Merge tag 'v0.149.0' into pmm-6401-read-prometheus-data-files

Release v1.149.0

README.md

@@ -2,7 +2,6 @@
 
 [![Latest Release](https://img.shields.io/github/v/release/VictoriaMetrics/VictoriaMetrics?sort=semver&label=&filter=!*-victorialogs&logo=github&labelColor=gray&color=gray&link=https%3A%2F%2Fgithub.com%2FVictoriaMetrics%2FVictoriaMetrics%2Freleases%2Flatest)](https://github.com/VictoriaMetrics/VictoriaMetrics/releases)
 [![Docker Pulls](https://img.shields.io/docker/pulls/victoriametrics/victoria-metrics?label=&logo=docker&logoColor=white&labelColor=2496ED&color=2496ED&link=https%3A%2F%2Fhub.docker.com%2Fr%2Fvictoriametrics%2Fvictoria-metrics)](https://hub.docker.com/u/victoriametrics)
-[![Go Report](https://goreportcard.com/badge/github.com/VictoriaMetrics/VictoriaMetrics?link=https%3A%2F%2Fgoreportcard.com%2Freport%2Fgithub.com%2FVictoriaMetrics%2FVictoriaMetrics)](https://goreportcard.com/report/github.com/VictoriaMetrics/VictoriaMetrics)
 [![Build Status](https://github.com/VictoriaMetrics/VictoriaMetrics/actions/workflows/build.yml/badge.svg?branch=master&link=https%3A%2F%2Fgithub.com%2FVictoriaMetrics%2FVictoriaMetrics%2Factions)](https://github.com/VictoriaMetrics/VictoriaMetrics/actions/workflows/build.yml)
 [![License](https://img.shields.io/github/license/VictoriaMetrics/VictoriaMetrics?labelColor=green&label=&link=https%3A%2F%2Fgithub.com%2FVictoriaMetrics%2FVictoriaMetrics%2Fblob%2Fmaster%2FLICENSE)](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/LICENSE)
 [![Join Slack](https://img.shields.io/badge/Join%20Slack-4A154B?logo=slack)](https://slack.victoriametrics.com)

app/victoria-metrics/main.go

@@ -41,6 +41,12 @@ var (
 		"limit is reached; see also -search.maxQueryDuration")
 )
 
+// custom api help links [["/api","doc"]] without http.pathPrefix.
+var customAPIPathList = [][]string{
+	{"/graph/explore", "explore metrics grafana page"},
+	{"/graph/d/prometheus-advanced/advanced-data-exploration", "PMM grafana dashboard"},
+}
+
 func getDefaultMaxConcurrentRequests() int {
 	// 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
@@ -148,6 +154,10 @@ func requestHandler(w http.ResponseWriter, r *http.Request) bool {
 			{"api/v1/status/active_queries", "active queries"},
 			{"-/reload", "reload configuration"},
 		})
+		for _, p := range customAPIPathList {
+			p, doc := p[0], p[1]
+			fmt.Fprintf(w, "<a href=%q>%s</a> - %s<br/>", p, p, doc)
+		}
 		return true
 	}
 	if vminsert.RequestHandler(w, r) {

app/vmalert/datasource/client.go

@@ -11,6 +11,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/netutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promauth"
@@ -267,7 +268,11 @@ func (c *Client) do(req *http.Request) (*http.Response, error) {
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		_ = resp.Body.Close()
-		return nil, fmt.Errorf("unexpected response code %d for %s. Response body %s", resp.StatusCode, ru, body)
+		err = &httpserver.ErrorWithStatusCode{
+			StatusCode: resp.StatusCode,
+			Err:        fmt.Errorf("unexpected response code %d for %s. Response body %s", resp.StatusCode, ru, body),
+		}
+		return nil, err
 	}
 	return resp, nil
 }

app/vmalert/notifier/alert.go

@@ -31,6 +31,8 @@ type Alert struct {
 	State AlertState
 	// Expr contains expression that was executed to generate the Alert
 	Expr string
+	// Interval contains the evaluation interval of the Alert's group
+	Interval time.Duration
 	// ActiveAt defines the moment of time when Alert has become active
 	ActiveAt time.Time
 	// Start defines the moment of time when Alert has become firing
@@ -84,6 +86,7 @@ type AlertTplData struct {
 	Labels    map[string]string
 	Value     float64
 	Expr      string
+	Interval  time.Duration
 	AlertID   uint64
 	GroupID   uint64
 	ActiveAt  time.Time
@@ -96,6 +99,7 @@ var tplHeaders = []string{
 	"{{ $type := .Type }}",
 	"{{ $labels := .Labels }}",
 	"{{ $expr := .Expr }}",
+	"{{ $interval := .Interval }}",
 	"{{ $externalLabels := .ExternalLabels }}",
 	"{{ $externalURL := .ExternalURL }}",
 	"{{ $alertID := .AlertID }}",
@@ -115,6 +119,7 @@ func (a *Alert) ExecTemplate(q templates.QueryFn, labels, annotations map[string
 		Type:     a.Type,
 		Labels:   labels,
 		Expr:     a.Expr,
+		Interval: a.Interval,
 		AlertID:  a.ID,
 		GroupID:  a.GroupID,
 		ActiveAt: a.ActiveAt,

app/vmalert/notifier/alert_test.go

@@ -129,6 +129,17 @@ func TestAlertExecTemplate(t *testing.T) {
 		"exprEscapedHTML":   "vm_rows{"label"="bar"}<0",
 	})
 
+	// interval-template
+	f(&Alert{
+		Interval: 10 * time.Second,
+	}, map[string]string{
+		"interval":         "{{ .Interval }}",
+		"intervalVariable": "{{ $interval }}",
+	}, map[string]string{
+		"interval":         "10s",
+		"intervalVariable": "10s",
+	})
+
 	// query
 	f(&Alert{
 		Expr: `vm_rows{"label"="bar"}>0`,

app/vmalert/replay.go

@@ -25,11 +25,12 @@ var (
 	replayMaxDatapoints = flag.Int("replay.maxDatapointsPerQuery", 1e3,
 		"Max number of data points expected in one request. It affects the max time range for every '/query_range' request during the replay. The higher the value, the less requests will be made during replay.")
 	replayRuleRetryAttempts = flag.Int("replay.ruleRetryAttempts", 5,
-		"Defines how many retries to make before giving up on rule if request for it returns an error.")
+		"Defines how many retries to make before giving up on rule if request for it returns a retriable error.")
 	disableProgressBar = flag.Bool("replay.disableProgressBar", false, "Whether to disable rendering progress bars during the replay. "+
 		"Progress bar rendering might be verbose or break the logs parsing, so it is recommended to be disabled when not used in interactive mode.")
 	ruleEvaluationConcurrency = flag.Int("replay.ruleEvaluationConcurrency", 1, "The maximum number of concurrent '/query_range' requests when replay recording rule or alerting rule with for=0. "+
 		"Increasing this value when replaying for a long time, since each request is limited by -replay.maxDatapointsPerQuery.")
+	continueWithExecutionErr = flag.Bool("replay.continueWithExecutionErr", false, "Whether to continue replaying other rules if a rule execution fails with a 422 response code, which can happen due to an expression syntax error or a resource limit being hit.")
 )
 
 func replay(groupsCfg []config.Group, qb datasource.QuerierBuilder, rw remotewrite.RWClient) (totalRows, droppedRows int, err error) {
@@ -73,7 +74,7 @@ func replay(groupsCfg []config.Group, qb datasource.QuerierBuilder, rw remotewri
 
 	for _, cfg := range groupsCfg {
 		ng := rule.NewGroup(cfg, qb, *evaluationInterval, labels)
-		totalRows += ng.Replay(tFrom, tTo, rw, *replayMaxDatapoints, *replayRuleRetryAttempts, *replayRulesDelay, *disableProgressBar, *ruleEvaluationConcurrency)
+		totalRows += ng.Replay(tFrom, tTo, rw, *replayMaxDatapoints, *replayRuleRetryAttempts, *replayRulesDelay, *disableProgressBar, *ruleEvaluationConcurrency, *continueWithExecutionErr)
 	}
 	logger.Infof("replay evaluation finished, generated %d samples", totalRows)
 	if err := rw.Close(); err != nil {

app/vmalert/replay_test.go

@@ -8,8 +8,10 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/datasource"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promutil"
+	"github.com/VictoriaMetrics/metricsql"
 )
 
 type fakeReplayQuerier struct {
@@ -32,6 +34,14 @@ func (fc *fakeRWClient) Close() error {
 }
 
 func (fr *fakeReplayQuerier) QueryRange(_ context.Context, q string, from, to time.Time) (res datasource.Result, err error) {
+	_, err = metricsql.Parse(q)
+	if err != nil {
+		return res, &httpserver.ErrorWithStatusCode{
+			StatusCode: 422,
+			Err:        err,
+		}
+	}
+
 	key := fmt.Sprintf("%s+%s", from.Format("15:04:05"), to.Format("15:04:05"))
 	dps, ok := fr.registry[q]
 	if !ok {
@@ -275,4 +285,28 @@ func TestReplay(t *testing.T) {
 			},
 		},
 	}, 10)
+
+	// rule with wrong expression won't break the other rule with continueWithExecutionErr
+	continueWithExecutionErrOld := *continueWithExecutionErr
+	defer func() {
+		*continueWithExecutionErr = continueWithExecutionErrOld
+	}()
+	*continueWithExecutionErr = true
+	f("2021-01-01T12:00:00.000Z", "2021-01-01T12:02:30.000Z", 1, 1, time.Millisecond, []config.Group{
+		{Rules: []config.Rule{{Record: "foo", Expr: "sum(up)"}}},
+		{Rules: []config.Rule{{Record: "bar", Expr: "up ++"}}},
+	}, &fakeReplayQuerier{
+		registry: map[string]map[string][]datasource.Metric{
+			"sum(up)": {
+				"12:00:00+12:01:00": {
+					{
+						Timestamps: []int64{1, 2},
+						Values:     []float64{1, 2},
+					},
+				},
+				"12:01:00+12:02:00": {},
+				"12:02:00+12:02:30": {},
+			},
+		},
+	}, 2)
 }

app/vmalert/rule/alerting.go

@@ -530,6 +530,7 @@ func (ar *AlertingRule) exec(ctx context.Context, ts time.Time, limit int) ([]pr
 				ar.logDebugf(ts, a, "INACTIVE => PENDING")
 			}
 			a.Value = m.Values[0]
+			a.Interval = ar.EvalInterval
 			a.Annotations = annotations
 			a.KeepFiringSince = time.Time{}
 			continue
@@ -612,6 +613,7 @@ func (ar *AlertingRule) expandAnnotationTemplates(m datasource.Metric, qFn templ
 		Type:      ar.Type.String(),
 		Labels:    ls.origin,
 		Expr:      ar.Expr,
+		Interval:  ar.EvalInterval,
 		AlertID:   hash(ls.processed),
 		GroupID:   ar.GroupID,
 		ActiveAt:  activeAt,
@@ -673,6 +675,7 @@ func (ar *AlertingRule) newAlert(m datasource.Metric, start time.Time, labels, a
 		Name:        ar.Name,
 		Type:        ar.Type.String(),
 		Expr:        ar.Expr,
+		Interval:    ar.EvalInterval,
 		For:         ar.For,
 		ActiveAt:    start,
 		Value:       m.Values[0],

app/vmalert/rule/alerting_test.go

@@ -662,6 +662,7 @@ func TestAlertingRuleExecRange(t *testing.T) {
 			GroupID:     fakeGroup.GetID(),
 			Name:        "for-pending",
 			Type:        config.NewPrometheusType().String(),
+			Interval:    time.Second,
 			Labels:      map[string]string{"alertname": "for-pending"},
 			Annotations: map[string]string{},
 			State:       notifier.StatePending,
@@ -682,6 +683,7 @@ func TestAlertingRuleExecRange(t *testing.T) {
 			GroupID:     fakeGroup.GetID(),
 			Name:        "for-firing",
 			Type:        config.NewPrometheusType().String(),
+			Interval:    3 * time.Second,
 			Labels:      map[string]string{"alertname": "for-firing"},
 			Annotations: map[string]string{},
 			State:       notifier.StateFiring,
@@ -703,6 +705,7 @@ func TestAlertingRuleExecRange(t *testing.T) {
 			GroupID:     fakeGroup.GetID(),
 			Name:        "for-hold-pending",
 			Type:        config.NewPrometheusType().String(),
+			Interval:    time.Second,
 			Labels:      map[string]string{"alertname": "for-hold-pending"},
 			Annotations: map[string]string{},
 			State:       notifier.StatePending,
@@ -759,6 +762,7 @@ func TestAlertingRuleExecRange(t *testing.T) {
 			GroupID:     fakeGroup.GetID(),
 			Name:        "multi-series",
 			Type:        config.NewPrometheusType().String(),
+			Interval:    3 * time.Second,
 			Labels:      map[string]string{"alertname": "multi-series"},
 			Annotations: map[string]string{},
 			State:       notifier.StateFiring,
@@ -771,6 +775,7 @@ func TestAlertingRuleExecRange(t *testing.T) {
 			GroupID:     fakeGroup.GetID(),
 			Name:        "multi-series",
 			Type:        config.NewPrometheusType().String(),
+			Interval:    3 * time.Second,
 			Labels:      map[string]string{"alertname": "multi-series", "foo": "bar"},
 			Annotations: map[string]string{},
 			State:       notifier.StatePending,
@@ -1134,7 +1139,8 @@ func TestAlertingRule_Template(t *testing.T) {
 		fq.Add(metrics...)
 		fq.SetPartialResponse(isResponsePartial)
 
-		if _, err := rule.exec(context.TODO(), time.Now(), 0); err != nil {
+		ts := time.Unix(3600, 0)
+		if _, err := rule.exec(context.TODO(), ts, 0); err != nil {
 			t.Fatalf("unexpected error: %s", err)
 		}
 		for hash, expAlert := range alertsExpected {
@@ -1152,12 +1158,14 @@ func TestAlertingRule_Template(t *testing.T) {
 	}
 
 	f(&AlertingRule{
-		Name: "common",
+		Name:         "common",
+		EvalInterval: time.Hour,
 		Labels: map[string]string{
 			"region": "east",
 		},
 		Annotations: map[string]string{
-			"summary": `{{ $labels.alertname }}: Too high connection number for "{{ $labels.instance }}"`,
+			"summary":   `{{ $labels.alertname }}: Too high connection number for "{{ $labels.instance }}"`,
+			"dashboard": `&from={{ ($activeAt.Add (parseDurationTime (printf "-%s" .Interval))).UnixMilli }}&to={{ $activeAt.UnixMilli }}`,
 		},
 		alerts: make(map[uint64]*notifier.Alert),
 	}, []datasource.Metric{
@@ -1166,7 +1174,8 @@ func TestAlertingRule_Template(t *testing.T) {
 	}, false, map[uint64]*notifier.Alert{
 		hash(map[string]string{alertNameLabel: "common", "region": "east", "instance": "foo"}): {
 			Annotations: map[string]string{
-				"summary": `common: Too high connection number for "foo"`,
+				"summary":   `common: Too high connection number for "foo"`,
+				"dashboard": "&from=0&to=3600000",
 			},
 			Labels: map[string]string{
 				alertNameLabel: "common",
@@ -1176,7 +1185,8 @@ func TestAlertingRule_Template(t *testing.T) {
 		},
 		hash(map[string]string{alertNameLabel: "common", "region": "east", "instance": "bar"}): {
 			Annotations: map[string]string{
-				"summary": `common: Too high connection number for "bar"`,
+				"summary":   `common: Too high connection number for "bar"`,
+				"dashboard": "&from=0&to=3600000",
 			},
 			Labels: map[string]string{
 				alertNameLabel: "common",
@@ -1388,7 +1398,7 @@ func TestAlertingRule_ToLabels(t *testing.T) {
 		"alertname":     "ConfigurationReloadFailure",
 		"alertgroup":    "vmalert",
 		"pod":           "vmalert-0",
-		"invalid_label": `error evaluating template: template: :1:298: executing "" at <.Values.mustRuntimeFail>: can't evaluate field Values in type notifier.tplData`,
+		"invalid_label": `error evaluating template: template: :1:326: executing "" at <.Values.mustRuntimeFail>: can't evaluate field Values in type notifier.tplData`,
 	}
 
 	expectedProcessedLabels := map[string]string{
@@ -1398,7 +1408,7 @@ func TestAlertingRule_ToLabels(t *testing.T) {
 		"exported_alertname": "ConfigurationReloadFailure",
 		"group":              "vmalert",
 		"alertgroup":         "vmalert",
-		"invalid_label":      `error evaluating template: template: :1:298: executing "" at <.Values.mustRuntimeFail>: can't evaluate field Values in type notifier.tplData`,
+		"invalid_label":      `error evaluating template: template: :1:326: executing "" at <.Values.mustRuntimeFail>: can't evaluate field Values in type notifier.tplData`,
 	}
 
 	ls, err := ar.toLabels(metric, nil)

app/vmalert/rule/group.go

@@ -548,7 +548,7 @@ func (g *Group) infof(format string, args ...any) {
 }
 
 // Replay performs group replay
-func (g *Group) Replay(start, end time.Time, rw remotewrite.RWClient, maxDataPoint, replayRuleRetryAttempts int, replayDelay time.Duration, disableProgressBar bool, ruleEvaluationConcurrency int) int {
+func (g *Group) Replay(start, end time.Time, rw remotewrite.RWClient, maxDataPoint, replayRuleRetryAttempts int, replayDelay time.Duration, disableProgressBar bool, ruleEvaluationConcurrency int, continueWithExecutionErr bool) int {
 	var total int
 	step := g.Interval * time.Duration(maxDataPoint)
 	ri := rangeIterator{start: start, end: end, step: step}
@@ -576,7 +576,7 @@ func (g *Group) Replay(start, end time.Time, rw remotewrite.RWClient, maxDataPoi
 			if !disableProgressBar {
 				bar = pb.StartNew(iterations)
 			}
-			total += replayRuleRange(rule, ri, bar, rw, replayRuleRetryAttempts, ruleEvaluationConcurrency)
+			total += replayRuleRange(rule, ri, bar, rw, replayRuleRetryAttempts, ruleEvaluationConcurrency, continueWithExecutionErr)
 			if bar != nil {
 				bar.Finish()
 			}
@@ -598,7 +598,7 @@ func (g *Group) Replay(start, end time.Time, rw remotewrite.RWClient, maxDataPoi
 		rule := g.Rules[i]
 		sem <- struct{}{}
 		wg.Go(func() {
-			res <- replayRuleRange(rule, ri, bar, rw, replayRuleRetryAttempts, ruleEvaluationConcurrency)
+			res <- replayRuleRange(rule, ri, bar, rw, replayRuleRetryAttempts, ruleEvaluationConcurrency, continueWithExecutionErr)
 			<-sem
 		})
 	}
@@ -618,7 +618,7 @@ func (g *Group) Replay(start, end time.Time, rw remotewrite.RWClient, maxDataPoi
 	return total
 }
 
-func replayRuleRange(r Rule, ri rangeIterator, bar *pb.ProgressBar, rw remotewrite.RWClient, replayRuleRetryAttempts, ruleEvaluationConcurrency int) int {
+func replayRuleRange(r Rule, ri rangeIterator, bar *pb.ProgressBar, rw remotewrite.RWClient, replayRuleRetryAttempts, ruleEvaluationConcurrency int, continueWithExecutionErr bool) int {
 	fmt.Printf("> Rule %q (ID: %d)\n", r, r.ID())
 	// alerting rule with for>0 can't be replayed concurrently, since the status change might depend on the previous evaluation
 	// see https://github.com/VictoriaMetrics/VictoriaMetrics/commit/abcb21aa5ee918ba9a4e9cde495dba06e1e9564c
@@ -633,7 +633,7 @@ func replayRuleRange(r Rule, ri rangeIterator, bar *pb.ProgressBar, rw remotewri
 		start := ri.s
 		end := ri.e
 		wg.Go(func() {
-			n, err := replayRule(r, start, end, rw, replayRuleRetryAttempts)
+			n, err := replayRule(r, start, end, rw, replayRuleRetryAttempts, continueWithExecutionErr)
 			if err != nil {
 				logger.Fatalf("rule %q: %s", r, err)
 			}

app/vmalert/rule/rule.go

@@ -4,12 +4,14 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"net/http"
 	"sync"
 	"time"
 
 	"github.com/VictoriaMetrics/metrics"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/remotewrite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompb"
 )
@@ -118,7 +120,7 @@ func (s *ruleState) add(e StateEntry) {
 	s.entries[s.cur] = e
 }
 
-func replayRule(r Rule, start, end time.Time, rw remotewrite.RWClient, replayRuleRetryAttempts int) (int, error) {
+func replayRule(r Rule, start, end time.Time, rw remotewrite.RWClient, replayRuleRetryAttempts int, continueWithExecutionErr bool) (int, error) {
 	var err error
 	var tss []prompb.TimeSeries
 	for i := range replayRuleRetryAttempts {
@@ -126,6 +128,21 @@ func replayRule(r Rule, start, end time.Time, rw remotewrite.RWClient, replayRul
 		if err == nil {
 			break
 		}
+		// retry request if possible to tolerate temporary network or datasource unavailability issues
+		var esc *httpserver.ErrorWithStatusCode
+		if errors.As(err, &esc) {
+			statusCode := esc.StatusCode
+			// if the status code is 422, it means that the query was executed but failed due to an expression syntax error or a the resource limit being hit,
+			// continue replaying but skip the problematic execution if continueWithExecutionErr is true, otherwise, return the error without retry.
+			if statusCode == http.StatusUnprocessableEntity {
+				if continueWithExecutionErr {
+					logger.Errorf("rule %q: %s", r, err)
+					return 0, nil
+				} else {
+					return 0, err
+				}
+			}
+		}
 		logger.Errorf("attempt %d to execute rule %q failed: %s", i+1, r, err)
 		time.Sleep(time.Second)
 	}

app/vmctl/remoteread/remoteread.go

@@ -8,12 +8,14 @@ import (
 	"io"
 	"net/http"
 	"net/url"
+	"strconv"
 	"strings"
 	"time"
 
 	"github.com/gogo/protobuf/proto"
 	"github.com/golang/snappy"
 	"github.com/prometheus/prometheus/config"
+	"github.com/prometheus/prometheus/model/histogram"
 	"github.com/prometheus/prometheus/prompb"
 	"github.com/prometheus/prometheus/storage/remote"
 	"github.com/prometheus/prometheus/tsdb/chunkenc"
@@ -234,9 +236,29 @@ func processResponse(body io.ReadCloser, callback StreamCallback) error {
 	// shouldn't be accounted as an error.
 	for _, res := range readResp.Results {
 		for _, ts := range res.Timeseries {
-			vmTs := convertSamples(ts.Samples, ts.Labels)
-			if err := callback(vmTs); err != nil {
-				return err
+			// A series contains either float samples or native histogram samples.
+			// Both fields are processed independently, since a series may switch
+			// from float to native histogram representation at some point in time,
+			// so the requested time range may contain samples of both types.
+			if len(ts.Samples) > 0 {
+				vmTs := convertSamples(ts.Samples, ts.Labels)
+				if err := callback(vmTs); err != nil {
+					return err
+				}
+			}
+			if len(ts.Histograms) > 0 {
+				hSamples := make([]histogramSample, 0, len(ts.Histograms))
+				for _, h := range ts.Histograms {
+					hSamples = append(hSamples, histogramSample{
+						timestamp: h.Timestamp,
+						fh:        h.ToFloatHistogram(),
+					})
+				}
+				for _, vmTs := range convertHistograms(hSamples, ts.Labels) {
+					if err := callback(vmTs); err != nil {
+						return err
+					}
+				}
 			}
 		}
 	}
@@ -263,17 +285,45 @@ func processStreamResponse(body io.ReadCloser, callback StreamCallback) error {
 
 		for _, series := range res.ChunkedSeries {
 			samples := make([]prompb.Sample, 0)
+			var hSamples []histogramSample
 			for _, chunk := range series.Chunks {
-				s, err := parseSamples(chunk.Data)
-				if err != nil {
-					return err
+				switch chunk.Type {
+				case prompb.Chunk_XOR, prompb.Chunk_UNKNOWN:
+					// In proto3 the `type` field may be left unset (UNKNOWN) for XOR chunks.
+					// Prometheus remote.proto: "REQUIREMENT: when using proto3, this field
+					// MUST be set when using anything else than XOR". Senders before native
+					// histograms support (Prometheus < 2.40) do not set this field at all,
+					// so UNKNOWN chunks must be parsed as XOR ones.
+					s, err := parseSamples(chunk.Data)
+					if err != nil {
+						return err
+					}
+					samples = append(samples, s...)
+				case prompb.Chunk_HISTOGRAM, prompb.Chunk_FLOAT_HISTOGRAM:
+					hs, err := parseHistograms(chunk.Type, chunk.Data)
+					if err != nil {
+						return err
+					}
+					hSamples = append(hSamples, hs...)
+				default:
+					return fmt.Errorf("unsupported chunk encoding %q", chunk.Type)
 				}
-				samples = append(samples, s...)
 			}
 
-			ts := convertSamples(samples, series.Labels)
-			if err := callback(ts); err != nil {
-				return err
+			// A series contains either XOR chunks or native histogram chunks.
+			// Both are processed independently, since a series may switch
+			// from float to native histogram representation at some point in time,
+			// so the requested time range may contain chunks of both types.
+			if len(samples) > 0 {
+				ts := convertSamples(samples, series.Labels)
+				if err := callback(ts); err != nil {
+					return err
+				}
+			}
+			for _, ts := range convertHistograms(hSamples, series.Labels) {
+				if err := callback(ts); err != nil {
+					return err
+				}
 			}
 		}
 	}
@@ -312,6 +362,151 @@ func parseSamples(chunk []byte) ([]prompb.Sample, error) {
 	return samples, it.Err()
 }
 
+// histogramSample represents a single native histogram sample.
+type histogramSample struct {
+	timestamp int64
+	fh        *histogram.FloatHistogram
+}
+
+func parseHistograms(encoding prompb.Chunk_Encoding, chunk []byte) ([]histogramSample, error) {
+	var enc chunkenc.Encoding
+	switch encoding {
+	case prompb.Chunk_HISTOGRAM:
+		enc = chunkenc.EncHistogram
+	case prompb.Chunk_FLOAT_HISTOGRAM:
+		enc = chunkenc.EncFloatHistogram
+	default:
+		return nil, fmt.Errorf("unsupported histogram chunk encoding %q", encoding)
+	}
+	c, err := chunkenc.FromData(enc, chunk)
+	if err != nil {
+		return nil, fmt.Errorf("error read chunk: %w", err)
+	}
+
+	var hSamples []histogramSample
+	it := c.Iterator(nil)
+	for {
+		typ := it.Next()
+		if typ == chunkenc.ValNone {
+			break
+		}
+		switch typ {
+		case chunkenc.ValHistogram:
+			ts, h := it.AtHistogram(nil)
+			hSamples = append(hSamples, histogramSample{
+				timestamp: ts,
+				fh:        h.ToFloat(nil),
+			})
+		case chunkenc.ValFloatHistogram:
+			ts, fh := it.AtFloatHistogram(nil)
+			hSamples = append(hSamples, histogramSample{
+				timestamp: ts,
+				fh:        fh,
+			})
+		default:
+			// Skip unsupported values
+			continue
+		}
+	}
+	if err := it.Err(); err != nil {
+		return nil, fmt.Errorf("error iterate over chunks: %w", err)
+	}
+
+	return hSamples, nil
+}
+
+// convertHistograms converts native histogram samples into VictoriaMetrics histogram
+// time series in the same way as VictoriaMetrics converts native histograms
+// received via Prometheus remote write protocol: every native histogram sample
+// is converted into `<name>_count` and `<name>_sum` series plus a set of
+// `<name>_bucket` series with `vmrange` labels containing non-cumulative bucket counts.
+// The only difference is that for native histograms with custom buckets (NHCB)
+// bucket bounds are taken from the custom values, while the remote write protocol
+// parser ignores custom values and estimates the bounds with the exponential formula.
+// See https://prometheus.io/docs/specs/native_histograms/#data-model
+func convertHistograms(hSamples []histogramSample, labels []prompb.Label) []*vm.TimeSeries {
+	if len(hSamples) == 0 {
+		return nil
+	}
+
+	labelPairs := make([]vm.LabelPair, 0, len(labels))
+	nameValue := ""
+	for _, label := range labels {
+		if label.Name == "__name__" {
+			nameValue = label.Value
+			continue
+		}
+		labelPairs = append(labelPairs, vm.LabelPair{Name: label.Name, Value: label.Value})
+	}
+	// the metric has no name, skip it in the same way as VictoriaMetrics does
+	// when it receives a native histogram without the metric name via remote write protocol.
+	if nameValue == "" {
+		return nil
+	}
+
+	countSeries := &vm.TimeSeries{
+		Name:       nameValue + "_count",
+		LabelPairs: labelPairs,
+	}
+	sumSeries := &vm.TimeSeries{
+		Name:       nameValue + "_sum",
+		LabelPairs: labelPairs,
+	}
+	bucketSeries := make(map[string]*vm.TimeSeries)
+	// vmranges preserves the order of bucketSeries creation
+	// in order to get deterministic results.
+	var vmranges []string
+
+	for _, hs := range hSamples {
+		fh := hs.fh
+		countSeries.Timestamps = append(countSeries.Timestamps, hs.timestamp)
+		countSeries.Values = append(countSeries.Values, fh.Count)
+		sumSeries.Timestamps = append(sumSeries.Timestamps, hs.timestamp)
+		sumSeries.Values = append(sumSeries.Values, fh.Sum)
+
+		it := fh.AllBucketIterator()
+		for it.Next() {
+			b := it.At()
+			if b.Count <= 0 {
+				continue
+			}
+			vmrange := formatVmrange(b.Lower, b.Upper)
+			s := bucketSeries[vmrange]
+			if s == nil {
+				bucketLabelPairs := make([]vm.LabelPair, len(labelPairs), len(labelPairs)+1)
+				copy(bucketLabelPairs, labelPairs)
+				bucketLabelPairs = append(bucketLabelPairs, vm.LabelPair{Name: "vmrange", Value: vmrange})
+				s = &vm.TimeSeries{
+					Name:       nameValue + "_bucket",
+					LabelPairs: bucketLabelPairs,
+				}
+				bucketSeries[vmrange] = s
+				vmranges = append(vmranges, vmrange)
+			}
+			s.Timestamps = append(s.Timestamps, hs.timestamp)
+			s.Values = append(s.Values, b.Count)
+		}
+	}
+
+	tss := make([]*vm.TimeSeries, 0, 2+len(vmranges))
+	tss = append(tss, countSeries, sumSeries)
+	for _, vmrange := range vmranges {
+		tss = append(tss, bucketSeries[vmrange])
+	}
+	return tss
+}
+
+// formatVmrange formats the given bucket bounds into `vmrange` label value
+// in the same way as VictoriaMetrics does for native histograms
+// received via Prometheus remote write protocol.
+func formatVmrange(lower, upper float64) string {
+	b := make([]byte, 0, 24)
+	b = strconv.AppendFloat(b, lower, 'e', 3, 64)
+	b = append(b, "..."...)
+	b = strconv.AppendFloat(b, upper, 'e', 3, 64)
+	return string(b)
+}
+
 type keyValue struct {
 	key   string
 	value string

app/vmctl/remoteread/remoteread_test.go

@@ -0,0 +1,334 @@
+package remoteread
+
+import (
+	"bytes"
+	"io"
+	"reflect"
+	"testing"
+
+	"github.com/gogo/protobuf/proto"
+	"github.com/golang/snappy"
+	"github.com/prometheus/prometheus/model/histogram"
+	"github.com/prometheus/prometheus/prompb"
+	"github.com/prometheus/prometheus/storage/remote"
+	"github.com/prometheus/prometheus/tsdb/chunkenc"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmctl/vm"
+)
+
+func testHistogram(mul int64) *histogram.Histogram {
+	return &histogram.Histogram{
+		Schema:          0,
+		Count:           uint64(10 * mul),
+		Sum:             25.5 * float64(mul),
+		ZeroThreshold:   0.001,
+		ZeroCount:       uint64(2 * mul),
+		PositiveSpans:   []histogram.Span{{Offset: 0, Length: 2}},
+		PositiveBuckets: []int64{1 * mul, 2 * mul},
+		NegativeSpans:   []histogram.Span{{Offset: 0, Length: 1}},
+		NegativeBuckets: []int64{4 * mul},
+	}
+}
+
+func TestConvertHistograms(t *testing.T) {
+	f := func(hSamples []histogramSample, labels []prompb.Label, expected []*vm.TimeSeries) {
+		t.Helper()
+
+		tss := convertHistograms(hSamples, labels)
+		if !reflect.DeepEqual(tss, expected) {
+			t.Fatalf("unexpected result\ngot:\n%v\nwant:\n%v", tss, expected)
+		}
+	}
+
+	// series without samples
+	f(nil, []prompb.Label{{Name: "__name__", Value: "foo"}}, nil)
+
+	// series without the metric name must be skipped
+	f([]histogramSample{
+		{timestamp: 1000, fh: testHistogram(1).ToFloat(nil)},
+	}, []prompb.Label{{Name: "job", Value: "bar"}}, nil)
+
+	// native histogram must be converted to _count, _sum and _bucket series
+	// in the same way as VictoriaMetrics does for Prometheus remote write protocol
+	labels := []prompb.Label{
+		{Name: "__name__", Value: "request_duration_seconds"},
+		{Name: "job", Value: "bar"},
+	}
+	jobLabel := []vm.LabelPair{{Name: "job", Value: "bar"}}
+	bucketLabels := func(vmrange string) []vm.LabelPair {
+		return []vm.LabelPair{
+			{Name: "job", Value: "bar"},
+			{Name: "vmrange", Value: vmrange},
+		}
+	}
+	f([]histogramSample{
+		{timestamp: 1000, fh: testHistogram(1).ToFloat(nil)},
+		{timestamp: 2000, fh: testHistogram(2).ToFloat(nil)},
+	}, labels, []*vm.TimeSeries{
+		{
+			Name:       "request_duration_seconds_count",
+			LabelPairs: jobLabel,
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{10, 20},
+		},
+		{
+			Name:       "request_duration_seconds_sum",
+			LabelPairs: jobLabel,
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{25.5, 51},
+		},
+		{
+			Name:       "request_duration_seconds_bucket",
+			LabelPairs: bucketLabels("-1.000e+00...-5.000e-01"),
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{4, 8},
+		},
+		{
+			Name:       "request_duration_seconds_bucket",
+			LabelPairs: bucketLabels("-1.000e-03...1.000e-03"),
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{2, 4},
+		},
+		{
+			Name:       "request_duration_seconds_bucket",
+			LabelPairs: bucketLabels("5.000e-01...1.000e+00"),
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{1, 2},
+		},
+		{
+			Name:       "request_duration_seconds_bucket",
+			LabelPairs: bucketLabels("1.000e+00...2.000e+00"),
+			Timestamps: []int64{1000, 2000},
+			Values:     []float64{3, 6},
+		},
+	})
+}
+
+func TestParseHistograms(t *testing.T) {
+	c := chunkenc.NewHistogramChunk()
+	app, err := c.Appender()
+	if err != nil {
+		t.Fatalf("cannot create chunk appender: %s", err)
+	}
+	if _, _, _, err := app.AppendHistogram(nil, 0, 1000, testHistogram(1), true); err != nil {
+		t.Fatalf("cannot append histogram: %s", err)
+	}
+	if _, _, _, err := app.AppendHistogram(nil, 0, 2000, testHistogram(2), true); err != nil {
+		t.Fatalf("cannot append histogram: %s", err)
+	}
+
+	hSamples, err := parseHistograms(prompb.Chunk_HISTOGRAM, c.Bytes())
+	if err != nil {
+		t.Fatalf("cannot parse histogram chunk: %s", err)
+	}
+	if len(hSamples) != 2 {
+		t.Fatalf("unexpected number of histogram samples; got %d; want 2", len(hSamples))
+	}
+	for i, expected := range []struct {
+		timestamp int64
+		count     float64
+		sum       float64
+	}{
+		{timestamp: 1000, count: 10, sum: 25.5},
+		{timestamp: 2000, count: 20, sum: 51},
+	} {
+		if hSamples[i].timestamp != expected.timestamp {
+			t.Fatalf("unexpected timestamp; got %d; want %d", hSamples[i].timestamp, expected.timestamp)
+		}
+		if hSamples[i].fh.Count != expected.count {
+			t.Fatalf("unexpected count; got %f; want %f", hSamples[i].fh.Count, expected.count)
+		}
+		if hSamples[i].fh.Sum != expected.sum {
+			t.Fatalf("unexpected sum; got %f; want %f", hSamples[i].fh.Sum, expected.sum)
+		}
+	}
+
+	// unsupported chunk encoding must return error
+	if _, err := parseHistograms(prompb.Chunk_XOR, c.Bytes()); err == nil {
+		t.Fatalf("expecting non-nil error for unsupported chunk encoding")
+	}
+}
+
+func TestProcessResponse(t *testing.T) {
+	readResp := &prompb.ReadResponse{
+		Results: []*prompb.QueryResult{
+			{
+				Timeseries: []*prompb.TimeSeries{
+					{
+						Labels: []prompb.Label{
+							{Name: "__name__", Value: "cpu_usage"},
+							{Name: "job", Value: "bar"},
+						},
+						Samples: []prompb.Sample{
+							{Timestamp: 1000, Value: 1.5},
+						},
+					},
+					{
+						Labels: []prompb.Label{
+							{Name: "__name__", Value: "request_duration_seconds"},
+							{Name: "job", Value: "bar"},
+						},
+						Histograms: []prompb.Histogram{
+							prompb.FromIntHistogram(1000, testHistogram(1)),
+						},
+					},
+				},
+			},
+		},
+	}
+	data, err := proto.Marshal(readResp)
+	if err != nil {
+		t.Fatalf("cannot marshal ReadResponse: %s", err)
+	}
+	compressed := snappy.Encode(nil, data)
+
+	var tss []*vm.TimeSeries
+	err = processResponse(io.NopCloser(bytes.NewReader(compressed)), func(ts *vm.TimeSeries) error {
+		tss = append(tss, ts)
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("cannot process response: %s", err)
+	}
+
+	// 1 float series + _count + _sum + 4 buckets
+	if len(tss) != 7 {
+		t.Fatalf("unexpected number of time series; got %d; want 7", len(tss))
+	}
+	if tss[0].Name != "cpu_usage" || !reflect.DeepEqual(tss[0].Values, []float64{1.5}) {
+		t.Fatalf("unexpected float series: %v", tss[0])
+	}
+	if tss[1].Name != "request_duration_seconds_count" || !reflect.DeepEqual(tss[1].Values, []float64{10}) {
+		t.Fatalf("unexpected _count series: %v", tss[1])
+	}
+	if tss[2].Name != "request_duration_seconds_sum" || !reflect.DeepEqual(tss[2].Values, []float64{25.5}) {
+		t.Fatalf("unexpected _sum series: %v", tss[2])
+	}
+	for _, ts := range tss[3:] {
+		if ts.Name != "request_duration_seconds_bucket" {
+			t.Fatalf("unexpected bucket series name %q", ts.Name)
+		}
+	}
+}
+
+type nopFlusher struct{}
+
+func (nopFlusher) Flush() {}
+
+func TestProcessStreamResponse(t *testing.T) {
+	// build a histogram chunk
+	hc := chunkenc.NewHistogramChunk()
+	hApp, err := hc.Appender()
+	if err != nil {
+		t.Fatalf("cannot create histogram chunk appender: %s", err)
+	}
+	if _, _, _, err := hApp.AppendHistogram(nil, 0, 1000, testHistogram(1), true); err != nil {
+		t.Fatalf("cannot append histogram: %s", err)
+	}
+
+	// build a float chunk
+	xc := chunkenc.NewXORChunk()
+	xApp, err := xc.Appender()
+	if err != nil {
+		t.Fatalf("cannot create xor chunk appender: %s", err)
+	}
+	xApp.Append(0, 1000, 1.5)
+
+	res := &prompb.ChunkedReadResponse{
+		ChunkedSeries: []*prompb.ChunkedSeries{
+			{
+				Labels: []prompb.Label{
+					{Name: "__name__", Value: "request_duration_seconds"},
+					{Name: "job", Value: "bar"},
+				},
+				Chunks: []prompb.Chunk{
+					{Type: prompb.Chunk_HISTOGRAM, Data: hc.Bytes()},
+				},
+			},
+			{
+				Labels: []prompb.Label{
+					{Name: "__name__", Value: "cpu_usage"},
+				},
+				Chunks: []prompb.Chunk{
+					{Type: prompb.Chunk_XOR, Data: xc.Bytes()},
+				},
+			},
+			{
+				Labels: []prompb.Label{
+					{Name: "__name__", Value: "memory_usage"},
+				},
+				Chunks: []prompb.Chunk{
+					// the `type` field may be unset for XOR chunks,
+					// such chunks must be parsed as XOR ones
+					{Type: prompb.Chunk_UNKNOWN, Data: xc.Bytes()},
+				},
+			},
+		},
+	}
+	data, err := proto.Marshal(res)
+	if err != nil {
+		t.Fatalf("cannot marshal ChunkedReadResponse: %s", err)
+	}
+	var buf bytes.Buffer
+	cw := remote.NewChunkedWriter(&buf, nopFlusher{})
+	if _, err := cw.Write(data); err != nil {
+		t.Fatalf("cannot write chunked response: %s", err)
+	}
+
+	var tss []*vm.TimeSeries
+	err = processStreamResponse(io.NopCloser(&buf), func(ts *vm.TimeSeries) error {
+		tss = append(tss, ts)
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("cannot process stream response: %s", err)
+	}
+
+	// _count + _sum + 4 buckets + 1 float series + 1 float series from UNKNOWN chunk
+	if len(tss) != 8 {
+		t.Fatalf("unexpected number of time series; got %d; want 8", len(tss))
+	}
+	if tss[0].Name != "request_duration_seconds_count" || !reflect.DeepEqual(tss[0].Values, []float64{10}) {
+		t.Fatalf("unexpected _count series: %v", tss[0])
+	}
+	if tss[1].Name != "request_duration_seconds_sum" || !reflect.DeepEqual(tss[1].Values, []float64{25.5}) {
+		t.Fatalf("unexpected _sum series: %v", tss[1])
+	}
+	for _, ts := range tss[2:6] {
+		if ts.Name != "request_duration_seconds_bucket" {
+			t.Fatalf("unexpected bucket series name %q", ts.Name)
+		}
+	}
+	if tss[6].Name != "cpu_usage" || !reflect.DeepEqual(tss[6].Values, []float64{1.5}) {
+		t.Fatalf("unexpected float series: %v", tss[6])
+	}
+	if tss[7].Name != "memory_usage" || !reflect.DeepEqual(tss[7].Values, []float64{1.5}) {
+		t.Fatalf("unexpected float series from UNKNOWN chunk: %v", tss[7])
+	}
+}
+
+func TestParseFloatHistograms(t *testing.T) {
+	c := chunkenc.NewFloatHistogramChunk()
+	app, err := c.Appender()
+	if err != nil {
+		t.Fatalf("cannot create chunk appender: %s", err)
+	}
+	fh := testHistogram(1).ToFloat(nil)
+	if _, _, _, err := app.AppendFloatHistogram(nil, 0, 1000, fh, true); err != nil {
+		t.Fatalf("cannot append float histogram: %s", err)
+	}
+
+	hSamples, err := parseHistograms(prompb.Chunk_FLOAT_HISTOGRAM, c.Bytes())
+	if err != nil {
+		t.Fatalf("cannot parse float histogram chunk: %s", err)
+	}
+	if len(hSamples) != 1 {
+		t.Fatalf("unexpected number of histogram samples; got %d; want 1", len(hSamples))
+	}
+	if hSamples[0].timestamp != 1000 {
+		t.Fatalf("unexpected timestamp; got %d; want 1000", hSamples[0].timestamp)
+	}
+	if hSamples[0].fh.Count != 10 {
+		t.Fatalf("unexpected count; got %f; want 10", hSamples[0].fh.Count)
+	}
+}

app/vmselect/netstorage/netstorage.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/searchutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage/promdb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/cgroup"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
@@ -320,14 +321,21 @@ func (rss *Results) runParallel(qt *querytracer.Tracer, f func(rs *Result, worke
 }
 
 var (
-	rowsReadPerSeries  = metrics.NewHistogram(`vm_rows_read_per_series`)
-	rowsReadPerQuery   = metrics.NewHistogram(`vm_rows_read_per_query`)
-	seriesReadPerQuery = metrics.NewHistogram(`vm_series_read_per_query`)
+	rowsReadPerSeries            = metrics.NewHistogram(`vm_rows_read_per_series`)
+	rowsReadPerQuery             = metrics.NewHistogram(`vm_rows_read_per_query`)
+	seriesReadPerQuery           = metrics.NewHistogram(`vm_series_read_per_query`)
+	seriesPrometheusReadPerQuery = metrics.NewHistogram(`vm_promdb_series_read_per_query`)
 )
 
 type packedTimeseries struct {
 	metricName string
 	brs        []blockRef
+	pd         *promData
+}
+
+type promData struct {
+	values     []float64
+	timestamps []int64
 }
 
 type unpackWork struct {
@@ -431,9 +439,21 @@ func (pts *packedTimeseries) Unpack(dst *Result, tbf *tmpBlocksFile, tr storage.
 		putSortBlocksHeap(sbh)
 		return err
 	}
-	dedupInterval := storage.GetDedupInterval()
+	if pts.pd != nil {
+		// Add data from Prometheus to dst.
+		// It usually has smaller timestamps than the data from sbs, so put it first.
+		dst.Values = append(dst.Values, pts.pd.values...)
+		dst.Timestamps = append(dst.Timestamps, pts.pd.timestamps...)
+	}
+	dedupInterval := storage.GetDedupInterval(tr.MinTimestamp)
 	mergeSortBlocks(dst, sbh, dedupInterval)
 	putSortBlocksHeap(sbh)
+	if pts.pd != nil {
+		if !sort.IsSorted(dst) {
+			sort.Sort(dst)
+		}
+		pts.pd = nil
+	}
 	return nil
 }
 
@@ -542,6 +562,27 @@ func (pts *packedTimeseries) unpackTo(dst []*sortBlock, tbf *tmpBlocksFile, tr s
 	return dst, firstErr
 }
 
+// sort.Interface implementation for Result
+
+// Len implements sort.Interface
+func (r *Result) Len() int {
+	return len(r.Timestamps)
+}
+
+// Less implements sort.Interface
+func (r *Result) Less(i, j int) bool {
+	timestamps := r.Timestamps
+	return timestamps[i] < timestamps[j]
+}
+
+// Swap implements sort.Interface
+func (r *Result) Swap(i, j int) {
+	timestamps := r.Timestamps
+	values := r.Values
+	timestamps[i], timestamps[j] = timestamps[j], timestamps[i]
+	values[i], values[j] = values[j], values[i]
+}
+
 func getSortBlock() *sortBlock {
 	v := sbPool.Get()
 	if v == nil {
@@ -768,6 +809,15 @@ func LabelNames(qt *querytracer.Tracer, sq *storage.SearchQuery, maxLabelNames i
 	if err != nil {
 		return nil, fmt.Errorf("error during labels search on time range: %w", err)
 	}
+
+	// Merge labels obtained from Prometheus storage.
+	promLabels, err := promdb.GetLabelNamesOnTimeRange(sq.GetTimeRange(), deadline)
+	if err != nil {
+		return nil, fmt.Errorf("cannot obtain labels from Prometheus storage: %w", err)
+	}
+	qt.Printf("get %d label names from Prometheus storage", len(promLabels))
+	labels = mergeStrings(labels, promLabels)
+
 	// Sort labels like Prometheus does
 	sort.Strings(labels)
 	qt.Printf("sort %d labels", len(labels))
@@ -826,14 +876,44 @@ func LabelValues(qt *querytracer.Tracer, labelName string, sq *storage.SearchQue
 	}
 	labelValues, err := vmstorage.VMSelectAPI.LabelValues(qt, sq, labelName, maxLabelValues, deadline.Deadline())
 	if err != nil {
-		return nil, fmt.Errorf("error during label values search on time range for labelName=%q: %w", labelName, err)
+		return nil, fmt.Errorf("error during label values search on time range: %w", err)
 	}
+
+	// Merge label values obtained from Prometheus storage.
+	promLabelValues, err := promdb.GetLabelValuesOnTimeRange(labelName, sq.GetTimeRange(), deadline)
+	if err != nil {
+		return nil, fmt.Errorf("cannot obtain label values on time range for %q from Prometheus storage: %w", labelName, err)
+	}
+	qt.Printf("get %d label values from Prometheus storage", len(promLabelValues))
+	labelValues = mergeStrings(labelValues, promLabelValues)
+
 	// Sort labelValues like Prometheus does
 	sort.Strings(labelValues)
 	qt.Printf("sort %d label values", len(labelValues))
 	return labelValues, nil
 }
 
+func mergeStrings(a, b []string) []string {
+	if len(a) == 0 {
+		return b
+	}
+	if len(b) == 0 {
+		return a
+	}
+	m := make(map[string]struct{}, len(a)+len(b))
+	for _, s := range a {
+		m[s] = struct{}{}
+	}
+	for _, s := range b {
+		m[s] = struct{}{}
+	}
+	result := make([]string, 0, len(m))
+	for s := range m {
+		result = append(result, s)
+	}
+	return result
+}
+
 // GetMetricsMetadata returns time series metric names metadata for the given args
 func GetMetricsMetadata(qt *querytracer.Tracer, limit int, metricName string) ([]*metricsmetadata.Row, error) {
 	qt = qt.NewChild("get metrics metadata: limit=%d, metric_name=%q", limit, metricName)
@@ -1193,6 +1273,26 @@ func ProcessSearchQuery(qt *querytracer.Tracer, sq *storage.SearchQuery, deadlin
 	}
 	qt.Printf("fetch unique series=%d, blocks=%d, samples=%d, bytes=%d", len(m), blocksRead, samples, tbf.Len())
 
+	// Fetch data from promdb.
+	pm := make(map[string]*promData)
+	err = promdb.VisitSeries(sq, deadline, func(metricName []byte, values []float64, timestamps []int64) {
+		pd := pm[string(metricName)]
+		if pd == nil {
+			if _, ok := m[string(metricName)]; !ok {
+				orderedMetricNames = append(orderedMetricNames, string(metricName))
+			}
+			pd = &promData{}
+			pm[string(metricName)] = pd
+		}
+		pd.values = append(pd.values, values...)
+		pd.timestamps = append(pd.timestamps, timestamps...)
+	})
+	if err != nil {
+		putTmpBlocksFile(tbf)
+		vmstorage.PutSearch(sr)
+		return nil, fmt.Errorf("error when searching in Prometheus data: %w", err)
+	}
+	seriesPrometheusReadPerQuery.Update(float64(len(pm)))
 	var rss Results
 	rss.tr = sq.GetTimeRange()
 	rss.deadline = deadline
@@ -1201,6 +1301,7 @@ func ProcessSearchQuery(qt *querytracer.Tracer, sq *storage.SearchQuery, deadlin
 		pts[i] = packedTimeseries{
 			metricName: metricName,
 			brs:        brssPool[m[metricName]].brs,
+			pd:         pm[metricName],
 		}
 	}
 	rss.packedTimeseries = pts

app/vmselect/vmui/index.html

@@ -37,7 +37,7 @@
   <meta property="og:title" content="UI for VictoriaMetrics">
   <meta property="og:url" content="https://victoriametrics.com/">
   <meta property="og:description" content="Explore and troubleshoot your VictoriaMetrics data">
-  <script type="module" crossorigin src="./assets/index-D5egN2id.js"></script>
+  <script type="module" crossorigin src="./assets/index-B1dXK3k7.js"></script>
   <link rel="modulepreload" crossorigin href="./assets/rolldown-runtime-CNC7AqOf.js">
   <link rel="modulepreload" crossorigin href="./assets/vendor-DwJYpOdw.js">
   <link rel="stylesheet" crossorigin href="./assets/vendor-CnsZ1jie.css">

app/vmstorage/main.go

@@ -10,6 +10,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage/promdb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
@@ -112,6 +113,9 @@ var (
 
 	metadataStorageSize = flagutil.NewBytes("storage.maxMetadataStorageSize", 0, "Overrides max size for metrics metadata entries in-memory storage. "+
 		"If set to 0 or a negative value, defaults to 1% of allowed memory.")
+
+	downsamplingPeriods = flagutil.NewArrayString("downsampling.period", "Comma-separated downsampling periods in the format 'offset:period'. For example, '30d:10m' instructs "+
+		"to leave a single sample per 10 minutes for samples older than 30 days. See https://docs.victoriametrics.com/#downsampling for details")
 )
 
 func DataPath() string {
@@ -120,7 +124,12 @@ func DataPath() string {
 
 // Init initializes vmstorage.
 func Init(vmselectMaxConcurrentRequests int, vmselectMaxQueueDuration time.Duration, resetCacheIfNeeded func(mrs []storage.MetricRow)) {
-	storage.SetDedupInterval(*minScrapeInterval)
+
+	err := storage.SetDownsamplingPeriods(*downsamplingPeriods, *minScrapeInterval)
+	if err != nil {
+		logger.Fatalf("cannot parse -downsampling.period: %s", err)
+	}
+
 	storage.SetDataFlushInterval(*inmemoryDataFlushInterval)
 	storage.LegacySetRetentionTimezoneOffset(*retentionTimezoneOffset)
 	storage.SetFreeDiskSpaceLimit(minFreeDiskSpaceBytes.N)
@@ -176,6 +185,8 @@ func Init(vmselectMaxConcurrentRequests int, vmselectMaxQueueDuration time.Durat
 	logger.Infof("successfully opened storage %q in %.3f seconds; partsCount: %d; blocksCount: %d; rowsCount: %d; sizeBytes: %d",
 		*storageDataPath, time.Since(startTime).Seconds(), partsCount, blocksCount, rowsCount, sizeBytes)
 
+	promdb.Init(retentionPeriod.Milliseconds())
+
 	// register storage metrics
 	storageMetrics = metrics.NewSet()
 	storageMetrics.RegisterMetricsWriter(vmStorage.writeStorageMetrics)
@@ -230,6 +241,7 @@ func Stop() {
 
 	logger.Infof("gracefully closing the storage at %s", *storageDataPath)
 	startTime := time.Now()
+	promdb.MustClose()
 
 	if vmselectSrv != nil {
 		vmselectSrv.MustStop()

app/vmstorage/promdb/promdb.go

@@ -0,0 +1,268 @@
+package promdb
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"log/slog"
+	"time"
+
+	"github.com/oklog/ulid/v2"
+	"github.com/prometheus/prometheus/model/labels"
+	promstorage "github.com/prometheus/prometheus/storage"
+	"github.com/prometheus/prometheus/tsdb"
+	"github.com/prometheus/prometheus/tsdb/chunkenc"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/searchutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+)
+
+var prometheusDataPath = flag.String("prometheusDataPath", "", "Optional path to readonly historical Prometheus data")
+
+var prometheusRetentionMsecs int64
+
+// Init must be called after flag.Parse and before using the package.
+//
+// See also MustClose.
+func Init(retentionMsecs int64) {
+	if promDB != nil {
+		logger.Fatalf("BUG: promdb.Init is called multiple times without promdb.MustClose call")
+	}
+	prometheusRetentionMsecs = retentionMsecs
+	if *prometheusDataPath == "" {
+		return
+	}
+	l := slog.New(slog.Default().Handler())
+	opts := tsdb.DefaultOptions()
+	opts.RetentionDuration = retentionMsecs
+
+	// Set max block duration to 10% of retention period or 31 days
+	// according to https://prometheus.io/docs/prometheus/latest/storage/#compaction
+	maxBlockDuration := int64((31 * 24 * time.Hour) / time.Millisecond)
+	if maxBlockDuration > retentionMsecs/10 {
+		maxBlockDuration = retentionMsecs / 10
+	}
+	if maxBlockDuration < opts.MinBlockDuration {
+		maxBlockDuration = opts.MinBlockDuration
+	}
+	opts.MaxBlockDuration = maxBlockDuration
+
+	// Custom delete function is needed, because Prometheus by default doesn't delete
+	// blocks outside the retention if no new blocks are created with samples with the current timestamps.
+	// See https://github.com/prometheus/prometheus/blob/997bb7134fcfd7279f250e183e78681e48a56aff/tsdb/db.go#L1116
+	opts.BlocksToDelete = func(blocks []*tsdb.Block) map[ulid.ULID]struct{} {
+		m := make(map[ulid.ULID]struct{})
+		minRetentionTime := time.Now().Unix()*1000 - retentionMsecs
+		for _, block := range blocks {
+			meta := block.Meta()
+			// delete block marked for deletion by compaction code.
+			if meta.Compaction.Deletable {
+				m[meta.ULID] = struct{}{}
+				continue
+			}
+			if block.MaxTime() < minRetentionTime {
+				m[meta.ULID] = struct{}{}
+			}
+		}
+		return m
+	}
+	pdb, err := tsdb.Open(*prometheusDataPath, l, nil, opts, nil)
+	if err != nil {
+		logger.Panicf("FATAL: cannot open Prometheus data at -prometheusDataPath=%q: %s", *prometheusDataPath, err)
+	}
+	promDB = pdb
+	logger.Infof("successfully opened historical Prometheus data at -prometheusDataPath=%q with retentionMsecs=%d", *prometheusDataPath, retentionMsecs)
+}
+
+// MustClose must be called on graceful shutdown.
+//
+// Package functionality cannot be used after this call.
+func MustClose() {
+	if *prometheusDataPath == "" {
+		return
+	}
+	if promDB == nil {
+		logger.Panicf("BUG: promdb.MustClose is called without promdb.Init call")
+	}
+	if err := promDB.Close(); err != nil {
+		logger.Panicf("FATAL: cannot close promDB: %s", err)
+	}
+	promDB = nil
+	logger.Infof("successfully closed historical Prometheus data at -prometheusDataPath=%q", *prometheusDataPath)
+}
+
+var promDB *tsdb.DB
+
+// GetLabelNamesOnTimeRange returns label names.
+func GetLabelNamesOnTimeRange(tr storage.TimeRange, deadline searchutil.Deadline) ([]string, error) {
+	if *prometheusDataPath == "" {
+		return nil, nil
+	}
+	d := time.Unix(int64(deadline.Deadline()), 0)
+	ctx, cancel := context.WithDeadline(context.Background(), d)
+	defer cancel()
+	q, err := promDB.Querier(tr.MinTimestamp, tr.MaxTimestamp)
+	if err != nil {
+		return nil, err
+	}
+	defer mustCloseQuerier(q)
+
+	names, _, err := q.LabelNames(ctx, nil)
+	// Make full copy of names, since they cannot be used after q is closed.
+	names = copyStringsWithMemory(names)
+	return names, err
+}
+
+// GetLabelValuesOnTimeRange returns values for the given labelName on the given tr.
+func GetLabelValuesOnTimeRange(labelName string, tr storage.TimeRange, deadline searchutil.Deadline) ([]string, error) {
+	if *prometheusDataPath == "" {
+		return nil, nil
+	}
+	d := time.Unix(int64(deadline.Deadline()), 0)
+	ctx, cancel := context.WithDeadline(context.Background(), d)
+	defer cancel()
+	q, err := promDB.Querier(tr.MinTimestamp, tr.MaxTimestamp)
+	if err != nil {
+		return nil, err
+	}
+	defer mustCloseQuerier(q)
+
+	values, _, err := q.LabelValues(ctx, labelName, nil)
+	// Make full copy of values, since they cannot be used after q is closed.
+	values = copyStringsWithMemory(values)
+	return values, err
+}
+
+func copyStringsWithMemory(a []string) []string {
+	result := make([]string, len(a))
+	for i, s := range a {
+		result[i] = string(append([]byte{}, s...))
+	}
+	return result
+}
+
+// SeriesVisitor is called by VisitSeries for each matching time series.
+//
+// The caller shouldn't hold references to metricName, values and timestamps after returning.
+type SeriesVisitor func(metricName []byte, values []float64, timestamps []int64)
+
+// VisitSeries calls f for each series found in the pdb.
+func VisitSeries(sq *storage.SearchQuery, deadline searchutil.Deadline, f SeriesVisitor) error {
+	if *prometheusDataPath == "" {
+		return nil
+	}
+	d := time.Unix(int64(deadline.Deadline()), 0)
+	ctx, cancel := context.WithDeadline(context.Background(), d)
+	defer cancel()
+	minTime, maxTime := getSearchTimeRange(sq)
+	q, err := promDB.Querier(minTime, maxTime)
+	if err != nil {
+		return err
+	}
+	defer mustCloseQuerier(q)
+	var seriesSet []promstorage.SeriesSet
+	for _, tf := range sq.TagFilterss {
+		ms, err := convertTagFiltersToMatchers(tf)
+		if err != nil {
+			return fmt.Errorf("cannot convert tag filters to matchers: %w", err)
+		}
+		s := q.Select(ctx, false, nil, ms...)
+		seriesSet = append(seriesSet, s)
+	}
+	ss := promstorage.NewMergeSeriesSet(seriesSet, 0, promstorage.ChainedSeriesMerge)
+	var (
+		mn         storage.MetricName
+		metricName []byte
+		values     []float64
+		timestamps []int64
+	)
+	var it chunkenc.Iterator
+	for ss.Next() {
+		s := ss.At()
+		convertPromLabelsToMetricName(&mn, s.Labels())
+		metricName = mn.SortAndMarshal(metricName[:0])
+		values = values[:0]
+		timestamps = timestamps[:0]
+		it = s.Iterator(it)
+		for {
+			typ := it.Next()
+			if typ == chunkenc.ValNone {
+				break
+			}
+			if typ != chunkenc.ValFloat {
+				// Skip unsupported values
+				continue
+			}
+			ts, v := it.At()
+			values = append(values, v)
+			timestamps = append(timestamps, ts)
+		}
+		if err := it.Err(); err != nil {
+			return fmt.Errorf("error when iterating Prometheus series: %w", err)
+		}
+		f(metricName, values, timestamps)
+	}
+	return ss.Err()
+}
+
+func getSearchTimeRange(sq *storage.SearchQuery) (int64, int64) {
+	maxTime := sq.MaxTimestamp
+	minTime := sq.MinTimestamp
+	minRetentionTime := time.Now().Unix()*1000 - prometheusRetentionMsecs
+	if maxTime < minRetentionTime {
+		maxTime = minRetentionTime
+	}
+	if minTime < minRetentionTime {
+		minTime = minRetentionTime
+	}
+	return minTime, maxTime
+}
+
+func convertPromLabelsToMetricName(dst *storage.MetricName, ls labels.Labels) {
+	dst.Reset()
+	ls.Range(func(label labels.Label) {
+		if label.Name == "__name__" {
+			dst.MetricGroup = append(dst.MetricGroup[:0], label.Value...)
+		} else {
+			dst.AddTag(label.Name, label.Value)
+		}
+	})
+}
+
+func convertTagFiltersToMatchers(tfs []storage.TagFilter) ([]*labels.Matcher, error) {
+	ms := make([]*labels.Matcher, 0, len(tfs))
+	for _, tf := range tfs {
+		var mt labels.MatchType
+		if tf.IsNegative {
+			if tf.IsRegexp {
+				mt = labels.MatchNotRegexp
+			} else {
+				mt = labels.MatchNotEqual
+			}
+		} else {
+			if tf.IsRegexp {
+				mt = labels.MatchRegexp
+			} else {
+				mt = labels.MatchEqual
+			}
+		}
+		key := string(tf.Key)
+		if key == "" {
+			key = "__name__"
+		}
+		value := string(tf.Value)
+		m, err := labels.NewMatcher(mt, key, value)
+		if err != nil {
+			return nil, err
+		}
+		ms = append(ms, m)
+	}
+	return ms, nil
+}
+
+func mustCloseQuerier(q promstorage.Querier) {
+	if err := q.Close(); err != nil {
+		logger.Panicf("FATAL: cannot close querier: %s", err)
+	}
+}

apptest/tests/per_day_index_test.go

@@ -26,33 +26,20 @@ func TestClusterSearchWithDisabledPerDayIndex(t *testing.T) {
 	defer tc.Stop()
 
 	testSearchWithDisabledPerDayIndex(tc, func(name string, disablePerDayIndex bool) apptest.PrometheusWriteQuerier {
-		// Using static ports for vmstorage because random ports may cause
-		// changes in how data is sharded.
-		vmstorage1 := tc.MustStartVmstorage("vmstorage1-"+name, []string{
-			"-storageDataPath=" + tc.Dir() + "/vmstorage1",
+		vmstorage := tc.MustStartVmstorage("vmstorage-"+name, []string{
+			"-storageDataPath=" + tc.Dir() + "/vmstorage",
 			"-retentionPeriod=100y",
-			"-httpListenAddr=127.0.0.1:61001",
-			"-vminsertAddr=127.0.0.1:61002",
-			"-vmselectAddr=127.0.0.1:61003",
-			fmt.Sprintf("-disablePerDayIndex=%t", disablePerDayIndex),
-		})
-		vmstorage2 := tc.MustStartVmstorage("vmstorage2-"+name, []string{
-			"-storageDataPath=" + tc.Dir() + "/vmstorage2",
-			"-retentionPeriod=100y",
-			"-httpListenAddr=127.0.0.1:62001",
-			"-vminsertAddr=127.0.0.1:62002",
-			"-vmselectAddr=127.0.0.1:62003",
 			fmt.Sprintf("-disablePerDayIndex=%t", disablePerDayIndex),
 		})
 		vminsert := tc.MustStartVminsert("vminsert-"+name, []string{
-			"-storageNode=" + vmstorage1.VminsertAddr() + "," + vmstorage2.VminsertAddr(),
+			"-storageNode=" + vmstorage.VminsertAddr(),
 		})
 		vmselect := tc.MustStartVmselect("vmselect"+name, []string{
-			"-storageNode=" + vmstorage1.VmselectAddr() + "," + vmstorage2.VmselectAddr(),
+			"-storageNode=" + vmstorage.VmselectAddr(),
 			"-search.maxStalenessInterval=1m",
 		})
 		return &apptest.Vmcluster{
-			Vmstorages: []*apptest.Vmstorage{vmstorage1, vmstorage2},
+			Vmstorages: []*apptest.Vmstorage{vmstorage},
 			Vminsert:   vminsert,
 			Vmselect:   vmselect,
 		}

apptest/tests/prometheus_mock_storage.go

@@ -59,7 +59,7 @@ func (ms *PrometheusMockStorage) Read(_ context.Context, query *prompb.Query, so
 		}
 
 		if !notMatch {
-			q.Timeseries = append(q.Timeseries, &prompb.TimeSeries{Labels: s.Labels, Samples: s.Samples})
+			q.Timeseries = append(q.Timeseries, &prompb.TimeSeries{Labels: s.Labels, Samples: s.Samples, Histograms: s.Histograms})
 		}
 	}
 

apptest/tests/remote_read_server.go

@@ -12,6 +12,7 @@ import (
 
 	"github.com/gogo/protobuf/proto"
 	"github.com/golang/snappy"
+	"github.com/prometheus/prometheus/model/histogram"
 	"github.com/prometheus/prometheus/model/labels"
 	"github.com/prometheus/prometheus/prompb"
 	"github.com/prometheus/prometheus/storage/remote"
@@ -86,10 +87,17 @@ func (rrs *RemoteReadServer) getReadHandler(t *testing.T) http.Handler {
 						samples = append(samples, sample)
 					}
 				}
+				var histograms []prompb.Histogram
+				for _, h := range s.Histograms {
+					if h.Timestamp >= startTs && h.Timestamp < endTs {
+						histograms = append(histograms, h)
+					}
+				}
 				var series prompb.TimeSeries
-				if len(samples) > 0 {
+				if len(samples) > 0 || len(histograms) > 0 {
 					series.Labels = s.Labels
 					series.Samples = samples
+					series.Histograms = histograms
 				}
 				ts[i] = &series
 			}
@@ -317,6 +325,37 @@ func generateRemoteReadSamples(idx int, startTime, endTime, numOfSamples int64)
 	return samples
 }
 
+// GenerateRemoteReadHistogramSeries generates a remote read series
+// with native histogram samples within the given time range.
+func GenerateRemoteReadHistogramSeries(start, end, numOfSamples int64) []*prompb.TimeSeries {
+	timeSeries := &prompb.TimeSeries{
+		Labels: []prompb.Label{
+			{Name: labels.MetricName, Value: "vm_histogram_metric"},
+			{Name: "job", Value: "0"},
+		},
+	}
+
+	delta := (end - start) / numOfSamples
+	mul := int64(0)
+	for t := start; t != end; t += delta {
+		mul++
+		h := &histogram.Histogram{
+			Schema:          0,
+			Count:           uint64(10 * mul),
+			Sum:             25.5 * float64(mul),
+			ZeroThreshold:   0.001,
+			ZeroCount:       uint64(2 * mul),
+			PositiveSpans:   []histogram.Span{{Offset: 0, Length: 2}},
+			PositiveBuckets: []int64{mul, 2 * mul},
+			NegativeSpans:   []histogram.Span{{Offset: 0, Length: 1}},
+			NegativeBuckets: []int64{4 * mul},
+		}
+		timeSeries.Histograms = append(timeSeries.Histograms, prompb.FromIntHistogram(t*1000, h))
+	}
+
+	return []*prompb.TimeSeries{timeSeries}
+}
+
 func labelsToLabelsProto(ls labels.Labels) []prompb.Label {
 	result := make([]prompb.Label, 0, ls.Len())
 	ls.Range(func(l labels.Label) {

apptest/tests/vmctl_remote_read_mogration_test.go

@@ -75,6 +75,118 @@ func TestClusterVmctlRemoteReadProtocol(t *testing.T) {
 	testRemoteReadProtocol(tc, clusterDst, newRemoteReadServer, vmctlFlags)
 }
 
+func TestSingleVmctlRemoteReadNativeHistograms(t *testing.T) {
+	fs.MustRemoveDir(t.Name())
+
+	tc := apptest.NewTestCase(t)
+	defer tc.Stop()
+
+	vmsingleDst := tc.MustStartDefaultVmsingle()
+	vmAddr := fmt.Sprintf("http://%s/", vmsingleDst.HTTPAddr())
+	vmctlFlags := []string{
+		`remote-read`,
+		`--remote-read-filter-time-start=2025-06-11T15:31:10Z`,
+		`--remote-read-filter-time-end=2025-06-11T15:31:20Z`,
+		`--remote-read-step-interval=minute`,
+		`--vm-addr=` + vmAddr,
+		`--disable-progress-bar=true`,
+	}
+
+	testRemoteReadNativeHistograms(tc, vmsingleDst, NewRemoteReadServer, vmctlFlags)
+}
+
+func TestSingleVmctlRemoteReadStreamNativeHistograms(t *testing.T) {
+	fs.MustRemoveDir(t.Name())
+
+	tc := apptest.NewTestCase(t)
+	defer tc.Stop()
+
+	vmsingleDst := tc.MustStartDefaultVmsingle()
+	vmAddr := fmt.Sprintf("http://%s/", vmsingleDst.HTTPAddr())
+	vmctlFlags := []string{
+		`remote-read`,
+		`--remote-read-filter-time-start=2025-06-11T15:31:10Z`,
+		`--remote-read-filter-time-end=2025-06-11T15:31:20Z`,
+		`--remote-read-step-interval=minute`,
+		`--vm-addr=` + vmAddr,
+		`--remote-read-use-stream=true`,
+		`--disable-progress-bar=true`,
+	}
+
+	testRemoteReadNativeHistograms(tc, vmsingleDst, NewRemoteReadStreamServer, vmctlFlags)
+}
+
+// testRemoteReadNativeHistograms verifies that native histograms are migrated
+// as _count, _sum and _bucket series with vmrange labels in the same way
+// as VictoriaMetrics converts native histograms received via Prometheus remote write protocol.
+func testRemoteReadNativeHistograms(tc *apptest.TestCase, sut apptest.PrometheusWriteQuerier, newRemoteReadServer func(t *testing.T, series []*prompb.TimeSeries) *RemoteReadServer, vmctlFlags []string) {
+	t := tc.T()
+	t.Helper()
+
+	series := GenerateRemoteReadHistogramSeries(1749655870, 1749655880, 2)
+
+	rrs := newRemoteReadServer(t, series)
+	defer rrs.Close()
+
+	vmctlFlags = append(vmctlFlags, `--remote-read-src-addr=`+rrs.HTTPAddr())
+	tc.MustStartVmctl("vmctl", vmctlFlags)
+
+	sut.ForceFlush(t)
+
+	tc.Assert(&apptest.AssertOptions{
+		Retries: 300,
+		Msg:     `unexpected native histogram metrics stored on vmsingle via the prometheus protocol`,
+		Got: func() any {
+			got := sut.PrometheusAPIV1Export(t, `{__name__=~".*"}`, apptest.QueryOpts{
+				Start: "2025-06-11T15:31:10Z",
+				End:   "2025-06-11T15:32:20Z",
+			})
+			got.Sort()
+			return got.Data.Result
+		},
+		Want: expectedNativeHistogramQueryResult(),
+		CmpOpts: []cmp.Option{
+			cmpopts.IgnoreFields(apptest.PrometheusAPIV1QueryResponse{}, "Status", "Data.ResultType"),
+		},
+	})
+}
+
+// expectedNativeHistogramQueryResult returns the series expected to be stored in VictoriaMetrics
+// after migrating the series generated by GenerateRemoteReadHistogramSeries(1749655870, 1749655880, 2).
+func expectedNativeHistogramQueryResult() []*apptest.QueryResult {
+	metric := func(name, vmrange string) map[string]string {
+		m := map[string]string{
+			"__name__": name,
+			"job":      "0",
+		}
+		if vmrange != "" {
+			m["vmrange"] = vmrange
+		}
+		return m
+	}
+	samples := func(v1, v2 float64) []*apptest.Sample {
+		return []*apptest.Sample{
+			{Timestamp: 1749655870000, Value: v1},
+			{Timestamp: 1749655875000, Value: v2},
+		}
+	}
+	resp := &apptest.PrometheusAPIV1QueryResponse{
+		Data: &apptest.QueryData{
+			Result: []*apptest.QueryResult{
+				{Metric: metric("vm_histogram_metric_count", ""), Samples: samples(10, 20)},
+				{Metric: metric("vm_histogram_metric_sum", ""), Samples: samples(25.5, 51)},
+				{Metric: metric("vm_histogram_metric_bucket", "-1.000e+00...-5.000e-01"), Samples: samples(4, 8)},
+				{Metric: metric("vm_histogram_metric_bucket", "-1.000e-03...1.000e-03"), Samples: samples(2, 4)},
+				{Metric: metric("vm_histogram_metric_bucket", "5.000e-01...1.000e+00"), Samples: samples(1, 2)},
+				{Metric: metric("vm_histogram_metric_bucket", "1.000e+00...2.000e+00"), Samples: samples(3, 6)},
+			},
+		},
+	}
+	// sort in the same way as the exported result
+	resp.Sort()
+	return resp.Data.Result
+}
+
 func testRemoteReadProtocol(tc *apptest.TestCase, sut apptest.PrometheusWriteQuerier, newRemoteReadServer func(t *testing.T) *RemoteReadServer, vmctlFlags []string) {
 	t := tc.T()
 	t.Helper()

docs/victoriametrics/Cluster-VictoriaMetrics.md

@@ -829,7 +829,7 @@ See also [minimum downtime strategy](#minimum-downtime-strategy).
 
 ## Slowness-based re-routing
 
-By default{{% available_from "#" %}}, `vminsert` automatically re-routes writes away from the slowest `vmstorage` node
+By default{{% available_from "v1.149.0" %}}, `vminsert` automatically re-routes writes away from the slowest `vmstorage` node
 to preserve maximum ingestion throughput. This prevents a single slow `vmstorage` node
 from throttling the entire cluster.
 
@@ -843,7 +843,7 @@ Disable slowness-based re-routing with `-disableRerouting=true` when keeping met
 perfectly balanced across nodes or minimizing the number of [active time series](https://docs.victoriametrics.com/victoriametrics/faq/#what-is-an-active-time-series) 
 matters more than peak write throughput.
 
-Slowness-based re-routing is automatically disabled{{% available_from "#" %}} when `-replicationFactor` is greater than `1`,
+Slowness-based re-routing is automatically disabled{{% available_from "v1.149.0" %}} when `-replicationFactor` is greater than `1`,
 because rerouting does not guarantee that replicated copies land on distinct storage nodes,
 which violates the replication contract.
 

docs/victoriametrics/README.md

@@ -39,6 +39,8 @@ VictoriaMetrics has the following prominent features:
   * Easy and fast backups from [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
     can be done with [vmbackup](https://docs.victoriametrics.com/victoriametrics/vmbackup/) / [vmrestore](https://docs.victoriametrics.com/victoriametrics/vmrestore/) tools.
     See [this article](https://medium.com/@valyala/speeding-up-backups-for-big-time-series-databases-533c1a927883) for more details.
+* It supports storage and retrieval of samples with timestamps that fall within the `[1970-01-02T00:00:00.000Z, 2262-03-31T23:59:59.999Z]` time range with millisecond precision.
+  See [Retention](#retention) for details.
 * It implements a PromQL-like query language - [MetricsQL](https://docs.victoriametrics.com/victoriametrics/metricsql/), which provides improved functionality on top of PromQL.
 * It provides a global query view. Multiple Prometheus instances or any other data sources may ingest data into VictoriaMetrics. Later this data may be queried via a single query.
 * It provides high performance and good vertical and horizontal scalability for both
@@ -1540,6 +1542,9 @@ It is safe to extend `-retentionPeriod` on existing data. If `-retentionPeriod`
 value than before, then data outside the configured period will be eventually deleted.
 
 VictoriaMetrics does not support indefinite retention, but you can specify an arbitrarily high duration, e.g. `-retentionPeriod=100y`.
+Just keep in mind that VictoriaMetrics does not support samples with negative timestamps. Timestamps at `1970-01-01` are also not
+supported because this date has a special meaning internally. It therefore rejects samples with timestamps before
+`1970-01-02T00:00:00.000Z`.
 
 By default, VictoriaMetrics doesn't accept samples with timestamps bigger than `now+2d`, e.g. 2 days in the future.
 If you need accepting samples with bigger timestamps, then specify the desired "future retention" via `-futureRetention` command-line flag.
@@ -1551,6 +1556,9 @@ For example, the following command starts VictoriaMetrics, which accepts samples
 /path/to/victoria-metrics -futureRetention=1y
 ```
 
+VictoriaMetrics does not support stamples after `2262-03-31T23:59:59.999Z`. If the future retention includes dates after this timestamp,
+the samples for those dates will be rejected.
+
 By default, VictoriaMetrics accepts samples with timestamps as old as the configured `-retentionPeriod` allows, e.g. it accepts backfilled
 historical data as long as it fits into the retention. If you need rejecting samples with historical timestamps older than the specified
 duration, then specify the desired duration via the `-maxBackfillAge` command-line flag. This can be useful for limiting ingestion of

docs/victoriametrics/changelog/CHANGELOG.md

@@ -26,22 +26,32 @@ See also [LTS releases](https://docs.victoriametrics.com/victoriametrics/lts-rel
 
 ## tip
 
+## [v1.149.0](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/tag/v1.149.0)
+
+Release candidate
+
 **Update Note 1:** `vminsert` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/): the default value of `-disableRerouting` flag has changed from `true` to `false`, enabling [slowness-based re-routing](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/#slowness-based-re-routing) by default. Slowness re-routing is automatically disabled when `-replicationFactor` is greater than 1. If you rely on the old behavior, pass `-disableRerouting` command-line flag to `vminsert`. See [#11287](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11287).
 
 **Update Note 2:** [vmsingle](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/) and `vmselect` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/): the `/api/v1/admin/tsdb/delete_series`, `/tags/delSeries` endpoints now require `POST` method. Previously, it also accepted `GET` requests. If you use `GET` requests for this endpoint, update your scripts or tooling to use `POST` instead. See [#5552](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5552).
 
 * SECURITY: [vmsingle](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/) and `vmselect` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/): restrict `/api/v1/admin/tsdb/delete_series`, `/tags/delSeries` endpoints to `POST` method only to prevent some [SSRF](https://en.wikipedia.org/wiki/Server-side_request_forgery)-based data deletion attacks. See [#5552](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5552).
 
+* FEATURE: [dashboards](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/master/dashboards): add `Fsync avg duration` panel to the Troubleshooting section of the single-node, cluster, and vmagent dashboards. This panel surfaces degradation of IO operation for faster incident triage. See [#10432](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10432).
 * FEATURE: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): add `name` label identifying the corresponding `-remoteWrite.url` target to the `vm_persistentqueue_*` metrics exposed by persistent queue. See [#7944](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7944). Thanks to @tIGO for contribution.
+* FEATURE: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): add `-remoteWrite.obfuscateLabels` flag for hashing values of the specified labels before sending metrics to the corresponding `-remoteWrite.url`. This allows sharing metrics with external systems while keeping sensitive label values hidden. See [#10599](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10599).
+* FEATURE: [vmalert](https://docs.victoriametrics.com/victoriametrics/vmalert/): add `-replay.continueWithExecutionErr` flag to allow continuing to replay other rules when a rule execution fails with a 422 response code, which can happen due to an expression syntax error or a resource limit being hit. See [11313](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11313).
+* FEATURE: [vmalert](https://docs.victoriametrics.com/victoriametrics/vmalert/): add template variable `$interval` to expose the alerting rule group's evaluation interval. This allows generating dashboard links with a lookback window relative to the rule's interval, for example `&from={{ ($activeAt.Add (parseDurationTime (printf "-%s" .Interval))).UnixMilli }}&to={{ $activeAt.UnixMilli }}`. See [#11232](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11232). Thanks to @1solomonwakhungu for contribution.
+* FEATURE: [vmbackupmanager](https://docs.victoriametrics.com/victoriametrics/vmbackupmanager/) and [alerts](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/rules): introduce `vm_backup_last_success_at` metric to track the last successful backup by type. Add [alerting rules](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/rules/alerts-vmbackupmanager.yml) `NoLatestBackupWithinLastDay`, `NoHourlyBackupWithinLastDay`, `NoDailyBackupWithinLast3Days`, `NoWeeklyBackupWithinLast14Days` and `NoMonthlyBackupWithinLast62Days` to remind users about the missing backups. See [#11217](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11217).
+* FEATURE: [vmctl](https://docs.victoriametrics.com/victoriametrics/vmctl/): support [Prometheus native histograms](https://prometheus.io/docs/specs/native_histograms/) migration in [remote read mode](https://docs.victoriametrics.com/victoriametrics/vmctl/remoteread/). Native histograms are converted into `_count`, `_sum` and `_bucket` series with `vmrange` labels in the same way as VictoriaMetrics [converts native histograms received via Prometheus remote write protocol](https://docs.victoriametrics.com/victoriametrics/integrations/prometheus/#native-histograms), except that for native histograms with custom buckets the original bucket bounds are preserved instead of being estimated with the exponential formula. Previously native histograms were silently ignored in `SAMPLES` mode, while in stream mode the migration failed with `EOF` error. See [#11292](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11292). Thanks to @liuxu623 for contribution.
 * FEATURE: `vminsert` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/): enable [slowness-based re-routing](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/#slowness-based-re-routing) by default. Previously, `-disableRerouting` defaulted to `true`, which limited ingestion throughput to the slowest `vmstorage` node. Now `-disableRerouting` defaults to `false`, so `vminsert` automatically routes data away from the slowest `vmstorage` node, improving overall ingestion performance. Slowness re-routing is automatically disabled when `-replicationFactor` is greater than 1. See [#11287](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11287).
 * FEATURE: [vmui](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/#vmui): persist the selected auto-refresh interval in the URL. See [VictoriaLogs#1310](https://github.com/VictoriaMetrics/VictoriaLogs/issues/1310).
-* FEATURE: [dashboards](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/master/dashboards): add `Fsync avg duration` panel to the Troubleshooting section of the single-node, cluster, and vmagent dashboards. This panel surfaces degradation of IO operation for faster incident triage. See [#10432](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10432).
-* FEATURE: [vmbackupmanager](https://docs.victoriametrics.com/victoriametrics/vmbackupmanager/) and [alerts](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/rules): introduce `vm_backup_last_success_at` metric to track the last successful backup by type. Add [alerting rules](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/deployment/docker/rules/alerts-vmbackupmanager.yml) `NoLatestBackupWithinLastDay`, `NoHourlyBackupWithinLastDay`, `NoDailyBackupWithinLast3Days`, `NoWeeklyBackupWithinLast14Days` and `NoMonthlyBackupWithinLast62Days` to remind users about the missing backups. See [#11217](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11217).
-* FEATURE: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): add `-remoteWrite.obfuscateLabels` flag for hashing values of the specified labels before sending metrics to the corresponding `-remoteWrite.url`. This allows sharing metrics with external systems while keeping sensitive label values hidden. See [#10599](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10599).
 
 * BUGFIX: [MetricsQL](https://docs.victoriametrics.com/victoriametrics/metricsql/): properly drop data points filtered out by an inner [comparison operation](https://prometheus.io/docs/prometheus/latest/querying/operators/#comparison-binary-operators) when its result is used on the right side of another comparison. Previously, queries like `foo != (bar > 100)` could return unexpected results because filtered-out data points are represented internally as `NaN`, and `value != NaN` evaluates to `true`. Comparisons against explicitly present `NaN` values keep the previous behavior. See [#10018](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10018). Thanks to @zasdaym for contribution.
+* BUGFIX: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/) and [vmsingle](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/): ignore HTTP proxy environment variables when scraping targets over Unix domain sockets. See [#11318](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11318). Thanks to @lwmacct for contribution.
 * BUGFIX: [vmalert](https://docs.victoriametrics.com/victoriametrics/vmalert/): fixed the display of rule state badges on the `Groups` page in the web UI. See [#11160](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11160).
 * BUGFIX: [vmbackupmanager](https://docs.victoriametrics.com/victoriametrics/vmbackupmanager/): previously, `vmbackupmanager` was crashing on startup when it failed to restore backup state from remote storage, causing a crash loop. Now it logs the error and continues running, retrying the state restore before each scheduled backup. Added `vm_backup_errors_total{type="restoreState"}` metric to track backup state restore failures. See [#11217](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11217).
+* BUGFIX: [stream aggregation](https://docs.victoriametrics.com/victoriametrics/stream-aggregation/): fix incorrect [sum_samples_total](https://docs.victoriametrics.com/victoriametrics/stream-aggregation/configuration/#sum_samples_total) results when `enable_windows: true` is set. See [#11261](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11261). Thanks to @beyond-infra for contribution.
+* BUGFIX: [vmsingle](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/), `vmselect` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/) and [vmctl](https://docs.victoriametrics.com/victoriametrics/vmctl/): accept scientific notation with sub-second precision (e.g. `1.784144612388E9`) for timestamp args such as `start` and `end` in `/api/v1/query_range` and `--vm-native-filter-time-start` and `--vm-native-filter-time-end` in `vmctl`. Previously, values with this pattern were rejected, which is incompatible with Prometheus. See [#11268](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11268). Thanks to @STiFLeR7 for contribution.
 
 ## [v1.148.0](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/tag/v1.148.0)
 

docs/victoriametrics/vmagent.md

@@ -688,7 +688,7 @@ Extra labels can be added to metrics collected by `vmagent` via the following me
 ## Obfuscating label values
 
 `vmagent` can obfuscate the values of specified labels before sending metrics to `-remoteWrite.url`
-via `-remoteWrite.obfuscateLabels`{{% available_from "#" %}}.
+via `-remoteWrite.obfuscateLabels`{{% available_from "v1.149.0" %}}.
 
 This is useful when one or more `-remoteWrite.url` endpoints point to external monitoring services
 outside the organization, and sensitive label values such as `ip`, `host`, `instance`, or `datacenter`

docs/victoriametrics/vmalert.md

@@ -341,6 +341,7 @@ The following variables are available in templating:
 | $externalLabels or .ExternalLabels | List of labels configured via `-external.label` command-line flag.                                                                                    | `Issues with {{ $labels.instance }} (datacenter-{{ $externalLabels.dc }})`                                                                                                             |
 | $externalURL or .ExternalURL       | URL configured via `-external.url` command-line flag. Used for cases when vmalert is hidden behind proxy.                                             | `Visit {{ $externalURL }} for more details`                                                                                                                                            |
 | $isPartial or .IsPartial           | Indicates whether the latest rule query response from the datasource(that supports returning `isPartial` option, such as vmcluster) could be partial. | `{{ if $isPartial }}WARNING: The latest alert state may be a false alarm due to a partial response from the datasource.{{ end }}`                                                      |
+| $interval or .Interval             | Alerting rule group's evaluation interval.                                                                                                            | `http://vm-grafana.com/<dashboard-id>?viewPanel=<panel-id>&from={{ ($activeAt.Add (parseDurationTime (printf "-%s" .Interval))).UnixMilli }}&to={{ $activeAt.UnixMilli }}` |
 
 Additionally, `vmalert` provides some extra templating functions listed in [template functions](#template-functions) and [reusable templates](#reusable-templates).
 

docs/victoriametrics/vmalert_common_flags.md

@@ -404,6 +404,8 @@ See the docs at https://docs.victoriametrics.com/victoriametrics/vmalert/ .
      Optional TLS server name to use for connections to -remoteWrite.url. By default, the server name from -remoteWrite.url is used
   -remoteWrite.url string
      Optional URL to persist alerts state and recording rules results in form of timeseries. It must support either VictoriaMetrics remote write protocol or Prometheus remote_write protocol. Supports address in the form of IP address with a port (e.g., http://127.0.0.1:8428) or DNS SRV record. For example, if -remoteWrite.url=http://127.0.0.1:8428 is specified, then the alerts state will be written to http://127.0.0.1:8428/api/v1/write . See also -remoteWrite.disablePathAppend, '-remoteWrite.showURL'.
+  -replay.continueWithExecutionErr bool
+     Whether to continue replaying other rules if a rule execution fails with a 422 response code, which can happen due to an expression syntax error or a resource limit being hit.
   -replay.disableProgressBar
      Whether to disable rendering progress bars during the replay. Progress bar rendering might be verbose or break the logs parsing, so it is recommended to be disabled when not used in interactive mode.
   -replay.maxDatapointsPerQuery int
@@ -411,7 +413,7 @@ See the docs at https://docs.victoriametrics.com/victoriametrics/vmalert/ .
   -replay.ruleEvaluationConcurrency int
      The maximum number of concurrent '/query_range' requests when replay recording rule or alerting rule with for=0. Increasing this value when replaying for a long time, since each request is limited by -replay.maxDatapointsPerQuery. (default 1)
   -replay.ruleRetryAttempts int
-     Defines how many retries to make before giving up on rule if request for it returns an error. (default 5)
+     Defines how many retries to make before giving up on rule if request for it returns a retriable error. (default 5)
   -replay.rulesDelay duration
      Delay before evaluating the next rule within the group. Is important for chained rules. Keep it equal or bigger than -remoteWrite.flushInterval. When set to >0, replay ignores group's concurrency setting. (default 1s)
   -replay.timeFrom string

docs/victoriametrics/vmbackupmanager.md

@@ -480,7 +480,7 @@ Clusters here are referred to as `source` and `destination`.
 
 To verify that `vmbackupmanager` is executing backup tasks normally, the following metrics can help:
 
-* `vm_backup_last_success_at{type="<backup_type>"}` - unix timestamp of the last successful backup{{% available_from "#" %}}. Remains `0` if no backup has completed successfully since startup. Check error logs and verify remote storage accessibility if this persists.
+* `vm_backup_last_success_at{type="<backup_type>"}` - unix timestamp of the last successful backup{{% available_from "v1.149.0" %}}. Remains `0` if no backup has completed successfully since startup. Check error logs and verify remote storage accessibility if this persists.
 * `vm_backup_last_run_failed{type="<backup_type>"}` - whether the last backup task for the given backup type failed. The value `1` means the last task failed. Check the error logs of `vmbackupmanager` for the root cause
 * `vm_backup_errors_total{type="<backup_type>"}` - total number of backup errors for the given backup type.
 

lib/promscrape/client.go

@@ -90,7 +90,11 @@ func newClient(ctx context.Context, sw *ScrapeWork) (*client, error) {
 	}
 
 	tr := httputil.NewTransport(false, "vm_promscrape")
-	if proxyURLFunc != nil {
+	if sw.UnixSocket != "" {
+		// Unix sockets are direct local endpoints and cannot be reached via HTTP proxies.
+		// Proxy could be set implicitly via global env variable HTTP_PROXY
+		tr.Proxy = nil
+	} else if proxyURLFunc != nil {
 		tr.Proxy = proxyURLFunc
 	}
 	tr.TLSHandshakeTimeout = 10 * time.Second

lib/storage/block.go

@@ -163,7 +163,8 @@ func (b *Block) deduplicateSamplesDuringMerge() {
 		// Nothing to dedup.
 		return
 	}
-	dedupInterval := GetDedupInterval()
+	maxTimestamp := srcTimestamps[len(srcTimestamps)-1]
+	dedupInterval := GetDedupInterval(maxTimestamp)
 	if dedupInterval <= 0 {
 		// Deduplication is disabled.
 		return

lib/storage/dedup.go

@@ -1,29 +1,11 @@
 package storage
 
 import (
-	"time"
-
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
 )
 
-// SetDedupInterval sets the deduplication interval, which is applied to raw samples during data ingestion and querying.
-//
-// De-duplication is disabled if dedupInterval is 0.
-//
-// This function must be called before initializing the storage.
-func SetDedupInterval(dedupInterval time.Duration) {
-	globalDedupInterval = dedupInterval.Milliseconds()
-}
-
-// GetDedupInterval returns the dedup interval in milliseconds, which has been set via SetDedupInterval.
-func GetDedupInterval() int64 {
-	return globalDedupInterval
-}
-
-var globalDedupInterval int64
-
 func isDedupEnabled() bool {
-	return globalDedupInterval > 0
+	return len(downsamplingPeriods) > 0
 }
 
 // DeduplicateSamples removes samples from src* if they are closer to each other than dedupInterval in milliseconds.

lib/storage/downsampling.go

@@ -0,0 +1,123 @@
+package storage
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
+	"github.com/VictoriaMetrics/metricsql"
+)
+
+// SetDownsamplingPeriods configures downsampling.
+//
+// The function must be called before opening or creating any storage.
+func SetDownsamplingPeriods(periods []string, dedupInterval time.Duration) error {
+	dsps, err := parseDownsamplingPeriods(periods)
+	if err != nil {
+		return err
+	}
+	dedupIntervalMs := dedupInterval.Milliseconds()
+	if dedupIntervalMs > 0 {
+		if len(dsps) > 0 && dsps[len(dsps)-1].Offset == 0 {
+			return fmt.Errorf("-dedup.minScrapeInterval=%s cannot be used if -downsampling.period=%s contains zero offset", dedupInterval, periods)
+		}
+		// Deduplication is a special case of downsampling with zero offset.
+		dsps = append(dsps, DownsamplingPeriod{
+			Offset:   0,
+			Interval: dedupIntervalMs,
+		})
+	}
+	downsamplingPeriods = dsps
+	return nil
+}
+
+// DownsamplingPeriod describes downsampling period
+type DownsamplingPeriod struct {
+	// Offset in milliseconds from the current time when the downsampling with the given interval must be applied
+	Offset int64
+	// Interval for downsampling - only a single sample is left per each interval
+	Interval int64
+}
+
+// String implements interface
+func (dsp DownsamplingPeriod) String() string {
+	offset := time.Duration(dsp.Offset) * time.Millisecond
+	interval := time.Duration(dsp.Interval) * time.Millisecond
+	return fmt.Sprintf("%s:%s", offset, interval)
+}
+
+func (dsp *DownsamplingPeriod) parse(s string) error {
+	idx := strings.Index(s, ":")
+	if idx <= 0 {
+		return fmt.Errorf("incorrect format for downsampling period: %s, want `offset:interval` format", s)
+	}
+	offsetStr, intervalStr := s[:idx], s[idx+1:]
+	interval, err := metricsql.DurationValue(intervalStr, 0)
+	if err != nil {
+		return fmt.Errorf("incorrect interval: %s format for downsampling interval: %s err: %w", intervalStr, s, err)
+	}
+	offset, err := metricsql.DurationValue(offsetStr, 0)
+	if err != nil {
+		return fmt.Errorf("incorrect duration: %s format for downsampling offset: %s err: %w", offsetStr, s, err)
+	}
+	dsp.Interval = interval
+	dsp.Offset = offset
+	// sanity check
+	if offset > 0 && interval > offset {
+		return fmt.Errorf("downsampling interval=%d cannot exceed offset=%d", dsp.Interval, dsp.Offset)
+	}
+	return nil
+}
+
+var downsamplingPeriods []DownsamplingPeriod
+
+// GetDedupInterval returns dedup interval, which must be applied to samples with the given timestamp.
+func GetDedupInterval(timestamp int64) int64 {
+	dsp := getDownsamplingPeriod(timestamp)
+	return dsp.Interval
+}
+
+// getDownsamplingPeriod returns downsampling period, which must be used for the given timestamp
+func getDownsamplingPeriod(timestamp int64) DownsamplingPeriod {
+	offset := int64(fasttime.UnixTimestamp())*1000 - timestamp
+	for _, dsp := range downsamplingPeriods {
+		if offset >= dsp.Offset {
+			return dsp
+		}
+	}
+	return DownsamplingPeriod{}
+}
+
+func parseDownsamplingPeriods(periods []string) ([]DownsamplingPeriod, error) {
+	if len(periods) == 0 {
+		return nil, nil
+	}
+	var dsps []DownsamplingPeriod
+	for _, period := range periods {
+		var dsp DownsamplingPeriod
+		if err := dsp.parse(period); err != nil {
+			return nil, fmt.Errorf("cannot parse downsampling period %q: %w", period, err)
+		}
+		dsps = append(dsps, dsp)
+	}
+	sort.Slice(dsps, func(i, j int) bool {
+		return dsps[i].Offset > dsps[j].Offset
+	})
+	dspPrev := dsps[0]
+	// sanity checks.
+	for _, dsp := range dsps[1:] {
+		if dspPrev.Interval <= dsp.Interval {
+			return nil, fmt.Errorf("prev downsampling interval %d must be bigger than the next interval %d", dspPrev.Interval, dsp.Interval)
+		}
+		if dspPrev.Offset == dsp.Offset {
+			return nil, fmt.Errorf("duplicate downsampling offset: %d", dsp.Offset)
+		}
+		if dspPrev.Interval%dsp.Interval != 0 {
+			return nil, fmt.Errorf("downsamping intervals must be multiples; prev: %d, current: %d", dspPrev.Interval, dsp.Interval)
+		}
+		dspPrev = dsp
+	}
+	return dsps, nil
+}

lib/storage/downsampling_test.go

@@ -0,0 +1,62 @@
+package storage
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestParseDownsamplingPeriodsFailure(t *testing.T) {
+	f := func(name string, src []string) {
+		t.Helper()
+		t.Run(name, func(t *testing.T) {
+			if _, err := parseDownsamplingPeriods(src); err == nil {
+				t.Fatalf("want fail for input: %s", strings.Join(src, ","))
+			}
+		})
+	}
+	f("empty duration", []string{"15d"})
+	f("empty interval", []string{":1m"})
+	f("incorrect duration decrease", []string{"30d:15h", "60d:1h"})
+	f("duplicate offset", []string{"30d:15h", "30d:1h"})
+	f("duplicate interval", []string{"60d:1h", "30d:1h"})
+	f("not multiple intervals", []string{"90d:12h", "60:9h", "30d:7h"})
+}
+
+func TestParseDownsamplingPeriodsSuccess(t *testing.T) {
+	f := func(name string, src []string, expected []DownsamplingPeriod) {
+		t.Helper()
+		t.Run(name, func(t *testing.T) {
+			dsps, err := parseDownsamplingPeriods(src)
+			if err != nil {
+				t.Fatalf("cannot parse downsampling configuration for: %s, err: %s", strings.Join(src, ","), err)
+			}
+			assertDownsamplingPeriods(t, expected, dsps)
+		})
+	}
+	f("one period", []string{"30d:1m"}, []DownsamplingPeriod{
+		{Offset: 30 * 24 * 3600 * 1000, Interval: 60 * 1000},
+	})
+	f("three periods", []string{"15d:30s", "30d:1m", "60d:15m"}, []DownsamplingPeriod{
+		{Offset: 60 * 24 * 3600 * 1000, Interval: 15 * 60 * 1000},
+		{Offset: 30 * 24 * 3600 * 1000, Interval: 60 * 1000},
+		{Offset: 15 * 24 * 3600 * 1000, Interval: 30 * 1000},
+	})
+	f("with the same divider periods", []string{"15d:1m", "30d:7m", "60d:14m", "90d:28m"}, []DownsamplingPeriod{
+		{Offset: 90 * 24 * 3600 * 1000, Interval: 28 * 60 * 1000},
+		{Offset: 60 * 24 * 3600 * 1000, Interval: 14 * 60 * 1000},
+		{Offset: 30 * 24 * 3600 * 1000, Interval: 7 * 60 * 1000},
+		{Offset: 15 * 24 * 3600 * 1000, Interval: 60 * 1000},
+	})
+}
+
+func assertDownsamplingPeriods(t *testing.T, want, got []DownsamplingPeriod) {
+	t.Helper()
+	if len(want) != len(got) {
+		t.Fatalf("len mismatch, want: %d, got: %d", len(want), len(got))
+	}
+	for i := 0; i < len(want); i++ {
+		if want[i] != got[i] {
+			t.Fatalf("want period: %s, got period: %s, idx: %d", want[i], got[i], i)
+		}
+	}
+}

lib/storage/metric_name.go

@@ -402,6 +402,12 @@ func (mn *MetricName) String() string {
 	return fmt.Sprintf("%s{%s}", mnCopy.MetricGroup, tagsStr)
 }
 
+// SortAndMarshal sorts mn tags and then marshals them to dst.
+func (mn *MetricName) SortAndMarshal(dst []byte) []byte {
+	mn.sortTags()
+	return mn.Marshal(dst)
+}
+
 // Marshal appends marshaled mn to dst and returns the result.
 //
 // mn.sortTags must be called before calling this function

lib/storage/partition.go

@@ -1196,7 +1196,7 @@ func (pt *partition) releasePartsToMerge(pws []*partWrapper) {
 }
 
 func (pt *partition) isFinalDedupNeeded() bool {
-	dedupInterval := GetDedupInterval()
+	dedupInterval := GetDedupInterval(pt.tr.MaxTimestamp)
 
 	pws := pt.GetParts(nil, false)
 	minDedupInterval := getMinDedupInterval(pws)
@@ -1432,7 +1432,7 @@ func (pt *partition) mergePartsInternal(dstPartPath string, bsw *blockStreamWrit
 		return nil, fmt.Errorf("cannot merge %d parts to %s: %w", len(bsrs), dstPartPath, err)
 	}
 	if dstPartPath != "" {
-		ph.MinDedupInterval = GetDedupInterval()
+		ph.MinDedupInterval = GetDedupInterval(ph.MaxTimestamp)
 		ph.MustWriteMetadata(dstPartPath)
 	}
 	return &ph, nil

lib/storage/storage.go

@@ -1118,6 +1118,14 @@ func searchAndMerge[T any](qt *querytracer.Tracer, s *Storage, tr TimeRange, sea
 	qt = qt.NewChild("search indexDBs: timeRange=%v", &tr)
 	defer qt.Done()
 
+	var zeroValue T
+	if tr.MinTimestamp < minUnixMilli {
+		tr.MinTimestamp = minUnixMilli
+	}
+	if tr.MaxTimestamp < tr.MinTimestamp {
+		return zeroValue, nil
+	}
+
 	var idbts []indexDBWithType
 
 	ptws := s.tb.GetPartitions(tr)

lib/storage/storage_test.go

@@ -402,6 +402,10 @@ func TestStorageDeletePendingSeries(t *testing.T) {
 	defer testRemoveAll(t)
 
 	const numMonths = 10
+	start := time.Date(1971, 1, 1, 0, 0, 0, 0, time.UTC)
+	middle := start.AddDate(0, (numMonths-1)/2, 0)
+	end := start.AddDate(0, numMonths-1, 0)
+
 	s := MustOpenStorage(t.Name(), OpenOptions{})
 
 	var metricGroupName = []byte("metric")
@@ -456,7 +460,7 @@ func TestStorageDeletePendingSeries(t *testing.T) {
 	assertCountMonthsWithLabels := func(count int) {
 		t.Helper()
 
-		ts := time.Unix(0, 0)
+		ts := start
 		n := 0
 		for range numMonths {
 			lns, err := s.SearchLabelNames(nil, nil, TimeRange{ts.UnixMilli(), ts.UnixMilli()}, 1e5, 1e9, noDeadline)
@@ -481,7 +485,7 @@ func TestStorageDeletePendingSeries(t *testing.T) {
 		var search Search
 		defer search.MustClose()
 
-		search.Init(nil, s, []*TagFilters{tfs}, TimeRange{0, math.MaxInt64}, 1e5, noDeadline)
+		search.Init(nil, s, []*TagFilters{tfs}, TimeRange{start.UnixMilli(), math.MaxInt64}, 1e5, noDeadline)
 		n := 0
 		for search.NextMetricBlock() {
 			var b Block
@@ -498,10 +502,6 @@ func TestStorageDeletePendingSeries(t *testing.T) {
 	// Verify no metrics exist
 	assertCountRows(0)
 
-	start := time.Unix(0, 0)
-	middle := start.AddDate(0, (numMonths-1)/2, 0)
-	end := start.AddDate(0, numMonths-1, 0)
-
 	// Add some rows and flush, so next DeleteSeries() can delete them
 	addRows(start, middle, false)
 	s.DebugFlush()
@@ -3385,53 +3385,190 @@ func TestStorageQueryWithoutIndex(t *testing.T) {
 	testStorageSearchWithoutIndex(t, &opts)
 }
 
-func TestStorageAddRows_SamplesWithZeroDate(t *testing.T) {
+func TestStorageAddRowsWithZeroDate(t *testing.T) {
 	defer testRemoveAll(t)
 
-	f := func(t *testing.T, disablePerDayIndex bool) {
-		t.Helper()
-
-		s := MustOpenStorage(t.Name(), OpenOptions{
-			DisablePerDayIndex: disablePerDayIndex,
-		})
-		defer s.MustClose()
-
-		mn := MetricName{MetricGroup: []byte("metric")}
-		mr := MetricRow{MetricNameRaw: mn.marshalRaw(nil)}
-		for range 10 {
-			mr.Timestamp = rand.Int63n(msecPerDay)
-			mr.Value = float64(rand.Intn(1000))
-			s.AddRows([]MetricRow{mr}, defaultPrecisionBits)
-			s.DebugFlush()
-			// Reset TSID cache so that insertion takes the path that involves
-			// checking whether the index contains metricName->TSID mapping.
-			s.resetAndSaveTSIDCache()
-		}
-
-		want := 1
-		firstUnixDay := TimeRange{
-			MinTimestamp: 0,
-			MaxTimestamp: msecPerDay - 1,
-		}
-		if got := s.newTimeseriesCreated.Load(); got != uint64(want) {
-			t.Errorf("unexpected new timeseries count: got %d, want %d", got, want)
-		}
-		if got := testCountAllMetricNames(s, firstUnixDay); got != want {
-			t.Errorf("unexpected metric name count: got %d, want %d", got, want)
-		}
-		if got := testCountAllMetricIDs(s, firstUnixDay); got != want {
-			t.Errorf("unexpected metric id count: got %d, want %d", got, want)
-		}
-	}
-
 	for _, disablePerDayIndex := range []bool{false, true} {
 		name := fmt.Sprintf("disablePerDayIndex=%t", disablePerDayIndex)
 		t.Run(name, func(t *testing.T) {
-			f(t, disablePerDayIndex)
+			testStorageAddRowsWithZeroDate(t, disablePerDayIndex)
 		})
 	}
 }
 
+func testStorageAddRowsWithZeroDate(t *testing.T, disablePerDayIndex bool) {
+	s := MustOpenStorage(t.Name(), OpenOptions{
+		DisablePerDayIndex: disablePerDayIndex,
+	})
+	defer s.MustClose()
+
+	const numDays = 4
+	var metricNamesAll []string
+	labelNamesAll := []string{"__name__", "label"}
+	var labelValuesAll []string
+	mrs := make([]MetricRow, numDays)
+	for day := range numDays {
+		metricName := fmt.Sprintf("metric_%02d", day)
+		labelName := fmt.Sprintf("label_%02d", day)
+		labelValue := fmt.Sprintf("value_%02d", day)
+
+		if day != 0 {
+			metricNamesAll = append(metricNamesAll, metricName)
+			labelNamesAll = append(labelNamesAll, labelName)
+			labelValuesAll = append(labelValuesAll, labelValue)
+		}
+
+		mn := MetricName{
+			MetricGroup: []byte(metricName),
+			Tags: []Tag{
+				{Key: []byte(labelName), Value: []byte("value")},
+				{Key: []byte("label"), Value: []byte(labelValue)},
+			},
+		}
+		mn.sortTags()
+
+		mrs[day].MetricNameRaw = mn.marshalRaw(nil)
+		mrs[day].Timestamp = int64(day * msecPerDay)
+	}
+
+	s.AddRows(mrs, defaultPrecisionBits)
+	s.DebugFlush()
+	if got, want := s.newTimeseriesCreated.Load(), uint64(numDays-1); got != want {
+		t.Fatalf("unexpected new timeseries count: got %d, want %d", got, want)
+	}
+	if got, want := s.tooSmallTimestampRows.Load(), uint64(1); got != want {
+		t.Fatalf("unexpected rows with too small timestamp: got %d, want %d", got, want)
+	}
+
+	assertMetricNames := func(tr TimeRange, want []string) {
+		t.Helper()
+		tfs := NewTagFilters()
+		if err := tfs.Add(nil, []byte("metric_.*"), false, true); err != nil {
+			t.Fatalf("unexpected error in TagFilters.Add: %v", err)
+		}
+		got, err := s.SearchMetricNames(nil, []*TagFilters{tfs}, tr, 1e9, noDeadline)
+		if err != nil {
+			t.Fatalf("SearchMetricNames(%v, %v) failed unexpectedly: %v", tfs, &tr, err)
+		}
+		for i, name := range got {
+			var mn MetricName
+			if err := mn.UnmarshalString(name); err != nil {
+				t.Fatalf("Could not unmarshal metric name %q: %v", name, err)
+			}
+			got[i] = string(mn.MetricGroup)
+		}
+		slices.Sort(got)
+		slices.Sort(want)
+		if diff := cmp.Diff(want, got); diff != "" {
+			t.Fatalf("unexpected metric names (-want, +got):\n%s", diff)
+		}
+	}
+	assertLabelNames := func(tr TimeRange, want []string) {
+		t.Helper()
+		tfs := NewTagFilters()
+		if err := tfs.Add(nil, []byte("metric_.*"), false, true); err != nil {
+			t.Fatalf("unexpected error in TagFilters.Add: %v", err)
+		}
+		got, err := s.SearchLabelNames(nil, []*TagFilters{tfs}, tr, 1e9, 1e9, noDeadline)
+		if err != nil {
+			t.Fatalf("SearchLabelNames(%v, %v) failed unexpectedly: %s", tfs, &tr, err)
+		}
+		slices.Sort(got)
+		slices.Sort(want)
+		if diff := cmp.Diff(want, got); diff != "" {
+			t.Fatalf("unexpected label names (-want, +got):\n%s", diff)
+		}
+	}
+	assertLabelValues := func(tr TimeRange, want []string) {
+		t.Helper()
+		tfs := NewTagFilters()
+		if err := tfs.Add([]byte("label"), []byte("value_.*"), false, true); err != nil {
+			t.Fatalf("unexpected error in TagFilters.Add: %v", err)
+		}
+		got, err := s.SearchLabelValues(nil, "label", []*TagFilters{tfs}, tr, 1e9, 1e9, noDeadline)
+		if err != nil {
+			t.Fatalf("SearchLabelValues(%v, %v) failed unexpectedly: %s", tfs, tr, err)
+		}
+		slices.Sort(got)
+		slices.Sort(want)
+		if diff := cmp.Diff(want, got); diff != "" {
+			t.Fatalf("unexpected label values (-want, +got):\n%s", diff)
+		}
+	}
+	assertData := func(tr TimeRange, want []MetricRow) {
+		t.Helper()
+		tfs := NewTagFilters()
+		if err := tfs.Add(nil, []byte("metric_.*"), false, true); err != nil {
+			t.Fatalf("TagFilters.Add() failed unexpectedly: %v", err)
+		}
+		if err := testAssertSearchResult(s, tr, tfs, want); err != nil {
+			t.Fatalf("Search(%v, %v) failed unexpectedly: %v", tfs, tr, err)
+		}
+	}
+
+	var tr TimeRange
+
+	// Empty time range.
+	// Expect empty search results
+	tr = TimeRange{}
+	assertMetricNames(tr, nil)
+	assertLabelNames(tr, []string{})
+	assertLabelValues(tr, []string{})
+	assertData(tr, nil)
+
+	// First day time range.
+	// Expect empty search results
+	tr = TimeRange{
+		MinTimestamp: 0,
+		MaxTimestamp: msecPerDay - 1,
+	}
+	assertMetricNames(tr, nil)
+	assertLabelNames(tr, []string{})
+	assertLabelValues(tr, []string{})
+	assertData(tr, nil)
+
+	// Second day time range.
+	tr = TimeRange{
+		MinTimestamp: msecPerDay,
+		MaxTimestamp: 2*msecPerDay - 1,
+	}
+	if disablePerDayIndex {
+		// Expect index search results for all days if per-day index is
+		// disabled.
+		assertMetricNames(tr, metricNamesAll)
+		assertLabelNames(tr, labelNamesAll)
+		assertLabelValues(tr, labelValuesAll)
+	} else {
+		// Expect index search results on second day only if per-day index is
+		// enabled.
+		assertMetricNames(tr, []string{"metric_01"})
+		assertLabelNames(tr, []string{"__name__", "label", "label_01"})
+		assertLabelValues(tr, []string{"value_01"})
+	}
+	assertData(tr, mrs[1:2])
+
+	// First two days time range.
+	// Expect results on second day only.
+	tr = TimeRange{
+		MinTimestamp: 0,
+		MaxTimestamp: 2*msecPerDay - 1,
+	}
+	if disablePerDayIndex {
+		// Expect index search results for all days if per-day index is
+		// disabled.
+		assertMetricNames(tr, metricNamesAll)
+		assertLabelNames(tr, labelNamesAll)
+		assertLabelValues(tr, labelValuesAll)
+	} else {
+		// Expect index search results on second day only if per-day index is
+		// enabled.
+		assertMetricNames(tr, []string{"metric_01"})
+		assertLabelNames(tr, []string{"__name__", "label", "label_01"})
+		assertLabelValues(tr, []string{"value_01"})
+	}
+	assertData(tr, mrs[1:2])
+}
+
 // testSearchMetricIDs returns metricIDs for the given tfss and tr.
 //
 // The returned metricIDs are sorted. The function panics in in case of error.

lib/storage/table.go

@@ -429,9 +429,8 @@ func (tb *table) getMinMaxIngestionTimestamps() (int64, int64) {
 func (tb *table) getMinMaxTimestampsForAge(minAgeMsecs int64) (int64, int64) {
 	now := int64(fasttime.UnixTimestamp() * 1000)
 	minTimestamp := now - minAgeMsecs
-	if minTimestamp < 0 {
-		// Negative timestamps aren't supported by the storage.
-		minTimestamp = 0
+	if minTimestamp < minUnixMilli {
+		minTimestamp = minUnixMilli
 	}
 	maxTimestamp := int64(maxUnixMilli)
 	if maxUnixMilli-now > tb.s.futureRetentionMsecs {

lib/storage/time.go

@@ -40,12 +40,6 @@ type TimeRange struct {
 	MaxTimestamp int64
 }
 
-// Zero time range and zero date are used to force global index search.
-var (
-	globalIndexTimeRange = TimeRange{}
-	globalIndexDate      = uint64(0)
-)
-
 // DateRange returns the date range for the given time range.
 func (tr *TimeRange) DateRange() (uint64, uint64) {
 	minDate := uint64(tr.MinTimestamp) / msecPerDay
@@ -117,10 +111,29 @@ func (tr *TimeRange) contains(timestamp int64) bool {
 	return tr.MinTimestamp <= timestamp && timestamp <= tr.MaxTimestamp
 }
 
+// Zero time range and zero date are used to force global index search.
+var (
+	globalIndexDate      = uint64(0)
+	globalIndexTimeRange = TimeRange{}
+)
+
 const (
 	msecPerDay  = 24 * 3600 * 1000
 	msecPerHour = 3600 * 1000
 
+	// minUnixMilli is the min millisecond that is allowed to be used as the
+	// sample timestamp.
+	//
+	// It corresponds to the first millisecond of the second day of the Unix
+	// Epoch, i.e. 1970-01-02T00:00:00.000Z.
+	//
+	// The first day of the Unix Epoch is reserved: zero date and zero time
+	// range are used for indicating that the the global index search is
+	// required. See globalIndexDate and globalIndexTimeRange above.
+	//
+	// Negative timestamps aren't supported.
+	minUnixMilli = msecPerDay
+
 	// maxUnixMilli is the max millisecond that is allowed to be used as the
 	// sample timestamp.
 	//
@@ -130,6 +143,6 @@ const (
 	// time.UnixMicro(math.MaxInt64/1000) == 2262-04-11 23:47:16.854775 UTC.
 	//
 	// Round it to the last millisecond of the last complete partition:
-	// 2262-03-31 23:59:59.999 UTC.
+	// 2262-03-31T23:59:59.999Z.
 	maxUnixMilli = 9222422399999
 )

lib/streamaggr/streamaggr_synctest_test.go

@@ -935,4 +935,25 @@ foo:1m_sum_samples{baz="qwe"} 10
   dedup_interval: 30s
   outputs: [sum_samples]
 `, "11111111")
+
+	// Reproduce issue #11261: sum_samples_total must be monotonic with enable_windows: true
+	// See https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11262
+	f([]string{`
+test_delta 1
+`, `
+test_delta 1
+`, `
+test_delta 1
+`, `
+test_delta 1
+`}, time.Minute, `test_delta 1
+test_delta 2
+test_delta 3
+test_delta 4
+`, `
+- interval: 1m
+  keep_metric_names: true
+  outputs: [sum_samples_total]
+  enable_windows: true
+`, "1111")
 }

lib/streamaggr/sum_samples.go

@@ -4,30 +4,39 @@ import (
 	"math"
 )
 
+type sumSamplesAggrValueShared struct {
+	total float64
+}
+
 type sumSamplesAggrValue struct {
-	sum float64
+	delta  float64
+	shared *sumSamplesAggrValueShared
 }
 
 func (av *sumSamplesAggrValue) pushSample(_ aggrConfig, sample *pushSample, _ string, _ int64) {
-	if math.Abs(av.sum) >= (1 << 53) {
-		// It is time to reset the entry, since it starts losing float64 precision
-		av.sum = 0
-	}
-	av.sum += sample.value
+	av.delta += sample.value
 }
 
 func (av *sumSamplesAggrValue) flush(c aggrConfig, ctx *flushCtx, key string, _ bool) {
 	ac := c.(*sumSamplesAggrConfig)
 	if ac.resetTotalOnFlush {
-		ctx.appendSeries(key, "sum_samples", av.sum)
-		av.sum = 0
+		ctx.appendSeries(key, "sum_samples", av.delta)
+		av.delta = 0
 		return
 	}
-	ctx.appendSeries(key, "sum_samples_total", av.sum)
+	total := av.shared.total + av.delta
+	av.delta = 0
+	if math.Abs(total) >= (1 << 53) {
+		// It is time to reset the entry, since it starts losing float64 precision
+		av.shared.total = 0
+	} else {
+		av.shared.total = total
+	}
+	ctx.appendSeries(key, "sum_samples_total", total)
 }
 
-func (*sumSamplesAggrValue) state() any {
-	return nil
+func (av *sumSamplesAggrValue) state() any {
+	return av.shared
 }
 
 func newSumSamplesAggrConfig(resetTotalOnFlush bool) aggrConfig {
@@ -40,6 +49,14 @@ type sumSamplesAggrConfig struct {
 	resetTotalOnFlush bool
 }
 
-func (*sumSamplesAggrConfig) getValue(_ any) aggrValue {
-	return &sumSamplesAggrValue{}
+func (*sumSamplesAggrConfig) getValue(s any) aggrValue {
+	var shared *sumSamplesAggrValueShared
+	if s == nil {
+		shared = &sumSamplesAggrValueShared{}
+	} else {
+		shared = s.(*sumSamplesAggrValueShared)
+	}
+	return &sumSamplesAggrValue{
+		shared: shared,
+	}
 }

lib/timeutil/time.go

@@ -206,17 +206,36 @@ func tryParseScientificNumberForUnixTimestamp(s string, decimalExp int64) (int64
 		return multiplyByDecimalExp(n, decimalExp)
 	}
 
-	intStr := s[:dotIdx]
-	fracStr := s[dotIdx+1:]
-	if decimalExp < int64(len(fracStr)) {
+	if decimalExp < 0 {
+		// Negative exponents on a fractional mantissa are intentionally not
+		// supported. See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11268
 		return 0, false
 	}
+
+	intStr := s[:dotIdx]
+	fracStr := s[dotIdx+1:]
 	n, ok := tryParseFractionalNumberForUnixTimestamp(intStr, fracStr)
 	if !ok {
 		return 0, false
 	}
-	decimalExp -= int64(len(fracStr))
-	return multiplyByDecimalExp(n, decimalExp)
+	if decimalExp >= int64(len(fracStr)) {
+		// The exponent shifts the decimal point past every fractional digit,
+		// so the value is an integer number of seconds (or coarser).
+		decimalExp -= int64(len(fracStr))
+		return multiplyByDecimalExp(n, decimalExp)
+	}
+
+	// The exponent leaves fractional digits, e.g. 1.784144612388E9 == 1784144612.388
+	// Pad n as plain fractional timestamps do.
+	fracDigits := int64(len(fracStr)) - decimalExp
+	for fracDigits%3 != 0 {
+		if n >= 0 && n > math.MaxInt64/10 || n < 0 && n < math.MinInt64/10 {
+			return 0, false
+		}
+		n *= 10
+		fracDigits++
+	}
+	return n, true
 }
 
 func tryParseFractionalNumberForUnixTimestamp(intStr, fracStr string) (int64, bool) {

lib/timeutil/time_test.go

@@ -69,6 +69,17 @@ func TestTryParseUnixTimestamp_Success(t *testing.T) {
 	f("1.23e2", 123000000000)
 	f("1.2e1", 12000000000)
 	f("1123.456789123456789E15", 1123456789123456789)
+
+	// scientific notation with sub-second precision, i.e. more fractional digits
+	// than the exponent shifts (https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11268).
+	// These must match the equivalent plain fractional form.
+	f("1.784144612388E9", 1784144612388000000) // == 1784144612.388
+	f("1.784144612388e9", 1784144612388000000)
+	f("-1.784144612388e9", -1784144612388000000)
+	f("1.5000000005e9", 1500000000500000000) // == 1500000000.5
+	f("1.23456789e9", 1234567890000000000)   // exponent consumes all frac digits (integer result)
+	f("1.23e1", 12300000000000)              // == 12.3
+	f("1.234e0", 1234000000000)              // == 1.234
 }
 
 func TestTryParseUnixTimestamp_Failure(t *testing.T) {
@@ -115,9 +126,7 @@ func TestTryParseUnixTimestamp_Failure(t *testing.T) {
 	f("1e19")
 	f("1.3e123456789090123")
 
-	// too small decimal exponent
-	f("1.23e1")
-	f("1.234e0")
+	// negative decimal exponent
 	f("1E-1")
 	f("1.3e-123456789090123")
 }