ddl/ingest: refactor checkpoint manager (#54747)

ref #54436

Author: tangenta

pkg/ddl/backfilling.go

@@ -30,6 +30,7 @@ import (
 	"github.com/pingcap/tidb/pkg/ddl/logutil"
 	sess "github.com/pingcap/tidb/pkg/ddl/session"
 	ddlutil "github.com/pingcap/tidb/pkg/ddl/util"
+	"github.com/pingcap/tidb/pkg/disttask/operator"
 	"github.com/pingcap/tidb/pkg/expression"
 	exprctx "github.com/pingcap/tidb/pkg/expression/context"
 	"github.com/pingcap/tidb/pkg/kv"
@@ -689,6 +690,7 @@ func (dc *ddlCtx) runAddIndexInLocalIngestMode(
 	cpMgr, err := ingest.NewCheckpointManager(
 		ctx,
 		sessPool,
+		reorgInfo.PhysicalTableID,
 		job.ID,
 		indexIDs,
 		ingest.LitBackCtxMgr.EncodeJobSortPath(job.ID),
@@ -700,7 +702,6 @@ func (dc *ddlCtx) runAddIndexInLocalIngestMode(
 			zap.Error(err))
 	} else {
 		defer cpMgr.Close()
-		cpMgr.Reset(t.GetPhysicalID(), reorgInfo.StartKey, reorgInfo.EndKey)
 		bcCtx.AttachCheckpointManager(cpMgr)
 	}
 
@@ -723,10 +724,6 @@ func (dc *ddlCtx) runAddIndexInLocalIngestMode(
 			zap.Int64s("index IDs", indexIDs))
 		return errors.Trace(err)
 	}
-	// in happy path FinishAndUnregisterEngines will be called in pipe.Close. We can
-	// ignore the error here.
-	//nolint: errcheck
-	defer bcCtx.FinishAndUnregisterEngines()
 
 	pipe, err := NewAddIndexIngestPipeline(
 		opCtx,
@@ -748,13 +745,31 @@ func (dc *ddlCtx) runAddIndexInLocalIngestMode(
 	if err != nil {
 		return err
 	}
-	err = pipe.Execute()
+	err = executeAndClosePipeline(opCtx, pipe)
+	if err != nil {
+		err1 := bcCtx.FinishAndUnregisterEngines(ingest.OptCloseEngines)
+		if err1 != nil {
+			logutil.DDLIngestLogger().Error("unregister engine failed",
+				zap.Int64("jobID", job.ID),
+				zap.Error(err1),
+				zap.Int64s("index IDs", indexIDs))
+		}
+		return err
+	}
+	if cpMgr != nil {
+		cpMgr.AdvanceWatermark(true, true)
+	}
+	return bcCtx.FinishAndUnregisterEngines(ingest.OptCleanData | ingest.OptCheckDup)
+}
+
+func executeAndClosePipeline(ctx *OperatorCtx, pipe *operator.AsyncPipeline) error {
+	err := pipe.Execute()
 	if err != nil {
 		return err
 	}
 	err = pipe.Close()
-	if opCtx.OperatorErr() != nil {
-		return opCtx.OperatorErr()
+	if opErr := ctx.OperatorErr(); opErr != nil {
+		return opErr
 	}
 	return err
 }

pkg/ddl/backfilling_operators.go

@@ -340,13 +340,13 @@ func (src *TableScanTaskSource) Open() error {
 
 // adjustStartKey adjusts the start key so that we can skip the ranges that have been processed
 // according to the information of checkpoint manager.
-func (src *TableScanTaskSource) adjustStartKey(start, end kv.Key) kv.Key {
+func (src *TableScanTaskSource) adjustStartKey(start, end kv.Key) (adjusted kv.Key, done bool) {
 	if src.cpMgr == nil {
-		return start
+		return start, false
 	}
 	cpKey := src.cpMgr.LastProcessedKey()
 	if len(cpKey) == 0 {
-		return start
+		return start, false
 	}
 	if cpKey.Cmp(start) < 0 || cpKey.Cmp(end) > 0 {
 		logutil.Logger(src.ctx).Error("invalid checkpoint key",
@@ -357,16 +357,23 @@ func (src *TableScanTaskSource) adjustStartKey(start, end kv.Key) kv.Key {
 		if intest.InTest {
 			panic("invalid checkpoint key")
 		}
-		return start
+		return start, false
 	}
-	return cpKey.Next()
+	if cpKey.Cmp(end) == 0 {
+		return cpKey, true
+	}
+	return cpKey.Next(), false
 }
 
 func (src *TableScanTaskSource) generateTasks() error {
 	taskIDAlloc := newTaskIDAllocator()
 	defer src.sink.Finish()
 
-	startKey := src.adjustStartKey(src.startKey, src.endKey)
+	startKey, done := src.adjustStartKey(src.startKey, src.endKey)
+	if done {
+		// All table data are done.
+		return nil
+	}
 	for {
 		kvRanges, err := loadTableRanges(
 			src.ctx,
@@ -931,27 +938,23 @@ func (s *indexWriteResultSink) flush() error {
 		failpoint.Return(errors.New("mock flush error"))
 	})
 	flushed, imported, err := s.backendCtx.Flush(ingest.FlushModeForceFlushAndImport)
-	if err != nil {
-		logutil.Logger(s.ctx).Error("flush error",
-			zap.String("category", "ddl"), zap.Error(err))
-		return err
-	}
 	if s.cpMgr != nil {
+		// Try to advance watermark even if there is an error.
 		s.cpMgr.AdvanceWatermark(flushed, imported)
 	}
+	if err != nil {
+		msg := "flush error"
+		if flushed {
+			msg = "import error"
+		}
+		logutil.Logger(s.ctx).Error(msg, zap.String("category", "ddl"), zap.Error(err))
+		return err
+	}
 	return nil
 }
 
 func (s *indexWriteResultSink) Close() error {
-	err := s.errGroup.Wait()
-	// for local pipeline
-	if bc := s.backendCtx; bc != nil {
-		err2 := bc.FinishAndUnregisterEngines()
-		if err == nil {
-			err = err2
-		}
-	}
-	return err
+	return s.errGroup.Wait()
 }
 
 func (*indexWriteResultSink) String() string {

pkg/ddl/backfilling_read_index.go

@@ -109,25 +109,29 @@ func (r *readIndexExecutor) RunSubtask(ctx context.Context, subtask *proto.Subta
 	defer opCtx.Cancel()
 	r.curRowCount.Store(0)
 
-	var pipe *operator.AsyncPipeline
 	if len(r.cloudStorageURI) > 0 {
-		pipe, err = r.buildExternalStorePipeline(opCtx, subtask.ID, sm, subtask.Concurrency)
-	} else {
-		pipe, err = r.buildLocalStorePipeline(opCtx, sm, subtask.Concurrency)
+		pipe, err := r.buildExternalStorePipeline(opCtx, subtask.ID, sm, subtask.Concurrency)
+		if err != nil {
+			return err
+		}
+		return executeAndClosePipeline(opCtx, pipe)
 	}
+
+	pipe, err := r.buildLocalStorePipeline(opCtx, sm, subtask.Concurrency)
 	if err != nil {
 		return err
 	}
-
-	err = pipe.Execute()
+	err = executeAndClosePipeline(opCtx, pipe)
 	if err != nil {
+		// For dist task local based ingest, checkpoint is unsupported.
+		// If there is an error we should keep local sort dir clean.
+		err1 := r.bc.FinishAndUnregisterEngines(ingest.OptCleanData)
+		if err1 != nil {
+			logutil.DDLLogger().Warn("read index executor unregister engine failed", zap.Error(err1))
+		}
 		return err
 	}
-	err = pipe.Close()
-	if opCtx.OperatorErr() != nil {
-		return opCtx.OperatorErr()
-	}
-	return err
+	return r.bc.FinishAndUnregisterEngines(ingest.OptCleanData | ingest.OptCheckDup)
 }
 
 func (r *readIndexExecutor) RealtimeSummary() *execute.SubtaskSummary {

pkg/ddl/index.go

@@ -960,28 +960,6 @@ func runIngestReorgJobDist(w *worker, d *ddlCtx, t *meta.Meta, job *model.Job,
 
 func runIngestReorgJob(w *worker, d *ddlCtx, t *meta.Meta, job *model.Job,
 	tbl table.Table, allIndexInfos []*model.IndexInfo) (done bool, ver int64, err error) {
-	bc, ok := ingest.LitBackCtxMgr.Load(job.ID)
-	if ok && bc.Done() {
-		return true, 0, nil
-	}
-	ctx := tidblogutil.WithCategory(w.ctx, "ddl-ingest")
-	var discovery pd.ServiceDiscovery
-	if d != nil {
-		//nolint:forcetypeassert
-		discovery = d.store.(tikv.Storage).GetRegionCache().PDClient().GetServiceDiscovery()
-	}
-	hasUnique := false
-	for _, indexInfo := range allIndexInfos {
-		if indexInfo.Unique {
-			hasUnique = true
-			break
-		}
-	}
-	bc, err = ingest.LitBackCtxMgr.Register(ctx, job.ID, hasUnique, nil, discovery, job.ReorgMeta.ResourceGroupName)
-	if err != nil {
-		ver, err = convertAddIdxJob2RollbackJob(d, t, job, tbl.Meta(), allIndexInfos, err)
-		return false, ver, errors.Trace(err)
-	}
 	done, ver, err = runReorgJobAndHandleErr(w, d, t, job, tbl, allIndexInfos, false)
 	if err != nil {
 		if kv.ErrKeyExists.Equal(err) {
@@ -996,11 +974,8 @@ func runIngestReorgJob(w *worker, d *ddlCtx, t *meta.Meta, job *model.Job,
 		}
 		return false, ver, errors.Trace(err)
 	}
-	if !done {
-		return false, ver, nil
-	}
-	bc.SetDone()
-	return true, ver, nil
+	failpoint.InjectCall("afterRunIngestReorgJob", job, done)
+	return done, ver, nil
 }
 
 func errorIsRetryable(err error, job *model.Job) bool {

pkg/ddl/ingest/BUILD.bazel

@@ -73,7 +73,7 @@ go_test(
     embed = [":ingest"],
     flaky = True,
     race = "on",
-    shard_count = 18,
+    shard_count = 20,
     deps = [
         "//pkg/config",
         "//pkg/ddl",

pkg/ddl/ingest/backend.go

@@ -58,11 +58,9 @@ type BackendCtx interface {
 	// are Register-ed before. It's safe to call it multiple times.
 	//
 	// FinishAndUnregisterEngines is only used in local disk based ingest.
-	FinishAndUnregisterEngines() error
+	FinishAndUnregisterEngines(opt UnregisterOpt) error
 
 	FlushController
-	Done() bool
-	SetDone()
 
 	AttachCheckpointManager(*CheckpointManager)
 	GetCheckpointManager() *CheckpointManager
@@ -99,7 +97,6 @@ type litBackendCtx struct {
 	ctx      context.Context
 	cfg      *lightning.Config
 	sysVars  map[string]string
-	done     bool
 
 	flushing        atomic.Bool
 	timeOfLastFlush atomicutil.Time
@@ -300,7 +297,6 @@ func (bc *litBackendCtx) unsafeImportAndReset(ei *engineInfo) error {
 				zap.Int64("job ID", ei.jobID), zap.Int64("index ID", ei.indexID))
 		}
 		ei.openedEngine = nil
-		ei.closedEngine = nil
 		return err
 	}
 	return nil
@@ -331,16 +327,6 @@ func (bc *litBackendCtx) checkFlush(mode FlushMode) (shouldFlush bool, shouldImp
 	return shouldFlush, shouldImport
 }
 
-// Done returns true if the lightning backfill is done.
-func (bc *litBackendCtx) Done() bool {
-	return bc.done
-}
-
-// SetDone sets the done flag.
-func (bc *litBackendCtx) SetDone() {
-	bc.done = true
-}
-
 // AttachCheckpointManager attaches a checkpoint manager to the backend context.
 func (bc *litBackendCtx) AttachCheckpointManager(mgr *CheckpointManager) {
 	bc.checkpointMgr = mgr

pkg/ddl/ingest/backend_mgr.go

@@ -231,7 +231,7 @@ func (m *litBackendCtxMgr) Unregister(jobID int64) {
 	if !exist {
 		return
 	}
-	_ = bc.FinishAndUnregisterEngines()
+	_ = bc.FinishAndUnregisterEngines(OptCloseEngines)
 	bc.backend.Close()
 	m.memRoot.Release(structSizeBackendCtx)
 	m.memRoot.ReleaseWithTag(encodeBackendTag(jobID))