VictoriaMetrics/lib/storage/storage.go

2367 lines
73 KiB
Go
Raw Normal View History

2019-05-22 23:16:55 +02:00
package storage
import (
"bytes"
"errors"
2019-05-22 23:16:55 +02:00
"fmt"
"io"
2019-05-22 23:16:55 +02:00
"math"
"os"
"path/filepath"
"regexp"
"sort"
"strings"
2019-05-22 23:16:55 +02:00
"sync"
"sync/atomic"
"time"
"unsafe"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/backup/backupnames"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bloomfilter"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
2019-05-22 23:16:55 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
2019-05-22 23:16:55 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/querytracer"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/snapshot"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/uint64set"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
2019-05-22 23:16:55 +02:00
"github.com/VictoriaMetrics/fastcache"
"github.com/VictoriaMetrics/metricsql"
2019-05-22 23:16:55 +02:00
)
const (
msecsPerMonth = 31 * 24 * 3600 * 1000
maxRetentionMsecs = 100 * 12 * msecsPerMonth
)
2019-05-22 23:16:55 +02:00
// Storage represents TSDB storage.
type Storage struct {
// Atomic counters must go at the top of the structure in order to properly align by 8 bytes on 32-bit archs.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212 .
tooSmallTimestampRows uint64
tooBigTimestampRows uint64
slowRowInserts uint64
slowPerDayIndexInserts uint64
slowMetricNameLoads uint64
hourlySeriesLimitRowsDropped uint64
dailySeriesLimitRowsDropped uint64
path string
cachePath string
retentionMsecs int64
2019-05-22 23:16:55 +02:00
2019-05-25 20:51:11 +02:00
// lock file for exclusive access to the storage on the given path.
2019-05-22 23:16:55 +02:00
flockF *os.File
idbCurr atomic.Value
tb *table
// Series cardinality limiters.
hourlySeriesLimiter *bloomfilter.Limiter
dailySeriesLimiter *bloomfilter.Limiter
2019-05-22 23:16:55 +02:00
// tsidCache is MetricName -> TSID cache.
tsidCache *workingsetcache.Cache
2019-05-22 23:16:55 +02:00
// metricIDCache is MetricID -> TSID cache.
metricIDCache *workingsetcache.Cache
2019-05-22 23:16:55 +02:00
// metricNameCache is MetricID -> MetricName cache.
metricNameCache *workingsetcache.Cache
2019-05-22 23:16:55 +02:00
// dateMetricIDCache is (Date, MetricID) cache.
dateMetricIDCache *dateMetricIDCache
2019-05-22 23:16:55 +02:00
2019-11-11 23:16:42 +01:00
// Fast cache for MetricID values occurred during the current hour.
currHourMetricIDs atomic.Value
2019-11-11 23:16:42 +01:00
// Fast cache for MetricID values occurred during the previous hour.
prevHourMetricIDs atomic.Value
// Fast cache for pre-populating per-day inverted index for the next day.
// This is needed in order to remove CPU usage spikes at 00:00 UTC
// due to creation of per-day inverted index for active time series.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/430 for details.
nextDayMetricIDs atomic.Value
// Pending MetricID values to be added to currHourMetricIDs.
2019-11-08 18:37:16 +01:00
pendingHourEntriesLock sync.Mutex
pendingHourEntries *uint64set.Set
// Pending MetricIDs to be added to nextDayMetricIDs.
pendingNextDayMetricIDsLock sync.Mutex
pendingNextDayMetricIDs *uint64set.Set
// prefetchedMetricIDs contains metricIDs for pre-fetched metricNames in the prefetchMetricNames function.
prefetchedMetricIDs atomic.Value
// prefetchedMetricIDsDeadline is used for periodic reset of prefetchedMetricIDs in order to limit its size under high rate of creating new series.
prefetchedMetricIDsDeadline uint64
// prefetchedMetricIDsLock is used for serializing updates of prefetchedMetricIDs from concurrent goroutines.
prefetchedMetricIDsLock sync.Mutex
stop chan struct{}
currHourMetricIDsUpdaterWG sync.WaitGroup
nextDayMetricIDsUpdaterWG sync.WaitGroup
retentionWatcherWG sync.WaitGroup
freeDiskSpaceWatcherWG sync.WaitGroup
// The snapshotLock prevents from concurrent creation of snapshots,
// since this may result in snapshots without recently added data,
// which may be in the process of flushing to disk by concurrently running
// snapshot process.
snapshotLock sync.Mutex
// The minimum timestamp when composite index search can be used.
minTimestampForCompositeIndex int64
// An inmemory set of deleted metricIDs.
//
// It is safe to keep the set in memory even for big number of deleted
// metricIDs, since it usually requires 1 bit per deleted metricID.
deletedMetricIDs atomic.Value
deletedMetricIDsUpdateLock sync.Mutex
isReadOnly uint32
2019-05-22 23:16:55 +02:00
}
// MustOpenStorage opens storage on the given path with the given retentionMsecs.
func MustOpenStorage(path string, retentionMsecs int64, maxHourlySeries, maxDailySeries int) *Storage {
2019-05-22 23:16:55 +02:00
path, err := filepath.Abs(path)
if err != nil {
logger.Panicf("FATAL: cannot determine absolute path for %q: %s", path, err)
2019-05-22 23:16:55 +02:00
}
if retentionMsecs <= 0 {
retentionMsecs = maxRetentionMsecs
}
if retentionMsecs > maxRetentionMsecs {
retentionMsecs = maxRetentionMsecs
}
2019-05-22 23:16:55 +02:00
s := &Storage{
path: path,
cachePath: filepath.Join(path, cacheDirname),
retentionMsecs: retentionMsecs,
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
stop: make(chan struct{}),
2019-05-22 23:16:55 +02:00
}
fs.MustMkdirIfNotExist(path)
2019-05-22 23:16:55 +02:00
// Check whether the cache directory must be removed
// It is removed if it contains resetCacheOnStartupFilename.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1447 for details.
if fs.IsPathExist(filepath.Join(s.cachePath, resetCacheOnStartupFilename)) {
logger.Infof("removing cache directory at %q, since it contains `%s` file...", s.cachePath, resetCacheOnStartupFilename)
all: add Windows build for VictoriaMetrics This commit changes background merge algorithm, so it becomes compatible with Windows file semantics. The previous algorithm for background merge: 1. Merge source parts into a destination part inside tmp directory. 2. Create a file in txn directory with instructions on how to atomically swap source parts with the destination part. 3. Perform instructions from the file. 4. Delete the file with instructions. This algorithm guarantees that either source parts or destination part is visible in the partition after unclean shutdown at any step above, since the remaining files with instructions is replayed on the next restart, after that the remaining contents of the tmp directory is deleted. Unfortunately this algorithm doesn't work under Windows because it disallows removing and moving files, which are in use. So the new algorithm for background merge has been implemented: 1. Merge source parts into a destination part inside the partition directory itself. E.g. now the partition directory may contain both complete and incomplete parts. 2. Atomically update the parts.json file with the new list of parts after the merge, e.g. remove the source parts from the list and add the destination part to the list before storing it to parts.json file. 3. Remove the source parts from disk when they are no longer used. This algorithm guarantees that either source parts or destination part is visible in the partition after unclean shutdown at any step above, since incomplete partitions from step 1 or old source parts from step 3 are removed on the next startup by inspecting parts.json file. This algorithm should work under Windows, since it doesn't remove or move files in use. This algorithm has also the following benefits: - It should work better for NFS. - It fits object storage semantics. The new algorithm changes data storage format, so it is impossible to downgrade to the previous versions of VictoriaMetrics after upgrading to this algorithm. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3236 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3821 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/70
2023-03-19 09:36:05 +01:00
// Do not use fs.MustRemoveAll() here, since the cache directory may be mounted
// to a separate filesystem. In this case the fs.MustRemoveAll() will fail while
// trying to remove the mount root.
fs.RemoveDirContents(s.cachePath)
logger.Infof("cache directory at %q has been successfully removed", s.cachePath)
}
// Protect from concurrent opens.
s.flockF = fs.MustCreateFlockFile(path)
2019-05-22 23:16:55 +02:00
// Check whether restore process finished successfully
restoreLockF := filepath.Join(path, backupnames.RestoreInProgressFilename)
if fs.IsPathExist(restoreLockF) {
logger.Panicf("FATAL: incomplete vmrestore run; run vmrestore again or remove lock file %q", restoreLockF)
}
// Pre-create snapshots directory if it is missing.
snapshotsPath := filepath.Join(path, snapshotsDirname)
fs.MustMkdirIfNotExist(snapshotsPath)
fs.MustRemoveTemporaryDirs(snapshotsPath)
// Initialize series cardinality limiter.
if maxHourlySeries > 0 {
s.hourlySeriesLimiter = bloomfilter.NewLimiter(maxHourlySeries, time.Hour)
}
if maxDailySeries > 0 {
s.dailySeriesLimiter = bloomfilter.NewLimiter(maxDailySeries, 24*time.Hour)
}
2019-05-22 23:16:55 +02:00
// Load caches.
mem := memory.Allowed()
s.tsidCache = s.mustLoadCache("metricName_tsid", getTSIDCacheSize())
s.metricIDCache = s.mustLoadCache("metricID_tsid", mem/16)
s.metricNameCache = s.mustLoadCache("metricID_metricName", mem/10)
s.dateMetricIDCache = newDateMetricIDCache()
2019-05-22 23:16:55 +02:00
hour := fasttime.UnixHour()
hmCurr := s.mustLoadHourMetricIDs(hour, "curr_hour_metric_ids")
hmPrev := s.mustLoadHourMetricIDs(hour-1, "prev_hour_metric_ids")
s.currHourMetricIDs.Store(hmCurr)
s.prevHourMetricIDs.Store(hmPrev)
s.pendingHourEntries = &uint64set.Set{}
date := fasttime.UnixDate()
nextDayMetricIDs := s.mustLoadNextDayMetricIDs(date)
s.nextDayMetricIDs.Store(nextDayMetricIDs)
s.pendingNextDayMetricIDs = &uint64set.Set{}
s.prefetchedMetricIDs.Store(&uint64set.Set{})
// Load metadata
metadataDir := filepath.Join(path, metadataDirname)
isEmptyDB := !fs.IsPathExist(filepath.Join(path, indexdbDirname))
fs.MustMkdirIfNotExist(metadataDir)
s.minTimestampForCompositeIndex = mustGetMinTimestampForCompositeIndex(metadataDir, isEmptyDB)
2019-05-22 23:16:55 +02:00
// Load indexdb
idbPath := filepath.Join(path, indexdbDirname)
idbSnapshotsPath := filepath.Join(idbPath, snapshotsDirname)
fs.MustMkdirIfNotExist(idbSnapshotsPath)
fs.MustRemoveTemporaryDirs(idbSnapshotsPath)
idbCurr, idbPrev := s.mustOpenIndexDBTables(idbPath)
2019-05-22 23:16:55 +02:00
idbCurr.SetExtDB(idbPrev)
s.idbCurr.Store(idbCurr)
// Load deleted metricIDs from idbCurr and idbPrev
dmisCurr, err := idbCurr.loadDeletedMetricIDs()
if err != nil {
logger.Panicf("FATAL: cannot load deleted metricIDs for the current indexDB at %q: %s", path, err)
}
dmisPrev, err := idbPrev.loadDeletedMetricIDs()
if err != nil {
logger.Panicf("FATAL: cannot load deleted metricIDs for the previous indexDB at %q: %s", path, err)
}
s.setDeletedMetricIDs(dmisCurr)
s.updateDeletedMetricIDs(dmisPrev)
// check for free disk space before opening the table
// to prevent unexpected part merges. See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4023
s.startFreeDiskSpaceWatcher()
2019-05-22 23:16:55 +02:00
// Load data
tablePath := filepath.Join(path, dataDirname)
tb := mustOpenTable(tablePath, s)
2019-05-22 23:16:55 +02:00
s.tb = tb
s.startCurrHourMetricIDsUpdater()
s.startNextDayMetricIDsUpdater()
2019-05-22 23:16:55 +02:00
s.startRetentionWatcher()
return s
2019-05-22 23:16:55 +02:00
}
var maxTSIDCacheSize int
// SetTSIDCacheSize overrides the default size of storage/tsid cache
func SetTSIDCacheSize(size int) {
maxTSIDCacheSize = size
}
func getTSIDCacheSize() int {
if maxTSIDCacheSize <= 0 {
return int(float64(memory.Allowed()) * 0.37)
}
return maxTSIDCacheSize
}
func (s *Storage) getDeletedMetricIDs() *uint64set.Set {
return s.deletedMetricIDs.Load().(*uint64set.Set)
}
func (s *Storage) setDeletedMetricIDs(dmis *uint64set.Set) {
s.deletedMetricIDs.Store(dmis)
}
func (s *Storage) updateDeletedMetricIDs(metricIDs *uint64set.Set) {
s.deletedMetricIDsUpdateLock.Lock()
dmisOld := s.getDeletedMetricIDs()
dmisNew := dmisOld.Clone()
dmisNew.Union(metricIDs)
s.setDeletedMetricIDs(dmisNew)
s.deletedMetricIDsUpdateLock.Unlock()
}
// DebugFlush makes sure all the recently added data is visible to search.
//
// Note: this function doesn't store all the in-memory data to disk - it just converts
// recently added items to searchable parts, which can be stored either in memory
// (if they are quite small) or to persistent disk.
//
// This function is for debugging and testing purposes only,
// since it may slow down data ingestion when used frequently.
func (s *Storage) DebugFlush() {
s.tb.flushPendingRows()
2019-05-22 23:16:55 +02:00
s.idb().tb.DebugFlush()
}
// CreateSnapshot creates snapshot for s and returns the snapshot name.
func (s *Storage) CreateSnapshot(deadline uint64) (string, error) {
2019-05-22 23:16:55 +02:00
logger.Infof("creating Storage snapshot for %q...", s.path)
startTime := time.Now()
s.snapshotLock.Lock()
defer s.snapshotLock.Unlock()
var dirsToRemoveOnError []string
defer func() {
for _, dir := range dirsToRemoveOnError {
fs.MustRemoveAll(dir)
}
}()
snapshotName := snapshot.NewName()
2019-05-22 23:16:55 +02:00
srcDir := s.path
dstDir := filepath.Join(srcDir, snapshotsDirname, snapshotName)
fs.MustMkdirFailIfExist(dstDir)
dirsToRemoveOnError = append(dirsToRemoveOnError, dstDir)
2019-05-22 23:16:55 +02:00
smallDir, bigDir, err := s.tb.CreateSnapshot(snapshotName, deadline)
2019-05-22 23:16:55 +02:00
if err != nil {
return "", fmt.Errorf("cannot create table snapshot: %w", err)
2019-05-22 23:16:55 +02:00
}
dirsToRemoveOnError = append(dirsToRemoveOnError, smallDir, bigDir)
dstDataDir := filepath.Join(dstDir, dataDirname)
fs.MustMkdirFailIfExist(dstDataDir)
dstSmallDir := filepath.Join(dstDataDir, smallDirname)
fs.MustSymlinkRelative(smallDir, dstSmallDir)
dstBigDir := filepath.Join(dstDataDir, bigDirname)
fs.MustSymlinkRelative(bigDir, dstBigDir)
fs.MustSyncPath(dstDataDir)
2019-05-22 23:16:55 +02:00
srcMetadataDir := filepath.Join(srcDir, metadataDirname)
dstMetadataDir := filepath.Join(dstDir, metadataDirname)
fs.MustCopyDirectory(srcMetadataDir, dstMetadataDir)
idbSnapshot := filepath.Join(srcDir, indexdbDirname, snapshotsDirname, snapshotName)
2019-05-22 23:16:55 +02:00
idb := s.idb()
currSnapshot := filepath.Join(idbSnapshot, idb.name)
if err := idb.tb.CreateSnapshotAt(currSnapshot, deadline); err != nil {
return "", fmt.Errorf("cannot create curr indexDB snapshot: %w", err)
2019-05-22 23:16:55 +02:00
}
dirsToRemoveOnError = append(dirsToRemoveOnError, idbSnapshot)
2019-05-22 23:16:55 +02:00
ok := idb.doExtDB(func(extDB *indexDB) {
prevSnapshot := filepath.Join(idbSnapshot, extDB.name)
err = extDB.tb.CreateSnapshotAt(prevSnapshot, deadline)
2019-05-22 23:16:55 +02:00
})
if ok && err != nil {
return "", fmt.Errorf("cannot create prev indexDB snapshot: %w", err)
2019-05-22 23:16:55 +02:00
}
dstIdbDir := filepath.Join(dstDir, indexdbDirname)
fs.MustSymlinkRelative(idbSnapshot, dstIdbDir)
2019-05-22 23:16:55 +02:00
fs.MustSyncPath(dstDir)
2019-05-22 23:16:55 +02:00
logger.Infof("created Storage snapshot for %q at %q in %.3f seconds", srcDir, dstDir, time.Since(startTime).Seconds())
dirsToRemoveOnError = nil
2019-05-22 23:16:55 +02:00
return snapshotName, nil
}
// ListSnapshots returns sorted list of existing snapshots for s.
func (s *Storage) ListSnapshots() ([]string, error) {
snapshotsPath := filepath.Join(s.path, snapshotsDirname)
2019-05-22 23:16:55 +02:00
d, err := os.Open(snapshotsPath)
if err != nil {
return nil, fmt.Errorf("cannot open snapshots directory: %w", err)
2019-05-22 23:16:55 +02:00
}
defer fs.MustClose(d)
fnames, err := d.Readdirnames(-1)
if err != nil {
return nil, fmt.Errorf("cannot read snapshots directory at %q: %w", snapshotsPath, err)
2019-05-22 23:16:55 +02:00
}
snapshotNames := make([]string, 0, len(fnames))
for _, fname := range fnames {
if err := snapshot.Validate(fname); err != nil {
2019-05-22 23:16:55 +02:00
continue
}
snapshotNames = append(snapshotNames, fname)
}
sort.Strings(snapshotNames)
return snapshotNames, nil
}
// DeleteSnapshot deletes the given snapshot.
func (s *Storage) DeleteSnapshot(snapshotName string) error {
if err := snapshot.Validate(snapshotName); err != nil {
return fmt.Errorf("invalid snapshotName %q: %w", snapshotName, err)
2019-05-22 23:16:55 +02:00
}
snapshotPath := filepath.Join(s.path, snapshotsDirname, snapshotName)
2019-05-22 23:16:55 +02:00
logger.Infof("deleting snapshot %q...", snapshotPath)
startTime := time.Now()
s.tb.MustDeleteSnapshot(snapshotName)
idbPath := filepath.Join(s.path, indexdbDirname, snapshotsDirname, snapshotName)
fs.MustRemoveDirAtomic(idbPath)
fs.MustRemoveDirAtomic(snapshotPath)
2019-05-22 23:16:55 +02:00
logger.Infof("deleted snapshot %q in %.3f seconds", snapshotPath, time.Since(startTime).Seconds())
2019-05-22 23:16:55 +02:00
return nil
}
// DeleteStaleSnapshots deletes snapshot older than given maxAge
func (s *Storage) DeleteStaleSnapshots(maxAge time.Duration) error {
list, err := s.ListSnapshots()
if err != nil {
return err
}
expireDeadline := time.Now().UTC().Add(-maxAge)
for _, snapshotName := range list {
t, err := snapshot.Time(snapshotName)
if err != nil {
return fmt.Errorf("cannot parse snapshot date from %q: %w", snapshotName, err)
}
if t.Before(expireDeadline) {
if err := s.DeleteSnapshot(snapshotName); err != nil {
return fmt.Errorf("cannot delete snapshot %q: %w", snapshotName, err)
}
}
}
return nil
}
2019-05-22 23:16:55 +02:00
func (s *Storage) idb() *indexDB {
return s.idbCurr.Load().(*indexDB)
}
// Metrics contains essential metrics for the Storage.
type Metrics struct {
RowsAddedTotal uint64
DedupsDuringMerge uint64
TooSmallTimestampRows uint64
TooBigTimestampRows uint64
SlowRowInserts uint64
SlowPerDayIndexInserts uint64
SlowMetricNameLoads uint64
HourlySeriesLimitRowsDropped uint64
HourlySeriesLimitMaxSeries uint64
HourlySeriesLimitCurrentSeries uint64
DailySeriesLimitRowsDropped uint64
DailySeriesLimitMaxSeries uint64
DailySeriesLimitCurrentSeries uint64
TimestampsBlocksMerged uint64
TimestampsBytesSaved uint64
TSIDCacheSize uint64
TSIDCacheSizeBytes uint64
TSIDCacheSizeMaxBytes uint64
TSIDCacheRequests uint64
TSIDCacheMisses uint64
TSIDCacheCollisions uint64
MetricIDCacheSize uint64
MetricIDCacheSizeBytes uint64
MetricIDCacheSizeMaxBytes uint64
MetricIDCacheRequests uint64
MetricIDCacheMisses uint64
MetricIDCacheCollisions uint64
MetricNameCacheSize uint64
MetricNameCacheSizeBytes uint64
MetricNameCacheSizeMaxBytes uint64
MetricNameCacheRequests uint64
MetricNameCacheMisses uint64
MetricNameCacheCollisions uint64
2019-05-22 23:16:55 +02:00
DateMetricIDCacheSize uint64
DateMetricIDCacheSizeBytes uint64
DateMetricIDCacheSyncsCount uint64
DateMetricIDCacheResetsCount uint64
2019-05-22 23:16:55 +02:00
HourMetricIDCacheSize uint64
HourMetricIDCacheSizeBytes uint64
NextDayMetricIDCacheSize uint64
NextDayMetricIDCacheSizeBytes uint64
PrefetchedMetricIDsSize uint64
PrefetchedMetricIDsSizeBytes uint64
NextRetentionSeconds uint64
2019-05-22 23:16:55 +02:00
IndexDBMetrics IndexDBMetrics
TableMetrics TableMetrics
}
// Reset resets m.
func (m *Metrics) Reset() {
*m = Metrics{}
}
// UpdateMetrics updates m with metrics from s.
func (s *Storage) UpdateMetrics(m *Metrics) {
m.RowsAddedTotal = atomic.LoadUint64(&rowsAddedTotal)
m.DedupsDuringMerge = atomic.LoadUint64(&dedupsDuringMerge)
m.TooSmallTimestampRows += atomic.LoadUint64(&s.tooSmallTimestampRows)
m.TooBigTimestampRows += atomic.LoadUint64(&s.tooBigTimestampRows)
m.SlowRowInserts += atomic.LoadUint64(&s.slowRowInserts)
m.SlowPerDayIndexInserts += atomic.LoadUint64(&s.slowPerDayIndexInserts)
m.SlowMetricNameLoads += atomic.LoadUint64(&s.slowMetricNameLoads)
if sl := s.hourlySeriesLimiter; sl != nil {
m.HourlySeriesLimitRowsDropped += atomic.LoadUint64(&s.hourlySeriesLimitRowsDropped)
m.HourlySeriesLimitMaxSeries += uint64(sl.MaxItems())
m.HourlySeriesLimitCurrentSeries += uint64(sl.CurrentItems())
}
if sl := s.dailySeriesLimiter; sl != nil {
m.DailySeriesLimitRowsDropped += atomic.LoadUint64(&s.dailySeriesLimitRowsDropped)
m.DailySeriesLimitMaxSeries += uint64(sl.MaxItems())
m.DailySeriesLimitCurrentSeries += uint64(sl.CurrentItems())
}
m.TimestampsBlocksMerged = atomic.LoadUint64(&timestampsBlocksMerged)
m.TimestampsBytesSaved = atomic.LoadUint64(&timestampsBytesSaved)
2019-05-22 23:16:55 +02:00
var cs fastcache.Stats
s.tsidCache.UpdateStats(&cs)
m.TSIDCacheSize += cs.EntriesCount
m.TSIDCacheSizeBytes += cs.BytesSize
m.TSIDCacheSizeMaxBytes += cs.MaxBytesSize
2019-05-22 23:16:55 +02:00
m.TSIDCacheRequests += cs.GetCalls
m.TSIDCacheMisses += cs.Misses
m.TSIDCacheCollisions += cs.Collisions
cs.Reset()
s.metricIDCache.UpdateStats(&cs)
m.MetricIDCacheSize += cs.EntriesCount
m.MetricIDCacheSizeBytes += cs.BytesSize
m.MetricIDCacheSizeMaxBytes += cs.MaxBytesSize
2019-05-22 23:16:55 +02:00
m.MetricIDCacheRequests += cs.GetCalls
m.MetricIDCacheMisses += cs.Misses
m.MetricIDCacheCollisions += cs.Collisions
cs.Reset()
s.metricNameCache.UpdateStats(&cs)
m.MetricNameCacheSize += cs.EntriesCount
m.MetricNameCacheSizeBytes += cs.BytesSize
m.MetricNameCacheSizeMaxBytes += cs.MaxBytesSize
2019-05-22 23:16:55 +02:00
m.MetricNameCacheRequests += cs.GetCalls
m.MetricNameCacheMisses += cs.Misses
m.MetricNameCacheCollisions += cs.Collisions
m.DateMetricIDCacheSize += uint64(s.dateMetricIDCache.EntriesCount())
m.DateMetricIDCacheSizeBytes += uint64(s.dateMetricIDCache.SizeBytes())
m.DateMetricIDCacheSyncsCount += atomic.LoadUint64(&s.dateMetricIDCache.syncsCount)
m.DateMetricIDCacheResetsCount += atomic.LoadUint64(&s.dateMetricIDCache.resetsCount)
2019-05-22 23:16:55 +02:00
hmCurr := s.currHourMetricIDs.Load().(*hourMetricIDs)
hmPrev := s.prevHourMetricIDs.Load().(*hourMetricIDs)
hourMetricIDsLen := hmPrev.m.Len()
if hmCurr.m.Len() > hourMetricIDsLen {
hourMetricIDsLen = hmCurr.m.Len()
}
m.HourMetricIDCacheSize += uint64(hourMetricIDsLen)
m.HourMetricIDCacheSizeBytes += hmCurr.m.SizeBytes()
m.HourMetricIDCacheSizeBytes += hmPrev.m.SizeBytes()
nextDayMetricIDs := &s.nextDayMetricIDs.Load().(*byDateMetricIDEntry).v
m.NextDayMetricIDCacheSize += uint64(nextDayMetricIDs.Len())
m.NextDayMetricIDCacheSizeBytes += nextDayMetricIDs.SizeBytes()
prefetchedMetricIDs := s.prefetchedMetricIDs.Load().(*uint64set.Set)
m.PrefetchedMetricIDsSize += uint64(prefetchedMetricIDs.Len())
m.PrefetchedMetricIDsSizeBytes += uint64(prefetchedMetricIDs.SizeBytes())
m.NextRetentionSeconds = uint64(nextRetentionDuration(s.retentionMsecs).Seconds())
2019-05-22 23:16:55 +02:00
s.idb().UpdateMetrics(&m.IndexDBMetrics)
s.tb.UpdateMetrics(&m.TableMetrics)
}
// SetFreeDiskSpaceLimit sets the minimum free disk space size of current storage path
//
// The function must be called before opening or creating any storage.
func SetFreeDiskSpaceLimit(bytes int64) {
freeDiskSpaceLimitBytes = uint64(bytes)
}
var freeDiskSpaceLimitBytes uint64
// IsReadOnly returns information is storage in read only mode
func (s *Storage) IsReadOnly() bool {
return atomic.LoadUint32(&s.isReadOnly) == 1
}
func (s *Storage) startFreeDiskSpaceWatcher() {
f := func() {
freeSpaceBytes := fs.MustGetFreeSpace(s.path)
if freeSpaceBytes < freeDiskSpaceLimitBytes {
// Switch the storage to readonly mode if there is no enough free space left at s.path
logger.Warnf("switching the storage at %s to read-only mode, since it has less than -storage.minFreeDiskSpaceBytes=%d of free space: %d bytes left",
s.path, freeDiskSpaceLimitBytes, freeSpaceBytes)
atomic.StoreUint32(&s.isReadOnly, 1)
return
}
if atomic.CompareAndSwapUint32(&s.isReadOnly, 1, 0) {
logger.Warnf("enabling writing to the storage at %s, since it has more than -storage.minFreeDiskSpaceBytes=%d of free space: %d bytes left",
s.path, freeDiskSpaceLimitBytes, freeSpaceBytes)
}
}
f()
s.freeDiskSpaceWatcherWG.Add(1)
go func() {
defer s.freeDiskSpaceWatcherWG.Done()
ticker := time.NewTicker(time.Second)
defer ticker.Stop()
for {
select {
case <-s.stop:
return
case <-ticker.C:
f()
}
}
}()
}
2019-05-22 23:16:55 +02:00
func (s *Storage) startRetentionWatcher() {
s.retentionWatcherWG.Add(1)
go func() {
s.retentionWatcher()
s.retentionWatcherWG.Done()
}()
}
func (s *Storage) retentionWatcher() {
for {
d := nextRetentionDuration(s.retentionMsecs)
2019-05-22 23:16:55 +02:00
select {
case <-s.stop:
return
case <-time.After(d):
s.mustRotateIndexDB()
}
}
}
func (s *Storage) startCurrHourMetricIDsUpdater() {
s.currHourMetricIDsUpdaterWG.Add(1)
go func() {
s.currHourMetricIDsUpdater()
s.currHourMetricIDsUpdaterWG.Done()
}()
}
func (s *Storage) startNextDayMetricIDsUpdater() {
s.nextDayMetricIDsUpdaterWG.Add(1)
go func() {
s.nextDayMetricIDsUpdater()
s.nextDayMetricIDsUpdaterWG.Done()
}()
}
var currHourMetricIDsUpdateInterval = time.Second * 10
func (s *Storage) currHourMetricIDsUpdater() {
ticker := time.NewTicker(currHourMetricIDsUpdateInterval)
defer ticker.Stop()
for {
select {
case <-s.stop:
hour := fasttime.UnixHour()
s.updateCurrHourMetricIDs(hour)
return
case <-ticker.C:
hour := fasttime.UnixHour()
s.updateCurrHourMetricIDs(hour)
}
}
}
var nextDayMetricIDsUpdateInterval = time.Second * 11
func (s *Storage) nextDayMetricIDsUpdater() {
ticker := time.NewTicker(nextDayMetricIDsUpdateInterval)
defer ticker.Stop()
for {
select {
case <-s.stop:
date := fasttime.UnixDate()
s.updateNextDayMetricIDs(date)
return
case <-ticker.C:
date := fasttime.UnixDate()
s.updateNextDayMetricIDs(date)
}
}
}
2019-05-22 23:16:55 +02:00
func (s *Storage) mustRotateIndexDB() {
// Create new indexdb table.
newTableName := nextIndexDBTableName()
idbNewPath := filepath.Join(s.path, indexdbDirname, newTableName)
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
rotationTimestamp := fasttime.UnixTimestamp()
idbNew := mustOpenIndexDB(idbNewPath, s, rotationTimestamp, &s.isReadOnly)
2019-05-22 23:16:55 +02:00
// Drop extDB
idbCurr := s.idb()
idbCurr.doExtDB(func(extDB *indexDB) {
extDB.scheduleToDrop()
})
idbCurr.SetExtDB(nil)
// Start using idbNew
idbNew.SetExtDB(idbCurr)
s.idbCurr.Store(idbNew)
// Persist changes on the file system.
fs.MustSyncPath(s.path)
2019-05-22 23:16:55 +02:00
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
// Do not flush tsidCache to avoid read/write path slowdown
// and slowly re-populate new idb with entries from the cache via maybeCreateIndexes().
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
2019-05-22 23:16:55 +02:00
// Flush metric id caches for the current and the previous hour,
// since they may contain entries missing in idbNew.
// This should prevent from missing data in queries when
// the following steps are performed for short -retentionPeriod (e.g. 1 day):
//
// 1. Add samples for some series between 3-4 UTC. These series are registered in currHourMetricIDs.
// 2. The indexdb rotation is performed at 4 UTC. currHourMetricIDs is moved to prevHourMetricIDs.
// 3. Continue adding samples for series from step 1 during time range 4-5 UTC.
// These series are already registered in prevHourMetricIDs, so VM doesn't add per-day entries to the current indexdb.
// 4. Stop adding new samples for these series just before 5 UTC.
// 5. The next indexdb rotation is performed at 4 UTC next day.
// The information about the series from step 5 disappears from indexdb, since the old indexdb from step 1 is deleted,
// while the current indexdb doesn't contain information about the series.
// So queries for the last 24 hours stop returning samples added at step 3.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
s.pendingHourEntriesLock.Lock()
s.pendingHourEntries = &uint64set.Set{}
s.pendingHourEntriesLock.Unlock()
s.currHourMetricIDs.Store(&hourMetricIDs{})
s.prevHourMetricIDs.Store(&hourMetricIDs{})
2019-05-22 23:16:55 +02:00
// Flush dateMetricIDCache, so idbNew can be populated with fresh data.
s.dateMetricIDCache.Reset()
// Do not flush metricIDCache and metricNameCache, since all the metricIDs
// from prev idb remain valid after the rotation.
// There is no need in resetting nextDayMetricIDs, since it should be automatically reset every day.
2019-05-22 23:16:55 +02:00
}
func (s *Storage) resetAndSaveTSIDCache() {
// Reset cache and then store the reset cache on disk in order to prevent
// from inconsistent behaviour after possible unclean shutdown.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1347
s.tsidCache.Reset()
s.mustSaveCache(s.tsidCache, "metricName_tsid")
}
2019-05-22 23:16:55 +02:00
// MustClose closes the storage.
//
// It is expected that the s is no longer used during the close.
2019-05-22 23:16:55 +02:00
func (s *Storage) MustClose() {
close(s.stop)
s.freeDiskSpaceWatcherWG.Wait()
2019-05-22 23:16:55 +02:00
s.retentionWatcherWG.Wait()
s.currHourMetricIDsUpdaterWG.Wait()
s.nextDayMetricIDsUpdaterWG.Wait()
2019-05-22 23:16:55 +02:00
s.tb.MustClose()
s.idb().MustClose()
// Save caches.
s.mustSaveCache(s.tsidCache, "metricName_tsid")
s.tsidCache.Stop()
s.mustSaveCache(s.metricIDCache, "metricID_tsid")
s.metricIDCache.Stop()
s.mustSaveCache(s.metricNameCache, "metricID_metricName")
s.metricNameCache.Stop()
2019-05-22 23:16:55 +02:00
hmCurr := s.currHourMetricIDs.Load().(*hourMetricIDs)
s.mustSaveHourMetricIDs(hmCurr, "curr_hour_metric_ids")
hmPrev := s.prevHourMetricIDs.Load().(*hourMetricIDs)
s.mustSaveHourMetricIDs(hmPrev, "prev_hour_metric_ids")
nextDayMetricIDs := s.nextDayMetricIDs.Load().(*byDateMetricIDEntry)
s.mustSaveNextDayMetricIDs(nextDayMetricIDs)
2019-05-22 23:16:55 +02:00
// Release lock file.
fs.MustClose(s.flockF)
s.flockF = nil
// Stop series limiters.
if sl := s.hourlySeriesLimiter; sl != nil {
sl.MustStop()
}
if sl := s.dailySeriesLimiter; sl != nil {
sl.MustStop()
}
2019-05-22 23:16:55 +02:00
}
func (s *Storage) mustLoadNextDayMetricIDs(date uint64) *byDateMetricIDEntry {
e := &byDateMetricIDEntry{
date: date,
}
name := "next_day_metric_ids"
path := filepath.Join(s.cachePath, name)
if !fs.IsPathExist(path) {
logger.Infof("nothing to load from %q", path)
return e
}
src, err := os.ReadFile(path)
if err != nil {
logger.Panicf("FATAL: cannot read %s: %s", path, err)
}
if len(src) < 16 {
logger.Errorf("discarding %s, since it has broken header; got %d bytes; want %d bytes", path, len(src), 16)
return e
}
// Unmarshal header
dateLoaded := encoding.UnmarshalUint64(src)
src = src[8:]
if dateLoaded != date {
logger.Infof("discarding %s, since it contains data for stale date; got %d; want %d", path, dateLoaded, date)
return e
}
// Unmarshal uint64set
m, tail, err := unmarshalUint64Set(src)
if err != nil {
logger.Infof("discarding %s because cannot load uint64set: %s", path, err)
return e
}
if len(tail) > 0 {
logger.Infof("discarding %s because non-empty tail left; len(tail)=%d", path, len(tail))
return e
}
e.v = *m
return e
}
func (s *Storage) mustLoadHourMetricIDs(hour uint64, name string) *hourMetricIDs {
hm := &hourMetricIDs{
hour: hour,
}
path := filepath.Join(s.cachePath, name)
if !fs.IsPathExist(path) {
logger.Infof("nothing to load from %q", path)
return hm
}
src, err := os.ReadFile(path)
if err != nil {
logger.Panicf("FATAL: cannot read %s: %s", path, err)
}
if len(src) < 16 {
logger.Errorf("discarding %s, since it has broken header; got %d bytes; want %d bytes", path, len(src), 16)
return hm
}
2019-11-08 18:37:16 +01:00
// Unmarshal header
hourLoaded := encoding.UnmarshalUint64(src)
src = src[8:]
if hourLoaded != hour {
logger.Infof("discarding %s, since it contains outdated hour; got %d; want %d", path, hourLoaded, hour)
return hm
}
2019-11-08 18:37:16 +01:00
// Unmarshal uint64set
m, tail, err := unmarshalUint64Set(src)
if err != nil {
logger.Infof("discarding %s because cannot load uint64set: %s", path, err)
return hm
}
if len(tail) > 0 {
logger.Infof("discarding %s because non-empty tail left; len(tail)=%d", path, len(tail))
return hm
}
hm.m = m
return hm
}
func (s *Storage) mustSaveNextDayMetricIDs(e *byDateMetricIDEntry) {
name := "next_day_metric_ids"
path := filepath.Join(s.cachePath, name)
dst := make([]byte, 0, e.v.Len()*8+16)
// Marshal header
dst = encoding.MarshalUint64(dst, e.date)
// Marshal e.v
dst = marshalUint64Set(dst, &e.v)
if err := os.WriteFile(path, dst, 0644); err != nil {
logger.Panicf("FATAL: cannot write %d bytes to %q: %s", len(dst), path, err)
}
}
func (s *Storage) mustSaveHourMetricIDs(hm *hourMetricIDs, name string) {
path := filepath.Join(s.cachePath, name)
dst := make([]byte, 0, hm.m.Len()*8+24)
2019-11-08 18:37:16 +01:00
// Marshal header
dst = encoding.MarshalUint64(dst, hm.hour)
2019-11-08 18:37:16 +01:00
// Marshal hm.m
dst = marshalUint64Set(dst, hm.m)
if err := os.WriteFile(path, dst, 0644); err != nil {
logger.Panicf("FATAL: cannot write %d bytes to %q: %s", len(dst), path, err)
}
}
func unmarshalUint64Set(src []byte) (*uint64set.Set, []byte, error) {
mLen := encoding.UnmarshalUint64(src)
src = src[8:]
if uint64(len(src)) < 8*mLen {
return nil, nil, fmt.Errorf("cannot unmarshal uint64set; got %d bytes; want at least %d bytes", len(src), 8*mLen)
}
m := &uint64set.Set{}
for i := uint64(0); i < mLen; i++ {
metricID := encoding.UnmarshalUint64(src)
src = src[8:]
m.Add(metricID)
}
return m, src, nil
}
func marshalUint64Set(dst []byte, m *uint64set.Set) []byte {
dst = encoding.MarshalUint64(dst, uint64(m.Len()))
m.ForEach(func(part []uint64) bool {
for _, metricID := range part {
dst = encoding.MarshalUint64(dst, metricID)
}
return true
})
return dst
}
func mustGetMinTimestampForCompositeIndex(metadataDir string, isEmptyDB bool) int64 {
path := filepath.Join(metadataDir, "minTimestampForCompositeIndex")
minTimestamp, err := loadMinTimestampForCompositeIndex(path)
if err == nil {
return minTimestamp
}
if !os.IsNotExist(err) {
logger.Errorf("cannot read minTimestampForCompositeIndex, so trying to re-create it; error: %s", err)
}
date := time.Now().UnixNano() / 1e6 / msecPerDay
if !isEmptyDB {
// The current and the next day can already contain non-composite indexes,
// so they cannot be queried with composite indexes.
date += 2
} else {
date = 0
}
minTimestamp = date * msecPerDay
dateBuf := encoding.MarshalInt64(nil, minTimestamp)
fs.MustWriteAtomic(path, dateBuf, true)
return minTimestamp
}
func loadMinTimestampForCompositeIndex(path string) (int64, error) {
data, err := os.ReadFile(path)
if err != nil {
return 0, err
}
if len(data) != 8 {
return 0, fmt.Errorf("unexpected length of %q; got %d bytes; want 8 bytes", path, len(data))
}
return encoding.UnmarshalInt64(data), nil
}
func (s *Storage) mustLoadCache(name string, sizeBytes int) *workingsetcache.Cache {
path := filepath.Join(s.cachePath, name)
return workingsetcache.Load(path, sizeBytes)
2019-05-22 23:16:55 +02:00
}
func (s *Storage) mustSaveCache(c *workingsetcache.Cache, name string) {
saveCacheLock.Lock()
defer saveCacheLock.Unlock()
path := filepath.Join(s.cachePath, name)
if err := c.Save(path); err != nil {
logger.Panicf("FATAL: cannot save cache to %q: %s", path, err)
2019-05-22 23:16:55 +02:00
}
}
// saveCacheLock prevents from data races when multiple concurrent goroutines save the same cache.
var saveCacheLock sync.Mutex
// SetRetentionTimezoneOffset sets the offset, which is used for calculating the time for indexdb rotation.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/pull/2574
func SetRetentionTimezoneOffset(offset time.Duration) {
retentionTimezoneOffsetMsecs = offset.Milliseconds()
}
var retentionTimezoneOffsetMsecs int64
func nextRetentionDuration(retentionMsecs int64) time.Duration {
nowMsecs := time.Now().UnixNano() / 1e6
return nextRetentionDurationAt(nowMsecs, retentionMsecs)
}
func nextRetentionDurationAt(atMsecs int64, retentionMsecs int64) time.Duration {
// Round retentionMsecs to days. This guarantees that per-day inverted index works as expected
retentionMsecs = ((retentionMsecs + msecPerDay - 1) / msecPerDay) * msecPerDay
// The effect of time zone on retention period is moved out.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/pull/2574
deadline := ((atMsecs + retentionMsecs + retentionTimezoneOffsetMsecs - 1) / retentionMsecs) * retentionMsecs
// Schedule the deadline to +4 hours from the next retention period start.
// This should prevent from possible double deletion of indexdb
// due to time drift - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/248 .
deadline += int64(4 * 3600 * 1000)
deadline -= retentionTimezoneOffsetMsecs
return time.Duration(deadline-atMsecs) * time.Millisecond
2019-05-22 23:16:55 +02:00
}
// SearchMetricNames returns marshaled metric names matching the given tfss on the given tr.
//
// The marshaled metric names must be unmarshaled via MetricName.UnmarshalString().
func (s *Storage) SearchMetricNames(qt *querytracer.Tracer, tfss []*TagFilters, tr TimeRange, maxMetrics int, deadline uint64) ([]string, error) {
qt = qt.NewChild("search for matching metric names: filters=%s, timeRange=%s", tfss, &tr)
defer qt.Done()
metricIDs, err := s.idb().searchMetricIDs(qt, tfss, tr, maxMetrics, deadline)
if err != nil {
return nil, err
}
if len(metricIDs) == 0 {
return nil, nil
}
if err = s.prefetchMetricNames(qt, metricIDs, deadline); err != nil {
return nil, err
}
idb := s.idb()
metricNames := make([]string, 0, len(metricIDs))
metricNamesSeen := make(map[string]struct{}, len(metricIDs))
var metricName []byte
for i, metricID := range metricIDs {
if i&paceLimiterSlowIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(deadline); err != nil {
return nil, err
}
}
var err error
metricName, err = idb.searchMetricNameWithCache(metricName[:0], metricID)
if err != nil {
if err == io.EOF {
// Skip missing metricName for metricID.
// It should be automatically fixed. See indexDB.searchMetricName for details.
continue
}
return nil, fmt.Errorf("error when searching metricName for metricID=%d: %w", metricID, err)
}
if _, ok := metricNamesSeen[string(metricName)]; ok {
// The given metric name was already seen; skip it
continue
}
metricNames = append(metricNames, string(metricName))
metricNamesSeen[metricNames[len(metricNames)-1]] = struct{}{}
}
qt.Printf("loaded %d metric names", len(metricNames))
return metricNames, nil
}
// prefetchMetricNames pre-fetches metric names for the given metricIDs into metricID->metricName cache.
//
// This should speed-up further searchMetricNameWithCache calls for srcMetricIDs from tsids.
func (s *Storage) prefetchMetricNames(qt *querytracer.Tracer, srcMetricIDs []uint64, deadline uint64) error {
qt = qt.NewChild("prefetch metric names for %d metricIDs", len(srcMetricIDs))
defer qt.Done()
if len(srcMetricIDs) == 0 {
qt.Printf("nothing to prefetch")
return nil
}
var metricIDs uint64Sorter
prefetchedMetricIDs := s.prefetchedMetricIDs.Load().(*uint64set.Set)
for _, metricID := range srcMetricIDs {
if prefetchedMetricIDs.Has(metricID) {
continue
}
metricIDs = append(metricIDs, metricID)
}
qt.Printf("%d out of %d metric names must be pre-fetched", len(metricIDs), len(srcMetricIDs))
if len(metricIDs) < 500 {
// It is cheaper to skip pre-fetching and obtain metricNames inline.
qt.Printf("skip pre-fetching metric names for low number of metrid ids=%d", len(metricIDs))
return nil
}
atomic.AddUint64(&s.slowMetricNameLoads, uint64(len(metricIDs)))
// Pre-fetch metricIDs.
sort.Sort(metricIDs)
var missingMetricIDs []uint64
var metricName []byte
var err error
idb := s.idb()
is := idb.getIndexSearch(deadline)
defer idb.putIndexSearch(is)
for loops, metricID := range metricIDs {
if loops&paceLimiterSlowIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
metricName, err = is.searchMetricNameWithCache(metricName[:0], metricID)
if err != nil {
if err == io.EOF {
missingMetricIDs = append(missingMetricIDs, metricID)
continue
}
return fmt.Errorf("error in pre-fetching metricName for metricID=%d: %w", metricID, err)
}
}
idb.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(deadline)
defer extDB.putIndexSearch(is)
for loops, metricID := range missingMetricIDs {
if loops&paceLimiterSlowIterationsMask == 0 {
if err = checkSearchDeadlineAndPace(is.deadline); err != nil {
return
}
}
metricName, err = is.searchMetricNameWithCache(metricName[:0], metricID)
if err != nil && err != io.EOF {
err = fmt.Errorf("error in pre-fetching metricName for metricID=%d in extDB: %w", metricID, err)
return
}
}
})
if err != nil && err != io.EOF {
return err
}
qt.Printf("pre-fetch metric names for %d metric ids", len(metricIDs))
// Store the pre-fetched metricIDs, so they aren't pre-fetched next time.
s.prefetchedMetricIDsLock.Lock()
var prefetchedMetricIDsNew *uint64set.Set
if fasttime.UnixTimestamp() < atomic.LoadUint64(&s.prefetchedMetricIDsDeadline) {
// Periodically reset the prefetchedMetricIDs in order to limit its size.
prefetchedMetricIDsNew = &uint64set.Set{}
atomic.StoreUint64(&s.prefetchedMetricIDsDeadline, fasttime.UnixTimestamp()+73*60)
} else {
prefetchedMetricIDsNew = prefetchedMetricIDs.Clone()
}
prefetchedMetricIDsNew.AddMulti(metricIDs)
if prefetchedMetricIDsNew.SizeBytes() > uint64(memory.Allowed())/32 {
// Reset prefetchedMetricIDsNew if it occupies too much space.
prefetchedMetricIDsNew = &uint64set.Set{}
}
s.prefetchedMetricIDs.Store(prefetchedMetricIDsNew)
s.prefetchedMetricIDsLock.Unlock()
qt.Printf("cache metric ids for pre-fetched metric names")
return nil
}
// ErrDeadlineExceeded is returned when the request times out.
var ErrDeadlineExceeded = fmt.Errorf("deadline exceeded")
// DeleteSeries deletes all the series matching the given tfss.
2019-05-22 23:16:55 +02:00
//
// Returns the number of metrics deleted.
func (s *Storage) DeleteSeries(qt *querytracer.Tracer, tfss []*TagFilters) (int, error) {
deletedCount, err := s.idb().DeleteTSIDs(qt, tfss)
2019-05-22 23:16:55 +02:00
if err != nil {
return deletedCount, fmt.Errorf("cannot delete tsids: %w", err)
2019-05-22 23:16:55 +02:00
}
// Do not reset MetricName->TSID cache in order to prevent from adding new data points
// to deleted time series in Storage.add, since it is already reset inside DeleteTSIDs.
// Do not reset MetricID->MetricName cache, since it must be used only
2019-05-22 23:16:55 +02:00
// after filtering out deleted metricIDs.
2019-05-22 23:16:55 +02:00
return deletedCount, nil
}
// SearchLabelNamesWithFiltersOnTimeRange searches for label names matching the given tfss on tr.
2023-04-10 19:16:36 +02:00
func (s *Storage) SearchLabelNamesWithFiltersOnTimeRange(qt *querytracer.Tracer, tfss []*TagFilters, tr TimeRange, maxLabelNames, maxMetrics int, deadline uint64,
) ([]string, error) {
return s.idb().SearchLabelNamesWithFiltersOnTimeRange(qt, tfss, tr, maxLabelNames, maxMetrics, deadline)
}
// SearchLabelValuesWithFiltersOnTimeRange searches for label values for the given labelName, filters and tr.
func (s *Storage) SearchLabelValuesWithFiltersOnTimeRange(qt *querytracer.Tracer, labelName string, tfss []*TagFilters,
2023-04-10 19:16:36 +02:00
tr TimeRange, maxLabelValues, maxMetrics int, deadline uint64,
) ([]string, error) {
return s.idb().SearchLabelValuesWithFiltersOnTimeRange(qt, labelName, tfss, tr, maxLabelValues, maxMetrics, deadline)
2019-05-22 23:16:55 +02:00
}
// SearchTagValueSuffixes returns all the tag value suffixes for the given tagKey and tagValuePrefix on the given tr.
//
// This allows implementing https://graphite-api.readthedocs.io/en/latest/api.html#metrics-find or similar APIs.
//
// If more than maxTagValueSuffixes suffixes is found, then only the first maxTagValueSuffixes suffixes is returned.
func (s *Storage) SearchTagValueSuffixes(qt *querytracer.Tracer, tr TimeRange, tagKey, tagValuePrefix string,
2023-04-10 19:16:36 +02:00
delimiter byte, maxTagValueSuffixes int, deadline uint64,
) ([]string, error) {
return s.idb().SearchTagValueSuffixes(qt, tr, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes, deadline)
}
// SearchGraphitePaths returns all the matching paths for the given graphite query on the given tr.
func (s *Storage) SearchGraphitePaths(qt *querytracer.Tracer, tr TimeRange, query []byte, maxPaths int, deadline uint64) ([]string, error) {
query = replaceAlternateRegexpsWithGraphiteWildcards(query)
return s.searchGraphitePaths(qt, tr, nil, query, maxPaths, deadline)
}
// replaceAlternateRegexpsWithGraphiteWildcards replaces (foo|..|bar) with {foo,...,bar} in b and returns the new value.
func replaceAlternateRegexpsWithGraphiteWildcards(b []byte) []byte {
var dst []byte
for {
n := bytes.IndexByte(b, '(')
if n < 0 {
if len(dst) == 0 {
// Fast path - b doesn't contain the openining brace.
return b
}
dst = append(dst, b...)
return dst
}
dst = append(dst, b[:n]...)
b = b[n+1:]
n = bytes.IndexByte(b, ')')
if n < 0 {
dst = append(dst, '(')
dst = append(dst, b...)
return dst
}
x := b[:n]
b = b[n+1:]
if string(x) == ".*" {
dst = append(dst, '*')
continue
}
dst = append(dst, '{')
for len(x) > 0 {
n = bytes.IndexByte(x, '|')
if n < 0 {
dst = append(dst, x...)
break
}
dst = append(dst, x[:n]...)
x = x[n+1:]
dst = append(dst, ',')
}
dst = append(dst, '}')
}
}
func (s *Storage) searchGraphitePaths(qt *querytracer.Tracer, tr TimeRange, qHead, qTail []byte, maxPaths int, deadline uint64) ([]string, error) {
n := bytes.IndexAny(qTail, "*[{")
if n < 0 {
// Verify that qHead matches a metric name.
qHead = append(qHead, qTail...)
suffixes, err := s.SearchTagValueSuffixes(qt, tr, "", bytesutil.ToUnsafeString(qHead), '.', 1, deadline)
if err != nil {
return nil, err
}
if len(suffixes) == 0 {
// The query doesn't match anything.
return nil, nil
}
if len(suffixes[0]) > 0 {
// The query matches a metric name with additional suffix.
return nil, nil
}
return []string{string(qHead)}, nil
}
qHead = append(qHead, qTail[:n]...)
suffixes, err := s.SearchTagValueSuffixes(qt, tr, "", bytesutil.ToUnsafeString(qHead), '.', maxPaths, deadline)
if err != nil {
return nil, err
}
if len(suffixes) == 0 {
return nil, nil
}
if len(suffixes) >= maxPaths {
return nil, fmt.Errorf("more than maxPaths=%d suffixes found", maxPaths)
}
qNode := qTail[n:]
qTail = nil
mustMatchLeafs := true
if m := bytes.IndexByte(qNode, '.'); m >= 0 {
qTail = qNode[m+1:]
qNode = qNode[:m+1]
mustMatchLeafs = false
}
re, err := getRegexpForGraphiteQuery(string(qNode))
if err != nil {
return nil, err
}
qHeadLen := len(qHead)
var paths []string
for _, suffix := range suffixes {
if len(paths) > maxPaths {
return nil, fmt.Errorf("more than maxPath=%d paths found", maxPaths)
}
if !re.MatchString(suffix) {
continue
}
if mustMatchLeafs {
qHead = append(qHead[:qHeadLen], suffix...)
paths = append(paths, string(qHead))
continue
}
qHead = append(qHead[:qHeadLen], suffix...)
ps, err := s.searchGraphitePaths(qt, tr, qHead, qTail, maxPaths, deadline)
if err != nil {
return nil, err
}
paths = append(paths, ps...)
}
return paths, nil
}
func getRegexpForGraphiteQuery(q string) (*regexp.Regexp, error) {
parts, tail := getRegexpPartsForGraphiteQuery(q)
if len(tail) > 0 {
return nil, fmt.Errorf("unexpected tail left after parsing %q: %q", q, tail)
}
reStr := "^" + strings.Join(parts, "") + "$"
return metricsql.CompileRegexp(reStr)
}
func getRegexpPartsForGraphiteQuery(q string) ([]string, string) {
var parts []string
for {
n := strings.IndexAny(q, "*{}[,")
if n < 0 {
parts = append(parts, regexp.QuoteMeta(q))
return parts, ""
}
parts = append(parts, regexp.QuoteMeta(q[:n]))
q = q[n:]
switch q[0] {
case ',', '}':
return parts, q
case '*':
parts = append(parts, "[^.]*")
q = q[1:]
case '{':
var tmp []string
for {
a, tail := getRegexpPartsForGraphiteQuery(q[1:])
tmp = append(tmp, strings.Join(a, ""))
if len(tail) == 0 {
parts = append(parts, regexp.QuoteMeta("{"))
parts = append(parts, strings.Join(tmp, ","))
return parts, ""
}
if tail[0] == ',' {
q = tail
continue
}
if tail[0] == '}' {
if len(tmp) == 1 {
parts = append(parts, tmp[0])
} else {
parts = append(parts, "(?:"+strings.Join(tmp, "|")+")")
}
q = tail[1:]
break
}
logger.Panicf("BUG: unexpected first char at tail %q; want `.` or `}`", tail)
}
case '[':
n := strings.IndexByte(q, ']')
if n < 0 {
parts = append(parts, regexp.QuoteMeta(q))
return parts, ""
}
parts = append(parts, q[:n+1])
q = q[n+1:]
}
}
}
2019-05-22 23:16:55 +02:00
// GetSeriesCount returns the approximate number of unique time series.
//
// It includes the deleted series too and may count the same series
// up to two times - in db and extDB.
func (s *Storage) GetSeriesCount(deadline uint64) (uint64, error) {
return s.idb().GetSeriesCount(deadline)
2019-05-22 23:16:55 +02:00
}
// GetTSDBStatus returns TSDB status data for /api/v1/status/tsdb
func (s *Storage) GetTSDBStatus(qt *querytracer.Tracer, tfss []*TagFilters, date uint64, focusLabel string, topN, maxMetrics int, deadline uint64) (*TSDBStatus, error) {
return s.idb().GetTSDBStatus(qt, tfss, date, focusLabel, topN, maxMetrics, deadline)
}
2019-05-22 23:16:55 +02:00
// MetricRow is a metric to insert into storage.
type MetricRow struct {
// MetricNameRaw contains raw metric name, which must be decoded
// with MetricName.UnmarshalRaw.
2019-05-22 23:16:55 +02:00
MetricNameRaw []byte
Timestamp int64
Value float64
}
// CopyFrom copies src to mr.
func (mr *MetricRow) CopyFrom(src *MetricRow) {
mr.MetricNameRaw = append(mr.MetricNameRaw[:0], src.MetricNameRaw...)
mr.Timestamp = src.Timestamp
mr.Value = src.Value
}
// String returns string representation of the mr.
func (mr *MetricRow) String() string {
metricName := string(mr.MetricNameRaw)
var mn MetricName
if err := mn.UnmarshalRaw(mr.MetricNameRaw); err == nil {
2019-05-22 23:16:55 +02:00
metricName = mn.String()
}
return fmt.Sprintf("%s (Timestamp=%d, Value=%f)", metricName, mr.Timestamp, mr.Value)
2019-05-22 23:16:55 +02:00
}
// Marshal appends marshaled mr to dst and returns the result.
func (mr *MetricRow) Marshal(dst []byte) []byte {
dst = encoding.MarshalBytes(dst, mr.MetricNameRaw)
dst = encoding.MarshalUint64(dst, uint64(mr.Timestamp))
dst = encoding.MarshalUint64(dst, math.Float64bits(mr.Value))
return dst
}
// UnmarshalX unmarshals mr from src and returns the remaining tail from src.
//
// mr refers to src, so it remains valid until src changes.
func (mr *MetricRow) UnmarshalX(src []byte) ([]byte, error) {
2019-05-22 23:16:55 +02:00
tail, metricNameRaw, err := encoding.UnmarshalBytes(src)
if err != nil {
return tail, fmt.Errorf("cannot unmarshal MetricName: %w", err)
2019-05-22 23:16:55 +02:00
}
mr.MetricNameRaw = metricNameRaw
2019-05-22 23:16:55 +02:00
if len(tail) < 8 {
return tail, fmt.Errorf("cannot unmarshal Timestamp: want %d bytes; have %d bytes", 8, len(tail))
}
timestamp := encoding.UnmarshalUint64(tail)
mr.Timestamp = int64(timestamp)
tail = tail[8:]
if len(tail) < 8 {
return tail, fmt.Errorf("cannot unmarshal Value: want %d bytes; have %d bytes", 8, len(tail))
}
value := encoding.UnmarshalUint64(tail)
mr.Value = math.Float64frombits(value)
tail = tail[8:]
return tail, nil
}
// ForceMergePartitions force-merges partitions in s with names starting from the given partitionNamePrefix.
//
// Partitions are merged sequentially in order to reduce load on the system.
func (s *Storage) ForceMergePartitions(partitionNamePrefix string) error {
return s.tb.ForceMergePartitions(partitionNamePrefix)
}
var rowsAddedTotal uint64
2019-05-22 23:16:55 +02:00
// AddRows adds the given mrs to s.
//
// The caller should limit the number of concurrent AddRows calls to the number
// of available CPU cores in order to limit memory usage.
2019-05-22 23:16:55 +02:00
func (s *Storage) AddRows(mrs []MetricRow, precisionBits uint8) error {
if len(mrs) == 0 {
return nil
}
// Add rows to the storage in blocks with limited size in order to reduce memory usage.
var firstErr error
ic := getMetricRowsInsertCtx()
maxBlockLen := len(ic.rrs)
for len(mrs) > 0 {
mrsBlock := mrs
if len(mrs) > maxBlockLen {
mrsBlock = mrs[:maxBlockLen]
mrs = mrs[maxBlockLen:]
} else {
mrs = nil
}
if err := s.add(ic.rrs, ic.tmpMrs, mrsBlock, precisionBits); err != nil {
if firstErr == nil {
firstErr = err
}
continue
}
atomic.AddUint64(&rowsAddedTotal, uint64(len(mrsBlock)))
}
putMetricRowsInsertCtx(ic)
2019-05-22 23:16:55 +02:00
return firstErr
2019-05-22 23:16:55 +02:00
}
type metricRowsInsertCtx struct {
rrs []rawRow
tmpMrs []*MetricRow
}
func getMetricRowsInsertCtx() *metricRowsInsertCtx {
v := metricRowsInsertCtxPool.Get()
if v == nil {
v = &metricRowsInsertCtx{
rrs: make([]rawRow, maxMetricRowsPerBlock),
tmpMrs: make([]*MetricRow, maxMetricRowsPerBlock),
}
}
return v.(*metricRowsInsertCtx)
}
func putMetricRowsInsertCtx(ic *metricRowsInsertCtx) {
tmpMrs := ic.tmpMrs
for i := range tmpMrs {
tmpMrs[i] = nil
}
metricRowsInsertCtxPool.Put(ic)
}
var metricRowsInsertCtxPool sync.Pool
const maxMetricRowsPerBlock = 8000
// RegisterMetricNames registers all the metric names from mns in the indexdb, so they can be queried later.
//
// The the MetricRow.Timestamp is used for registering the metric name starting from the given timestamp.
// Th MetricRow.Value field is ignored.
func (s *Storage) RegisterMetricNames(qt *querytracer.Tracer, mrs []MetricRow) error {
qt = qt.NewChild("registering %d series", len(mrs))
defer qt.Done()
var metricName []byte
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
var genTSID generationTSID
mn := GetMetricName()
defer PutMetricName(mn)
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
idb := s.idb()
is := idb.getIndexSearch(noDeadline)
defer idb.putIndexSearch(is)
for i := range mrs {
mr := &mrs[i]
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
if s.getTSIDFromCache(&genTSID, mr.MetricNameRaw) {
if err := s.registerSeriesCardinality(genTSID.TSID.MetricID, mr.MetricNameRaw); err != nil {
continue
}
if genTSID.generation == idb.generation {
// Fast path - mr.MetricNameRaw has been already registered in the current idb.
continue
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
}
}
// Slow path - register mr.MetricNameRaw.
if err := mn.UnmarshalRaw(mr.MetricNameRaw); err != nil {
return fmt.Errorf("cannot unmarshal MetricNameRaw %q: %w", mr.MetricNameRaw, err)
}
mn.sortTags()
metricName = mn.Marshal(metricName[:0])
date := uint64(mr.Timestamp) / msecPerDay
if err := is.GetOrCreateTSIDByName(&genTSID.TSID, metricName, mr.MetricNameRaw, date); err != nil {
if errors.Is(err, errSeriesCardinalityExceeded) {
continue
}
return fmt.Errorf("cannot create TSID for metricName %q: %w", metricName, err)
}
genTSID.generation = idb.generation
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
s.putTSIDToCache(&genTSID, mr.MetricNameRaw)
}
return nil
}
func (s *Storage) add(rows []rawRow, dstMrs []*MetricRow, mrs []MetricRow, precisionBits uint8) error {
2019-05-22 23:16:55 +02:00
idb := s.idb()
is := idb.getIndexSearch(noDeadline)
defer idb.putIndexSearch(is)
var (
2021-03-09 08:18:19 +01:00
// These vars are used for speeding up bulk imports of multiple adjacent rows for the same metricName.
prevTSID TSID
prevMetricNameRaw []byte
)
var pmrs *pendingMetricRows
minTimestamp, maxTimestamp := s.tb.getMinMaxTimestamps()
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
var genTSID generationTSID
// Return only the first error, since it has no sense in returning all errors.
var firstWarn error
j := 0
2019-05-22 23:16:55 +02:00
for i := range mrs {
mr := &mrs[i]
if math.IsNaN(mr.Value) {
if !decimal.IsStaleNaN(mr.Value) {
// Skip NaNs other than Prometheus staleness marker, since the underlying encoding
// doesn't know how to work with them.
continue
}
2019-05-22 23:16:55 +02:00
}
if mr.Timestamp < minTimestamp {
// Skip rows with too small timestamps outside the retention.
if firstWarn == nil {
metricName := getUserReadableMetricName(mr.MetricNameRaw)
firstWarn = fmt.Errorf("cannot insert row with too small timestamp %d outside the retention; minimum allowed timestamp is %d; "+
"probably you need updating -retentionPeriod command-line flag; metricName: %s",
mr.Timestamp, minTimestamp, metricName)
}
atomic.AddUint64(&s.tooSmallTimestampRows, 1)
continue
}
if mr.Timestamp > maxTimestamp {
// Skip rows with too big timestamps significantly exceeding the current time.
if firstWarn == nil {
metricName := getUserReadableMetricName(mr.MetricNameRaw)
firstWarn = fmt.Errorf("cannot insert row with too big timestamp %d exceeding the current time; maximum allowed timestamp is %d; metricName: %s",
mr.Timestamp, maxTimestamp, metricName)
}
atomic.AddUint64(&s.tooBigTimestampRows, 1)
continue
}
dstMrs[j] = mr
r := &rows[j]
2019-05-22 23:16:55 +02:00
j++
r.Timestamp = mr.Timestamp
r.Value = mr.Value
r.PrecisionBits = precisionBits
if string(mr.MetricNameRaw) == string(prevMetricNameRaw) {
// Fast path - the current mr contains the same metric name as the previous mr, so it contains the same TSID.
// This path should trigger on bulk imports when many rows contain the same MetricNameRaw.
r.TSID = prevTSID
continue
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
if s.getTSIDFromCache(&genTSID, mr.MetricNameRaw) {
if err := s.registerSeriesCardinality(r.TSID.MetricID, mr.MetricNameRaw); err != nil {
j--
continue
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
r.TSID = genTSID.TSID
// Fast path - the TSID for the given MetricNameRaw has been found in cache and isn't deleted.
// There is no need in checking whether r.TSID.MetricID is deleted, since tsidCache doesn't
// contain MetricName->TSID entries for deleted time series.
// See Storage.DeleteSeries code for details.
prevTSID = r.TSID
prevMetricNameRaw = mr.MetricNameRaw
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
if genTSID.generation != idb.generation {
// The found entry is from the previous cache generation,
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
// so attempt to re-populate the current generation with this entry.
// This is needed for https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
date := uint64(r.Timestamp) / msecPerDay
created, err := is.maybeCreateIndexes(&genTSID.TSID, mr.MetricNameRaw, date)
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
if err != nil {
return fmt.Errorf("cannot create indexes: %w", err)
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
}
if created {
genTSID.generation = idb.generation
s.putTSIDToCache(&genTSID, mr.MetricNameRaw)
}
}
continue
2019-05-22 23:16:55 +02:00
}
// Slow path - the TSID is missing in the cache.
// Postpone its search in the loop below.
j--
if pmrs == nil {
pmrs = getPendingMetricRows()
2019-05-22 23:16:55 +02:00
}
if err := pmrs.addRow(mr); err != nil {
// Do not stop adding rows on error - just skip invalid row.
// This guarantees that invalid rows don't prevent
// from adding valid rows into the storage.
if firstWarn == nil {
firstWarn = err
}
continue
2019-05-22 23:16:55 +02:00
}
}
if pmrs != nil {
// Sort pendingMetricRows by canonical metric name in order to speed up search via `is` in the loop below.
pendingMetricRows := pmrs.pmrs
sort.Slice(pendingMetricRows, func(i, j int) bool {
return string(pendingMetricRows[i].MetricName) < string(pendingMetricRows[j].MetricName)
})
prevMetricNameRaw = nil
var slowInsertsCount uint64
for i := range pendingMetricRows {
pmr := &pendingMetricRows[i]
mr := pmr.mr
dstMrs[j] = mr
r := &rows[j]
j++
r.Timestamp = mr.Timestamp
r.Value = mr.Value
r.PrecisionBits = precisionBits
if string(mr.MetricNameRaw) == string(prevMetricNameRaw) {
// Fast path - the current mr contains the same metric name as the previous mr, so it contains the same TSID.
// This path should trigger on bulk imports when many rows contain the same MetricNameRaw.
r.TSID = prevTSID
continue
}
slowInsertsCount++
date := uint64(r.Timestamp) / msecPerDay
if err := is.GetOrCreateTSIDByName(&r.TSID, pmr.MetricName, mr.MetricNameRaw, date); err != nil {
j--
if errors.Is(err, errSeriesCardinalityExceeded) {
continue
}
// Do not stop adding rows on error - just skip invalid row.
// This guarantees that invalid rows don't prevent
// from adding valid rows into the storage.
if firstWarn == nil {
firstWarn = fmt.Errorf("cannot obtain or create TSID for MetricName %q: %w", pmr.MetricName, err)
}
continue
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
genTSID.generation = idb.generation
genTSID.TSID = r.TSID
s.putTSIDToCache(&genTSID, mr.MetricNameRaw)
prevTSID = r.TSID
prevMetricNameRaw = mr.MetricNameRaw
2019-05-22 23:16:55 +02:00
}
putPendingMetricRows(pmrs)
atomic.AddUint64(&s.slowRowInserts, slowInsertsCount)
2019-05-22 23:16:55 +02:00
}
if firstWarn != nil {
storageAddRowsLogger.Warnf("warn occurred during rows addition: %s", firstWarn)
}
dstMrs = dstMrs[:j]
rows = rows[:j]
2019-05-22 23:16:55 +02:00
err := s.updatePerDateData(rows, dstMrs)
if err != nil {
err = fmt.Errorf("cannot update per-date data: %w", err)
} else {
s.tb.MustAddRows(rows)
2019-05-22 23:16:55 +02:00
}
if err != nil {
return fmt.Errorf("error occurred during rows addition: %w", err)
2019-10-20 22:38:51 +02:00
}
return nil
2019-05-22 23:16:55 +02:00
}
var storageAddRowsLogger = logger.WithThrottler("storageAddRows", 5*time.Second)
func (s *Storage) registerSeriesCardinality(metricID uint64, metricNameRaw []byte) error {
if sl := s.hourlySeriesLimiter; sl != nil && !sl.Add(metricID) {
atomic.AddUint64(&s.hourlySeriesLimitRowsDropped, 1)
logSkippedSeries(metricNameRaw, "-storage.maxHourlySeries", sl.MaxItems())
return errSeriesCardinalityExceeded
}
if sl := s.dailySeriesLimiter; sl != nil && !sl.Add(metricID) {
atomic.AddUint64(&s.dailySeriesLimitRowsDropped, 1)
logSkippedSeries(metricNameRaw, "-storage.maxDailySeries", sl.MaxItems())
return errSeriesCardinalityExceeded
}
return nil
}
var errSeriesCardinalityExceeded = fmt.Errorf("cannot create series because series cardinality limit exceeded")
func logSkippedSeries(metricNameRaw []byte, flagName string, flagValue int) {
select {
case <-logSkippedSeriesTicker.C:
// Do not use logger.WithThrottler() here, since this will result in increased CPU load
// because of getUserReadableMetricName() calls per each logSkippedSeries call.
userReadableMetricName := getUserReadableMetricName(metricNameRaw)
logger.Warnf("skip series %s because %s=%d reached", userReadableMetricName, flagName, flagValue)
default:
}
}
var logSkippedSeriesTicker = time.NewTicker(5 * time.Second)
func getUserReadableMetricName(metricNameRaw []byte) string {
mn := GetMetricName()
defer PutMetricName(mn)
if err := mn.UnmarshalRaw(metricNameRaw); err != nil {
return fmt.Sprintf("cannot unmarshal metricNameRaw %q: %s", metricNameRaw, err)
}
return mn.String()
}
type pendingMetricRow struct {
MetricName []byte
mr *MetricRow
}
type pendingMetricRows struct {
pmrs []pendingMetricRow
metricNamesBuf []byte
lastMetricNameRaw []byte
lastMetricName []byte
mn MetricName
}
func (pmrs *pendingMetricRows) reset() {
mrs := pmrs.pmrs
for i := range mrs {
pmr := &mrs[i]
pmr.MetricName = nil
pmr.mr = nil
}
pmrs.pmrs = mrs[:0]
pmrs.metricNamesBuf = pmrs.metricNamesBuf[:0]
pmrs.lastMetricNameRaw = nil
pmrs.lastMetricName = nil
pmrs.mn.Reset()
}
func (pmrs *pendingMetricRows) addRow(mr *MetricRow) error {
// Do not spend CPU time on re-calculating canonical metricName during bulk import
// of many rows for the same metric.
if string(mr.MetricNameRaw) != string(pmrs.lastMetricNameRaw) {
if err := pmrs.mn.UnmarshalRaw(mr.MetricNameRaw); err != nil {
return fmt.Errorf("cannot unmarshal MetricNameRaw %q: %w", mr.MetricNameRaw, err)
}
pmrs.mn.sortTags()
metricNamesBufLen := len(pmrs.metricNamesBuf)
pmrs.metricNamesBuf = pmrs.mn.Marshal(pmrs.metricNamesBuf)
pmrs.lastMetricName = pmrs.metricNamesBuf[metricNamesBufLen:]
pmrs.lastMetricNameRaw = mr.MetricNameRaw
}
mrs := pmrs.pmrs
if cap(mrs) > len(mrs) {
mrs = mrs[:len(mrs)+1]
} else {
mrs = append(mrs, pendingMetricRow{})
}
pmrs.pmrs = mrs
pmr := &mrs[len(mrs)-1]
pmr.MetricName = pmrs.lastMetricName
pmr.mr = mr
return nil
}
func getPendingMetricRows() *pendingMetricRows {
v := pendingMetricRowsPool.Get()
if v == nil {
v = &pendingMetricRows{}
}
return v.(*pendingMetricRows)
}
func putPendingMetricRows(pmrs *pendingMetricRows) {
pmrs.reset()
pendingMetricRowsPool.Put(pmrs)
}
var pendingMetricRowsPool sync.Pool
func (s *Storage) updatePerDateData(rows []rawRow, mrs []*MetricRow) error {
2019-05-22 23:16:55 +02:00
var date uint64
var hour uint64
2019-05-22 23:16:55 +02:00
var prevTimestamp int64
var (
2021-03-09 08:18:19 +01:00
// These vars are used for speeding up bulk imports when multiple adjacent rows
// contain the same (metricID, date) pairs.
prevDate uint64
prevMetricID uint64
)
hm := s.currHourMetricIDs.Load().(*hourMetricIDs)
hmPrev := s.prevHourMetricIDs.Load().(*hourMetricIDs)
hmPrevDate := hmPrev.hour / 24
nextDayMetricIDs := &s.nextDayMetricIDs.Load().(*byDateMetricIDEntry).v
ts := fasttime.UnixTimestamp()
// Start pre-populating the next per-day inverted index during the last hour of the current day.
// pMin linearly increases from 0 to 1 during the last hour of the day.
pMin := (float64(ts%(3600*24)) / 3600) - 23
type pendingDateMetricID struct {
date uint64
metricID uint64
mr *MetricRow
}
var pendingDateMetricIDs []pendingDateMetricID
var pendingNextDayMetricIDs []uint64
var pendingHourEntries []uint64
2019-05-22 23:16:55 +02:00
for i := range rows {
r := &rows[i]
if r.Timestamp != prevTimestamp {
date = uint64(r.Timestamp) / msecPerDay
hour = uint64(r.Timestamp) / msecPerHour
2019-05-22 23:16:55 +02:00
prevTimestamp = r.Timestamp
}
metricID := r.TSID.MetricID
2021-02-09 01:51:40 +01:00
if metricID == prevMetricID && date == prevDate {
// Fast path for bulk import of multiple rows with the same (date, metricID) pairs.
continue
}
prevDate = date
prevMetricID = metricID
if hour == hm.hour {
// The r belongs to the current hour. Check for the current hour cache.
if hm.m.Has(metricID) {
// Fast path: the metricID is in the current hour cache.
// This means the metricID has been already added to per-day inverted index.
// Gradually pre-populate per-day inverted index for the next day during the last hour of the current day.
// This should reduce CPU usage spike and slowdown at the beginning of the next day
// when entries for all the active time series must be added to the index.
// This should address https://github.com/VictoriaMetrics/VictoriaMetrics/issues/430 .
if pMin > 0 {
p := float64(uint32(fastHashUint64(metricID))) / (1 << 32)
if p < pMin && !nextDayMetricIDs.Has(metricID) {
pendingDateMetricIDs = append(pendingDateMetricIDs, pendingDateMetricID{
date: date + 1,
metricID: metricID,
mr: mrs[i],
})
pendingNextDayMetricIDs = append(pendingNextDayMetricIDs, metricID)
}
}
continue
}
pendingHourEntries = append(pendingHourEntries, metricID)
if date == hmPrevDate && hmPrev.m.Has(metricID) {
// The metricID is already registered for the current day on the previous hour.
continue
}
}
// Slower path: check global cache for (date, metricID) entry.
2021-02-09 01:51:40 +01:00
if s.dateMetricIDCache.Has(date, metricID) {
continue
}
2021-02-09 01:51:40 +01:00
// Slow path: store the (date, metricID) entry in the indexDB.
pendingDateMetricIDs = append(pendingDateMetricIDs, pendingDateMetricID{
date: date,
metricID: metricID,
mr: mrs[i],
2021-02-09 01:51:40 +01:00
})
}
if len(pendingNextDayMetricIDs) > 0 {
s.pendingNextDayMetricIDsLock.Lock()
s.pendingNextDayMetricIDs.AddMulti(pendingNextDayMetricIDs)
s.pendingNextDayMetricIDsLock.Unlock()
}
if len(pendingHourEntries) > 0 {
s.pendingHourEntriesLock.Lock()
s.pendingHourEntries.AddMulti(pendingHourEntries)
s.pendingHourEntriesLock.Unlock()
}
if len(pendingDateMetricIDs) == 0 {
2023-02-13 13:27:13 +01:00
// Fast path - there are no new (date, metricID) entries in rows.
return nil
}
// Slow path - add new (date, metricID) entries to indexDB.
atomic.AddUint64(&s.slowPerDayIndexInserts, uint64(len(pendingDateMetricIDs)))
// Sort pendingDateMetricIDs by (date, metricID) in order to speed up `is` search in the loop below.
sort.Slice(pendingDateMetricIDs, func(i, j int) bool {
a := pendingDateMetricIDs[i]
b := pendingDateMetricIDs[j]
if a.date != b.date {
return a.date < b.date
}
return a.metricID < b.metricID
})
idb := s.idb()
is := idb.getIndexSearch(noDeadline)
defer idb.putIndexSearch(is)
var firstError error
dateMetricIDsForCache := make([]dateMetricID, 0, len(pendingDateMetricIDs))
mn := GetMetricName()
for _, dmid := range pendingDateMetricIDs {
date := dmid.date
metricID := dmid.metricID
ok, err := is.hasDateMetricID(date, metricID)
if err != nil {
if firstError == nil {
firstError = fmt.Errorf("error when locating (date=%s, metricID=%d) in database: %w", dateToString(date), metricID, err)
}
2019-05-22 23:16:55 +02:00
continue
}
if !ok {
// The (date, metricID) entry is missing in the indexDB. Add it there together with per-day indexes.
2021-02-09 01:51:40 +01:00
// It is OK if the (date, metricID) entry is added multiple times to db
// by concurrent goroutines.
if err := mn.UnmarshalRaw(dmid.mr.MetricNameRaw); err != nil {
if firstError == nil {
firstError = fmt.Errorf("cannot unmarshal MetricNameRaw %q: %w", dmid.mr.MetricNameRaw, err)
}
continue
}
mn.sortTags()
is.createPerDayIndexes(date, metricID, mn)
}
dateMetricIDsForCache = append(dateMetricIDsForCache, dateMetricID{
date: date,
metricID: metricID,
})
2019-05-22 23:16:55 +02:00
}
PutMetricName(mn)
// The (date, metricID) entries must be added to cache only after they have been successfully added to indexDB.
s.dateMetricIDCache.Store(dateMetricIDsForCache)
return firstError
2019-05-22 23:16:55 +02:00
}
func fastHashUint64(x uint64) uint64 {
x ^= x >> 12 // a
x ^= x << 25 // b
x ^= x >> 27 // c
return x * 2685821657736338717
}
// dateMetricIDCache is fast cache for holding (date, metricID) entries.
//
// It should be faster than map[date]*uint64set.Set on multicore systems.
type dateMetricIDCache struct {
// 64-bit counters must be at the top of the structure to be properly aligned on 32-bit arches.
syncsCount uint64
resetsCount uint64
// Contains immutable map
byDate atomic.Value
// Contains mutable map protected by mu
byDateMutable *byDateMetricIDMap
nextSyncDeadline uint64
mu sync.Mutex
}
func newDateMetricIDCache() *dateMetricIDCache {
var dmc dateMetricIDCache
dmc.resetLocked()
return &dmc
}
func (dmc *dateMetricIDCache) Reset() {
dmc.mu.Lock()
dmc.resetLocked()
dmc.mu.Unlock()
}
func (dmc *dateMetricIDCache) resetLocked() {
// Do not reset syncsCount and resetsCount
dmc.byDate.Store(newByDateMetricIDMap())
dmc.byDateMutable = newByDateMetricIDMap()
dmc.nextSyncDeadline = 10 + fasttime.UnixTimestamp()
atomic.AddUint64(&dmc.resetsCount, 1)
}
func (dmc *dateMetricIDCache) EntriesCount() int {
byDate := dmc.byDate.Load().(*byDateMetricIDMap)
n := 0
for _, e := range byDate.m {
n += e.v.Len()
}
return n
}
func (dmc *dateMetricIDCache) SizeBytes() uint64 {
byDate := dmc.byDate.Load().(*byDateMetricIDMap)
n := uint64(0)
for _, e := range byDate.m {
n += e.v.SizeBytes()
}
return n
}
func (dmc *dateMetricIDCache) Has(date, metricID uint64) bool {
byDate := dmc.byDate.Load().(*byDateMetricIDMap)
v := byDate.get(date)
if v.Has(metricID) {
// Fast path.
// The majority of calls must go here.
return true
}
// Slow path. Check mutable map.
dmc.mu.Lock()
v = dmc.byDateMutable.get(date)
ok := v.Has(metricID)
dmc.syncLockedIfNeeded()
dmc.mu.Unlock()
return ok
}
type dateMetricID struct {
date uint64
metricID uint64
}
func (dmc *dateMetricIDCache) Store(dmids []dateMetricID) {
var prevDate uint64
metricIDs := make([]uint64, 0, len(dmids))
dmc.mu.Lock()
for _, dmid := range dmids {
if prevDate == dmid.date {
metricIDs = append(metricIDs, dmid.metricID)
continue
}
if len(metricIDs) > 0 {
v := dmc.byDateMutable.getOrCreate(prevDate)
v.AddMulti(metricIDs)
}
metricIDs = append(metricIDs[:0], dmid.metricID)
prevDate = dmid.date
}
if len(metricIDs) > 0 {
v := dmc.byDateMutable.getOrCreate(prevDate)
v.AddMulti(metricIDs)
}
dmc.mu.Unlock()
}
func (dmc *dateMetricIDCache) Set(date, metricID uint64) {
dmc.mu.Lock()
v := dmc.byDateMutable.getOrCreate(date)
v.Add(metricID)
dmc.mu.Unlock()
}
func (dmc *dateMetricIDCache) syncLockedIfNeeded() {
currentTime := fasttime.UnixTimestamp()
if currentTime >= dmc.nextSyncDeadline {
dmc.nextSyncDeadline = currentTime + 10
dmc.syncLocked()
}
}
func (dmc *dateMetricIDCache) syncLocked() {
if len(dmc.byDateMutable.m) == 0 {
// Nothing to sync.
return
}
byDate := dmc.byDate.Load().(*byDateMetricIDMap)
byDateMutable := dmc.byDateMutable
for date, e := range byDateMutable.m {
v := byDate.get(date)
if v == nil {
continue
}
v = v.Clone()
v.Union(&e.v)
dme := &byDateMetricIDEntry{
date: date,
v: *v,
}
if date == byDateMutable.hotEntry.Load().(*byDateMetricIDEntry).date {
byDateMutable.hotEntry.Store(dme)
}
byDateMutable.m[date] = dme
}
for date, e := range byDate.m {
v := byDateMutable.get(date)
if v != nil {
continue
}
byDateMutable.m[date] = e
}
dmc.byDate.Store(dmc.byDateMutable)
dmc.byDateMutable = newByDateMetricIDMap()
atomic.AddUint64(&dmc.syncsCount, 1)
if dmc.SizeBytes() > uint64(memory.Allowed())/256 {
dmc.resetLocked()
}
}
type byDateMetricIDMap struct {
hotEntry atomic.Value
m map[uint64]*byDateMetricIDEntry
}
func newByDateMetricIDMap() *byDateMetricIDMap {
dmm := &byDateMetricIDMap{
m: make(map[uint64]*byDateMetricIDEntry),
}
dmm.hotEntry.Store(&byDateMetricIDEntry{})
return dmm
}
func (dmm *byDateMetricIDMap) get(date uint64) *uint64set.Set {
hotEntry := dmm.hotEntry.Load().(*byDateMetricIDEntry)
if hotEntry.date == date {
// Fast path
return &hotEntry.v
}
// Slow path
e := dmm.m[date]
if e == nil {
return nil
}
dmm.hotEntry.Store(e)
return &e.v
}
func (dmm *byDateMetricIDMap) getOrCreate(date uint64) *uint64set.Set {
v := dmm.get(date)
if v != nil {
return v
}
e := &byDateMetricIDEntry{
date: date,
}
dmm.m[date] = e
return &e.v
}
type byDateMetricIDEntry struct {
date uint64
v uint64set.Set
}
func (s *Storage) updateNextDayMetricIDs(date uint64) {
e := s.nextDayMetricIDs.Load().(*byDateMetricIDEntry)
s.pendingNextDayMetricIDsLock.Lock()
pendingMetricIDs := s.pendingNextDayMetricIDs
s.pendingNextDayMetricIDs = &uint64set.Set{}
s.pendingNextDayMetricIDsLock.Unlock()
if pendingMetricIDs.Len() == 0 && e.date == date {
// Fast path: nothing to update.
return
}
// Slow path: union pendingMetricIDs with e.v
if e.date == date {
pendingMetricIDs.Union(&e.v)
} else {
// Do not add pendingMetricIDs from the previous day to the current day,
// since this may result in missing registration of the metricIDs in the per-day inverted index.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3309
pendingMetricIDs = &uint64set.Set{}
}
eNew := &byDateMetricIDEntry{
date: date,
v: *pendingMetricIDs,
}
s.nextDayMetricIDs.Store(eNew)
}
func (s *Storage) updateCurrHourMetricIDs(hour uint64) {
hm := s.currHourMetricIDs.Load().(*hourMetricIDs)
2019-11-08 18:37:16 +01:00
s.pendingHourEntriesLock.Lock()
newMetricIDs := s.pendingHourEntries
s.pendingHourEntries = &uint64set.Set{}
2019-11-08 18:37:16 +01:00
s.pendingHourEntriesLock.Unlock()
if newMetricIDs.Len() == 0 && hm.hour == hour {
// Fast path: nothing to update.
return
}
2019-11-08 18:37:16 +01:00
// Slow path: hm.m must be updated with non-empty s.pendingHourEntries.
var m *uint64set.Set
if hm.hour == hour {
m = hm.m.Clone()
m.Union(newMetricIDs)
} else {
m = newMetricIDs
if hour%24 == 0 {
// Do not add pending metricIDs from the previous hour to the current hour on the next day,
// since this may result in missing registration of the metricIDs in the per-day inverted index.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3309
m = &uint64set.Set{}
}
}
hmNew := &hourMetricIDs{
m: m,
hour: hour,
}
s.currHourMetricIDs.Store(hmNew)
if hm.hour != hour {
s.prevHourMetricIDs.Store(hm)
}
}
type hourMetricIDs struct {
m *uint64set.Set
hour uint64
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
type generationTSID struct {
TSID TSID
// generation stores the indexdb.generation value to identify to which indexdb belongs this TSID
generation uint64
}
func (s *Storage) getTSIDFromCache(dst *generationTSID, metricName []byte) bool {
2019-05-22 23:16:55 +02:00
buf := (*[unsafe.Sizeof(*dst)]byte)(unsafe.Pointer(dst))[:]
buf = s.tsidCache.Get(buf[:0], metricName)
return uintptr(len(buf)) == unsafe.Sizeof(*dst)
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 23:30:08 +01:00
func (s *Storage) putTSIDToCache(tsid *generationTSID, metricName []byte) {
2019-05-22 23:16:55 +02:00
buf := (*[unsafe.Sizeof(*tsid)]byte)(unsafe.Pointer(tsid))[:]
s.tsidCache.Set(metricName, buf)
}
func (s *Storage) mustOpenIndexDBTables(path string) (curr, prev *indexDB) {
fs.MustMkdirIfNotExist(path)
fs.MustRemoveTemporaryDirs(path)
2019-05-22 23:16:55 +02:00
// Search for the two most recent tables - the last one is active,
// the previous one contains backup data.
des := fs.MustReadDir(path)
2019-05-22 23:16:55 +02:00
var tableNames []string
for _, de := range des {
if !fs.IsDirOrSymlink(de) {
2019-05-22 23:16:55 +02:00
// Skip non-directories.
continue
}
tableName := de.Name()
2019-05-22 23:16:55 +02:00
if !indexDBTableNameRegexp.MatchString(tableName) {
// Skip invalid directories.
continue
}
tableNames = append(tableNames, tableName)
}
sort.Slice(tableNames, func(i, j int) bool {
return tableNames[i] < tableNames[j]
})
if len(tableNames) < 2 {
// Create missing tables
if len(tableNames) == 0 {
prevName := nextIndexDBTableName()
tableNames = append(tableNames, prevName)
}
currName := nextIndexDBTableName()
tableNames = append(tableNames, currName)
}
// Invariant: len(tableNames) >= 2
// Remove all the tables except two last tables.
for _, tn := range tableNames[:len(tableNames)-2] {
pathToRemove := filepath.Join(path, tn)
2019-05-22 23:16:55 +02:00
logger.Infof("removing obsolete indexdb dir %q...", pathToRemove)
all: add Windows build for VictoriaMetrics This commit changes background merge algorithm, so it becomes compatible with Windows file semantics. The previous algorithm for background merge: 1. Merge source parts into a destination part inside tmp directory. 2. Create a file in txn directory with instructions on how to atomically swap source parts with the destination part. 3. Perform instructions from the file. 4. Delete the file with instructions. This algorithm guarantees that either source parts or destination part is visible in the partition after unclean shutdown at any step above, since the remaining files with instructions is replayed on the next restart, after that the remaining contents of the tmp directory is deleted. Unfortunately this algorithm doesn't work under Windows because it disallows removing and moving files, which are in use. So the new algorithm for background merge has been implemented: 1. Merge source parts into a destination part inside the partition directory itself. E.g. now the partition directory may contain both complete and incomplete parts. 2. Atomically update the parts.json file with the new list of parts after the merge, e.g. remove the source parts from the list and add the destination part to the list before storing it to parts.json file. 3. Remove the source parts from disk when they are no longer used. This algorithm guarantees that either source parts or destination part is visible in the partition after unclean shutdown at any step above, since incomplete partitions from step 1 or old source parts from step 3 are removed on the next startup by inspecting parts.json file. This algorithm should work under Windows, since it doesn't remove or move files in use. This algorithm has also the following benefits: - It should work better for NFS. - It fits object storage semantics. The new algorithm changes data storage format, so it is impossible to downgrade to the previous versions of VictoriaMetrics after upgrading to this algorithm. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3236 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3821 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/70
2023-03-19 09:36:05 +01:00
fs.MustRemoveAll(pathToRemove)
2019-05-22 23:16:55 +02:00
logger.Infof("removed obsolete indexdb dir %q", pathToRemove)
}
// Persist changes on the file system.
fs.MustSyncPath(path)
2019-05-22 23:16:55 +02:00
// Open the last two tables.
currPath := filepath.Join(path, tableNames[len(tableNames)-1])
2019-05-22 23:16:55 +02:00
curr = mustOpenIndexDB(currPath, s, 0, &s.isReadOnly)
prevPath := filepath.Join(path, tableNames[len(tableNames)-2])
prev = mustOpenIndexDB(prevPath, s, 0, &s.isReadOnly)
2019-05-22 23:16:55 +02:00
return curr, prev
2019-05-22 23:16:55 +02:00
}
var indexDBTableNameRegexp = regexp.MustCompile("^[0-9A-F]{16}$")
func nextIndexDBTableName() string {
n := atomic.AddUint64(&indexDBTableIdx, 1)
return fmt.Sprintf("%016X", n)
}
var indexDBTableIdx = uint64(time.Now().UnixNano())