sink, sorter: do not hog memory

cdc/processor/pipeline/sink.go

@@ -28,7 +28,7 @@ import (
 )
 
 const (
-	defaultSyncResolvedBatch = 1024
+	defaultSyncResolvedBatch = 64
 )
 
 // TableStatus is status of the table pipeline
@@ -132,7 +132,7 @@ func (n *sinkNode) flushSink(ctx pipeline.NodeContext, resolvedTs model.Ts) (err
 	if resolvedTs <= n.checkpointTs {
 		return nil
 	}
-	if err := n.flushRow2Sink(ctx); err != nil {
+	if err := n.emitRow2Sink(ctx); err != nil {
 		return errors.Trace(err)
 	}
 	checkpointTs, err := n.sink.FlushRowChangedEvents(ctx, resolvedTs)
@@ -178,7 +178,7 @@ func (n *sinkNode) emitEvent(ctx pipeline.NodeContext, event *model.PolymorphicE
 	}
 
 	if len(n.eventBuffer) >= defaultSyncResolvedBatch {
-		if err := n.flushRow2Sink(ctx); err != nil {
+		if err := n.emitRow2Sink(ctx); err != nil {
 			return errors.Trace(err)
 		}
 	}
@@ -249,7 +249,30 @@ func splitUpdateEvent(updateEvent *model.PolymorphicEvent) (*model.PolymorphicEv
 	return &deleteEvent, &insertEvent, nil
 }
 
-func (n *sinkNode) flushRow2Sink(ctx pipeline.NodeContext) error {
+// clear event buffer and row buffer.
+// Also, it unrefs data that are holded by buffers.
+func (n *sinkNode) clearBuffers() {
+	// Do not hog memory.
+	if cap(n.rowBuffer) > defaultSyncResolvedBatch {
+		n.rowBuffer = make([]*model.RowChangedEvent, 0, defaultSyncResolvedBatch)
+	} else {
+		for i := range n.rowBuffer {
+			n.rowBuffer[i] = nil
+		}
+		n.rowBuffer = n.rowBuffer[:0]
+	}
+
+	if cap(n.eventBuffer) > defaultSyncResolvedBatch {
+		n.eventBuffer = make([]*model.PolymorphicEvent, 0, defaultSyncResolvedBatch)
+	} else {
+		for i := range n.eventBuffer {
+			n.eventBuffer[i] = nil
+		}
+		n.eventBuffer = n.eventBuffer[:0]
+	}
+}
+
+func (n *sinkNode) emitRow2Sink(ctx pipeline.NodeContext) error {
 	for _, ev := range n.eventBuffer {
 		err := ev.WaitPrepare(ctx)
 		if err != nil {
@@ -270,8 +293,7 @@ func (n *sinkNode) flushRow2Sink(ctx pipeline.NodeContext) error {
 	if err != nil {
 		return errors.Trace(err)
 	}
-	n.rowBuffer = n.rowBuffer[:0]
-	n.eventBuffer = n.eventBuffer[:0]
+	n.clearBuffers()
 	return nil
 }
 

cdc/puller/sorter/file_backend.go

@@ -29,7 +29,7 @@ import (
 )
 
 const (
-	fileBufferSize       = 32 * 1024 // 32KB
+	fileBufferSize       = 4 * 1024 // 4KB
 	fileMagic            = 0x12345678
 	numFileEntriesOffset = 4
 	blockMagic           = 0xbeefbeef
@@ -160,11 +160,10 @@ func (f *fileBackEnd) cleanStats() {
 }
 
 type fileBackEndReader struct {
-	backEnd     *fileBackEnd
-	f           *os.File
-	reader      *bufio.Reader
-	rawBytesBuf []byte
-	isEOF       bool
+	backEnd *fileBackEnd
+	f       *os.File
+	reader  *bufio.Reader
+	isEOF   bool
 
 	// to prevent truncation-like corruption
 	totalEvents uint64
@@ -237,14 +236,11 @@ func (r *fileBackEndReader) readNext() (*model.PolymorphicEvent, error) {
 		return nil, errors.Trace(wrapIOError(err))
 	}
 
-	if cap(r.rawBytesBuf) < int(size) {
-		r.rawBytesBuf = make([]byte, size)
-	} else {
-		r.rawBytesBuf = r.rawBytesBuf[:size]
-	}
+	// Note, do not hold the buffer in reader to avoid hogging memory.
+	rawBytesBuf := make([]byte, size)
 
 	// short reads are possible with bufio, hence the need for io.ReadFull
-	n, err := io.ReadFull(r.reader, r.rawBytesBuf)
+	n, err := io.ReadFull(r.reader, rawBytesBuf)
 	if err != nil {
 		return nil, errors.Trace(wrapIOError(err))
 	}
@@ -254,12 +250,12 @@ func (r *fileBackEndReader) readNext() (*model.PolymorphicEvent, error) {
 	}
 
 	event := new(model.PolymorphicEvent)
-	_, err = r.backEnd.serde.unmarshal(event, r.rawBytesBuf)
+	_, err = r.backEnd.serde.unmarshal(event, rawBytesBuf)
 	if err != nil {
 		return nil, errors.Trace(err)
 	}
 
-	r.readEvents += 1
+	r.readEvents++
 
 	failpoint.Inject("sorterDebug", func() {
 		r.readBytes += int64(4 + 4 + int(size))