app/vmselect: reduce memory usage when query touches big number of time series

CHANGELOG.md

@@ -1,6 +1,9 @@
 # tip
 
 
+* FEATURE: reduce memory usage when query touches big number of time series.
+
+
 # [v1.45.0](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/tag/v1.45.0)
 
 * FEATURE: allow setting `-retentionPeriod` smaller than one month. I.e. `-retentionPeriod=3d`, `-retentionPeriod=2w`, etc. is supported now.

app/vmselect/main.go

@@ -10,10 +10,12 @@ import (
 	"time"
 
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/graphite"
+	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/prometheus"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/promql"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/searchutils"
 	"github.com/VictoriaMetrics/VictoriaMetrics/app/vmstorage"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/httpserver"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
@@ -45,6 +47,9 @@ func getDefaultMaxConcurrentRequests() int {
 
 // Init initializes vmselect
 func Init() {
+	tmpDirPath := *vmstorage.DataPath + "/tmp"
+	fs.RemoveDirContents(tmpDirPath)
+	netstorage.InitTmpBlocksDir(tmpDirPath)
 	promql.InitRollupResultCache(*vmstorage.DataPath + "/cache/rollupResult")
 
 	concurrencyCh = make(chan struct{}, *maxConcurrentRequests)

app/vmselect/netstorage/netstorage.go

@@ -57,6 +57,7 @@ type Results struct {
 
 	packedTimeseries []packedTimeseries
 	sr               *storage.Search
+	tbf              *tmpBlocksFile
 }
 
 // Len returns the number of results in rss.
@@ -72,6 +73,8 @@ func (rss *Results) Cancel() {
 func (rss *Results) mustClose() {
 	putStorageSearch(rss.sr)
 	rss.sr = nil
+	putTmpBlocksFile(rss.tbf)
+	rss.tbf = nil
 }
 
 var timeseriesWorkCh = make(chan *timeseriesWork, gomaxprocs*16)
@@ -105,7 +108,7 @@ func timeseriesWorker(workerID uint) {
 			tsw.doneCh <- nil
 			continue
 		}
-		if err := tsw.pts.Unpack(&rs, rss.tr, rss.fetchData); err != nil {
+		if err := tsw.pts.Unpack(&rs, rss.tbf, rss.tr, rss.fetchData); err != nil {
 			tsw.doneCh <- fmt.Errorf("error during time series unpacking: %w", err)
 			continue
 		}
@@ -179,27 +182,29 @@ var gomaxprocs = runtime.GOMAXPROCS(-1)
 
 type packedTimeseries struct {
 	metricName string
-	brs        []storage.BlockRef
+	brs        []blockRef
 }
 
 var unpackWorkCh = make(chan *unpackWork, gomaxprocs*128)
 
 type unpackWorkItem struct {
-	br storage.BlockRef
+	br blockRef
 	tr storage.TimeRange
 }
 
 type unpackWork struct {
+	tbf    *tmpBlocksFile
 	ws     []unpackWorkItem
 	sbs    []*sortBlock
 	doneCh chan error
 }
 
 func (upw *unpackWork) reset() {
+	upw.tbf = nil
 	ws := upw.ws
 	for i := range ws {
 		w := &ws[i]
-		w.br = storage.BlockRef{}
+		w.br = blockRef{}
 		w.tr = storage.TimeRange{}
 	}
 	upw.ws = upw.ws[:0]
@@ -216,7 +221,7 @@ func (upw *unpackWork) reset() {
 func (upw *unpackWork) unpack(tmpBlock *storage.Block) {
 	for _, w := range upw.ws {
 		sb := getSortBlock()
-		if err := sb.unpackFrom(tmpBlock, w.br, w.tr); err != nil {
+		if err := sb.unpackFrom(tmpBlock, upw.tbf, w.br, w.tr); err != nil {
 			putSortBlock(sb)
 			upw.doneCh <- fmt.Errorf("cannot unpack block: %w", err)
 			return
@@ -262,7 +267,7 @@ func unpackWorker() {
 var unpackBatchSize = 8 * runtime.GOMAXPROCS(-1)
 
 // Unpack unpacks pts to dst.
-func (pts *packedTimeseries) Unpack(dst *Result, tr storage.TimeRange, fetchData bool) error {
+func (pts *packedTimeseries) Unpack(dst *Result, tbf *tmpBlocksFile, tr storage.TimeRange, fetchData bool) error {
 	dst.reset()
 	if err := dst.MetricName.Unmarshal(bytesutil.ToUnsafeBytes(pts.metricName)); err != nil {
 		return fmt.Errorf("cannot unmarshal metricName %q: %w", pts.metricName, err)
@@ -276,11 +281,13 @@ func (pts *packedTimeseries) Unpack(dst *Result, tr storage.TimeRange, fetchData
 	brsLen := len(pts.brs)
 	upws := make([]*unpackWork, 0, 1+brsLen/unpackBatchSize)
 	upw := getUnpackWork()
+	upw.tbf = tbf
 	for _, br := range pts.brs {
 		if len(upw.ws) >= unpackBatchSize {
 			unpackWorkCh <- upw
 			upws = append(upws, upw)
 			upw = getUnpackWork()
+			upw.tbf = tbf
 		}
 		upw.ws = append(upw.ws, unpackWorkItem{
 			br: br,
@@ -397,9 +404,10 @@ func (sb *sortBlock) reset() {
 	sb.NextIdx = 0
 }
 
-func (sb *sortBlock) unpackFrom(tmpBlock *storage.Block, br storage.BlockRef, tr storage.TimeRange) error {
+func (sb *sortBlock) unpackFrom(tmpBlock *storage.Block, tbf *tmpBlocksFile, br blockRef, tr storage.TimeRange) error {
 	tmpBlock.Reset()
-	br.MustReadBlock(tmpBlock, true)
+	brReal := tbf.MustReadBlockRefAt(br.partRef, br.addr)
+	brReal.MustReadBlock(tmpBlock, true)
 	if err := tmpBlock.UnmarshalData(); err != nil {
 		return fmt.Errorf("cannot unmarshal block: %w", err)
 	}
@@ -709,19 +717,31 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline search
 
 	sr := getStorageSearch()
 	maxSeriesCount := sr.Init(vmstorage.Storage, tfss, tr, *maxMetricsPerSearch, deadline.Deadline())
-
-	m := make(map[string][]storage.BlockRef, maxSeriesCount)
+	m := make(map[string][]blockRef, maxSeriesCount)
 	orderedMetricNames := make([]string, 0, maxSeriesCount)
 	blocksRead := 0
+	tbf := getTmpBlocksFile()
+	var buf []byte
 	for sr.NextMetricBlock() {
 		blocksRead++
 		if deadline.Exceeded() {
+			putTmpBlocksFile(tbf)
 			putStorageSearch(sr)
 			return nil, fmt.Errorf("timeout exceeded while fetching data block #%d from storage: %s", blocksRead, deadline.String())
 		}
+		buf = sr.MetricBlockRef.BlockRef.Marshal(buf[:0])
+		addr, err := tbf.WriteBlockRefData(buf)
+		if err != nil {
+			putTmpBlocksFile(tbf)
+			putStorageSearch(sr)
+			return nil, fmt.Errorf("cannot write %d bytes to temporary file: %w", len(buf), err)
+		}
 		metricName := sr.MetricBlockRef.MetricName
 		brs := m[string(metricName)]
-		brs = append(brs, *sr.MetricBlockRef.BlockRef)
+		brs = append(brs, blockRef{
+			partRef: sr.MetricBlockRef.BlockRef.PartRef(),
+			addr:    addr,
+		})
 		if len(brs) > 1 {
 			// An optimization: do not allocate a string for already existing metricName key in m
 			m[string(metricName)] = brs
@@ -733,12 +753,18 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline search
 		}
 	}
 	if err := sr.Error(); err != nil {
+		putTmpBlocksFile(tbf)
 		putStorageSearch(sr)
 		if errors.Is(err, storage.ErrDeadlineExceeded) {
 			return nil, fmt.Errorf("timeout exceeded during the query: %s", deadline.String())
 		}
 		return nil, fmt.Errorf("search error after reading %d data blocks: %w", blocksRead, err)
 	}
+	if err := tbf.Finalize(); err != nil {
+		putTmpBlocksFile(tbf)
+		putStorageSearch(sr)
+		return nil, fmt.Errorf("cannot finalize temporary file: %w", err)
+	}
 
 	var rss Results
 	rss.tr = tr
@@ -753,9 +779,15 @@ func ProcessSearchQuery(sq *storage.SearchQuery, fetchData bool, deadline search
 	}
 	rss.packedTimeseries = pts
 	rss.sr = sr
+	rss.tbf = tbf
 	return &rss, nil
 }
 
+type blockRef struct {
+	partRef storage.PartRef
+	addr    tmpBlockAddr
+}
+
 func setupTfss(tagFilterss [][]storage.TagFilter) ([]*storage.TagFilters, error) {
 	tfss := make([]*storage.TagFilters, 0, len(tagFilterss))
 	for _, tagFilters := range tagFilterss {

app/vmselect/netstorage/tmp_blocks_file.go

@@ -0,0 +1,188 @@
+package netstorage
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"sync"
+
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
+	"github.com/VictoriaMetrics/metrics"
+)
+
+// InitTmpBlocksDir initializes directory to store temporary search results.
+//
+// It stores data in system-defined temporary directory if tmpDirPath is empty.
+func InitTmpBlocksDir(tmpDirPath string) {
+	if len(tmpDirPath) == 0 {
+		tmpDirPath = os.TempDir()
+	}
+	tmpBlocksDir = tmpDirPath + "/searchResults"
+	fs.MustRemoveAll(tmpBlocksDir)
+	if err := fs.MkdirAllIfNotExist(tmpBlocksDir); err != nil {
+		logger.Panicf("FATAL: cannot create %q: %s", tmpBlocksDir, err)
+	}
+}
+
+var tmpBlocksDir string
+
+func maxInmemoryTmpBlocksFile() int {
+	mem := memory.Allowed()
+	maxLen := mem / 1024
+	if maxLen < 64*1024 {
+		return 64 * 1024
+	}
+	if maxLen > 4*1024*1024 {
+		return 4 * 1024 * 1024
+	}
+	return maxLen
+}
+
+var _ = metrics.NewGauge(`vm_tmp_blocks_max_inmemory_file_size_bytes`, func() float64 {
+	return float64(maxInmemoryTmpBlocksFile())
+})
+
+type tmpBlocksFile struct {
+	buf []byte
+
+	f *os.File
+	r *fs.ReaderAt
+
+	offset uint64
+}
+
+func getTmpBlocksFile() *tmpBlocksFile {
+	v := tmpBlocksFilePool.Get()
+	if v == nil {
+		return &tmpBlocksFile{
+			buf: make([]byte, 0, maxInmemoryTmpBlocksFile()),
+		}
+	}
+	return v.(*tmpBlocksFile)
+}
+
+func putTmpBlocksFile(tbf *tmpBlocksFile) {
+	tbf.MustClose()
+	tbf.buf = tbf.buf[:0]
+	tbf.f = nil
+	tbf.r = nil
+	tbf.offset = 0
+	tmpBlocksFilePool.Put(tbf)
+}
+
+var tmpBlocksFilePool sync.Pool
+
+type tmpBlockAddr struct {
+	offset uint64
+	size   int
+}
+
+func (addr tmpBlockAddr) String() string {
+	return fmt.Sprintf("offset %d, size %d", addr.offset, addr.size)
+}
+
+var (
+	tmpBlocksFilesCreated = metrics.NewCounter(`vm_tmp_blocks_files_created_total`)
+	_                     = metrics.NewGauge(`vm_tmp_blocks_files_directory_free_bytes`, func() float64 {
+		return float64(fs.MustGetFreeSpace(tmpBlocksDir))
+	})
+)
+
+// WriteBlockRefData writes br to tbf.
+//
+// It returns errors since the operation may fail on space shortage
+// and this must be handled.
+func (tbf *tmpBlocksFile) WriteBlockRefData(b []byte) (tmpBlockAddr, error) {
+	var addr tmpBlockAddr
+	addr.offset = tbf.offset
+	addr.size = len(b)
+	tbf.offset += uint64(addr.size)
+	if len(tbf.buf)+len(b) <= cap(tbf.buf) {
+		// Fast path - the data fits tbf.buf
+		tbf.buf = append(tbf.buf, b...)
+		return addr, nil
+	}
+
+	// Slow path: flush the data from tbf.buf to file.
+	if tbf.f == nil {
+		f, err := ioutil.TempFile(tmpBlocksDir, "")
+		if err != nil {
+			return addr, err
+		}
+		tbf.f = f
+		tmpBlocksFilesCreated.Inc()
+	}
+	_, err := tbf.f.Write(tbf.buf)
+	tbf.buf = append(tbf.buf[:0], b...)
+	if err != nil {
+		return addr, fmt.Errorf("cannot write block to %q: %w", tbf.f.Name(), err)
+	}
+	return addr, nil
+}
+
+func (tbf *tmpBlocksFile) Finalize() error {
+	if tbf.f == nil {
+		return nil
+	}
+	fname := tbf.f.Name()
+	if _, err := tbf.f.Write(tbf.buf); err != nil {
+		return fmt.Errorf("cannot write the remaining %d bytes to %q: %w", len(tbf.buf), fname, err)
+	}
+	tbf.buf = tbf.buf[:0]
+	r, err := fs.OpenReaderAt(fname)
+	if err != nil {
+		logger.Panicf("FATAL: cannot open %q: %s", fname, err)
+	}
+	// Hint the OS that the file is read almost sequentiallly.
+	// This should reduce the number of disk seeks, which is important
+	// for HDDs.
+	r.MustFadviseSequentialRead(true)
+	tbf.r = r
+	return nil
+}
+
+func (tbf *tmpBlocksFile) MustReadBlockRefAt(partRef storage.PartRef, addr tmpBlockAddr) storage.BlockRef {
+	var buf []byte
+	if tbf.f == nil {
+		buf = tbf.buf[addr.offset : addr.offset+uint64(addr.size)]
+	} else {
+		bb := tmpBufPool.Get()
+		defer tmpBufPool.Put(bb)
+		bb.B = bytesutil.Resize(bb.B, addr.size)
+		tbf.r.MustReadAt(bb.B, int64(addr.offset))
+		buf = bb.B
+	}
+	var br storage.BlockRef
+	if err := br.Init(partRef, buf); err != nil {
+		logger.Panicf("FATAL: cannot initialize BlockRef: %s", err)
+	}
+	return br
+}
+
+var tmpBufPool bytesutil.ByteBufferPool
+
+func (tbf *tmpBlocksFile) MustClose() {
+	if tbf.f == nil {
+		return
+	}
+	if tbf.r != nil {
+		// tbf.r could be nil if Finalize wasn't called.
+		tbf.r.MustClose()
+	}
+	fname := tbf.f.Name()
+
+	// Remove the file at first, then close it.
+	// This way the OS shouldn't try to flush file contents to storage
+	// on close.
+	if err := os.Remove(fname); err != nil {
+		logger.Panicf("FATAL: cannot remove %q: %s", fname, err)
+	}
+	if err := tbf.f.Close(); err != nil {
+		logger.Panicf("FATAL: cannot close %q: %s", fname, err)
+	}
+	tbf.f = nil
+}

lib/fs/fadvise_darwin.go

@@ -0,0 +1,10 @@
+package fs
+
+import (
+	"os"
+)
+
+func fadviseSequentialRead(f *os.File, prefetch bool) error {
+	// TODO: implement this properly
+	return nil
+}

lib/fs/fadvise_openbsd.go

@@ -0,0 +1,10 @@
+package fs
+
+import (
+	"os"
+)
+
+func fadviseSequentialRead(f *os.File, prefetch bool) error {
+	// TODO: implement this properly
+	return nil
+}

lib/fs/fadvise_unix.go

@@ -0,0 +1,22 @@
+// +build linux freebsd
+
+package fs
+
+import (
+	"fmt"
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+func fadviseSequentialRead(f *os.File, prefetch bool) error {
+	fd := int(f.Fd())
+	mode := unix.FADV_SEQUENTIAL
+	if prefetch {
+		mode |= unix.FADV_WILLNEED
+	}
+	if err := unix.Fadvise(int(fd), 0, 0, mode); err != nil {
+		return fmt.Errorf("error returned from unix.Fadvise(%d): %w", mode, err)
+	}
+	return nil
+}

lib/fs/reader_at.go

@@ -149,6 +149,15 @@ func (r *ReaderAt) MustClose() {
 	readersCount.Dec()
 }
 
+// MustFadviseSequentialRead hints the OS that f is read mostly sequentially.
+//
+// if prefetch is set, then the OS is hinted to prefetch f data.
+func (r *ReaderAt) MustFadviseSequentialRead(prefetch bool) {
+	if err := fadviseSequentialRead(r.f, prefetch); err != nil {
+		logger.Panicf("FATAL: error in fadviseSequentialRead(%q, %v): %s", r.f.Name(), prefetch, err)
+	}
+}
+
 // OpenReaderAt opens ReaderAt for reading from filename.
 //
 // MustClose must be called on the returned ReaderAt when it is no longer needed.

lib/storage/search.go

@@ -30,6 +30,36 @@ func (br *BlockRef) init(p *part, bh *blockHeader) {
 	br.bh = *bh
 }
 
+// Init initializes br from pr and data
+func (br *BlockRef) Init(pr PartRef, data []byte) error {
+	br.p = pr.p
+	tail, err := br.bh.Unmarshal(data)
+	if err != nil {
+		return err
+	}
+	if len(tail) > 0 {
+		return fmt.Errorf("unexpected non-empty tail left after unmarshaling blockHeader; len(tail)=%d; tail=%q", len(tail), tail)
+	}
+	return nil
+}
+
+// Marshal marshals br to dst.
+func (br *BlockRef) Marshal(dst []byte) []byte {
+	return br.bh.Marshal(dst)
+}
+
+// PartRef returns PartRef from br.
+func (br *BlockRef) PartRef() PartRef {
+	return PartRef{
+		p: br.p,
+	}
+}
+
+// PartRef is Part reference.
+type PartRef struct {
+	p *part
+}
+
 // MustReadBlock reads block from br to dst.
 //
 // if fetchData is false, then only block header is read, otherwise all the data is read.