Files
VictoriaMetrics/lib/mergeset/merge.go
Aliaksandr Valialkin 23fd269ccf lib/{storage,mergeset}: reduce the multi-CPU contention on global stats vars, which are updated during background merge
Background merge updates the global stats on the number of merged / deleted items. This may result in slowdown
when multiple goroutines update these global stats at frequent rate, since every goroutine must fetch the actual value
for the updated stats from slow memory on every update. It is much faster to count the needed stats locally per every goroutine
and then periodically updating the global stats (once per ~second).

Thanks to @tIGO for the intial implementation of this idea at https://github.com/VictoriaMetrics/VictoriaMetrics/pull/8683/files#diff-95e28ae911944708f94f3bb31fa9ba8bc185dedc23ae6fb02a272c34b8f83244

This should help improving scalability of background merges on multi-CPU systems.
See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8682
2025-06-05 12:24:03 +02:00

273 lines
7.2 KiB
Go

package mergeset
import (
"container/heap"
"fmt"
"sync"
"sync/atomic"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
)
// PrepareBlockCallback can transform the passed items allocated at the given data.
//
// The callback is called during merge before flushing full block of the given items
// to persistent storage.
//
// The callback must return sorted items. The first and the last item must be unchanged.
// The callback can reuse data and items for storing the result.
type PrepareBlockCallback func(data []byte, items []Item) ([]byte, []Item)
// mergeBlockStreams merges bsrs and writes result to bsw.
//
// It also fills ph.
//
// prepareBlock is optional.
//
// The function immediately returns when stopCh is closed.
//
// It also atomically adds the number of items merged to itemsMerged.
func mergeBlockStreams(ph *partHeader, bsw *blockStreamWriter, bsrs []*blockStreamReader, prepareBlock PrepareBlockCallback, stopCh <-chan struct{},
itemsMerged *atomic.Uint64) error {
bsm := bsmPool.Get().(*blockStreamMerger)
if err := bsm.Init(bsrs, prepareBlock); err != nil {
return fmt.Errorf("cannot initialize blockStreamMerger: %w", err)
}
err := bsm.Merge(bsw, ph, stopCh, itemsMerged)
bsm.reset()
bsmPool.Put(bsm)
bsw.MustClose()
if err == nil {
return nil
}
return fmt.Errorf("cannot merge %d block streams: %s: %w", len(bsrs), bsrs, err)
}
var bsmPool = &sync.Pool{
New: func() any {
return &blockStreamMerger{}
},
}
type blockStreamMerger struct {
prepareBlock PrepareBlockCallback
bsrHeap bsrHeap
// ib is a scratch block with pending items.
ib inmemoryBlock
phFirstItemCaught bool
// This are auxiliary buffers used in flushIB
// for consistency checks after prepareBlock call.
firstItem []byte
lastItem []byte
}
func (bsm *blockStreamMerger) reset() {
bsm.prepareBlock = nil
for i := range bsm.bsrHeap {
bsm.bsrHeap[i] = nil
}
bsm.bsrHeap = bsm.bsrHeap[:0]
bsm.ib.Reset()
bsm.phFirstItemCaught = false
}
func (bsm *blockStreamMerger) Init(bsrs []*blockStreamReader, prepareBlock PrepareBlockCallback) error {
bsm.reset()
bsm.prepareBlock = prepareBlock
for _, bsr := range bsrs {
if bsr.Next() {
bsm.bsrHeap = append(bsm.bsrHeap, bsr)
}
if err := bsr.Error(); err != nil {
return fmt.Errorf("cannot obtain the next block from blockStreamReader %q: %w", bsr.path, err)
}
}
heap.Init(&bsm.bsrHeap)
if len(bsm.bsrHeap) == 0 {
return fmt.Errorf("bsrHeap cannot be empty")
}
return nil
}
var errForciblyStopped = fmt.Errorf("forcibly stopped")
func (bsm *blockStreamMerger) Merge(bsw *blockStreamWriter, ph *partHeader, stopCh <-chan struct{}, itemsMerged *atomic.Uint64) error {
// Use local variables for tracking the number of merged items
// and periodically propagate the collected stats to the caller, so it could be reflected in the exposed metrics.
//
// This minimizes expensive updates of itemsMerged var from concurrently running goroutines,
// and improves concurrent merge scalability on multi-CPU systems - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8682 .
var updateStatsDeadline uint64
var localItemsMerged uint64
updateStats := func() {
itemsMerged.Add(localItemsMerged)
localItemsMerged = 0
}
defer updateStats()
again:
ct := fasttime.UnixTimestamp()
if ct > updateStatsDeadline {
updateStats()
// Update the external stats once per second
updateStatsDeadline = ct + 1
}
if len(bsm.bsrHeap) == 0 {
// Write the last (maybe incomplete) inmemoryBlock to bsw.
bsm.flushIB(bsw, ph, &localItemsMerged)
return nil
}
select {
case <-stopCh:
return errForciblyStopped
default:
}
bsr := bsm.bsrHeap[0]
var nextItem string
hasNextItem := false
if len(bsm.bsrHeap) > 1 {
bsr := bsm.bsrHeap.getNextReader()
nextItem = bsr.CurrItem()
hasNextItem = true
}
items := bsr.Block.items
data := bsr.Block.data
compareEveryItem := true
if bsr.currItemIdx < len(items) {
// An optimization, which allows skipping costly comparison for every merged item in the loop below.
// Thanks to @ahfuzhang for the suggestion at https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5651
lastItem := items[len(items)-1].String(data)
compareEveryItem = hasNextItem && lastItem > nextItem
}
for bsr.currItemIdx < len(items) {
item := items[bsr.currItemIdx].Bytes(data)
if compareEveryItem && string(item) > nextItem {
break
}
if !bsm.ib.Add(item) {
// The bsm.ib is full. Flush it to bsw and continue.
bsm.flushIB(bsw, ph, &localItemsMerged)
continue
}
bsr.currItemIdx++
}
if bsr.currItemIdx == len(items) {
// bsr.Block is fully read. Proceed to the next block.
if bsr.Next() {
heap.Fix(&bsm.bsrHeap, 0)
goto again
}
if err := bsr.Error(); err != nil {
return fmt.Errorf("cannot read storageBlock: %w", err)
}
heap.Pop(&bsm.bsrHeap)
goto again
}
// The next item in the bsr.Block exceeds nextItem.
// Return bsr to heap.
heap.Fix(&bsm.bsrHeap, 0)
goto again
}
func (bsm *blockStreamMerger) flushIB(bsw *blockStreamWriter, ph *partHeader, itemsMerged *uint64) {
items := bsm.ib.items
data := bsm.ib.data
if len(items) == 0 {
// Nothing to flush.
return
}
*itemsMerged += uint64(len(items))
if bsm.prepareBlock != nil {
bsm.firstItem = append(bsm.firstItem[:0], items[0].String(data)...)
bsm.lastItem = append(bsm.lastItem[:0], items[len(items)-1].String(data)...)
data, items = bsm.prepareBlock(data, items)
bsm.ib.data = data
bsm.ib.items = items
if len(items) == 0 {
// Nothing to flush
return
}
// Consistency checks after prepareBlock call.
firstItem := items[0].String(data)
if firstItem < string(bsm.firstItem) {
logger.Panicf("BUG: prepareBlock must return the first item bigger or equal to the original first item;\ngot\n%X\nwant\n%X", firstItem, bsm.firstItem)
}
lastItem := items[len(items)-1].String(data)
if lastItem > string(bsm.lastItem) {
logger.Panicf("BUG: prepareBlock must return the last item smaller or equal to the original last item;\ngot\n%X\nwant\n%X", lastItem, bsm.lastItem)
}
// Verify whether the bsm.ib.items are sorted only in tests, since this
// can be expensive check in prod for items with long common prefix.
if isInTest && !bsm.ib.isSorted() {
logger.Panicf("BUG: prepareBlock must return sorted items;\ngot\n%s", bsm.ib.debugItemsString())
}
}
ph.itemsCount += uint64(len(items))
if !bsm.phFirstItemCaught {
ph.firstItem = append(ph.firstItem[:0], items[0].String(data)...)
bsm.phFirstItemCaught = true
}
ph.lastItem = append(ph.lastItem[:0], items[len(items)-1].String(data)...)
bsw.WriteBlock(&bsm.ib)
bsm.ib.Reset()
ph.blocksCount++
}
type bsrHeap []*blockStreamReader
func (bh bsrHeap) getNextReader() *blockStreamReader {
if len(bh) < 2 {
return nil
}
if len(bh) < 3 {
return bh[1]
}
a := bh[1]
b := bh[2]
if a.CurrItem() <= b.CurrItem() {
return a
}
return b
}
func (bh *bsrHeap) Len() int {
return len(*bh)
}
func (bh *bsrHeap) Swap(i, j int) {
x := *bh
x[i], x[j] = x[j], x[i]
}
func (bh *bsrHeap) Less(i, j int) bool {
x := *bh
return x[i].CurrItem() < x[j].CurrItem()
}
func (bh *bsrHeap) Pop() any {
a := *bh
v := a[len(a)-1]
*bh = a[:len(a)-1]
return v
}
func (bh *bsrHeap) Push(x any) {
v := x.(*blockStreamReader)
*bh = append(*bh, v)
}