lib/mergeset: remove string allocation and copying when unmarshaling blockHeader

This should reduce CPU usage for the case from https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3343

lib/mergeset/block_header.go

@@ -16,6 +16,9 @@ type blockHeader struct {
 	// The first item.
 	firstItem []byte
 
+	// Whether commonPrefix and firstItem point to external data.
+	noCopy bool
+
 	// Marshal type used for block compression.
 	marshalType marshalType
 
@@ -40,8 +43,13 @@ func (bh *blockHeader) SizeBytes() int {
 }
 
 func (bh *blockHeader) Reset() {
+	if bh.noCopy {
+		bh.commonPrefix = nil
+		bh.firstItem = nil
+	} else {
 		bh.commonPrefix = bh.commonPrefix[:0]
 		bh.firstItem = bh.firstItem[:0]
+	}
 	bh.marshalType = marshalTypePlain
 	bh.itemsCount = 0
 	bh.itemsBlockOffset = 0
@@ -62,13 +70,17 @@ func (bh *blockHeader) Marshal(dst []byte) []byte {
 	return dst
 }
 
-func (bh *blockHeader) Unmarshal(src []byte) ([]byte, error) {
+// UnmarshalNoCopy unmarshals bh from src without copying the data from src.
+//
+// The src must remain unchanged while bh is in use.
+func (bh *blockHeader) UnmarshalNoCopy(src []byte) ([]byte, error) {
+	bh.noCopy = true
 	// Unmarshal commonPrefix
 	tail, cp, err := encoding.UnmarshalBytes(src)
 	if err != nil {
 		return tail, fmt.Errorf("cannot unmarshal commonPrefix: %w", err)
 	}
-	bh.commonPrefix = append(bh.commonPrefix[:0], cp...)
+	bh.commonPrefix = cp[:len(cp):len(cp)]
 	src = tail
 
 	// Unmarshal firstItem
@@ -76,7 +88,7 @@ func (bh *blockHeader) Unmarshal(src []byte) ([]byte, error) {
 	if err != nil {
 		return tail, fmt.Errorf("cannot unmarshal firstItem: %w", err)
 	}
-	bh.firstItem = append(bh.firstItem[:0], fi...)
+	bh.firstItem = fi[:len(fi):len(fi)]
 	src = tail
 
 	// Unmarshal marshalType
@@ -137,11 +149,13 @@ func (bh *blockHeader) Unmarshal(src []byte) ([]byte, error) {
 	return src, nil
 }
 
-// unmarshalBlockHeaders unmarshals all the block headers from src,
+// unmarshalBlockHeadersNoCopy unmarshals all the block headers from src,
 // appends them to dst and returns the appended result.
 //
 // Block headers must be sorted by bh.firstItem.
-func unmarshalBlockHeaders(dst []blockHeader, src []byte, blockHeadersCount int) ([]blockHeader, error) {
+//
+// It is expected that src remains unchanged while rhe returned blocks are in use.
+func unmarshalBlockHeadersNoCopy(dst []blockHeader, src []byte, blockHeadersCount int) ([]blockHeader, error) {
 	if blockHeadersCount <= 0 {
 		logger.Panicf("BUG: blockHeadersCount must be greater than 0; got %d", blockHeadersCount)
 	}
@@ -151,7 +165,7 @@ func unmarshalBlockHeaders(dst []blockHeader, src []byte, blockHeadersCount int)
 	}
 	dst = dst[:dstLen+blockHeadersCount]
 	for i := 0; i < blockHeadersCount; i++ {
-		tail, err := dst[dstLen+i].Unmarshal(src)
+		tail, err := dst[dstLen+i].UnmarshalNoCopy(src)
 		if err != nil {
 			return dst, fmt.Errorf("cannot unmarshal block header: %w", err)
 		}

lib/mergeset/block_stream_reader.go

@@ -295,22 +295,10 @@ func (bsr *blockStreamReader) readNextBHS() error {
 	}
 
 	// Unmarshal the unpacked index block into bsr.bhs.
-	if n := int(mr.blockHeadersCount) - cap(bsr.bhs); n > 0 {
-		bsr.bhs = append(bsr.bhs[:cap(bsr.bhs)], make([]blockHeader, n)...)
-	}
-	bsr.bhs = bsr.bhs[:mr.blockHeadersCount]
-	bsr.bhIdx = 0
-	b := bsr.unpackedBuf
-	for i := 0; i < int(mr.blockHeadersCount); i++ {
-		tail, err := bsr.bhs[i].Unmarshal(b)
+	bsr.bhs, err = unmarshalBlockHeadersNoCopy(bsr.bhs[:0], bsr.unpackedBuf, int(mr.blockHeadersCount))
 	if err != nil {
 		return fmt.Errorf("cannot unmarshal blockHeader #%d in the index block #%d: %w", len(bsr.bhs), bsr.mrIdx, err)
 	}
-		b = tail
-	}
-	if len(b) > 0 {
-		return fmt.Errorf("unexpected non-empty tail left after unmarshaling block headers; len(tail)=%d", len(b))
-	}
 	return nil
 }
 

lib/mergeset/encoding.go

@@ -334,8 +334,7 @@ func (ib *inmemoryBlock) marshalData(sb *storageBlock, firstItemDst, commonPrefi
 	return firstItemDst, commonPrefixDst, uint32(len(ib.items)), marshalTypeZSTD
 }
 
-// UnmarshalData decodes itemsCount items from sb and firstItem and stores
-// them to ib.
+// UnmarshalData decodes itemsCount items from sb and firstItem and stores them to ib.
 func (ib *inmemoryBlock) UnmarshalData(sb *storageBlock, firstItem, commonPrefix []byte, itemsCount uint32, mt marshalType) error {
 	ib.Reset()
 

lib/mergeset/part_search.go

@@ -286,7 +286,7 @@ func (ps *partSearch) readIndexBlock(mr *metaindexRow) (*indexBlock, error) {
 		return nil, fmt.Errorf("cannot decompress index block: %w", err)
 	}
 	idxb := &indexBlock{}
-	idxb.bhs, err = unmarshalBlockHeaders(idxb.bhs[:0], ps.indexBuf, int(mr.blockHeadersCount))
+	idxb.bhs, err = unmarshalBlockHeadersNoCopy(idxb.bhs[:0], ps.indexBuf, int(mr.blockHeadersCount))
 	if err != nil {
 		return nil, fmt.Errorf("cannot unmarshal block headers from index block (offset=%d, size=%d): %w", mr.indexBlockOffset, mr.indexBlockSize, err)
 	}