Merge tag 'v1.113.0' into pmm-6401-read-prometheus-data-files

v1.113.0

Signed-off-by: Artem Fetishev <rtm@victoriametrics.com>

app/victoria-metrics/main.go

@@ -40,6 +40,8 @@ var (
 		"The saved data survives unclean shutdowns such as OOM crash, hardware reset, SIGKILL, etc. "+
 		"Bigger intervals may help increase the lifetime of flash storage with limited write cycles (e.g. Raspberry PI). "+
 		"Smaller intervals increase disk IO load. Minimum supported value is 1s")
+	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")
 	maxIngestionRate = flag.Int("maxIngestionRate", 0, "The maximum number of samples vmsingle can receive per second. Data ingestion is paused when the limit is exceeded. "+
 		"By default there are no limits on samples ingestion rate.")
 	finalDedupScheduleInterval = flag.Duration("storage.finalDedupScheduleCheckInterval", time.Hour, "The interval for checking when final deduplication process should be started."+
@@ -48,6 +50,12 @@ var (
 		" See also https://docs.victoriametrics.com/#deduplication")
 )
 
+// 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 main() {
 	// VictoriaMetrics is optimized for reduced memory allocations,
 	// so it can run with the reduced GOGC in order to reduce the used memory,
@@ -88,7 +96,10 @@ func main() {
 	}
 	logger.Infof("starting VictoriaMetrics at %q...", listenAddrs)
 	startTime := time.Now()
-	storage.SetDedupInterval(*minScrapeInterval)
+	err := storage.SetDownsamplingPeriods(*downsamplingPeriods, *minScrapeInterval)
+	if err != nil {
+		logger.Fatalf("cannot parse -downsampling.period: %s", err)
+	}
 	storage.SetDataFlushInterval(*inmemoryDataFlushInterval)
 	if *finalDedupScheduleInterval < time.Hour {
 		logger.Fatalf("-dedup.finalDedupScheduleCheckInterval cannot be smaller than 1 hour; got %s", *finalDedupScheduleInterval)
@@ -152,6 +163,10 @@ 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/vmselect/netstorage/netstorage.go

@@ -15,6 +15,7 @@ import (
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/searchutils"
 	"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"
@@ -336,6 +337,12 @@ var (
 type packedTimeseries struct {
 	metricName string
 	brs        []blockRef
+	pd         *promData
+}
+
+type promData struct {
+	values     []float64
+	timestamps []int64
 }
 
 type unpackWork struct {
@@ -439,9 +446,21 @@ 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
 }
 
@@ -558,6 +577,27 @@ 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 {
@@ -795,6 +835,15 @@ 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(tr, 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))
@@ -866,14 +915,44 @@ func LabelValues(qt *querytracer.Tracer, labelName string, sq *storage.SearchQue
 	}
 	labelValues, err := vmstorage.SearchLabelValues(qt, labelName, tfss, tr, maxLabelValues, sq.MaxMetrics, 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, tr, 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
+}
+
 // GraphiteTagValues returns tag values for the given tagName until the given deadline.
 func GraphiteTagValues(qt *querytracer.Tracer, tagName, filter string, limit int, deadline searchutils.Deadline) ([]string, error) {
 	qt = qt.NewChild("get graphite tag values for tagName=%s, filter=%s, limit=%d", tagName, filter, limit)
@@ -1275,6 +1354,26 @@ 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)
+		putStorageSearch(sr)
+		return nil, fmt.Errorf("error when searching in Prometheus data: %w", err)
+	}
+
 	var rss Results
 	rss.tr = tr
 	rss.deadline = deadline
@@ -1283,6 +1382,7 @@ 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/vmstorage/main.go

@@ -12,6 +12,7 @@ import (
 
 	"github.com/VictoriaMetrics/metrics"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage/promdb"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
@@ -145,9 +146,11 @@ func Init(resetCacheIfNeeded func(mrs []storage.MetricRow)) {
 	// register storage metrics
 	storageMetrics = metrics.NewSet()
 	storageMetrics.RegisterMetricsWriter(func(w io.Writer) {
-		writeStorageMetrics(w, strg)
+		writeStorageMetrics(w, Storage)
 	})
 	metrics.RegisterSet(storageMetrics)
+
+	promdb.Init(retentionPeriod.Milliseconds())
 }
 
 var storageMetrics *metrics.Set
@@ -282,6 +285,7 @@ func Stop() {
 	logger.Infof("gracefully closing the storage at %s", *DataPath)
 	startTime := time.Now()
 	WG.WaitAndBlock()
+	promdb.MustClose()
 	stopStaleSnapshotsRemover()
 	Storage.MustClose()
 	logger.Infof("successfully closed the storage in %.3f seconds", time.Since(startTime).Seconds())

app/vmstorage/promdb/promdb.go

@@ -0,0 +1,268 @@
+package promdb
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"log/slog"
+	"time"
+
+	"github.com/oklog/ulid"
+	"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/searchutils"
+	"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 searchutils.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 searchutils.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 searchutils.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, labels []labels.Label) {
+	dst.Reset()
+	for _, label := range labels {
+		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)
+	}
+}

go.mod

@@ -2,6 +2,9 @@ module github.com/VictoriaMetrics/VictoriaMetrics
 
 go 1.24.0
 
+// See // See https://github.com/googleapis/google-cloud-go/issues/11283#issuecomment-2558515586
+exclude google.golang.org/grpc/stats/opentelemetry v0.0.0-20240907200651-3ffb98b2c93a
+
 // This is needed in order to avoid vmbackup and vmrestore binary size increase by 20MB
 // See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8008
 // TODO: remove this entry after https://github.com/googleapis/google-cloud-go/issues/11448 is fixed
@@ -31,6 +34,7 @@ require (
 	github.com/influxdata/influxdb v1.11.8
 	github.com/klauspost/compress v1.17.11
 	github.com/mattn/go-isatty v0.0.20
+	github.com/oklog/ulid v1.3.1
 	github.com/prometheus/prometheus v0.301.0
 	github.com/urfave/cli/v2 v2.27.5
 	github.com/valyala/fastjson v1.6.4
@@ -78,7 +82,8 @@ require (
 	github.com/cpuguy83/go-md2man/v2 v2.0.6 // indirect
 	github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc // indirect
 	github.com/dennwc/varint v1.0.0 // indirect
-	github.com/envoyproxy/go-control-plane/envoy v1.32.3 // indirect
+	github.com/envoyproxy/go-control-plane/envoy v1.32.4 // indirect
+	github.com/envoyproxy/protoc-gen-validate v1.2.1 // indirect
 	github.com/fatih/color v1.18.0 // indirect
 	github.com/felixge/httpsnoop v1.0.4 // indirect
 	github.com/go-logr/logr v1.4.2 // indirect
@@ -98,7 +103,6 @@ require (
 	github.com/modern-go/reflect2 v1.0.2 // indirect
 	github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 // indirect
 	github.com/mwitkow/go-conntrack v0.0.0-20190716064945-2f068394615f // indirect
-	github.com/oklog/ulid v1.3.1 // indirect
 	github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c // indirect
 	github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 // indirect
 	github.com/prometheus/client_golang v1.20.5 // indirect

go.sum

@@ -130,10 +130,10 @@ github.com/edsrzf/mmap-go v1.2.0 h1:hXLYlkbaPzt1SaQk+anYwKSRNhufIDCchSPkUD6dD84=
 github.com/edsrzf/mmap-go v1.2.0/go.mod h1:19H/e8pUPLicwkyNgOykDXkJ9F0MHE+Z52B8EIth78Q=
 github.com/emicklei/go-restful/v3 v3.11.0 h1:rAQeMHw1c7zTmncogyy8VvRZwtkmkZ4FxERmMY4rD+g=
 github.com/emicklei/go-restful/v3 v3.11.0/go.mod h1:6n3XBCmQQb25CM2LCACGz8ukIrRry+4bhvbpWn3mrbc=
-github.com/envoyproxy/go-control-plane/envoy v1.32.3 h1:c1EIw4vwYCaovxRZtyycws8aX6dJ9W2p+4bCi7mcDgw=
-github.com/envoyproxy/go-control-plane/envoy v1.32.3/go.mod h1:c955gQjaXHsMxMjHjEZ7nwIzMJYxXpN+sJIGufsSbg4=
-github.com/envoyproxy/protoc-gen-validate v1.1.0 h1:tntQDh69XqOCOZsDz0lVJQez/2L6Uu2PdjCQwWCJ3bM=
-github.com/envoyproxy/protoc-gen-validate v1.1.0/go.mod h1:sXRDRVmzEbkM7CVcM06s9shE/m23dg3wzjl0UWqJ2q4=
+github.com/envoyproxy/go-control-plane/envoy v1.32.4 h1:jb83lalDRZSpPWW2Z7Mck/8kXZ5CQAFYVjQcdVIr83A=
+github.com/envoyproxy/go-control-plane/envoy v1.32.4/go.mod h1:Gzjc5k8JcJswLjAx1Zm+wSYE20UrLtt7JZMWiWQXQEw=
+github.com/envoyproxy/protoc-gen-validate v1.2.1 h1:DEo3O99U8j4hBFwbJfrz9VtgcDfUKS7KJ7spH3d86P8=
+github.com/envoyproxy/protoc-gen-validate v1.2.1/go.mod h1:d/C80l/jxXLdfEIhX1W2TmLfsJ31lvEjwamM4DxlWXU=
 github.com/ergochat/readline v0.1.3 h1:/DytGTmwdUJcLAe3k3VJgowh5vNnsdifYT6uVaf4pSo=
 github.com/ergochat/readline v0.1.3/go.mod h1:o3ux9QLHLm77bq7hDB21UTm6HlV2++IPDMfIfKDuOgY=
 github.com/facette/natsort v0.0.0-20181210072756-2cd4dd1e2dcb h1:IT4JYU7k4ikYg1SCxNI1/Tieq/NFvh6dzLdgi7eu0tM=

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

@@ -15,15 +15,10 @@ 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

@@ -404,6 +404,12 @@ 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

@@ -1342,7 +1342,7 @@ func (pt *partition) runFinalDedup(stopCh <-chan struct{}) error {
 }
 
 func (pt *partition) isFinalDedupNeeded() bool {
-	dedupInterval := GetDedupInterval()
+	dedupInterval := GetDedupInterval(pt.tr.MaxTimestamp)
 
 	pws := pt.GetParts(nil, false)
 	minDedupInterval := getMinDedupInterval(pws)
@@ -1575,7 +1575,7 @@ 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

vendor/modules.txt

@@ -330,8 +330,10 @@ github.com/davecgh/go-spew/spew
 # github.com/dennwc/varint v1.0.0
 ## explicit; go 1.12
 github.com/dennwc/varint
-# github.com/envoyproxy/go-control-plane/envoy v1.32.3
+# github.com/envoyproxy/go-control-plane/envoy v1.32.4
 ## explicit; go 1.22
+# github.com/envoyproxy/protoc-gen-validate v1.2.1
+## explicit; go 1.21.1
 # github.com/ergochat/readline v0.1.3
 ## explicit; go 1.19
 github.com/ergochat/readline