sink(ticdc): make mysql sink support split transactions (#5281)

close #5280

cdc/model/mounter.go

@@ -13,7 +13,7 @@
 
 package model
 
-// PolymorphicEvent describes an event can be in multiple states
+// PolymorphicEvent describes an event can be in multiple states.
 type PolymorphicEvent struct {
 	StartTs uint64
 	// Commit or resolved TS
@@ -23,7 +23,16 @@ type PolymorphicEvent struct {
 	Row   *RowChangedEvent
 }
 
-// NewPolymorphicEvent creates a new PolymorphicEvent with a raw KV
+// NewEmptyPolymorphicEvent creates a new empty PolymorphicEvent.
+func NewEmptyPolymorphicEvent(ts uint64) *PolymorphicEvent {
+	return &PolymorphicEvent{
+		CRTs:  ts,
+		RawKV: &RawKVEntry{},
+		Row:   &RowChangedEvent{},
+	}
+}
+
+// NewPolymorphicEvent creates a new PolymorphicEvent with a raw KV.
 func NewPolymorphicEvent(rawKV *RawKVEntry) *PolymorphicEvent {
 	if rawKV.OpType == OpTypeResolved {
 		return NewResolvedPolymorphicEvent(rawKV.RegionID, rawKV.CRTs)

cdc/model/sink.go

@@ -267,6 +267,9 @@ type RowChangedEvent struct {
 	// ApproximateDataSize is the approximate size of protobuf binary
 	// representation of this event.
 	ApproximateDataSize int64 `json:"-" msg:"-"`
+
+	// SplitTxn marks this RowChangedEvent as the first line of a new txn.
+	SplitTxn bool `json:"-" msg:"-"`
 }
 
 // IsDelete returns true if the row is a delete event

cdc/processor/pipeline/sink_test.go

@@ -40,7 +40,11 @@ type mockSink struct {
 // we are testing sinkNode by itself.
 type mockFlowController struct{}
 
-func (c *mockFlowController) Consume(commitTs uint64, size uint64, blockCallBack func() error) error {
+func (c *mockFlowController) Consume(
+	msg *model.PolymorphicEvent,
+	size uint64,
+	blockCallBack func(bool) error,
+) error {
 	return nil
 }
 

cdc/processor/pipeline/sorter.go

@@ -206,8 +206,10 @@ func (n *sorterNode) start(
 					size := uint64(msg.Row.ApproximateBytes())
 					// NOTE we allow the quota to be exceeded if blocking means interrupting a transaction.
 					// Otherwise the pipeline would deadlock.
-					err = n.flowController.Consume(commitTs, size, func() error {
-						if lastCRTs > lastSentResolvedTs {
+					err = n.flowController.Consume(msg, size, func(batch bool) error {
+						if batch {
+							log.Panic("cdc does not support the batch resolve mechanism at this time")
+						} else if lastCRTs > lastSentResolvedTs {
 							// If we are blocking, we send a Resolved Event here to elicit a sink-flush.
 							// Not sending a Resolved Event here will very likely deadlock the pipeline.
 							lastSentResolvedTs = lastCRTs

cdc/processor/pipeline/table.go

@@ -78,7 +78,7 @@ type tablePipelineImpl struct {
 // TODO find a better name or avoid using an interface
 // We use an interface here for ease in unit testing.
 type tableFlowController interface {
-	Consume(commitTs uint64, size uint64, blockCallBack func() error) error
+	Consume(msg *model.PolymorphicEvent, size uint64, blockCallBack func(batch bool) error) error
 	Release(resolvedTs uint64)
 	Abort()
 	GetConsumption() uint64

cdc/sink/flowcontrol/flow_control.go

@@ -20,9 +20,16 @@ import (
 	"github.com/edwingeng/deque"
 	"github.com/pingcap/errors"
 	"github.com/pingcap/log"
+	"github.com/pingcap/tiflow/cdc/model"
 	"go.uber.org/zap"
 )
 
+const (
+	maxRowsPerTxn = 1024
+	maxSizePerTxn = 1024 * 1024 /* 1MB */
+	batchSize     = 100
+)
+
 // TableFlowController provides a convenient interface to control the memory consumption of a per table event stream
 type TableFlowController struct {
 	memoryQuota *tableMemoryQuota
@@ -31,13 +38,20 @@ type TableFlowController struct {
 		sync.Mutex
 		queue deque.Deque
 	}
+	// batchGroupCount is the number of txnSizeEntries with same commitTs, which could be:
+	// 1. Different txns with same commitTs but different startTs
+	// 2. TxnSizeEntry split from the same txns which exceeds max rows or max size
+	batchGroupCount uint
 
 	lastCommitTs uint64
 }
 
-type commitTsSizeEntry struct {
+type txnSizeEntry struct {
+	// txn id
+	startTs  uint64
 	commitTs uint64
 	size     uint64
+	rowCount uint64
 }
 
 // NewTableFlowController creates a new TableFlowController
@@ -55,7 +69,12 @@ func NewTableFlowController(quota uint64) *TableFlowController {
 
 // Consume is called when an event has arrived for being processed by the sink.
 // It will handle transaction boundaries automatically, and will not block intra-transaction.
-func (c *TableFlowController) Consume(commitTs uint64, size uint64, blockCallBack func() error) error {
+func (c *TableFlowController) Consume(
+	msg *model.PolymorphicEvent,
+	size uint64,
+	callBack func(batch bool) error,
+) error {
+	commitTs := msg.CRTs
 	lastCommitTs := atomic.LoadUint64(&c.lastCommitTs)
 
 	if commitTs < lastCommitTs {
@@ -65,8 +84,7 @@ func (c *TableFlowController) Consume(commitTs uint64, size uint64, blockCallBac
 	}
 
 	if commitTs > lastCommitTs {
-		atomic.StoreUint64(&c.lastCommitTs, commitTs)
-		err := c.memoryQuota.consumeWithBlocking(size, blockCallBack)
+		err := c.memoryQuota.consumeWithBlocking(size, callBack)
 		if err != nil {
 			return errors.Trace(err)
 		}
@@ -82,13 +100,7 @@ func (c *TableFlowController) Consume(commitTs uint64, size uint64, blockCallBac
 		}
 	}
 
-	c.queueMu.Lock()
-	defer c.queueMu.Unlock()
-	c.queueMu.queue.PushBack(&commitTsSizeEntry{
-		commitTs: commitTs,
-		size:     size,
-	})
-
+	c.enqueueSingleMsg(msg, size)
 	return nil
 }
 
@@ -98,7 +110,7 @@ func (c *TableFlowController) Release(resolvedTs uint64) {
 
 	c.queueMu.Lock()
 	for c.queueMu.queue.Len() > 0 {
-		if peeked := c.queueMu.queue.Front().(*commitTsSizeEntry); peeked.commitTs <= resolvedTs {
+		if peeked := c.queueMu.queue.Front().(*txnSizeEntry); peeked.commitTs <= resolvedTs {
 			nBytesToRelease += peeked.size
 			c.queueMu.queue.PopFront()
 		} else {
@@ -110,6 +122,62 @@ func (c *TableFlowController) Release(resolvedTs uint64) {
 	c.memoryQuota.release(nBytesToRelease)
 }
 
+// Note that msgs received by enqueueSingleMsg must be sorted by commitTs_startTs order.
+func (c *TableFlowController) enqueueSingleMsg(msg *model.PolymorphicEvent, size uint64) {
+	commitTs := msg.CRTs
+	lastCommitTs := atomic.LoadUint64(&c.lastCommitTs)
+
+	c.queueMu.Lock()
+	defer c.queueMu.Unlock()
+
+	var e deque.Elem
+	// 1. Processing a new txn with different commitTs.
+	if e = c.queueMu.queue.Back(); e == nil || lastCommitTs < commitTs {
+		atomic.StoreUint64(&c.lastCommitTs, commitTs)
+		c.queueMu.queue.PushBack(&txnSizeEntry{
+			startTs:  msg.StartTs,
+			commitTs: commitTs,
+			size:     size,
+			rowCount: 1,
+		})
+		c.batchGroupCount = 1
+		msg.Row.SplitTxn = true
+		return
+	}
+
+	// Processing txns with the same commitTs.
+	txnEntry := e.(*txnSizeEntry)
+	if txnEntry.commitTs != lastCommitTs {
+		log.Panic("got wrong commitTs from deque, report a bug",
+			zap.Uint64("lastCommitTs", c.lastCommitTs),
+			zap.Uint64("commitTsInDeque", txnEntry.commitTs))
+	}
+
+	// 2. Append row to current txn entry.
+	if txnEntry.startTs == msg.Row.StartTs &&
+		txnEntry.rowCount < maxRowsPerTxn && txnEntry.size < maxSizePerTxn {
+		txnEntry.size += size
+		txnEntry.rowCount++
+		return
+	}
+
+	// 3. Split the txn or handle a new txn with the same commitTs.
+	c.queueMu.queue.PushBack(&txnSizeEntry{
+		startTs:  msg.StartTs,
+		commitTs: commitTs,
+		size:     size,
+		rowCount: 1,
+	})
+	c.batchGroupCount++
+	msg.Row.SplitTxn = true
+
+	if c.batchGroupCount >= batchSize {
+		c.batchGroupCount = 0
+		// TODO(CharlesCheung): add batch resolve mechanism to mitigate oom problem
+		log.Debug("emit batch resolve event throw callback")
+	}
+}
+
 // Abort interrupts any ongoing Consume call
 func (c *TableFlowController) Abort() {
 	c.memoryQuota.abort()