt.Errorf("cannot obtain labels for match[]=%q, start=%d, end=%d: %s", matches, start, end, err)
+		}
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -354,12 +425,56 @@ func LabelsHandler(w http.ResponseWriter, r *http.Request) error {
 	return nil
 }
 
+func labelsWithMatches(matches []string, start, end int64, deadline netstorage.Deadline) ([]string, error) {
+	if len(matches) == 0 {
+		logger.Panicf("BUG: matches must be non-empty")
+	}
+	tagFilterss, err := getTagFilterssFromMatches(matches)
+	if err != nil {
+		return nil, err
+	}
+	if start >= end {
+		end = start + defaultStep
+	}
+	sq := &storage.SearchQuery{
+		MinTimestamp: start,
+		MaxTimestamp: end,
+		TagFilterss:  tagFilterss,
+	}
+	rss, err := netstorage.ProcessSearchQuery(sq, false, deadline)
+	if err != nil {
+		return nil, fmt.Errorf("cannot fetch data for %q: %s", sq, err)
+	}
+
+	m := make(map[string]struct{})
+	var mLock sync.Mutex
+	err = rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
+		mLock.Lock()
+		tags := rs.MetricName.Tags
+		for i := range tags {
+			t := &tags[i]
+			m[string(t.Key)] = struct{}{}
+		}
+		m["__name__"] = struct{}{}
+		mLock.Unlock()
+	})
+	if err != nil {
+		return nil, fmt.Errorf("error when data fetching: %s", err)
+	}
+
+	labels := make([]string, 0, len(m))
+	for label := range m {
+		labels = append(labels, label)
+	}
+	sort.Strings(labels)
+	return labels, nil
+}
+
 var labelsDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/labels"}`)
 
 // SeriesCountHandler processes /api/v1/series/count request.
-func SeriesCountHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
-	deadline := getDeadline(r)
+func SeriesCountHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
+	deadline := getDeadlineForQuery(r)
 	n, err := netstorage.GetSeriesCount(deadline)
 	if err != nil {
 		return fmt.Errorf("cannot obtain series count: %s", err)
@@ -375,8 +490,7 @@ var seriesCountDuration = metrics.NewSummary(`vm_request_duration_seconds{path="
 // SeriesHandler processes /api/v1/series request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#finding-series-by-label-matchers
-func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
+func SeriesHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 
 	if err := r.ParseForm(); err != nil {
@@ -399,14 +513,14 @@ func SeriesHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	deadline := getDeadlineForQuery(r)
 
 	tagFilterss, err := getTagFilterssFromMatches(matches)
 	if err != nil {
 		return err
 	}
 	if start >= end {
-		start = end - defaultStep
+		end = start + defaultStep
 	}
 	sq := &storage.SearchQuery{
 		MinTimestamp: start,
@@ -451,8 +565,7 @@ var seriesDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/
 // QueryHandler processes /api/v1/query request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#instant-queries
-func QueryHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
+func QueryHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 
 	query := r.FormValue("query")
@@ -463,54 +576,79 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	step, err := getDuration(r, "step", latencyOffset)
+	queryOffset := getLatencyOffsetMilliseconds()
+	step, err := getDuration(r, "step", queryOffset)
+	if err != nil {
+		return err
+	}
+	deadline := getDeadlineForQuery(r)
+	lookbackDelta, err := getMaxLookback(r)
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
 
 	if len(query) > *maxQueryLen {
-		return fmt.Errorf(`too long query; got %d bytes; mustn't exceed %d bytes`, len(query), *maxQueryLen)
+		return fmt.Errorf("too long query; got %d bytes; mustn't exceed `-search.maxQueryLen=%d` bytes", len(query), *maxQueryLen)
 	}
-	if ct-start < latencyOffset {
-		start -= latencyOffset
+	if !getBool(r, "nocache") && ct-start < queryOffset {
+		// Adjust start time only if `nocache` arg isn't set.
+		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/241
+		start = ct - queryOffset
 	}
 	if childQuery, windowStr, offsetStr := promql.IsMetricSelectorWithRollup(query); childQuery != "" {
-		var window int64
-		if len(windowStr) > 0 {
-			var err error
-			window, err = promql.DurationValue(windowStr, step)
-			if err != nil {
-				return err
-			}
+		window, err := parsePositiveDuration(windowStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse window: %s", err)
 		}
-		var offset int64
-		if len(offsetStr) > 0 {
-			var err error
-			offset, err = promql.DurationValue(offsetStr, step)
-			if err != nil {
-				return err
-			}
+		offset, err := parseDuration(offsetStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse offset: %s", err)
 		}
 		start -= offset
 		end := start
 		start = end - window
-		if err := exportHandler(w, []string{childQuery}, start, end, "promapi", deadline); err != nil {
-			return err
+		if err := exportHandler(w, []string{childQuery}, start, end, "promapi", 0, deadline); err != nil {
+			return fmt.Errorf("error when exporting data for query=%q on the time range (start=%d, end=%d): %s", childQuery, start, end, err)
+		}
+		queryDuration.UpdateDuration(startTime)
+		return nil
+	}
+	if childQuery, windowStr, stepStr, offsetStr := promql.IsRollup(query); childQuery != "" {
+		newStep, err := parsePositiveDuration(stepStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse step: %s", err)
+		}
+		if newStep > 0 {
+			step = newStep
+		}
+		window, err := parsePositiveDuration(windowStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse window: %s", err)
+		}
+		offset, err := parseDuration(offsetStr, step)
+		if err != nil {
+			return fmt.Errorf("cannot parse offset: %s", err)
+		}
+		start -= offset
+		end := start
+		start = end - window
+		if err := queryRangeHandler(w, childQuery, start, end, step, r, ct); err != nil {
+			return fmt.Errorf("error when executing query=%q on the time range (start=%d, end=%d, step=%d): %s", childQuery, start, end, step, err)
 		}
 		queryDuration.UpdateDuration(startTime)
 		return nil
 	}
 
 	ec := promql.EvalConfig{
-		Start:    start,
-		End:      start,
-		Step:     step,
-		Deadline: deadline,
+		Start:         start,
+		End:           start,
+		Step:          step,
+		Deadline:      deadline,
+		LookbackDelta: lookbackDelta,
 	}
 	result, err := promql.Exec(&ec, query, true)
 	if err != nil {
-		return fmt.Errorf("cannot execute %q: %s", query, err)
+		return fmt.Errorf("error when executing query=%q for (time=%d, step=%d): %s", query, start, step, err)
 	}
 
 	w.Header().Set("Content-Type", "application/json")
@@ -521,11 +659,24 @@ func QueryHandler(w http.ResponseWriter, r *http.Request) error {
 
 var queryDuration = metrics.NewSummary(`vm_request_duration_seconds{path="/api/v1/query"}`)
 
+func parseDuration(s string, step int64) (int64, error) {
+	if len(s) == 0 {
+		return 0, nil
+	}
+	return metricsql.DurationValue(s, step)
+}
+
+func parsePositiveDuration(s string, step int64) (int64, error) {
+	if len(s) == 0 {
+		return 0, nil
+	}
+	return metricsql.PositiveDurationValue(s, step)
+}
+
 // QueryRangeHandler processes /api/v1/query_range request.
 //
 // See https://prometheus.io/docs/prometheus/latest/querying/api/#range-queries
-func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
-	startTime := time.Now()
+func QueryRangeHandler(startTime time.Time, w http.ResponseWriter, r *http.Request) error {
 	ct := currentTime()
 
 	query := r.FormValue("query")
@@ -544,15 +695,27 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	if err != nil {
 		return err
 	}
-	deadline := getDeadline(r)
+	if err := queryRangeHandler(w, query, start, end, step, r, ct); err != nil {
+		return fmt.Errorf("error when executing query=%q on the time range (start=%d, end=%d, step=%d): %s", query, start, end, step, err)
+	}
+	queryRangeDuration.UpdateDuration(startTime)
+	return nil
+}
+
+func queryRangeHandler(w http.ResponseWriter, query string, start, end, step int64, r *http.Request, ct int64) error {
+	deadline := getDeadlineForQuery(r)
 	mayCache := !getBool(r, "nocache")
+	lookbackDelta, err := getMaxLookback(r)
+	if err != nil {
+		return err
+	}
 
 	// Validate input args.
 	if len(query) > *maxQueryLen {
-		return fmt.Errorf(`too long query; got %d bytes; mustn't exceed %d bytes`, len(query), *maxQueryLen)
+		return fmt.Errorf("too long query; got %d bytes; mustn't exceed `-search.maxQueryLen=%d` bytes", len(query), *maxQueryLen)
 	}
 	if start > end {
-		start = end
+		end = start + defaultStep
 	}
 	if err := promql.ValidateMaxPointsPerTimeseries(start, end, step); err != nil {
 		return err
@@ -562,17 +725,19 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 	}
 
 	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)
+		return fmt.Errorf("cannot execute query: %s", err)
 	}
-	if ct-end < latencyOffset {
+	queryOffset := getLatencyOffsetMilliseconds()
+	if ct-end < queryOffset {
 		result = adjustLastPoints(result)
 	}
 
@@ -582,7 +747,6 @@ func QueryRangeHandler(w http.ResponseWriter, r *http.Request) error {
 
 	w.Header().Set("Content-Type", "application/json")
 	WriteQueryRangeResponse(w, result)
-	queryRangeDuration.UpdateDuration(startTime)
 	return nil
 }
 
@@ -726,17 +890,31 @@ func getDuration(r *http.Request, argKey string, defaultValue int64) (int64, err
 
 const maxDurationMsecs = 100 * 365 * 24 * 3600 * 1000
 
-func getDeadline(r *http.Request) netstorage.Deadline {
+func getMaxLookback(r *http.Request) (int64, error) {
+	d := int64(*maxLookback / time.Millisecond)
+	return getDuration(r, "max_lookback", d)
+}
+
+func getDeadlineForQuery(r *http.Request) netstorage.Deadline {
+	dMax := int64(maxQueryDuration.Seconds() * 1e3)
+	return getDeadlineWithMaxDuration(r, dMax, "-search.maxQueryDuration")
+}
+
+func getDeadlineForExport(r *http.Request) netstorage.Deadline {
+	dMax := int64(maxExportDuration.Seconds() * 1e3)
+	return getDeadlineWithMaxDuration(r, dMax, "-search.maxExportDuration")
+}
+
+func getDeadlineWithMaxDuration(r *http.Request, dMax int64, flagHint string) netstorage.Deadline {
 	d, err := getDuration(r, "timeout", 0)
 	if err != nil {
 		d = 0
 	}
-	dMax := int64(maxQueryDuration.Seconds() * 1e3)
 	if d <= 0 || d > dMax {
 		d = dMax
 	}
 	timeout := time.Duration(d) * time.Millisecond
-	return netstorage.NewDeadline(timeout)
+	return netstorage.NewDeadline(timeout, flagHint)
 }
 
 func getBool(r *http.Request, argKey string) bool {
@@ -764,3 +942,11 @@ 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/series_count_response.qtpl

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

app/vmselect/prometheus/series_count_response.qtpl.go

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

app/vmselect/promql/aggr.go

@@ -8,7 +8,10 @@ import (
 	"strings"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/metrics"
+	"github.com/valyala/histogram"
 )
 
 var aggrFuncs = map[string]aggrFunc{
@@ -25,19 +28,28 @@ var aggrFuncs = map[string]aggrFunc{
 	"topk":         newAggrFuncTopK(false),
 	"quantile":     aggrFuncQuantile,
 
-	// Extended PromQL funcs
-	"median":   aggrFuncMedian,
-	"limitk":   aggrFuncLimitK,
-	"distinct": newAggrFunc(aggrFuncDistinct),
-	"sum2":     newAggrFunc(aggrFuncSum2),
-	"geomean":  newAggrFunc(aggrFuncGeomean),
+	// PromQL extension funcs
+	"median":         aggrFuncMedian,
+	"limitk":         aggrFuncLimitK,
+	"distinct":       newAggrFunc(aggrFuncDistinct),
+	"sum2":           newAggrFunc(aggrFuncSum2),
+	"geomean":        newAggrFunc(aggrFuncGeomean),
+	"histogram":      newAggrFunc(aggrFuncHistogram),
+	"topk_min":       newAggrFuncRangeTopK(minValue, false),
+	"topk_max":       newAggrFuncRangeTopK(maxValue, false),
+	"topk_avg":       newAggrFuncRangeTopK(avgValue, false),
+	"topk_median":    newAggrFuncRangeTopK(medianValue, false),
+	"bottomk_min":    newAggrFuncRangeTopK(minValue, true),
+	"bottomk_max":    newAggrFuncRangeTopK(maxValue, true),
+	"bottomk_avg":    newAggrFuncRangeTopK(avgValue, true),
+	"bottomk_median": newAggrFuncRangeTopK(medianValue, true),
 }
 
 type aggrFunc func(afa *aggrFuncArg) ([]*timeseries, error)
 
 type aggrFuncArg struct {
 	args [][]*timeseries
-	ae   *aggrFuncExpr
+	ae   *metricsql.AggrFuncExpr
 	ec   *EvalConfig
 }
 
@@ -46,20 +58,6 @@ func getAggrFunc(s string) aggrFunc {
 	return aggrFuncs[s]
 }
 
-func isAggrFunc(s string) bool {
-	return getAggrFunc(s) != nil
-}
-
-func isAggrFuncModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	case "by", "without":
-		return true
-	default:
-		return false
-	}
-}
-
 func newAggrFunc(afe func(tss []*timeseries) []*timeseries) aggrFunc {
 	return func(afa *aggrFuncArg) ([]*timeseries, error) {
 		args := afa.args
@@ -70,7 +68,7 @@ func newAggrFunc(afe func(tss []*timeseries) []*timeseries) aggrFunc {
 	}
 }
 
-func removeGroupTags(metricName *storage.MetricName, modifier *modifierExpr) {
+func removeGroupTags(metricName *storage.MetricName, modifier *metricsql.ModifierExpr) {
 	groupOp := strings.ToLower(modifier.Op)
 	switch groupOp {
 	case "", "by":
@@ -82,7 +80,7 @@ func removeGroupTags(metricName *storage.MetricName, modifier *modifierExpr) {
 	}
 }
 
-func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeseries, modifier *modifierExpr, keepOriginal bool) ([]*timeseries, error) {
+func aggrFuncExt(afe func(tss []*timeseries) []*timeseries, argOrig []*timeseries, modifier *metricsql.ModifierExpr, keepOriginal bool) ([]*timeseries, error) {
 	arg := copyTimeseriesMetricNames(argOrig)
 
 	// Perform grouping.
@@ -184,6 +182,38 @@ func aggrFuncGeomean(tss []*timeseries) []*timeseries {
 	return tss[:1]
 }
 
+func aggrFuncHistogram(tss []*timeseries) []*timeseries {
+	var h metrics.Histogram
+	m := make(map[string]*timeseries)
+	for i := range tss[0].Values {
+		h.Reset()
+		for _, ts := range tss {
+			v := ts.Values[i]
+			h.Update(v)
+		}
+		h.VisitNonZeroBuckets(func(vmrange string, count uint64) {
+			ts := m[vmrange]
+			if ts == nil {
+				ts = &timeseries{}
+				ts.CopyFromShallowTimestamps(tss[0])
+				ts.MetricName.RemoveTag("vmrange")
+				ts.MetricName.AddTag("vmrange", vmrange)
+				values := ts.Values
+				for k := range values {
+					values[k] = 0
+				}
+				m[vmrange] = ts
+			}
+			ts.Values[i] = float64(count)
+		})
+	}
+	rvs := make([]*timeseries, 0, len(m))
+	for _, ts := range m {
+		rvs = append(rvs, ts)
+	}
+	return vmrangeBucketsToLE(rvs)
+}
+
 func aggrFuncMin(tss []*timeseries) []*timeseries {
 	if len(tss) == 1 {
 		// Fast path - nothing to min.
@@ -425,37 +455,138 @@ func newAggrFuncTopK(isReverse bool) aggrFunc {
 			return nil, err
 		}
 		afe := func(tss []*timeseries) []*timeseries {
-			rvs := tss
-			for n := range rvs[0].Values {
-				sort.Slice(rvs, func(i, j int) bool {
-					a := rvs[i].Values[n]
-					b := rvs[j].Values[n]
-					cmp := lessWithNaNs(a, b)
+			for n := range tss[0].Values {
+				sort.Slice(tss, func(i, j int) bool {
+					a := tss[i].Values[n]
+					b := tss[j].Values[n]
 					if isReverse {
-						cmp = !cmp
+						a, b = b, a
 					}
-					return cmp
+					return lessWithNaNs(a, b)
 				})
-				if math.IsNaN(ks[n]) {
-					ks[n] = 0
-				}
-				k := int(ks[n])
-				if k < 0 {
-					k = 0
-				}
-				if k > len(rvs) {
-					k = len(rvs)
-				}
-				for _, ts := range rvs[:len(rvs)-k] {
-					ts.Values[n] = nan
-				}
+				fillNaNsAtIdx(n, ks[n], tss)
 			}
-			return removeNaNs(rvs)
+			return removeNaNs(tss)
 		}
 		return aggrFuncExt(afe, args[1], &afa.ae.Modifier, true)
 	}
 }
 
+type tsWithValue struct {
+	ts    *timeseries
+	value float64
+}
+
+func newAggrFuncRangeTopK(f func(values []float64) float64, isReverse bool) aggrFunc {
+	return func(afa *aggrFuncArg) ([]*timeseries, error) {
+		args := afa.args
+		if err := expectTransformArgsNum(args, 2); err != nil {
+			return nil, err
+		}
+		ks, err := getScalar(args[0], 0)
+		if err != nil {
+			return nil, err
+		}
+		afe := func(tss []*timeseries) []*timeseries {
+			maxs := make([]tsWithValue, len(tss))
+			for i, ts := range tss {
+				value := f(ts.Values)
+				maxs[i] = tsWithValue{
+					ts:    ts,
+					value: value,
+				}
+			}
+			sort.Slice(maxs, func(i, j int) bool {
+				a := maxs[i].value
+				b := maxs[j].value
+				if isReverse {
+					a, b = b, a
+				}
+				return lessWithNaNs(a, b)
+			})
+			for i := range maxs {
+				tss[i] = maxs[i].ts
+			}
+			for i, k := range ks {
+				fillNaNsAtIdx(i, k, tss)
+			}
+			return removeNaNs(tss)
+		}
+		return aggrFuncExt(afe, args[1], &afa.ae.Modifier, true)
+	}
+}
+
+func fillNaNsAtIdx(idx int, k float64, tss []*timeseries) {
+	if math.IsNaN(k) {
+		k = 0
+	}
+	kn := int(k)
+	if kn < 0 {
+		kn = 0
+	}
+	if kn > len(tss) {
+		kn = len(tss)
+	}
+	for _, ts := range tss[:len(tss)-kn] {
+		ts.Values[idx] = nan
+	}
+}
+
+func minValue(values []float64) float64 {
+	if len(values) == 0 {
+		return nan
+	}
+	min := values[0]
+	for _, v := range values[1:] {
+		if v < min {
+			min = v
+		}
+	}
+	return min
+}
+
+func maxValue(values []float64) float64 {
+	if len(values) == 0 {
+		return nan
+	}
+	max := values[0]
+	for _, v := range values[1:] {
+		if v > max {
+			max = v
+		}
+	}
+	return max
+}
+
+func avgValue(values []float64) float64 {
+	sum := float64(0)
+	count := 0
+	for _, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		count++
+		sum += v
+	}
+	if count == 0 {
+		return nan
+	}
+	return sum / float64(count)
+}
+
+func medianValue(values []float64) float64 {
+	h := histogram.GetFast()
+	for _, v := range values {
+		if math.IsNaN(v) {
+			continue
+		}
+		h.Update(v)
+	}
+	value := h.Quantile(0.5)
+	histogram.PutFast(h)
+	return value
+}
+
 func aggrFuncLimitK(afa *aggrFuncArg) ([]*timeseries, error) {
 	args := afa.args
 	if err := expectTransformArgsNum(args, 2); err != nil {

app/vmselect/promql/aggr_incremental.go

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

app/vmselect/promql/aggr_incremental_test.go

@@ -7,6 +7,8 @@ import (
 	"runtime"
 	"sync"
 	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 )
 
 func TestIncrementalAggr(t *testing.T) {
@@ -42,7 +44,7 @@ func TestIncrementalAggr(t *testing.T) {
 	f := func(name string, valuesExpected []float64) {
 		t.Helper()
 		callbacks := getIncrementalAggrFuncCallbacks(name)
-		ae := &aggrFuncExpr{
+		ae := &metricsql.AggrFuncExpr{
 			Name: name,
 		}
 		tssExpected := []*timeseries{{

app/vmselect/promql/arch_amd64.go

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

app/vmselect/promql/arch_arm.go

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

app/vmselect/promql/arch_arm64.go

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

app/vmselect/promql/binary_op.go

@@ -6,63 +6,36 @@ import (
 	"strings"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql/binaryop"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 )
 
 var binaryOpFuncs = map[string]binaryOpFunc{
-	"+": newBinaryOpArithFunc(binaryOpPlus),
-	"-": newBinaryOpArithFunc(binaryOpMinus),
-	"*": newBinaryOpArithFunc(binaryOpMul),
-	"/": newBinaryOpArithFunc(binaryOpDiv),
-	"%": newBinaryOpArithFunc(binaryOpMod),
-	"^": newBinaryOpArithFunc(binaryOpPow),
+	"+": newBinaryOpArithFunc(binaryop.Plus),
+	"-": newBinaryOpArithFunc(binaryop.Minus),
+	"*": newBinaryOpArithFunc(binaryop.Mul),
+	"/": newBinaryOpArithFunc(binaryop.Div),
+	"%": newBinaryOpArithFunc(binaryop.Mod),
+	"^": newBinaryOpArithFunc(binaryop.Pow),
 
 	// cmp ops
-	"==": newBinaryOpCmpFunc(binaryOpEq),
-	"!=": newBinaryOpCmpFunc(binaryOpNeq),
-	">":  newBinaryOpCmpFunc(binaryOpGt),
-	"<":  newBinaryOpCmpFunc(binaryOpLt),
-	">=": newBinaryOpCmpFunc(binaryOpGte),
-	"<=": newBinaryOpCmpFunc(binaryOpLte),
+	"==": newBinaryOpCmpFunc(binaryop.Eq),
+	"!=": newBinaryOpCmpFunc(binaryop.Neq),
+	">":  newBinaryOpCmpFunc(binaryop.Gt),
+	"<":  newBinaryOpCmpFunc(binaryop.Lt),
+	">=": newBinaryOpCmpFunc(binaryop.Gte),
+	"<=": newBinaryOpCmpFunc(binaryop.Lte),
 
 	// logical set ops
 	"and":    binaryOpAnd,
 	"or":     binaryOpOr,
 	"unless": binaryOpUnless,
 
-	// New op
-	"if":      newBinaryOpArithFunc(binaryOpIf),
-	"ifnot":   newBinaryOpArithFunc(binaryOpIfnot),
-	"default": newBinaryOpArithFunc(binaryOpDefault),
-}
-
-var binaryOpPriorities = map[string]int{
-	"default": -1,
-
-	"if":    0,
-	"ifnot": 0,
-
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#binary-operator-precedence
-	"or": 1,
-
-	"and":    2,
-	"unless": 2,
-
-	"==": 3,
-	"!=": 3,
-	"<":  3,
-	">":  3,
-	"<=": 3,
-	">=": 3,
-
-	"+": 4,
-	"-": 4,
-
-	"*": 5,
-	"/": 5,
-	"%": 5,
-
-	"^": 6,
+	// New ops
+	"if":      newBinaryOpArithFunc(binaryop.If),
+	"ifnot":   newBinaryOpArithFunc(binaryop.Ifnot),
+	"default": newBinaryOpArithFunc(binaryop.Default),
 }
 
 func getBinaryOpFunc(op string) binaryOpFunc {
@@ -70,144 +43,8 @@ func getBinaryOpFunc(op string) binaryOpFunc {
 	return binaryOpFuncs[op]
 }
 
-func isBinaryOp(op string) bool {
-	return getBinaryOpFunc(op) != nil
-}
-
-func binaryOpPriority(op string) int {
-	op = strings.ToLower(op)
-	return binaryOpPriorities[op]
-}
-
-func scanBinaryOpPrefix(s string) int {
-	n := 0
-	for op := range binaryOpFuncs {
-		if len(s) < len(op) {
-			continue
-		}
-		ss := strings.ToLower(s[:len(op)])
-		if ss == op && len(op) > n {
-			n = len(op)
-		}
-	}
-	return n
-}
-
-func isRightAssociativeBinaryOp(op string) bool {
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#binary-operator-precedence
-	return op == "^"
-}
-
-func isBinaryOpGroupModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	// See https://prometheus.io/docs/prometheus/latest/querying/operators/#vector-matching
-	case "on", "ignoring":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpJoinModifier(s string) bool {
-	s = strings.ToLower(s)
-	switch s {
-	case "group_left", "group_right":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpBoolModifier(s string) bool {
-	s = strings.ToLower(s)
-	return s == "bool"
-}
-
-func isBinaryOpCmp(op string) bool {
-	switch op {
-	case "==", "!=", ">", "<", ">=", "<=":
-		return true
-	default:
-		return false
-	}
-}
-
-func isBinaryOpLogicalSet(op string) bool {
-	op = strings.ToLower(op)
-	switch op {
-	case "and", "or", "unless":
-		return true
-	default:
-		return false
-	}
-}
-
-func binaryOpConstants(op string, left, right float64, isBool bool) float64 {
-	if isBinaryOpCmp(op) {
-		evalCmp := func(cf func(left, right float64) bool) float64 {
-			if isBool {
-				if cf(left, right) {
-					return 1
-				}
-				return 0
-			}
-			if cf(left, right) {
-				return left
-			}
-			return nan
-		}
-		switch op {
-		case "==":
-			left = evalCmp(binaryOpEq)
-		case "!=":
-			left = evalCmp(binaryOpNeq)
-		case ">":
-			left = evalCmp(binaryOpGt)
-		case "<":
-			left = evalCmp(binaryOpLt)
-		case ">=":
-			left = evalCmp(binaryOpGte)
-		case "<=":
-			left = evalCmp(binaryOpLte)
-		default:
-			logger.Panicf("BUG: unexpected comparison binaryOp: %q", op)
-		}
-	} else {
-		switch op {
-		case "+":
-			left = binaryOpPlus(left, right)
-		case "-":
-			left = binaryOpMinus(left, right)
-		case "*":
-			left = binaryOpMul(left, right)
-		case "/":
-			left = binaryOpDiv(left, right)
-		case "%":
-			left = binaryOpMod(left, right)
-		case "^":
-			left = binaryOpPow(left, right)
-		case "and":
-			// Nothing to do
-		case "or":
-			// Nothing to do
-		case "unless":
-			left = nan
-		case "default":
-			left = binaryOpDefault(left, right)
-		case "if":
-			left = binaryOpIf(left, right)
-		case "ifnot":
-			left = binaryOpIfnot(left, right)
-		default:
-			logger.Panicf("BUG: unexpected non-comparison binaryOp: %q", op)
-		}
-	}
-	return left
-}
-
 type binaryOpFuncArg struct {
-	be    *binaryOpExpr
+	be    *metricsql.BinaryOpExpr
 	left  []*timeseries
 	right []*timeseries
 }
@@ -267,7 +104,7 @@ func newBinaryOpFunc(bf func(left, right float64, isBool bool) float64) binaryOp
 	}
 }
 
-func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeseries, []*timeseries, []*timeseries, error) {
+func adjustBinaryOpTags(be *metricsql.BinaryOpExpr, left, right []*timeseries) ([]*timeseries, []*timeseries, []*timeseries, error) {
 	if len(be.GroupModifier.Op) == 0 && len(be.JoinModifier.Op) == 0 {
 		if isScalar(left) {
 			// Fast path: `scalar op vector`
@@ -348,7 +185,7 @@ func adjustBinaryOpTags(be *binaryOpExpr, left, right []*timeseries) ([]*timeser
 	return rvsLeft, rvsRight, dst, nil
 }
 
-func ensureSingleTimeseries(side string, be *binaryOpExpr, tss []*timeseries) error {
+func ensureSingleTimeseries(side string, be *metricsql.BinaryOpExpr, tss []*timeseries) error {
 	if len(tss) == 0 {
 		logger.Panicf("BUG: tss must contain at least one value")
 	}
@@ -362,7 +199,7 @@ func ensureSingleTimeseries(side string, be *binaryOpExpr, tss []*timeseries) er
 	return nil
 }
 
-func groupJoin(singleTimeseriesSide string, be *binaryOpExpr, rvsLeft, rvsRight, tssLeft, tssRight []*timeseries) ([]*timeseries, []*timeseries, error) {
+func groupJoin(singleTimeseriesSide string, be *metricsql.BinaryOpExpr, rvsLeft, rvsRight, tssLeft, tssRight []*timeseries) ([]*timeseries, []*timeseries, error) {
 	joinTags := be.JoinModifier.Args
 	var m map[string]*timeseries
 	for _, tsLeft := range tssLeft {
@@ -432,8 +269,8 @@ func mergeNonOverlappingTimeseries(dst, src *timeseries) bool {
 	return true
 }
 
-func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
-	if isBinaryOpCmp(be.Op) && !be.Bool {
+func resetMetricGroupIfRequired(be *metricsql.BinaryOpExpr, ts *timeseries) {
+	if metricsql.IsBinaryOpCmp(be.Op) && !be.Bool {
 		// Do not reset MetricGroup for non-boolean `compare` binary ops like Prometheus does.
 		return
 	}
@@ -445,97 +282,24 @@ func resetMetricGroupIfRequired(be *binaryOpExpr, ts *timeseries) {
 	ts.MetricName.ResetMetricGroup()
 }
 
-func binaryOpPlus(left, right float64) float64 {
-	return left + right
-}
-
-func binaryOpMinus(left, right float64) float64 {
-	return left - right
-}
-
-func binaryOpMul(left, right float64) float64 {
-	return left * right
-}
-
-func binaryOpDiv(left, right float64) float64 {
-	return left / right
-}
-
-func binaryOpMod(left, right float64) float64 {
-	return math.Mod(left, right)
-}
-
-func binaryOpPow(left, right float64) float64 {
-	return math.Pow(left, right)
-}
-
-func binaryOpDefault(left, right float64) float64 {
-	if math.IsNaN(left) {
-		return right
-	}
-	return left
-}
-
-func binaryOpIf(left, right float64) float64 {
-	if math.IsNaN(right) {
-		return nan
-	}
-	return left
-}
-
-func binaryOpIfnot(left, right float64) float64 {
-	if math.IsNaN(right) {
-		return left
-	}
-	return nan
-}
-
-func binaryOpEq(left, right float64) bool {
-	// Special handling for nan == nan.
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
-	if math.IsNaN(left) {
-		return math.IsNaN(right)
-	}
-
-	return left == right
-}
-
-func binaryOpNeq(left, right float64) bool {
-	// Special handling for comparison with nan.
-	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/150 .
-	if math.IsNaN(left) {
-		return !math.IsNaN(right)
-	}
-	if math.IsNaN(right) {
-		return true
-	}
-
-	return left != right
-}
-
-func binaryOpGt(left, right float64) bool {
-	return left > right
-}
-
-func binaryOpLt(left, right float64) bool {
-	return left < right
-}
-
-func binaryOpGte(left, right float64) bool {
-	return left >= right
-}
-
-func binaryOpLte(left, right float64) bool {
-	return left <= right
-}
-
 func binaryOpAnd(bfa *binaryOpFuncArg) ([]*timeseries, error) {
 	mLeft, mRight := createTimeseriesMapByTagSet(bfa.be, bfa.left, bfa.right)
 	var rvs []*timeseries
-	for k := range mRight {
-		if tss := mLeft[k]; tss != nil {
-			rvs = append(rvs, tss...)
+	for k, tssRight := range mRight {
+		tssLeft := mLeft[k]
+		if tssLeft == nil {
+			continue
 		}
+		for i := range tssLeft[0].Values {
+			if !isAllNaNs(tssRight, i) {
+				continue
+			}
+			for _, tsLeft := range tssLeft {
+				tsLeft.Values[i] = nan
+			}
+		}
+		tssLeft = removeNaNs(tssLeft)
+		rvs = append(rvs, tssLeft...)
 	}
 	return rvs, nil
 }
@@ -557,15 +321,36 @@ func binaryOpOr(bfa *binaryOpFuncArg) ([]*timeseries, error) {
 func binaryOpUnless(bfa *binaryOpFuncArg) ([]*timeseries, error) {
 	mLeft, mRight := createTimeseriesMapByTagSet(bfa.be, bfa.left, bfa.right)
 	var rvs []*timeseries
-	for k, tss := range mLeft {
-		if mRight[k] == nil {
-			rvs = append(rvs, tss...)
+	for k, tssLeft := range mLeft {
+		tssRight := mRight[k]
+		if tssRight == nil {
+			rvs = append(rvs, tssLeft...)
+			continue
 		}
+		for i := range tssLeft[0].Values {
+			if isAllNaNs(tssRight, i) {
+				continue
+			}
+			for _, tsLeft := range tssLeft {
+				tsLeft.Values[i] = nan
+			}
+		}
+		tssLeft = removeNaNs(tssLeft)
+		rvs = append(rvs, tssLeft...)
 	}
 	return rvs, nil
 }
 
-func createTimeseriesMapByTagSet(be *binaryOpExpr, left, right []*timeseries) (map[string][]*timeseries, map[string][]*timeseries) {
+func isAllNaNs(tss []*timeseries, idx int) bool {
+	for _, ts := range tss {
+		if !math.IsNaN(ts.Values[idx]) {
+			return false
+		}
+	}
+	return true
+}
+
+func createTimeseriesMapByTagSet(be *metricsql.BinaryOpExpr, left, right []*timeseries) (map[string][]*timeseries, map[string][]*timeseries) {
 	groupTags := be.GroupModifier.Args
 	groupOp := strings.ToLower(be.GroupModifier.Op)
 	if len(groupOp) == 0 {

app/vmselect/promql/eval.go

@@ -11,6 +11,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
 	"github.com/VictoriaMetrics/metrics"
 )
@@ -57,6 +58,14 @@ func AdjustStartEnd(start, end, step int64) (int64, int64) {
 	if adjust > 0 {
 		end += step - adjust
 	}
+
+	// Make sure that the new number of points is the same as the initial number of points.
+	newPoints := (end-start)/step + 1
+	for newPoints > points {
+		end -= step
+		newPoints--
+	}
+
 	return start, end
 }
 
@@ -70,6 +79,9 @@ type EvalConfig struct {
 
 	MayCache bool
 
+	// LookbackDelta is analog to `-query.lookback-delta` from Prometheus.
+	LookbackDelta int64
+
 	timestamps     []int64
 	timestampsOnce sync.Once
 }
@@ -82,6 +94,7 @@ func newEvalConfig(src *EvalConfig) *EvalConfig {
 	ec.Step = src.Step
 	ec.Deadline = src.Deadline
 	ec.MayCache = src.MayCache
+	ec.LookbackDelta = src.LookbackDelta
 
 	// do not copy src.timestamps - they must be generated again.
 	return &ec
@@ -140,25 +153,25 @@ func getTimestamps(start, end, step int64) []int64 {
 	return timestamps
 }
 
-func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
-	if me, ok := e.(*metricExpr); ok {
-		re := &rollupExpr{
+func evalExpr(ec *EvalConfig, e metricsql.Expr) ([]*timeseries, error) {
+	if me, ok := e.(*metricsql.MetricExpr); ok {
+		re := &metricsql.RollupExpr{
 			Expr: me,
 		}
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, me.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if re, ok := e.(*rollupExpr); ok {
-		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, re, nil)
+	if re, ok := e.(*metricsql.RollupExpr); ok {
+		rv, err := evalRollupFunc(ec, "default_rollup", rollupDefault, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, re.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if fe, ok := e.(*funcExpr); ok {
+	if fe, ok := e.(*metricsql.FuncExpr); ok {
 		nrf := getRollupFunc(fe.Name)
 		if nrf == nil {
 			args, err := evalExprs(ec, fe.Args)
@@ -188,17 +201,17 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		if err != nil {
 			return nil, err
 		}
-		rv, err := evalRollupFunc(ec, fe.Name, rf, re, nil)
+		rv, err := evalRollupFunc(ec, fe.Name, rf, e, re, nil)
 		if err != nil {
 			return nil, fmt.Errorf(`cannot evaluate %q: %s`, fe.AppendString(nil), err)
 		}
 		return rv, nil
 	}
-	if ae, ok := e.(*aggrFuncExpr); ok {
+	if ae, ok := e.(*metricsql.AggrFuncExpr); ok {
 		if callbacks := getIncrementalAggrFuncCallbacks(ae.Name); callbacks != nil {
 			fe, nrf := tryGetArgRollupFuncWithMetricExpr(ae)
 			if fe != nil {
-				// There is an optimized path for calculating aggrFuncExpr over rollupFunc over metricExpr.
+				// There is an optimized path for calculating metricsql.AggrFuncExpr over rollupFunc over metricsql.MetricExpr.
 				// The optimized path saves RAM for aggregates over big number of time series.
 				args, re, err := evalRollupFuncArgs(ec, fe)
 				if err != nil {
@@ -209,7 +222,7 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 					return nil, err
 				}
 				iafc := newIncrementalAggrFuncContext(ae, callbacks)
-				return evalRollupFunc(ec, fe.Name, rf, re, iafc)
+				return evalRollupFunc(ec, fe.Name, rf, e, re, iafc)
 			}
 		}
 		args, err := evalExprs(ec, ae.Args)
@@ -231,7 +244,7 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		}
 		return rv, nil
 	}
-	if be, ok := e.(*binaryOpExpr); ok {
+	if be, ok := e.(*metricsql.BinaryOpExpr); ok {
 		left, err := evalExpr(ec, be.Left)
 		if err != nil {
 			return nil, err
@@ -255,18 +268,18 @@ func evalExpr(ec *EvalConfig, e expr) ([]*timeseries, error) {
 		}
 		return rv, nil
 	}
-	if ne, ok := e.(*numberExpr); ok {
+	if ne, ok := e.(*metricsql.NumberExpr); ok {
 		rv := evalNumber(ec, ne.N)
 		return rv, nil
 	}
-	if se, ok := e.(*stringExpr); ok {
+	if se, ok := e.(*metricsql.StringExpr); ok {
 		rv := evalString(ec, se.S)
 		return rv, nil
 	}
 	return nil, fmt.Errorf("unexpected expression %q", e.AppendString(nil))
 }
 
-func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFunc) {
+func tryGetArgRollupFuncWithMetricExpr(ae *metricsql.AggrFuncExpr) (*metricsql.FuncExpr, newRollupFunc) {
 	if len(ae.Args) != 1 {
 		return nil, nil
 	}
@@ -277,31 +290,31 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 	// - rollupFunc(metricExpr)
 	// - rollupFunc(metricExpr[d])
 
-	if me, ok := e.(*metricExpr); ok {
+	if me, ok := e.(*metricsql.MetricExpr); ok {
 		// e = metricExpr
 		if me.IsEmpty() {
 			return nil, nil
 		}
-		fe := &funcExpr{
+		fe := &metricsql.FuncExpr{
 			Name: "default_rollup",
-			Args: []expr{me},
+			Args: []metricsql.Expr{me},
 		}
 		nrf := getRollupFunc(fe.Name)
 		return fe, nrf
 	}
-	if re, ok := e.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
+	if re, ok := e.(*metricsql.RollupExpr); ok {
+		if me, ok := re.Expr.(*metricsql.MetricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
 		// e = metricExpr[d]
-		fe := &funcExpr{
+		fe := &metricsql.FuncExpr{
 			Name: "default_rollup",
-			Args: []expr{re},
+			Args: []metricsql.Expr{re},
 		}
 		nrf := getRollupFunc(fe.Name)
 		return fe, nrf
 	}
-	fe, ok := e.(*funcExpr)
+	fe, ok := e.(*metricsql.FuncExpr)
 	if !ok {
 		return nil, nil
 	}
@@ -310,19 +323,23 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 		return nil, nil
 	}
 	rollupArgIdx := getRollupArgIdx(fe.Name)
+	if rollupArgIdx >= len(fe.Args) {
+		// Incorrect number of args for rollup func.
+		return nil, nil
+	}
 	arg := fe.Args[rollupArgIdx]
-	if me, ok := arg.(*metricExpr); ok {
+	if me, ok := arg.(*metricsql.MetricExpr); ok {
 		if me.IsEmpty() {
 			return nil, nil
 		}
 		// e = rollupFunc(metricExpr)
-		return &funcExpr{
+		return &metricsql.FuncExpr{
 			Name: fe.Name,
-			Args: []expr{me},
+			Args: []metricsql.Expr{me},
 		}, nrf
 	}
-	if re, ok := arg.(*rollupExpr); ok {
-		if me, ok := re.Expr.(*metricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
+	if re, ok := arg.(*metricsql.RollupExpr); ok {
+		if me, ok := re.Expr.(*metricsql.MetricExpr); !ok || me.IsEmpty() || re.ForSubquery() {
 			return nil, nil
 		}
 		// e = rollupFunc(metricExpr[d])
@@ -331,7 +348,7 @@ func tryGetArgRollupFuncWithMetricExpr(ae *aggrFuncExpr) (*funcExpr, newRollupFu
 	return nil, nil
 }
 
-func evalExprs(ec *EvalConfig, es []expr) ([][]*timeseries, error) {
+func evalExprs(ec *EvalConfig, es []metricsql.Expr) ([][]*timeseries, error) {
 	var rvs [][]*timeseries
 	for _, e := range es {
 		rv, err := evalExpr(ec, e)
@@ -343,9 +360,12 @@ func evalExprs(ec *EvalConfig, es []expr) ([][]*timeseries, error) {
 	return rvs, nil
 }
 
-func evalRollupFuncArgs(ec *EvalConfig, fe *funcExpr) ([]interface{}, *rollupExpr, error) {
-	var re *rollupExpr
+func evalRollupFuncArgs(ec *EvalConfig, fe *metricsql.FuncExpr) ([]interface{}, *metricsql.RollupExpr, error) {
+	var re *metricsql.RollupExpr
 	rollupArgIdx := getRollupArgIdx(fe.Name)
+	if len(fe.Args) <= rollupArgIdx {
+		return nil, nil, fmt.Errorf("expecting at least %d args to %q; got %d args; expr: %q", rollupArgIdx+1, fe.Name, len(fe.Args), fe.AppendString(nil))
+	}
 	args := make([]interface{}, len(fe.Args))
 	for i, arg := range fe.Args {
 		if i == rollupArgIdx {
@@ -362,11 +382,11 @@ func evalRollupFuncArgs(ec *EvalConfig, fe *funcExpr) ([]interface{}, *rollupExp
 	return args, re, nil
 }
 
-func getRollupExprArg(arg expr) *rollupExpr {
-	re, ok := arg.(*rollupExpr)
+func getRollupExprArg(arg metricsql.Expr) *metricsql.RollupExpr {
+	re, ok := arg.(*metricsql.RollupExpr)
 	if !ok {
-		// Wrap non-rollup arg into rollupExpr.
-		return &rollupExpr{
+		// Wrap non-rollup arg into metricsql.RollupExpr.
+		return &metricsql.RollupExpr{
 			Expr: arg,
 		}
 	}
@@ -374,45 +394,60 @@ func getRollupExprArg(arg expr) *rollupExpr {
 		// Return standard rollup if it doesn't contain subquery.
 		return re
 	}
-	me, ok := re.Expr.(*metricExpr)
+	me, ok := re.Expr.(*metricsql.MetricExpr)
 	if !ok {
 		// arg contains subquery.
 		return re
 	}
 	// Convert me[w:step] -> default_rollup(me)[w:step]
 	reNew := *re
-	reNew.Expr = &funcExpr{
+	reNew.Expr = &metricsql.FuncExpr{
 		Name: "default_rollup",
-		Args: []expr{
-			&rollupExpr{Expr: me},
+		Args: []metricsql.Expr{
+			&metricsql.RollupExpr{Expr: me},
 		},
 	}
 	return &reNew
 }
 
-func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr, iafc *incrementalAggrFuncContext) ([]*timeseries, error) {
+func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, expr metricsql.Expr, re *metricsql.RollupExpr, iafc *incrementalAggrFuncContext) ([]*timeseries, error) {
 	ecNew := ec
 	var offset int64
 	if len(re.Offset) > 0 {
 		var err error
-		offset, err = DurationValue(re.Offset, ec.Step)
+		offset, err = metricsql.DurationValue(re.Offset, ec.Step)
 		if err != nil {
 			return nil, err
 		}
-		ecNew = newEvalConfig(ec)
+		ecNew = newEvalConfig(ecNew)
 		ecNew.Start -= offset
 		ecNew.End -= offset
-		ecNew.Start, ecNew.End = AdjustStartEnd(ecNew.Start, ecNew.End, ecNew.Step)
+		if ecNew.MayCache {
+			start, end := AdjustStartEnd(ecNew.Start, ecNew.End, ecNew.Step)
+			offset += ecNew.Start - start
+			ecNew.Start = start
+			ecNew.End = end
+		}
+	}
+	if name == "rollup_candlestick" {
+		// Automatically apply `offset -step` to `rollup_candlestick` function
+		// in order to obtain expected OHLC results.
+		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/309#issuecomment-582113462
+		step := ecNew.Step
+		ecNew = newEvalConfig(ecNew)
+		ecNew.Start += step
+		ecNew.End += step
+		offset -= step
 	}
 	var rvs []*timeseries
 	var err error
-	if me, ok := re.Expr.(*metricExpr); ok {
-		rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, me, iafc, re.Window)
+	if me, ok := re.Expr.(*metricsql.MetricExpr); ok {
+		rvs, err = evalRollupFuncWithMetricExpr(ecNew, name, rf, expr, me, iafc, re.Window)
 	} else {
 		if iafc != nil {
 			logger.Panicf("BUG: iafc must be nil for rollup %q over subquery %q", name, re.AppendString(nil))
 		}
-		rvs, err = evalRollupFuncWithSubquery(ecNew, name, rf, re)
+		rvs, err = evalRollupFuncWithSubquery(ecNew, name, rf, expr, re)
 	}
 	if err != nil {
 		return nil, err
@@ -431,12 +466,12 @@ func evalRollupFunc(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr,
 	return rvs, nil
 }
 
-func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *rollupExpr) ([]*timeseries, error) {
-	// Do not use rollupResultCacheV here, since it works only with metricExpr.
+func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, expr metricsql.Expr, re *metricsql.RollupExpr) ([]*timeseries, error) {
+	// TODO: determine whether to use rollupResultCacheV here.
 	var step int64
 	if len(re.Step) > 0 {
 		var err error
-		step, err = DurationValue(re.Step, ec.Step)
+		step, err = metricsql.PositiveDurationValue(re.Step, ec.Step)
 		if err != nil {
 			return nil, err
 		}
@@ -446,7 +481,7 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	var window int64
 	if len(re.Window) > 0 {
 		var err error
-		window, err = DurationValue(re.Window, ec.Step)
+		window, err = metricsql.PositiveDurationValue(re.Window, ec.Step)
 		if err != nil {
 			return nil, err
 		}
@@ -463,9 +498,19 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 	if err != nil {
 		return nil, err
 	}
+	if len(tssSQ) == 0 {
+		if name == "absent_over_time" {
+			tss := evalNumber(ec, 1)
+			return tss, nil
+		}
+		return nil, nil
+	}
 
 	sharedTimestamps := getTimestamps(ec.Start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, ec.Start, ec.End, ec.Step, window, sharedTimestamps)
+	preFunc, rcs, err := getRollupConfigs(name, rf, expr, ec.Start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
+	if err != nil {
+		return nil, err
+	}
 	tss := make([]*timeseries, 0, len(tssSQ)*len(rcs))
 	var tssLock sync.Mutex
 	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
@@ -473,6 +518,13 @@ func evalRollupFuncWithSubquery(ec *EvalConfig, name string, rf rollupFunc, re *
 		values, timestamps = removeNanValues(values[:0], timestamps[:0], tsSQ.Values, tsSQ.Timestamps)
 		preFunc(values, timestamps)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &tsSQ.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, values, timestamps)
+				tssLock.Lock()
+				tss = tsm.AppendTimeseriesTo(tss)
+				tssLock.Unlock()
+				continue
+			}
 			var ts timeseries
 			doRollupForTimeseries(rc, &ts, &tsSQ.MetricName, values, timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
@@ -540,21 +592,22 @@ var (
 	rollupResultCacheMiss        = metrics.NewCounter(`vm_rollup_result_cache_miss_total`)
 )
 
-func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me *metricExpr, iafc *incrementalAggrFuncContext, windowStr string) ([]*timeseries, error) {
+func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc,
+	expr metricsql.Expr, me *metricsql.MetricExpr, iafc *incrementalAggrFuncContext, windowStr string) ([]*timeseries, error) {
 	if me.IsEmpty() {
 		return evalNumber(ec, nan), nil
 	}
 	var window int64
 	if len(windowStr) > 0 {
 		var err error
-		window, err = DurationValue(windowStr, ec.Step)
+		window, err = metricsql.PositiveDurationValue(windowStr, ec.Step)
 		if err != nil {
 			return nil, err
 		}
 	}
 
 	// Search for partial results in cache.
-	tssCached, start := rollupResultCacheV.Get(name, ec, me, iafc, window)
+	tssCached, start := rollupResultCacheV.Get(ec, expr, window)
 	if start > ec.End {
 		// The result is fully cached.
 		rollupResultCacheFullHits.Inc()
@@ -566,11 +619,26 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 		rollupResultCacheMiss.Inc()
 	}
 
+	// Obtain rollup configs before fetching data from db,
+	// so type errors can be caught earlier.
+	sharedTimestamps := getTimestamps(start, ec.End, ec.Step)
+	preFunc, rcs, err := getRollupConfigs(name, rf, expr, start, ec.End, ec.Step, window, ec.LookbackDelta, sharedTimestamps)
+	if err != nil {
+		return nil, err
+	}
+
 	// Fetch the remaining part of the result.
+	tfs := toTagFilters(me.LabelFilters)
+	minTimestamp := start - maxSilenceInterval
+	if window > ec.Step {
+		minTimestamp -= window
+	} else {
+		minTimestamp -= ec.Step
+	}
 	sq := &storage.SearchQuery{
-		MinTimestamp: start - window - maxSilenceInterval,
-		MaxTimestamp: ec.End + ec.Step,
-		TagFilterss:  [][]storage.TagFilter{me.TagFilters},
+		MinTimestamp: minTimestamp,
+		MaxTimestamp: ec.End,
+		TagFilterss:  [][]storage.TagFilter{tfs},
 	}
 	rss, err := netstorage.ProcessSearchQuery(sq, true, ec.Deadline)
 	if err != nil {
@@ -579,19 +647,32 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	rssLen := rss.Len()
 	if rssLen == 0 {
 		rss.Cancel()
+		var tss []*timeseries
+		if name == "absent_over_time" {
+			tss = getAbsentTimeseries(ec, me)
+		}
 		// Add missing points until ec.End.
 		// Do not cache the result, since missing points
 		// may be backfilled in the future.
-		tss := mergeTimeseries(tssCached, nil, start, ec)
+		tss = mergeTimeseries(tssCached, tss, start, ec)
 		return tss, nil
 	}
-	sharedTimestamps := getTimestamps(start, ec.End, ec.Step)
-	preFunc, rcs := getRollupConfigs(name, rf, start, ec.End, ec.Step, window, sharedTimestamps)
 
 	// Verify timeseries fit available memory after the rollup.
 	// Take into account points from tssCached.
 	pointsPerTimeseries := 1 + (ec.End-ec.Start)/ec.Step
-	rollupPoints := mulNoOverflow(pointsPerTimeseries, int64(rssLen*len(rcs)))
+	timeseriesLen := rssLen
+	if iafc != nil {
+		// Incremental aggregates require hold only GOMAXPROCS timeseries in memory.
+		timeseriesLen = runtime.GOMAXPROCS(-1)
+		if iafc.ae.Modifier.Op != "" {
+			// Increase the number of timeseries for non-empty group list: `aggr() by (something)`,
+			// since each group can have own set of time series in memory.
+			// Estimate the number of such groups is lower than 1000 :)
+			timeseriesLen *= 1000
+		}
+	}
+	rollupPoints := mulNoOverflow(pointsPerTimeseries, int64(timeseriesLen*len(rcs)))
 	rollupMemorySize := mulNoOverflow(rollupPoints, 16)
 	rml := getRollupMemoryLimiter()
 	if !rml.Get(uint64(rollupMemorySize)) {
@@ -607,16 +688,15 @@ func evalRollupFuncWithMetricExpr(ec *EvalConfig, name string, rf rollupFunc, me
 	removeMetricGroup := !rollupFuncsKeepMetricGroup[name]
 	var tss []*timeseries
 	if iafc != nil {
-		tss, err = evalRollupWithIncrementalAggregate(iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+		tss, err = evalRollupWithIncrementalAggregate(name, iafc, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	} else {
-		tss, err = evalRollupNoIncrementalAggregate(rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
+		tss, err = evalRollupNoIncrementalAggregate(name, rss, rcs, preFunc, sharedTimestamps, removeMetricGroup)
 	}
 	if err != nil {
 		return nil, err
 	}
 	tss = mergeTimeseries(tssCached, tss, start, ec)
-	rollupResultCacheV.Put(name, ec, me, iafc, window, tss)
-
+	rollupResultCacheV.Put(ec, expr, window, tss)
 	return tss, nil
 }
 
@@ -632,13 +712,20 @@ func getRollupMemoryLimiter() *memoryLimiter {
 	return &rollupMemoryLimiter
 }
 
-func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
+func evalRollupWithIncrementalAggregate(name string, iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
 	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		ts := getTimeseries()
 		defer putTimeseries(ts)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &rs.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, rs.Values, rs.Timestamps)
+				for _, ts := range tsm.m {
+					iafc.updateTimeseries(ts, workerID)
+				}
+				continue
+			}
 			ts.Reset()
 			doRollupForTimeseries(rc, ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
 			iafc.updateTimeseries(ts, workerID)
@@ -655,13 +742,20 @@ func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *n
 	return tss, nil
 }
 
-func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConfig,
+func evalRollupNoIncrementalAggregate(name string, rss *netstorage.Results, rcs []*rollupConfig,
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64, removeMetricGroup bool) ([]*timeseries, error) {
 	tss := make([]*timeseries, 0, rss.Len()*len(rcs))
 	var tssLock sync.Mutex
 	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		for _, rc := range rcs {
+			if tsm := newTimeseriesMap(name, sharedTimestamps, &rs.MetricName); tsm != nil {
+				rc.DoTimeseriesMap(tsm, rs.Values, rs.Timestamps)
+				tssLock.Lock()
+				tss = tsm.AppendTimeseriesTo(tss)
+				tssLock.Unlock()
+				continue
+			}
 			var ts timeseries
 			doRollupForTimeseries(rc, &ts, &rs.MetricName, rs.Values, rs.Timestamps, sharedTimestamps, removeMetricGroup)
 			tssLock.Lock()
@@ -689,60 +783,6 @@ func doRollupForTimeseries(rc *rollupConfig, tsDst *timeseries, mnSrc *storage.M
 	tsDst.denyReuse = true
 }
 
-func getRollupConfigs(name string, rf rollupFunc, start, end, step, window int64, sharedTimestamps []int64) (func(values []float64, timestamps []int64), []*rollupConfig) {
-	preFunc := func(values []float64, timestamps []int64) {}
-	if rollupFuncsRemoveCounterResets[name] {
-		preFunc = func(values []float64, timestamps []int64) {
-			removeCounterResets(values)
-		}
-	}
-	newRollupConfig := func(rf rollupFunc, tagValue string) *rollupConfig {
-		return &rollupConfig{
-			TagValue:        tagValue,
-			Func:            rf,
-			Start:           start,
-			End:             end,
-			Step:            step,
-			Window:          window,
-			MayAdjustWindow: rollupFuncsMayAdjustWindow[name],
-			Timestamps:      sharedTimestamps,
-		}
-	}
-	appendRollupConfigs := func(dst []*rollupConfig) []*rollupConfig {
-		dst = append(dst, newRollupConfig(rollupMin, "min"))
-		dst = append(dst, newRollupConfig(rollupMax, "max"))
-		dst = append(dst, newRollupConfig(rollupAvg, "avg"))
-		return dst
-	}
-	var rcs []*rollupConfig
-	switch name {
-	case "rollup":
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_rate", "rollup_deriv":
-		preFuncPrev := preFunc
-		preFunc = func(values []float64, timestamps []int64) {
-			preFuncPrev(values, timestamps)
-			derivValues(values, timestamps)
-		}
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_increase", "rollup_delta":
-		preFuncPrev := preFunc
-		preFunc = func(values []float64, timestamps []int64) {
-			preFuncPrev(values, timestamps)
-			deltaValues(values)
-		}
-		rcs = appendRollupConfigs(rcs)
-	case "rollup_candlestick":
-		rcs = append(rcs, newRollupConfig(rollupFirst, "open"))
-		rcs = append(rcs, newRollupConfig(rollupLast, "close"))
-		rcs = append(rcs, newRollupConfig(rollupMin, "low"))
-		rcs = append(rcs, newRollupConfig(rollupMax, "high"))
-	default:
-		rcs = append(rcs, newRollupConfig(rf, ""))
-	}
-	return preFunc, rcs
-}
-
 var bbPool bytesutil.ByteBufferPool
 
 func evalNumber(ec *EvalConfig, n float64) []*timeseries {
@@ -781,3 +821,23 @@ func mulNoOverflow(a, b int64) int64 {
 	}
 	return a * b
 }
+
+func toTagFilters(lfs []metricsql.LabelFilter) []storage.TagFilter {
+	tfs := make([]storage.TagFilter, len(lfs))
+	for i := range lfs {
+		toTagFilter(&tfs[i], &lfs[i])
+	}
+	return tfs
+}
+
+func toTagFilter(dst *storage.TagFilter, src *metricsql.LabelFilter) {
+	if src.Label != "__name__" {
+		dst.Key = []byte(src.Label)
+	} else {
+		// This is required for storage.Search.
+		dst.Key = nil
+	}
+	dst.Value = []byte(src.Value)
+	dst.IsRegexp = src.IsRegexp
+	dst.IsNegative = src.IsNegative
+}

app/vmselect/promql/exec.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/metrics"
 )
 
@@ -18,17 +19,6 @@ var logSlowQueryDuration = flag.Duration("search.logSlowQueryDuration", 5*time.S
 
 var slowQueries = metrics.NewCounter(`vm_slow_queries_total`)
 
-// ExpandWithExprs expands WITH expressions inside q and returns the resulting
-// PromQL without WITH expressions.
-func ExpandWithExprs(q string) (string, error) {
-	e, err := parsePromQLWithCache(q)
-	if err != nil {
-		return "", err
-	}
-	buf := e.AppendString(nil)
-	return string(buf), nil
-}
-
 // Exec executes q for the given ec.
 func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result, error) {
 	if *logSlowQueryDuration > 0 {
@@ -36,8 +26,8 @@ func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result,
 		defer func() {
 			d := time.Since(startTime)
 			if d >= *logSlowQueryDuration {
-				logger.Infof("slow query according to -search.logSlowQueryDuration=%s: duration=%s, start=%d, end=%d, step=%d, query=%q",
-					*logSlowQueryDuration, d, ec.Start/1000, ec.End/1000, ec.Step/1000, q)
+				logger.Infof("slow query according to -search.logSlowQueryDuration=%s: duration=%.3f seconds, start=%d, end=%d, step=%d, query=%q",
+					*logSlowQueryDuration, d.Seconds(), ec.Start/1000, ec.End/1000, ec.Step/1000, q)
 				slowQueries.Inc()
 			}
 		}()
@@ -50,25 +40,11 @@ func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result,
 		return nil, err
 	}
 
-	// Add an additional point to the end. This point is used
-	// in calculating the last value for rate, deriv, increase
-	// and delta funcs.
-	ec.End += ec.Step
-
 	rv, err := evalExpr(ec, e)
 	if err != nil {
 		return nil, err
 	}
 
-	// Remove the additional point at the end.
-	for _, ts := range rv {
-		ts.Values = ts.Values[:len(ts.Values)-1]
-
-		// ts.Timestamps may be shared between timeseries, so truncate it with len(ts.Values) instead of len(ts.Timestamps)-1
-		ts.Timestamps = ts.Timestamps[:len(ts.Values)]
-	}
-	ec.End -= ec.Step
-
 	if isFirstPointOnly {
 		// Remove all the points except the first one from every time series.
 		for _, ts := range rv {
@@ -85,17 +61,18 @@ func Exec(ec *EvalConfig, q string, isFirstPointOnly bool) ([]netstorage.Result,
 	return result, err
 }
 
-func maySortResults(e expr, tss []*timeseries) bool {
+func maySortResults(e metricsql.Expr, tss []*timeseries) bool {
 	if len(tss) > 100 {
 		// There is no sense in sorting a lot of results
 		return false
 	}
-	fe, ok := e.(*funcExpr)
+	fe, ok := e.(*metricsql.FuncExpr)
 	if !ok {
 		return true
 	}
 	switch fe.Name {
-	case "sort", "sort_desc":
+	case "sort", "sort_desc",
+		"sort_by_label", "sort_by_label_desc":
 		return false
 	default:
 		return true
@@ -110,7 +87,7 @@ func timeseriesToResult(tss []*timeseries, maySort bool) ([]netstorage.Result, e
 	for i, ts := range tss {
 		bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
 		if _, ok := m[string(bb.B)]; ok {
-			return nil, fmt.Errorf(`duplicate output timeseries: %s%s`, ts.MetricName.MetricGroup, stringMetricName(&ts.MetricName))
+			return nil, fmt.Errorf(`duplicate output timeseries: %s`, stringMetricName(&ts.MetricName))
 		}
 		m[string(bb.B)] = struct{}{}
 
@@ -154,10 +131,10 @@ func removeNaNs(tss []*timeseries) []*timeseries {
 	return rvs
 }
 
-func parsePromQLWithCache(q string) (expr, error) {
+func parsePromQLWithCache(q string) (metricsql.Expr, error) {
 	pcv := parseCacheV.Get(q)
 	if pcv == nil {
-		e, err := parsePromQL(q)
+		e, err := metricsql.Parse(q)
 		pcv = &parseCacheValue{
 			e:   e,
 			err: err,
@@ -189,16 +166,19 @@ var parseCacheV = func() *parseCache {
 const parseCacheMaxLen = 10e3
 
 type parseCacheValue struct {
-	e   expr
+	e   metricsql.Expr
 	err error
 }
 
 type parseCache struct {
-	m  map[string]*parseCacheValue
-	mu sync.RWMutex
+	// Move atomic counters to the top of struct for 8-byte alignment on 32-bit arch.
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
 
 	requests uint64
 	misses   uint64
+
+	m  map[string]*parseCacheValue
+	mu sync.RWMutex
 }
 
 func (pc *parseCache) Requests() uint64 {

app/vmselect/promql/parser_test.go

@@ -47,670 +47,3 @@ func TestParseMetricSelectorError(t *testing.T) {
 	f(`foo[5m]`)
 	f(`foo offset 5m`)
 }
-
-func TestParsePromQLSuccess(t *testing.T) {
-	another := func(s string, sExpected string) {
-		t.Helper()
-
-		e, err := parsePromQL(s)
-		if err != nil {
-			t.Fatalf("unexpected error when parsing %q: %s", s, err)
-		}
-		res := e.AppendString(nil)
-		if string(res) != sExpected {
-			t.Fatalf("unexpected string constructed;\ngot\n%q\nwant\n%q", res, sExpected)
-		}
-	}
-	same := func(s string) {
-		t.Helper()
-		another(s, s)
-	}
-
-	// metricExpr
-	same(`{}`)
-	same(`{}[5m]`)
-	same(`{}[5m:]`)
-	same(`{}[:]`)
-	another(`{}[: ]`, `{}[:]`)
-	same(`{}[:3s]`)
-	another(`{}[: 3s ]`, `{}[:3s]`)
-	same(`{}[5m:3s]`)
-	another(`{}[ 5m : 3s ]`, `{}[5m:3s]`)
-	same(`{} offset 5m`)
-	same(`{}[5m] offset 10y`)
-	same(`{}[5.3m:3.4s] offset 10y`)
-	same(`{}[:3.4s] offset 10y`)
-	same(`{Foo="bAR"}`)
-	same(`{foo="bar"}`)
-	same(`{foo="bar"}[5m]`)
-	same(`{foo="bar"}[5m:]`)
-	same(`{foo="bar"}[5m:3s]`)
-	same(`{foo="bar"} offset 10y`)
-	same(`{foo="bar"}[5m] offset 10y`)
-	same(`{foo="bar"}[5m:3s] offset 10y`)
-	another(`{foo="bar"}[5m] oFFSEt 10y`, `{foo="bar"}[5m] offset 10y`)
-	same("METRIC")
-	same("metric")
-	same("m_e:tri44:_c123")
-	another("-metric", "0 - metric")
-	same(`metric offset 10h`)
-	same("metric[5m]")
-	same("metric[5m:3s]")
-	same("metric[5m] offset 10h")
-	same("metric[5m:3s] offset 10h")
-	same("metric[5i:3i] offset 10i")
-	same(`metric{foo="bar"}`)
-	same(`metric{foo="bar"} offset 10h`)
-	same(`metric{foo!="bar"}[2d]`)
-	same(`metric{foo="bar"}[2d] offset 10h`)
-	same(`metric{foo="bar", b="sdfsdf"}[2d:3h] offset 10h`)
-	another(`  metric  {  foo  = "bar"  }  [  2d ]   offset   10h  `, `metric{foo="bar"}[2d] offset 10h`)
-	// metric name matching keywords
-	same("rate")
-	same("RATE")
-	same("by")
-	same("BY")
-	same("bool")
-	same("BOOL")
-	same("unless")
-	same("UNLESS")
-	same("Ignoring")
-	same("with")
-	same("WITH")
-	same("With")
-	same("alias")
-	same(`alias{foo="bar"}`)
-	same(`aLIas{alias="aa"}`)
-	another(`al\ias`, `alias`)
-	// identifiers with with escape chars
-	same(`foo\ bar`)
-	same(`foo\-bar\{{baz\+bar="aa"}`)
-	another(`\x2E\x2ef\oo{b\xEF\ar="aa"}`, `\x2e.foo{b\xefar="aa"}`)
-	// Duplicate filters
-	same(`foo{__name__="bar"}`)
-	same(`foo{a="b", a="c", __name__="aaa", b="d"}`)
-	// Metric filters ending with comma
-	another(`m{foo="bar",}`, `m{foo="bar"}`)
-	// String concat in tag value
-	another(`m{foo="bar" + "baz"}`, `m{foo="barbaz"}`)
-
-	// Valid regexp
-	same(`foo{bar=~"x"}`)
-	same(`foo{bar=~"^x"}`)
-	same(`foo{bar=~"^x$"}`)
-	same(`foo{bar=~"^(a[bc]|d)$"}`)
-	same(`foo{bar!~"x"}`)
-	same(`foo{bar!~"^x"}`)
-	same(`foo{bar!~"^x$"}`)
-	same(`foo{bar!~"^(a[bc]|d)$"}`)
-
-	// stringExpr
-	same(`""`)
-	same(`"\n\t\r 12:{}[]()44"`)
-	another(`''`, `""`)
-	another("``", `""`)
-	another("   `foo\"b'ar`  ", "\"foo\\\"b'ar\"")
-	another(`  'foo\'bar"BAZ'  `, `"foo'bar\"BAZ"`)
-	// string concat
-	another(`"foo"+'bar'`, `"foobar"`)
-
-	// numberExpr
-	same(`1`)
-	same(`1.23`)
-	same(`0.23`)
-	same(`1.2e+45`)
-	same(`1.2e-45`)
-	same(`-1`)
-	same(`-1.23`)
-	same(`-0.23`)
-	same(`-1.2e+45`)
-	same(`-1.2e-45`)
-	same(`-1.2e-45`)
-	another(`12.5E34`, `1.25e+35`)
-	another(`-.2`, `-0.2`)
-	another(`-.2E-2`, `-0.002`)
-	same(`NaN`)
-	another(`nan`, `NaN`)
-	another(`NAN`, `NaN`)
-	another(`nAN`, `NaN`)
-	another(`Inf`, `+Inf`)
-	another(`INF`, `+Inf`)
-	another(`inf`, `+Inf`)
-	another(`+Inf`, `+Inf`)
-	another(`-Inf`, `-Inf`)
-	another(`-inF`, `-Inf`)
-
-	// binaryOpExpr
-	another(`nan == nan`, `NaN`)
-	another(`nan ==bool nan`, `1`)
-	another(`nan !=bool nan`, `0`)
-	another(`nan !=bool 2`, `1`)
-	another(`2 !=bool nan`, `1`)
-	another(`nan >bool nan`, `0`)
-	another(`nan <bool nan`, `0`)
-	another(`1 ==bool nan`, `0`)
-	another(`NaN !=bool 1`, `1`)
-	another(`inf >=bool 2`, `1`)
-	another(`-1 >bool -inf`, `1`)
-	another(`-1 <bool -inf`, `0`)
-	another(`nan + 2 *3 * inf`, `NaN`)
-	another(`INF - Inf`, `NaN`)
-	another(`Inf + inf`, `+Inf`)
-	another(`1/0`, `+Inf`)
-	another(`0/0`, `NaN`)
-	another(`-m`, `0 - m`)
-	same(`m + ignoring () n[5m]`)
-	another(`M + IGNORING () N[5m]`, `M + ignoring () N[5m]`)
-	same(`m + on (foo) n[5m]`)
-	another(`m + ON (Foo) n[5m]`, `m + on (Foo) n[5m]`)
-	same(`m + ignoring (a, b) n[5m]`)
-	another(`1 or 2`, `1`)
-	another(`1 and 2`, `1`)
-	another(`1 unless 2`, `NaN`)
-	another(`1 default 2`, `1`)
-	another(`1 default NaN`, `1`)
-	another(`NaN default 2`, `2`)
-	another(`1 > 2`, `NaN`)
-	another(`1 > bool 2`, `0`)
-	another(`3 >= 2`, `3`)
-	another(`3 <= bool 2`, `0`)
-	another(`1 + -2 - 3`, `-4`)
-	another(`1 / 0 + 2`, `+Inf`)
-	another(`2 + -1 / 0`, `-Inf`)
-	another(`-1 ^ 0.5`, `NaN`)
-	another(`512.5 - (1 + 3) * (2 ^ 2) ^ 3`, `256.5`)
-	another(`1 == bool 1 != bool 24 < bool 4 > bool -1`, `1`)
-	another(`1 == bOOl 1 != BOOL 24 < Bool 4 > booL -1`, `1`)
-	another(`m1+on(foo)group_left m2`, `m1 + on (foo) group_left () m2`)
-	another(`M1+ON(FOO)GROUP_left M2`, `M1 + on (FOO) group_left () M2`)
-	same(`m1 + on (foo) group_right () m2`)
-	same(`m1 + on (foo, bar) group_right (x, y) m2`)
-	another(`m1 + on (foo, bar,) group_right (x, y,) m2`, `m1 + on (foo, bar) group_right (x, y) m2`)
-	same(`m1 == bool on (foo, bar) group_right (x, y) m2`)
-	another(`5 - 1 + 3 * 2 ^ 2 ^ 3 - 2  OR Metric {Bar= "Baz", aaa!="bb",cc=~"dd" ,zz !~"ff" } `,
-		`770 or Metric{Bar="Baz", aaa!="bb", cc=~"dd", zz!~"ff"}`)
-	same(`"foo" + bar()`)
-	same(`"foo" + bar{x="y"}`)
-	same(`("foo"[3s] + bar{x="y"})[5m:3s] offset 10s`)
-	same(`("foo"[3s] + bar{x="y"})[5i:3i] offset 10i`)
-	same(`bar + "foo" offset 3s`)
-	same(`bar + "foo" offset 3i`)
-	another(`1+2 if 2>3`, `NaN`)
-	another(`1+4 if 2<3`, `5`)
-	another(`2+6 default 3 if 2>3`, `8`)
-	another(`2+6 if 2>3 default NaN`, `NaN`)
-	another(`42 if 3>2 if 2+2<5`, `42`)
-	another(`42 if 3>2 if 2+2>=5`, `NaN`)
-	another(`1+2 ifnot 2>3`, `3`)
-	another(`1+4 ifnot 2<3`, `NaN`)
-	another(`2+6 default 3 ifnot 2>3`, `8`)
-	another(`2+6 ifnot 2>3 default NaN`, `8`)
-	another(`42 if 3>2 ifnot 2+2<5`, `NaN`)
-	another(`42 if 3>2 ifnot 2+2>=5`, `42`)
-
-	// parensExpr
-	another(`(-foo + ((bar) / (baz))) + ((23))`, `((0 - foo) + (bar / baz)) + 23`)
-	another(`(FOO + ((Bar) / (baZ))) + ((23))`, `(FOO + (Bar / baZ)) + 23`)
-	same(`(foo, bar)`)
-	another(`1+(foo, bar,)`, `1 + (foo, bar)`)
-	another(`((foo(bar,baz)), (1+(2)+(3,4)+()))`, `(foo(bar, baz), (3 + (3, 4)) + ())`)
-	same(`()`)
-
-	// funcExpr
-	same(`f()`)
-	another(`f(x,)`, `f(x)`)
-	another(`-f()-Ff()`, `(0 - f()) - Ff()`)
-	same(`F()`)
-	another(`+F()`, `F()`)
-	another(`++F()`, `F()`)
-	another(`--F()`, `0 - (0 - F())`)
-	same(`f(http_server_request)`)
-	same(`f(http_server_request)[4s:5m] offset 10m`)
-	same(`f(http_server_request)[4i:5i] offset 10i`)
-	same(`F(HttpServerRequest)`)
-	same(`f(job, foo)`)
-	same(`F(Job, Foo)`)
-	another(` FOO (bar) + f  (  m  (  ),ff(1 + (  2.5)) ,M[5m ]  , "ff"  )`, `FOO(bar) + f(m(), ff(3.5), M[5m], "ff")`)
-	// funcName matching keywords
-	same(`by(2)`)
-	same(`BY(2)`)
-	same(`or(2)`)
-	same(`OR(2)`)
-	same(`bool(2)`)
-	same(`BOOL(2)`)
-	same(`rate(rate(m))`)
-	same(`rate(rate(m[5m]))`)
-	same(`rate(rate(m[5m])[1h:])`)
-	same(`rate(rate(m[5m])[1h:3s])`)
-	// funcName with escape chars
-	same(`foo\(ba\-r()`)
-
-	// aggrFuncExpr
-	same(`sum(http_server_request) by ()`)
-	same(`sum(http_server_request) by (job)`)
-	same(`sum(http_server_request) without (job, foo)`)
-	another(`sum(x,y,) without (a,b,)`, `sum(x, y) without (a, b)`)
-	another(`sum by () (xx)`, `sum(xx) by ()`)
-	another(`sum by (s) (xx)[5s]`, `(sum(xx) by (s))[5s]`)
-	another(`SUM BY (ZZ, aa) (XX)`, `sum(XX) by (ZZ, aa)`)
-	another(`sum without (a, b) (xx,2+2)`, `sum(xx, 4) without (a, b)`)
-	another(`Sum WIthout (a, B) (XX,2+2)`, `sum(XX, 4) without (a, B)`)
-	same(`sum(a) or sum(b)`)
-	same(`sum(a) by () or sum(b) without (x, y)`)
-	same(`sum(a) + sum(b)`)
-	same(`sum(x) * (1 + sum(a))`)
-
-	// All the above
-	another(`Sum(Ff(M) * M{X=""}[5m] Offset 7m - 123, 35) BY (X, y) * F2("Test")`,
-		`sum((Ff(M) * M{X=""}[5m] offset 7m) - 123, 35) by (X, y) * F2("Test")`)
-	another(`# comment
-		Sum(Ff(M) * M{X=""}[5m] Offset 7m - 123, 35) BY (X, y) # yet another comment
-		* F2("Test")`,
-		`sum((Ff(M) * M{X=""}[5m] offset 7m) - 123, 35) by (X, y) * F2("Test")`)
-
-	// withExpr
-	another(`with () x`, `x`)
-	another(`with (x=1,) x`, `1`)
-	another(`with (x = m offset 5h) x + x`, `m offset 5h + m offset 5h`)
-	another(`with (x = m offset 5i) x + x`, `m offset 5i + m offset 5i`)
-	another(`with (foo = bar{x="x"}) 1`, `1`)
-	another(`with (foo = bar{x="x"}) "x"`, `"x"`)
-	another(`with (f="x") f`, `"x"`)
-	another(`with (foo = bar{x="x"}) x{x="y"}`, `x{x="y"}`)
-	another(`with (foo = bar{x="x"}) 1+1`, `2`)
-	another(`with (foo = bar{x="x"}) f()`, `f()`)
-	another(`with (foo = bar{x="x"}) sum(x)`, `sum(x)`)
-	another(`with (foo = bar{x="x"}) baz{foo="bar"}`, `baz{foo="bar"}`)
-	another(`with (foo = bar) baz`, `baz`)
-	another(`with (foo = bar) foo + foo{a="b"}`, `bar + bar{a="b"}`)
-	another(`with (foo = bar, bar=baz + f()) test`, `test`)
-	another(`with (ct={job="test"}) a{ct} + ct() + f({ct="x"})`, `(a{job="test"} + {job="test"}) + f({ct="x"})`)
-	another(`with (ct={job="test", i="bar"}) ct + {ct, x="d"} + foo{ct, ct} + ctx(1)`,
-		`(({job="test", i="bar"} + {job="test", i="bar", x="d"}) + foo{job="test", i="bar"}) + ctx(1)`)
-	another(`with (foo = bar) {__name__=~"foo"}`, `{__name__=~"foo"}`)
-	another(`with (foo = bar) foo{__name__="foo"}`, `bar`)
-	another(`with (foo = bar) {__name__="foo", x="y"}`, `bar{x="y"}`)
-	another(`with (foo(bar) = {__name__!="bar"}) foo(x)`, `{__name__!="bar"}`)
-	another(`with (foo(bar) = bar{__name__="bar"}) foo(x)`, `x`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar((x,y))`, `(x, y) + (x, y)`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x*y)`, `(x * y) + (x * y)`)
-	another(`with (foo\-bar(baz) = baz + baz) foo\-bar(x\*y)`, `x\*y + x\*y`)
-	another(`with (foo\-bar(b\ az) = b\ az + b\ az) foo\-bar(x\*y)`, `x\*y + x\*y`)
-	// override ttf to something new.
-	another(`with (ttf = a) ttf + b`, `a + b`)
-	// override ttf to ru
-	another(`with (ttf = ru(m, n)) ttf`, `(clamp_min(n - clamp_min(m, 0), 0) / clamp_min(n, 0)) * 100`)
-
-	// Verify withExpr recursion and forward reference
-	another(`with (x = x+y, y = x+x) y ^ 2`, `((x + y) + (x + y)) ^ 2`)
-	another(`with (f1(x)=f2(x), f2(x)=f1(x)^2) f1(foobar)`, `f2(foobar)`)
-	another(`with (f1(x)=f2(x), f2(x)=f1(x)^2) f2(foobar)`, `f2(foobar) ^ 2`)
-
-	// Verify withExpr funcs
-	another(`with (x() = y+1) x`, `y + 1`)
-	another(`with (x(foo) = foo+1) x(a)`, `a + 1`)
-	another(`with (x(a, b) = a + b) x(foo, bar)`, `foo + bar`)
-	another(`with (x(a, b) = a + b) x(foo, x(1, 2))`, `foo + 3`)
-	another(`with (x(a) = sum(a) by (b)) x(xx) / x(y)`, `sum(xx) by (b) / sum(y) by (b)`)
-	another(`with (f(a,f,x)=ff(x,f,a)) f(f(x,y,z),1,2)`, `ff(2, 1, ff(z, y, x))`)
-	another(`with (f(x)=1+f(x)) f(foo{bar="baz"})`, `1 + f(foo{bar="baz"})`)
-	another(`with (a=foo, y=bar, f(a)= a+a+y) f(x)`, `(x + x) + bar`)
-	another(`with (f(a, b) = m{a, b}) f({a="x", b="y"}, {c="d"})`, `m{a="x", b="y", c="d"}`)
-	another(`with (xx={a="x"}, f(a, b) = m{a, b}) f({xx, b="y"}, {c="d"})`, `m{a="x", b="y", c="d"}`)
-	another(`with (x() = {b="c"}) foo{x}`, `foo{b="c"}`)
-	another(`with (f(x)=x{foo="bar"} offset 5m) f(m offset 10m)`, `(m{foo="bar"} offset 10m) offset 5m`)
-	another(`with (f(x)=x{foo="bar",bas="a"}[5m]) f(m[10m] offset 3s)`, `(m{foo="bar", bas="a"}[10m] offset 3s)[5m]`)
-	another(`with (f(x)=x{foo="bar"}[5m] offset 10m) f(m{x="y"})`, `m{x="y", foo="bar"}[5m] offset 10m`)
-	another(`with (f(x)=x{foo="bar"}[5m] offset 10m) f({x="y", foo="bar", foo="bar"})`, `{x="y", foo="bar"}[5m] offset 10m`)
-	another(`with (f(m, x)=m{x}[5m] offset 10m) f(foo, {})`, `foo[5m] offset 10m`)
-	another(`with (f(m, x)=m{x, bar="baz"}[5m] offset 10m) f(foo, {})`, `foo{bar="baz"}[5m] offset 10m`)
-	another(`with (f(x)=x[5m] offset 3s) f(foo[3m]+bar)`, `(foo[3m] + bar)[5m] offset 3s`)
-	another(`with (f(x)=x[5m:3s] oFFsEt 1.5m) f(sum(s) by (a,b))`, `(sum(s) by (a, b))[5m:3s] offset 1.5m`)
-	another(`with (x="a", y=x) y+"bc"`, `"abc"`)
-	another(`with (x="a", y="b"+x) "we"+y+"z"+f()`, `"webaz" + f()`)
-	another(`with (f(x) = m{foo=x+"y", bar="y"+x, baz=x} + x) f("qwe")`, `m{foo="qwey", bar="yqwe", baz="qwe"} + "qwe"`)
-	another(`with (f(a)=a) f`, `f`)
-	another(`with (f\q(a)=a) f\q`, `fq`)
-
-	// Verify withExpr for aggr func modifiers
-	another(`with (f(x) = x, y = sum(m) by (f)) y`, `sum(m) by (f)`)
-	another(`with (f(x) = sum(m) by (x)) f(foo)`, `sum(m) by (foo)`)
-	another(`with (f(x) = sum(m) by (x)) f((foo, bar, foo))`, `sum(m) by (foo, bar)`)
-	another(`with (f(x) = sum(m) without (x,y)) f((a, b))`, `sum(m) without (a, b, y)`)
-	another(`with (f(x) = sum(m) without (y,x)) f((a, y))`, `sum(m) without (y, a)`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f(foo,())`, `a + on (foo) group_left (bar) b`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f((foo),())`, `a + on (foo) group_left (bar) b`)
-	another(`with (f(x,y) = a + on (x,y) group_left (y,bar) b) f((foo,xx),())`, `a + on (foo, xx) group_left (bar) b`)
-
-	// Verify nested with exprs
-	another(`with (f(x) = (with(x=y) x) + x) f(z)`, `y + z`)
-	another(`with (x=foo) f(a, with (y=x) y)`, `f(a, foo)`)
-	another(`with (x=foo) a * x + (with (y=x) y) / y`, `(a * foo) + (foo / y)`)
-	another(`with (x = with (y = foo) y + x) x/x`, `(foo + x) / (foo + x)`)
-	another(`with (
-		x = {foo="bar"},
-		q = m{x, y="1"},
-		f(x) =
-			with (
-				z(y) = x + y * q
-			)
-			z(foo) / f(x)
-	)
-	f(a)`, `(a + (foo * m{foo="bar", y="1"})) / f(a)`)
-
-	// complex withExpr
-	another(`WITH (
-		treshold = (0.9),
-		commonFilters = {job="cacher", instance=~"1.2.3.4"},
-		hits = rate(cache{type="hit", commonFilters}[5m]),
-		miss = rate(cache{type="miss", commonFilters}[5m]),
-		sumByInstance(arg) = sum(arg) by (instance),
-		hitRatio = sumByInstance(hits) / sumByInstance(hits + miss)
-	)
-	hitRatio < treshold`,
-		`(sum(rate(cache{type="hit", job="cacher", instance=~"1.2.3.4"}[5m])) by (instance) / sum(rate(cache{type="hit", job="cacher", instance=~"1.2.3.4"}[5m]) + rate(cache{type="miss", job="cacher", instance=~"1.2.3.4"}[5m])) by (instance)) < 0.9`)
-	another(`WITH (
-		x2(x) = x^2,
-		f(x, y) = x2(x) + x*y + x2(y)
-	)
-	f(a, 3)
-	`, `((a ^ 2) + (a * 3)) + 9`)
-	another(`WITH (
-		x2(x) = x^2,
-		f(x, y) = x2(x) + x*y + x2(y)
-	)
-	f(2, 3)
-	`, `19`)
-	another(`WITH (
-		commonFilters = {instance="foo"},
-		timeToFuckup(currv, maxv) = (maxv - currv) / rate(currv)
-	)
-	timeToFuckup(diskUsage{commonFilters}, maxDiskSize{commonFilters})`,
-		`(maxDiskSize{instance="foo"} - diskUsage{instance="foo"}) / rate(diskUsage{instance="foo"})`)
-	another(`WITH (
-	       commonFilters = {job="foo", instance="bar"},
-	       sumRate(m, cf) = sum(rate(m{cf})) by (job, instance),
-	       hitRate(hits, misses) = sumRate(hits, commonFilters) / (sumRate(hits, commonFilters) + sumRate(misses, commonFilters))
-	   )
-	   hitRate(cacheHits, cacheMisses)`,
-		`sum(rate(cacheHits{job="foo", instance="bar"})) by (job, instance) / (sum(rate(cacheHits{job="foo", instance="bar"})) by (job, instance) + sum(rate(cacheMisses{job="foo", instance="bar"})) by (job, instance))`)
-	another(`with(y=123,z=5) union(with(y=3,f(x)=x*y) f(2) + f(3), with(x=5,y=2) x*y*z)`, `union(15, 50)`)
-}
-
-func TestParsePromQLError(t *testing.T) {
-	f := func(s string) {
-		t.Helper()
-
-		e, err := parsePromQL(s)
-		if err == nil {
-			t.Fatalf("expecting non-nil error when parsing %q", s)
-		}
-		if e != nil {
-			t.Fatalf("expecting nil expr when parsing %q", s)
-		}
-	}
-
-	// an empty string
-	f("")
-	f("  \t\b\r\n  ")
-
-	// invalid metricExpr
-	f(`{__name__="ff"} offset 55`)
-	f(`{__name__="ff"} offset -5m`)
-	f(`foo[55]`)
-	f(`m[-5m]`)
-	f(`{`)
-	f(`foo{`)
-	f(`foo{bar`)
-	f(`foo{bar=`)
-	f(`foo{bar="baz"`)
-	f(`foo{bar="baz",  `)
-	f(`foo{123="23"}`)
-	f(`foo{foo}`)
-	f(`foo{,}`)
-	f(`foo{,foo="bar"}`)
-	f(`foo{foo=}`)
-	f(`foo{foo="ba}`)
-	f(`foo{"foo"="bar"}`)
-	f(`foo{$`)
-	f(`foo{a $`)
-	f(`foo{a="b",$`)
-	f(`foo{a="b"}$`)
-	f(`[`)
-	f(`[]`)
-	f(`f[5m]$`)
-	f(`[5m]`)
-	f(`[5m] offset 4h`)
-	f(`m[5m] offset $`)
-	f(`m[5m] offset 5h $`)
-	f(`m[]`)
-	f(`m[-5m]`)
-	f(`m[5m:`)
-	f(`m[5m:-`)
-	f(`m[5m:-1`)
-	f(`m[5m:-1]`)
-	f(`m[:`)
-	f(`m[:-`)
-	f(`m[:1]`)
-	f(`m[:-1m]`)
-	f(`m[5]`)
-	f(`m[[5m]]`)
-	f(`m[foo]`)
-	f(`m["ff"]`)
-	f(`m[10m`)
-	f(`m[123`)
-	f(`m["ff`)
-	f(`m[(f`)
-	f(`fd}`)
-	f(`]`)
-	f(`m $`)
-	f(`m{,}`)
-	f(`m{x=y}`)
-	f(`m{x=y/5}`)
-	f(`m{x=y+5}`)
-
-	// Invalid regexp
-	f(`foo{bar=~"x["}`)
-	f(`foo{bar=~"x("}`)
-	f(`foo{bar=~"x)"}`)
-	f(`foo{bar!~"x["}`)
-	f(`foo{bar!~"x("}`)
-	f(`foo{bar!~"x)"}`)
-
-	// invalid stringExpr
-	f(`'`)
-	f(`"`)
-	f("`")
-	f(`"foo`)
-	f(`'foo`)
-	f("`foo")
-	f(`"foo\"bar`)
-	f(`'foo\'bar`)
-	f("`foo\\`bar")
-	f(`"" $`)
-	f(`"foo" +`)
-	f(`n{"foo" + m`)
-
-	// invalid numberExpr
-	f(`12.`)
-	f(`1.2e`)
-	f(`23e-`)
-	f(`23E+`)
-	f(`.`)
-	f(`-12.`)
-	f(`-1.2e`)
-	f(`-23e-`)
-	f(`-23E+`)
-	f(`-.`)
-	f(`-1$$`)
-	f(`-$$`)
-	f(`+$$`)
-	f(`23 $$`)
-
-	// invalid binaryOpExpr
-	f(`+`)
-	f(`1 +`)
-	f(`1 + 2.`)
-	f(`3 unless`)
-	f(`23 + on (foo)`)
-	f(`m + on (,) m`)
-	f(`3 * ignoring`)
-	f(`m * on (`)
-	f(`m * on (foo`)
-	f(`m * on (foo,`)
-	f(`m * on (foo,)`)
-	f(`m * on (,foo)`)
-	f(`m * on (,)`)
-	f(`m == bool (bar) baz`)
-	f(`m == bool () baz`)
-	f(`m * by (baz) n`)
-	f(`m + bool group_left m2`)
-	f(`m + on () group_left (`)
-	f(`m + on () group_left (,`)
-	f(`m + on () group_left (,foo`)
-	f(`m + on () group_left (foo,)`)
-	f(`m + on () group_left (,foo)`)
-	f(`m + on () group_left (foo)`)
-	f(`m + on () group_right (foo) (m`)
-	f(`m or ignoring () group_left () n`)
-	f(`1 + bool 2`)
-	f(`m % bool n`)
-	f(`m * bool baz`)
-	f(`M * BOoL BaZ`)
-	f(`foo unless ignoring (bar) group_left xxx`)
-	f(`foo or bool bar`)
-	f(`foo == bool $$`)
-	f(`"foo" + bar`)
-
-	// invalid parensExpr
-	f(`(`)
-	f(`($`)
-	f(`(+`)
-	f(`(1`)
-	f(`(m+`)
-	f(`1)`)
-	f(`(,)`)
-	f(`(1)$`)
-
-	// invalid funcExpr
-	f(`f $`)
-	f(`f($)`)
-	f(`f[`)
-	f(`f()$`)
-	f(`f(`)
-	f(`f(foo`)
-	f(`f(f,`)
-	f(`f(,`)
-	f(`f(,)`)
-	f(`f(,foo)`)
-	f(`f(,foo`)
-	f(`f(foo,$`)
-	f(`f() by (a)`)
-	f(`f without (x) (y)`)
-	f(`f() foo (a)`)
-	f(`f bar (x) (b)`)
-	f(`f bar (x)`)
-
-	// invalid aggrFuncExpr
-	f(`sum(`)
-	f(`sum $`)
-	f(`sum [`)
-	f(`sum($)`)
-	f(`sum()$`)
-	f(`sum(foo) ba`)
-	f(`sum(foo) ba()`)
-	f(`sum(foo) by`)
-	f(`sum(foo) without x`)
-	f(`sum(foo) aaa`)
-	f(`sum(foo) aaa x`)
-	f(`sum() by $`)
-	f(`sum() by (`)
-	f(`sum() by ($`)
-	f(`sum() by (a`)
-	f(`sum() by (a $`)
-	f(`sum() by (a ]`)
-	f(`sum() by (a)$`)
-	f(`sum() by (,`)
-	f(`sum() by (a,$`)
-	f(`sum() by (,)`)
-	f(`sum() by (,a`)
-	f(`sum() by (,a)`)
-	f(`sum() on (b)`)
-	f(`sum() bool`)
-	f(`sum() group_left`)
-	f(`sum() group_right(x)`)
-	f(`sum ba`)
-	f(`sum ba ()`)
-	f(`sum by (`)
-	f(`sum by (a`)
-	f(`sum by (,`)
-	f(`sum by (,)`)
-	f(`sum by (,a`)
-	f(`sum by (,a)`)
-	f(`sum by (a)`)
-	f(`sum by (a) (`)
-	f(`sum by (a) [`)
-	f(`sum by (a) {`)
-	f(`sum by (a) (b`)
-	f(`sum by (a) (b,`)
-	f(`sum by (a) (,)`)
-	f(`avg by (a) (,b)`)
-	f(`sum by (x) (y) by (z)`)
-	f(`sum(m) by (1)`)
-
-	// invalid withExpr
-	f(`with $`)
-	f(`with a`)
-	f(`with a=b c`)
-	f(`with (`)
-	f(`with (x=b)$`)
-	f(`with ($`)
-	f(`with (foo`)
-	f(`with (foo $`)
-	f(`with (x y`)
-	f(`with (x =`)
-	f(`with (x = $`)
-	f(`with (x= y`)
-	f(`with (x= y $`)
-	f(`with (x= y)`)
-	f(`with (x=(`)
-	f(`with (x=[)`)
-	f(`with (x=() x)`)
-	f(`with ($$)`)
-	f(`with (x $$`)
-	f(`with (x = $$)`)
-	f(`with (x = foo) bar{x}`)
-	f(`with (x = {foo="bar"}[5m]) bar{x}`)
-	f(`with (x = {foo="bar"} offset 5m) bar{x}`)
-	f(`with (x = a, x = b) c`)
-	f(`with (x(a, a) = b) c`)
-	f(`with (x=m{f="x"}) foo{x}`)
-	f(`with (sum = x) y`)
-	f(`with (rate(a) = b) c`)
-	f(`with (clamp_min=x) y`)
-	f(`with (f()`)
-	f(`with (a=b c=d) e`)
-	f(`with (f(x)=x^2) m{x}`)
-	f(`with (f(x)=ff()) m{x}`)
-	f(`with (f(x`)
-	f(`with (x=m) a{x} + b`)
-	f(`with (x=m) b + a{x}`)
-	f(`with (x=m) f(b, a{x})`)
-	f(`with (x=m) sum(a{x})`)
-	f(`with (x=m) (a{x})`)
-	f(`with (f(a)=a) f(1, 2)`)
-	f(`with (f(x)=x{foo="bar"}) f(1)`)
-	f(`with (f(x)=x{foo="bar"}) f(m + n)`)
-	f(`with (f = with`)
-	f(`with (,)`)
-	f(`with (1) 2`)
-	f(`with (f(1)=2) 3`)
-	f(`with (f(,)=x) x`)
-	f(`with (x(a) = {b="c"}) foo{x}`)
-	f(`with (f(x) = m{foo=xx}) f("qwe")`)
-	f(`a + with(f(x)=x) f(1,2)`)
-	f(`with (f(x) = sum(m) by (x)) f({foo="bar"})`)
-	f(`with (f(x) = sum(m) by (x)) f((xx(), {foo="bar"}))`)
-	f(`with (f(x) = m + on (x) n) f(xx())`)
-	f(`with (f(x) = m + on (a) group_right (x) n) f(xx())`)
-}

app/vmselect/promql/rollup_result_cache.go

@@ -12,12 +12,18 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
 	"github.com/VictoriaMetrics/fastcache"
 	"github.com/VictoriaMetrics/metrics"
 )
 
-var disableCache = flag.Bool("search.disableCache", false, "Whether to disable response caching. This may be useful during data backfilling")
+var (
+	disableCache         = flag.Bool("search.disableCache", false, "Whether to disable response caching. This may be useful during data backfilling")
+	cacheTimestampOffset = flag.Duration("search.cacheTimestampOffset", 5*time.Minute, "The maximum duration since the current time for response data, "+
+		"which is always queried from the original raw data, without using the response cache. Increase this value if you see gaps in responses "+
+		"due to time synchronization issues between VictoriaMetrics and data sources")
+)
 
 var rollupResultCacheV = &rollupResultCache{
 	c: workingsetcache.New(1024*1024, time.Hour), // This is a cache for testing.
@@ -73,8 +79,8 @@ func InitRollupResultCache(cachePath string) {
 		return stats
 	}
 	if len(rollupResultCachePath) > 0 {
-		logger.Infof("loaded rollupResult cache from %q in %s; entriesCount: %d, sizeBytes: %d",
-			rollupResultCachePath, time.Since(startTime), fcs().EntriesCount, fcs().BytesSize)
+		logger.Infof("loaded rollupResult cache from %q in %.3f seconds; entriesCount: %d, sizeBytes: %d",
+			rollupResultCachePath, time.Since(startTime).Seconds(), fcs().EntriesCount, fcs().BytesSize)
 	}
 
 	metrics.NewGauge(`vm_cache_entries{type="promql/rollupResult"}`, func() float64 {
@@ -112,8 +118,8 @@ func StopRollupResultCache() {
 	rollupResultCacheV.c.UpdateStats(&fcs)
 	rollupResultCacheV.c.Stop()
 	rollupResultCacheV.c = nil
-	logger.Infof("saved rollupResult cache to %q in %s; entriesCount: %d, sizeBytes: %d",
-		rollupResultCachePath, time.Since(startTime), fcs.EntriesCount, fcs.BytesSize)
+	logger.Infof("saved rollupResult cache to %q in %.3f seconds; entriesCount: %d, sizeBytes: %d",
+		rollupResultCachePath, time.Since(startTime).Seconds(), fcs.EntriesCount, fcs.BytesSize)
 }
 
 type rollupResultCache struct {
@@ -126,9 +132,10 @@ var rollupResultCacheResets = metrics.NewCounter(`vm_cache_resets_total{type="pr
 func ResetRollupResultCache() {
 	rollupResultCacheResets.Inc()
 	rollupResultCacheV.c.Reset()
+	logger.Infof("rollupResult cache has been cleared")
 }
 
-func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64) (tss []*timeseries, newStart int64) {
+func (rrc *rollupResultCache) Get(ec *EvalConfig, expr metricsql.Expr, window int64) (tss []*timeseries, newStart int64) {
 	if *disableCache || !ec.mayCache() {
 		return nil, ec.Start
 	}
@@ -137,7 +144,7 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	bb := bbPool.Get()
 	defer bbPool.Put(bb)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	if len(metainfoBuf) == 0 {
 		return nil, ec.Start
@@ -157,7 +164,7 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 	if len(compressedResultBuf.B) == 0 {
 		mi.RemoveKey(key)
 		metainfoBuf = mi.Marshal(metainfoBuf[:0])
-		bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+		bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 		rrc.c.Set(bb.B, metainfoBuf)
 		return nil, ec.Start
 	}
@@ -209,15 +216,15 @@ func (rrc *rollupResultCache) Get(funcName string, ec *EvalConfig, me *metricExp
 
 var resultBufPool bytesutil.ByteBufferPool
 
-func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExpr, iafc *incrementalAggrFuncContext, window int64, tss []*timeseries) {
+func (rrc *rollupResultCache) Put(ec *EvalConfig, expr metricsql.Expr, window int64, tss []*timeseries) {
 	if *disableCache || len(tss) == 0 || !ec.mayCache() {
 		return
 	}
 
-	// Remove values up to currentTime - step - maxSilenceInterval,
+	// Remove values up to currentTime - step - cacheTimestampOffset,
 	// since these values may be added later.
 	timestamps := tss[0].Timestamps
-	deadline := (time.Now().UnixNano() / 1e6) - ec.Step - maxSilenceInterval
+	deadline := (time.Now().UnixNano() / 1e6) - ec.Step - cacheTimestampOffset.Milliseconds()
 	i := len(timestamps) - 1
 	for i >= 0 && timestamps[i] > deadline {
 		i--
@@ -260,7 +267,7 @@ func (rrc *rollupResultCache) Put(funcName string, ec *EvalConfig, me *metricExp
 	bb.B = key.Marshal(bb.B[:0])
 	rrc.c.SetBig(bb.B, compressedResultBuf.B)
 
-	bb.B = marshalRollupResultCacheKey(bb.B[:0], funcName, me, iafc, window, ec.Step)
+	bb.B = marshalRollupResultCacheKey(bb.B[:0], expr, window, ec.Step)
 	metainfoBuf := rrc.c.Get(nil, bb.B)
 	var mi rollupResultCacheMetainfo
 	if len(metainfoBuf) > 0 {
@@ -288,23 +295,13 @@ var (
 var tooBigRollupResults = metrics.NewCounter("vm_too_big_rollup_results_total")
 
 // Increment this value every time the format of the cache changes.
-const rollupResultCacheVersion = 6
+const rollupResultCacheVersion = 7
 
-func marshalRollupResultCacheKey(dst []byte, funcName string, me *metricExpr, iafc *incrementalAggrFuncContext, window, step int64) []byte {
+func marshalRollupResultCacheKey(dst []byte, expr metricsql.Expr, window, step int64) []byte {
 	dst = append(dst, rollupResultCacheVersion)
-	if iafc == nil {
-		dst = append(dst, 0)
-	} else {
-		dst = append(dst, 1)
-		dst = iafc.ae.AppendString(dst)
-	}
-	dst = encoding.MarshalUint64(dst, uint64(len(funcName)))
-	dst = append(dst, funcName...)
 	dst = encoding.MarshalInt64(dst, window)
 	dst = encoding.MarshalInt64(dst, step)
-	for i := range me.TagFilters {
-		dst = me.TagFilters[i].Marshal(dst)
-	}
+	dst = expr.AppendString(dst)
 	return dst
 }
 

app/vmselect/promql/rollup_result_cache_test.go

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

app/vmselect/promql/rollup_test.go

@@ -3,6 +3,8 @@ package promql
 import (
 	"math"
 	"testing"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/metricsql"
 )
 
 var (
@@ -57,7 +59,7 @@ func TestRollupIderivDuplicateTimestamps(t *testing.T) {
 	}
 	n = rollupIderiv(rfa)
 	if n != 500 {
-		t.Fatalf("unexpected value; got %v; want %v", n, 0.5)
+		t.Fatalf("unexpected value; got %v; want %v", n, 500)
 	}
 
 	rfa = &rollupFuncArg{
@@ -157,7 +159,7 @@ func TestDerivValues(t *testing.T) {
 	testRowsEqual(t, values, timestamps, valuesExpected, timestamps)
 }
 
-func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpected *metricExpr, vExpected float64) {
+func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpected *metricsql.MetricExpr, vExpected float64) {
 	t.Helper()
 	nrf := getRollupFunc(funcName)
 	if nrf == nil {
@@ -190,6 +192,52 @@ func testRollupFunc(t *testing.T, funcName string, args []interface{}, meExpecte
 	}
 }
 
+func TestRollupShareLEOverTime(t *testing.T) {
+	f := func(le, vExpected float64) {
+		t.Helper()
+		les := []*timeseries{{
+			Values:     []float64{le},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, les}
+		testRollupFunc(t, "share_le_over_time", args, &me, vExpected)
+	}
+
+	f(-123, 0)
+	f(0, 0)
+	f(10, 0)
+	f(12, 0.08333333333333333)
+	f(30, 0.16666666666666666)
+	f(50, 0.75)
+	f(100, 0.9166666666666666)
+	f(123, 1)
+	f(1000, 1)
+}
+
+func TestRollupShareGTOverTime(t *testing.T) {
+	f := func(gt, vExpected float64) {
+		t.Helper()
+		gts := []*timeseries{{
+			Values:     []float64{gt},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, gts}
+		testRollupFunc(t, "share_gt_over_time", args, &me, vExpected)
+	}
+
+	f(-123, 1)
+	f(0, 1)
+	f(10, 1)
+	f(12, 0.9166666666666666)
+	f(30, 0.8333333333333334)
+	f(50, 0.25)
+	f(100, 0.08333333333333333)
+	f(123, 0)
+	f(1000, 0)
+}
+
 func TestRollupQuantileOverTime(t *testing.T) {
 	f := func(phi, vExpected float64) {
 		t.Helper()
@@ -197,8 +245,8 @@ func TestRollupQuantileOverTime(t *testing.T) {
 			Values:     []float64{phi},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{phis, &rollupExpr{Expr: &me}}
+		var me metricsql.MetricExpr
+		args := []interface{}{phis, &metricsql.RollupExpr{Expr: &me}}
 		testRollupFunc(t, "quantile_over_time", args, &me, vExpected)
 	}
 
@@ -219,8 +267,8 @@ func TestRollupPredictLinear(t *testing.T) {
 			Values:     []float64{sec},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}, secs}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, secs}
 		testRollupFunc(t, "predict_linear", args, &me, vExpected)
 	}
 
@@ -241,8 +289,8 @@ func TestRollupHoltWinters(t *testing.T) {
 			Values:     []float64{tf},
 			Timestamps: []int64{123},
 		}}
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}, sfs, tfs}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}, sfs, tfs}
 		testRollupFunc(t, "holt_winters", args, &me, vExpected)
 	}
 
@@ -262,27 +310,72 @@ func TestRollupHoltWinters(t *testing.T) {
 	f(0.9, 0.9, 33.99637566941818)
 }
 
+func TestRollupHoeffdingBoundLower(t *testing.T) {
+	f := func(phi, vExpected float64) {
+		t.Helper()
+		phis := []*timeseries{{
+			Values:     []float64{phi},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{phis, &metricsql.RollupExpr{Expr: &me}}
+		testRollupFunc(t, "hoeffding_bound_lower", args, &me, vExpected)
+	}
+
+	f(0.5, 28.21949401521037)
+	f(-1, 47.083333333333336)
+	f(0, 47.083333333333336)
+	f(1, -inf)
+	f(2, -inf)
+	f(0.1, 39.72878000047643)
+	f(0.9, 12.701803086472331)
+}
+
+func TestRollupHoeffdingBoundUpper(t *testing.T) {
+	f := func(phi, vExpected float64) {
+		t.Helper()
+		phis := []*timeseries{{
+			Values:     []float64{phi},
+			Timestamps: []int64{123},
+		}}
+		var me metricsql.MetricExpr
+		args := []interface{}{phis, &metricsql.RollupExpr{Expr: &me}}
+		testRollupFunc(t, "hoeffding_bound_upper", args, &me, vExpected)
+	}
+
+	f(0.5, 65.9471726514563)
+	f(-1, 47.083333333333336)
+	f(0, 47.083333333333336)
+	f(1, inf)
+	f(2, inf)
+	f(0.1, 54.43788666619024)
+	f(0.9, 81.46486358019433)
+}
+
 func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f := func(funcName string, vExpected float64) {
 		t.Helper()
-		var me metricExpr
-		args := []interface{}{&rollupExpr{Expr: &me}}
+		var me metricsql.MetricExpr
+		args := []interface{}{&metricsql.RollupExpr{Expr: &me}}
 		testRollupFunc(t, funcName, args, &me, vExpected)
 	}
 
 	f("default_rollup", 34)
 	f("changes", 11)
-	f("delta", -89)
+	f("delta", 34)
 	f("deriv", -266.85860231406065)
 	f("deriv_fast", -712)
 	f("idelta", 0)
-	f("increase", 275)
+	f("increase", 398)
 	f("irate", 0)
 	f("rate", 2200)
 	f("resets", 5)
+	f("range_over_time", 111)
 	f("avg_over_time", 47.083333333333336)
 	f("min_over_time", 12)
 	f("max_over_time", 123)
+	f("tmin_over_time", 0.08)
+	f("tmax_over_time", 0.005)
 	f("sum_over_time", 565)
 	f("sum2_over_time", 37951)
 	f("geomean_over_time", 39.33466603189148)
@@ -291,7 +384,7 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
 	f("stdvar_over_time", 945.7430555555555)
 	f("first_over_time", 123)
 	f("last_over_time", 34)
-	f("integrate", 61.0275)
+	f("integrate", 5.4705)
 	f("distinct_over_time", 8)
 	f("ideriv", 0)
 	f("decreases_over_time", 5)
@@ -327,7 +420,7 @@ func TestRollupNewRollupFuncError(t *testing.T) {
 		Values:     []float64{321},
 		Timestamps: []int64{123},
 	}}
-	me := &metricExpr{}
+	me := &metricsql.MetricExpr{}
 	f("holt_winters", []interface{}{123, 123, 321})
 	f("holt_winters", []interface{}{me, 123, 321})
 	f("holt_winters", []interface{}{me, scalarTs, 321})
@@ -409,7 +502,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 123, 123, 123, 34, 34}
+		valuesExpected := []float64{nan, 123, nan, 34, nan, 44}
 		timestampsExpected := []int64{0, 5, 10, 15, 20, 25}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -423,7 +516,7 @@ 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, nan}
+		valuesExpected := []float64{44, 32, 34, nan}
 		timestampsExpected := []int64{100, 120, 140, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -437,7 +530,7 @@ func TestRollupNoWindowPartialPoints(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, nan, 123, 54, 44}
+		valuesExpected := []float64{nan, nan, 123, 34, 32}
 		timestampsExpected := []int64{-50, 0, 50, 100, 150}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -488,6 +581,51 @@ func TestRollupWindowPartialPoints(t *testing.T) {
 	})
 }
 
+func TestRollupFuncsLookbackDelta(t *testing.T) {
+	t.Run("1", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 1,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("7", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 7,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+	t.Run("0", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:          rollupFirst,
+			Start:         80,
+			End:           140,
+			Step:          10,
+			LookbackDelta: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{99, nan, 44, nan, 32, 34, nan}
+		timestampsExpected := []int64{80, 90, 100, 110, 120, 130, 140}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
+}
+
 func TestRollupFuncsNoWindow(t *testing.T) {
 	t.Run("first", func(t *testing.T) {
 		rc := rollupConfig{
@@ -499,7 +637,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 123, 21, 12, 34}
+		valuesExpected := []float64{nan, 123, 54, 44, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -527,7 +665,7 @@ 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, 12, 34}
+		valuesExpected := []float64{nan, 21, 12, 32, 34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -569,7 +707,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, -102, -9, 22, 0}
+		valuesExpected := []float64{nan, nan, -9, 22, 0}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -587,6 +725,20 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{10, 50, 90, 130}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("lag", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupLag,
+			Start:  0,
+			End:    160,
+			Step:   40,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, 0.004, 0, 0, 0.03}
+		timestampsExpected := []int64{0, 40, 80, 120, 160}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("lifetime_1", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupLifetime,
@@ -713,6 +865,20 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
+	t.Run("deriv_fast", func(t *testing.T) {
+		rc := rollupConfig{
+			Func:   rollupDerivFast,
+			Start:  0,
+			End:    20,
+			Step:   4,
+			Window: 0,
+		}
+		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+		values := rc.Do(nil, testValues, testTimestamps)
+		valuesExpected := []float64{nan, nan, nan, 0, -8900, 0}
+		timestampsExpected := []int64{0, 4, 8, 12, 16, 20}
+		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+	})
 	t.Run("ideriv", func(t *testing.T) {
 		rc := rollupConfig{
 			Func:   rollupIderiv,
@@ -751,7 +917,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 		}
 		rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
 		values := rc.Do(nil, testValues, testTimestamps)
-		valuesExpected := []float64{nan, 4.6035, 4.3934999999999995, 2.166, 0.34}
+		valuesExpected := []float64{nan, 1.526, 2.2795, 1.325, 0.34}
 		timestampsExpected := []int64{0, 40, 80, 120, 160}
 		testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
 	})
@@ -785,6 +951,27 @@ func TestRollupFuncsNoWindow(t *testing.T) {
 	})
 }
 
+func TestRollupBigNumberOfValues(t *testing.T) {
+	const srcValuesCount = 1e4
+	rc := rollupConfig{
+		Func:   rollupDefault,
+		End:    srcValuesCount,
+		Step:   srcValuesCount / 5,
+		Window: srcValuesCount / 4,
+	}
+	rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
+	srcValues := make([]float64, srcValuesCount)
+	srcTimestamps := make([]int64, srcValuesCount)
+	for i := 0; i < srcValuesCount; i++ {
+		srcValues[i] = float64(i)
+		srcTimestamps[i] = int64(i / 2)
+	}
+	values := rc.Do(nil, srcValues, srcTimestamps)
+	valuesExpected := []float64{1, 4001, 8001, 9999, nan, nan}
+	timestampsExpected := []int64{0, 2000, 4000, 6000, 8000, 10000}
+	testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
+}
+
 func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExpected []float64, timestampsExpected []int64) {
 	t.Helper()
 	if len(values) != len(valuesExpected) {
@@ -813,7 +1000,7 @@ 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)
 		}