cherry-pick: Fix affect of online DDL #466

drainer/schema.go

@@ -21,6 +21,8 @@ type Schema struct {
 	schemas map[int64]*model.DBInfo
 	tables  map[int64]*model.TableInfo
 
+	truncateTableID map[int64]struct{}
+
 	schemaMetaVersion int64
 
 	hasImplicitCol bool
@@ -41,6 +43,7 @@ func NewSchema(jobs []*model.Job, hasImplicitCol bool) (*Schema, error) {
 	s := &Schema{
 		hasImplicitCol:      hasImplicitCol,
 		version2SchemaTable: make(map[int64]TableName),
+		truncateTableID:     make(map[int64]struct{}),
 		jobs:                jobs,
 	}
 
@@ -253,12 +256,12 @@ func (s *Schema) handlePreviousDDLJobIfNeed(version int64) error {
 // the second value[string]: the table name
 // the third value[string]: the sql that is corresponding to the job
 // the fourth value[error]: the handleDDL execution's err
-func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
+func (s *Schema) handleDDL(job *model.Job) (schemaName string, tableName string, sql string, err error) {
 	if skipJob(job) {
 		return "", "", "", nil
 	}
 
-	sql := job.Query
+	sql = job.Query
 	if sql == "" {
 		return "", "", "", errors.Errorf("[ddl job sql miss]%+v", job)
 	}
@@ -275,17 +278,16 @@ func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, ""}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, "", sql, nil
+		schemaName = schema.Name.O
 
 	case model.ActionDropSchema:
-		schemaName, err := s.DropSchema(job.SchemaID)
+		schemaName, err = s.DropSchema(job.SchemaID)
 		if err != nil {
 			return "", "", "", errors.Trace(err)
 		}
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schemaName, ""}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schemaName, "", sql, nil
 
 	case model.ActionRenameTable:
 		// ignore schema doesn't support reanme ddl
@@ -312,7 +314,8 @@ func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, table.Name.O}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, table.Name.O, sql, nil
+		schemaName = schema.Name.O
+		tableName = table.Name.O
 
 	case model.ActionCreateTable:
 		table := job.BinlogInfo.TableInfo
@@ -332,29 +335,31 @@ func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, table.Name.O}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, table.Name.O, sql, nil
+		schemaName = schema.Name.O
+		tableName = table.Name.O
 
 	case model.ActionDropTable:
 		schema, ok := s.SchemaByID(job.SchemaID)
 		if !ok {
 			return "", "", "", errors.NotFoundf("schema %d", job.SchemaID)
 		}
 
-		tableName, err := s.DropTable(job.TableID)
+		tableName, err = s.DropTable(job.TableID)
 		if err != nil {
 			return "", "", "", errors.Trace(err)
 		}
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, tableName}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, tableName, sql, nil
+		schemaName = schema.Name.O
 
 	case model.ActionTruncateTable:
 		schema, ok := s.SchemaByID(job.SchemaID)
 		if !ok {
 			return "", "", "", errors.NotFoundf("schema %d", job.SchemaID)
 		}
 
+		// job.TableID is the old table id, different from table.ID
 		_, err := s.DropTable(job.TableID)
 		if err != nil {
 			return "", "", "", errors.Trace(err)
@@ -372,7 +377,9 @@ func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, table.Name.O}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, table.Name.O, sql, nil
+		schemaName = schema.Name.O
+		tableName = table.Name.O
+		s.truncateTableID[job.TableID] = struct{}{}
 
 	default:
 		binlogInfo := job.BinlogInfo
@@ -396,8 +403,17 @@ func (s *Schema) handleDDL(job *model.Job) (string, string, string, error) {
 
 		s.version2SchemaTable[job.BinlogInfo.SchemaVersion] = TableName{schema.Name.O, tbInfo.Name.O}
 		s.currentVersion = job.BinlogInfo.SchemaVersion
-		return schema.Name.O, tbInfo.Name.O, sql, nil
+		schemaName = schema.Name.O
+		tableName = tbInfo.Name.O
 	}
+
+	return
+}
+
+// IsTruncateTableID returns true if the table id have been truncated by truncate table DDL
+func (s *Schema) IsTruncateTableID(id int64) bool {
+	_, ok := s.truncateTableID[id]
+	return ok
 }
 
 func (s *Schema) getSchemaTableAndDelete(version int64) (string, string, error) {
@@ -425,11 +441,10 @@ func addImplicitColumn(table *model.TableInfo) {
 	table.Indices = []*model.IndexInfo{newIndex}
 }
 
-// there's only two status will be in HistoryDDLJob(we fetch at start time):
-// JobStateSynced and JobStateRollbackDone
-// If it fail to commit(to tikv) in 2pc phrase (when changing JobStateDone -> JobStateSynced and add to HistoryDDLJob),
-// then is would't not be add to HistoryDDLJob, and we may get (prewrite + rollback binlog),
-// this binlog event would reach drainer, finally we will get a (p + commit binlog) when tidb retry and successfully commit
+// TiDB write DDL Binlog for every DDL Job, we must ignore jobs that are cancelled or rollback
+// For older version TiDB, it write DDL Binlog in the txn that the state of job is changed to *synced*
+// Now, it write DDL Binlog in the txn that the state of job is changed to *done* (before change to *synced*)
+// At state *done*, it will be always and only changed to *synced*.
 func skipJob(job *model.Job) bool {
-	return !job.IsSynced()
+	return !job.IsSynced() && !job.IsDone()
 }

drainer/syncer.go

@@ -48,9 +48,6 @@ type Syncer struct {
 	positions    map[string]int64
 	initCommitTS int64
 
-	// because TiDB is case-insensitive, only lower-case here.
-	ignoreSchemaNames map[string]struct{}
-
 	ctx    context.Context
 	cancel context.CancelFunc
 
@@ -418,6 +415,9 @@ func (s *Syncer) run(jobs []*model.Job) error {
 		}
 	}
 	s.schema, err = NewSchema(jobs, false)
+	if err != nil {
+		return errors.Trace(err)
+	}
 
 	s.executors, err = createExecutors(s.cfg.DestDBType, s.cfg.To, s.cfg.WorkerCount)
 	if err != nil {
@@ -436,6 +436,7 @@ func (s *Syncer) run(jobs []*model.Job) error {
 		go s.sync(s.executors[i], s.jobCh[i], i)
 	}
 
+	var lastDDLSchemaVersion int64
 	var b *binlogItem
 	for {
 		select {
@@ -463,7 +464,16 @@ func (s *Syncer) run(jobs []*model.Job) error {
 				return errors.Errorf("prewrite %s unmarshal error %v", preWriteValue, err)
 			}
 
+			err = s.rewriteForOldVersion(preWrite)
+			if err != nil {
+				return errors.Annotate(err, "rewrite for old version fail")
+			}
+
 			log.Debug("DML SchemaVersion: ", preWrite.SchemaVersion)
+			if preWrite.SchemaVersion < lastDDLSchemaVersion {
+				log.Debug("encounter older schema dml")
+			}
+
 			err = s.schema.handlePreviousDDLJobIfNeed(preWrite.SchemaVersion)
 			if err != nil {
 				return errors.Trace(err)
@@ -481,6 +491,8 @@ func (s *Syncer) run(jobs []*model.Job) error {
 			s.schema.addJob(b.job)
 
 			log.Debug("DDL SchemaVersion: ", b.job.BinlogInfo.SchemaVersion)
+			lastDDLSchemaVersion = b.job.BinlogInfo.SchemaVersion
+
 			err = s.schema.handlePreviousDDLJobIfNeed(b.job.BinlogInfo.SchemaVersion)
 			if err != nil {
 				return errors.Trace(err)
@@ -645,3 +657,23 @@ func (s *Syncer) GetLastSyncTime() time.Time {
 func (s *Syncer) GetLatestCommitTS() int64 {
 	return s.cp.TS()
 }
+
+// see https://github.com/pingcap/tidb/issues/9304
+// currently, we only drop the data which table id is truncated.
+// because of online DDL, different TiDB instance may see the different schema,
+// it can't be treated simply as one timeline consider both DML and DDL,
+// we must carefully handle every DDL type now and need to find a better design.
+func (s *Syncer) rewriteForOldVersion(pv *pb.PrewriteValue) (err error) {
+	var mutations = make([]pb.TableMutation, 0, len(pv.GetMutations()))
+	for _, mutation := range pv.GetMutations() {
+		if s.schema.IsTruncateTableID(mutation.TableId) {
+			log.Infof("skip old version truncate dml, table id: %d", mutation.TableId)
+			continue
+		}
+
+		mutations = append(mutations, mutation)
+	}
+	pv.Mutations = mutations
+
+	return nil
+}

drainer/syncer_test.go

@@ -24,7 +24,7 @@ func (t *testDrainerSuite) TestHandleDDL(c *C) {
 	c.Assert(sql, Equals, "")
 
 	// check job.Query is empty
-	job = &model.Job{ID: 1, State: model.JobStateSynced}
+	job = &model.Job{ID: 1, State: model.JobStateDone}
 	_, _, sql, err = s.schema.handleDDL(job)
 	c.Assert(sql, Equals, "")
 	c.Assert(err, NotNil, Commentf("should return not found job.Query"))
@@ -82,7 +82,7 @@ func (t *testDrainerSuite) TestHandleDDL(c *C) {
 
 		job = &model.Job{
 			ID:         testCase.jobID,
-			State:      model.JobStateSynced,
+			State:      model.JobStateDone,
 			SchemaID:   testCase.schemaID,
 			TableID:    testCase.tableID,
 			Type:       testCase.jobType,

drainer/translator/flash.go

@@ -16,7 +16,6 @@ import (
 	"github.com/pingcap/tidb/mysql"
 	"github.com/pingcap/tidb/tablecodec"
 	"github.com/pingcap/tidb/types"
-	"github.com/pingcap/tidb/util/codec"
 )
 
 // flashTranslator translates TiDB binlog to flash sqls
@@ -42,45 +41,26 @@ func (f *flashTranslator) GenInsertSQLs(schema string, table *model.TableInfo, r
 	version := makeInternalVersionValue(uint64(commitTS))
 	delFlag := makeInternalDelmarkValue(false)
 
-	colsTypeMap := toFlashColumnTypeMap(columns)
 	columnList := genColumnList(columns)
 	// addition 2 holder is for del flag and version
 	columnPlaceholders := dml.GenColumnPlaceholders(len(columns) + 2)
 	sql := fmt.Sprintf("IMPORT INTO `%s`.`%s` (%s) values (%s);", schema, table.Name.L, columnList, columnPlaceholders)
 
 	for _, row := range rows {
 		//decode the pk value
-		remain, pk, err := codec.DecodeOne(row)
-		hashKey := pk.GetInt64()
+		pk, columnValues, err := insertRowToDatums(table, row)
 		if err != nil {
 			return nil, nil, nil, errors.Trace(err)
 		}
 
-		columnValues, err := tablecodec.DecodeRow(remain, colsTypeMap, gotime.Local)
-		if err != nil {
-			return nil, nil, nil, errors.Trace(err)
-		}
-
-		if columnValues == nil {
-			columnValues = make(map[int64]types.Datum)
-		}
+		hashKey := pk.GetInt64()
 
 		var vals []interface{}
 		vals = append(vals, hashKey)
 		for _, col := range columns {
-			if IsPKHandleColumn(table, col) {
-				columnValues[col.ID] = pk
-				pkVal, err := formatFlashData(&pk, &col.FieldType)
-				if err != nil {
-					return nil, nil, nil, errors.Trace(err)
-				}
-				vals = append(vals, pkVal)
-				continue
-			}
-
 			val, ok := columnValues[col.ID]
 			if !ok {
-				vals = append(vals, col.DefaultValue)
+				vals = append(vals, col.GetDefaultValue())
 			} else {
 				value, err := formatFlashData(&val, &col.FieldType)
 				if err != nil {

drainer/translator/flash_util.go

@@ -87,15 +87,6 @@ func isHandleTypeColumn(colDef *ast.ColumnDef) bool {
 		tp == mysql.TypeLonglong
 }
 
-func toFlashColumnTypeMap(columns []*model.ColumnInfo) map[int64]*types.FieldType {
-	colTypeMap := make(map[int64]*types.FieldType)
-	for _, col := range columns {
-		colTypeMap[col.ID] = &col.FieldType
-	}
-
-	return colTypeMap
-}
-
 func makeRow(pk int64, values []interface{}, version uint64, delFlag uint8) []interface{} {
 	var row []interface{}
 	row = append(row, pk)