REP-5318 Make change stream restartable.

This entails a small refactor of the change stream code so that the
change stream’s creation and iteration both happen under a retryer.

Author: FGasper

go.mod

@@ -12,6 +12,7 @@ require (
 	github.com/pkg/errors v0.9.1
 	github.com/rs/zerolog v1.28.0
 	github.com/samber/lo v1.47.0
+	github.com/samber/mo v1.13.0
 	github.com/stretchr/testify v1.8.0
 	github.com/urfave/cli v1.22.9
 	go.mongodb.org/mongo-driver v1.17.1

go.sum

@@ -84,6 +84,8 @@ github.com/russross/blackfriday/v2 v2.1.0 h1:JIOH55/0cWyOuilr9/qlrm0BSXldqnqwMsf
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
 github.com/samber/lo v1.47.0 h1:z7RynLwP5nbyRscyvcD043DWYoOcYRv3mV8lBeqOCLc=
 github.com/samber/lo v1.47.0/go.mod h1:RmDH9Ct32Qy3gduHQuKJ3gW1fMHAnE/fAzQuf6He5cU=
+github.com/samber/mo v1.13.0 h1:LB1OwfJMju3a6FjghH+AIvzMG0ZPOzgTWj1qaHs1IQ4=
+github.com/samber/mo v1.13.0/go.mod h1:BfkrCPuYzVG3ZljnZB783WIJIGk1mcZr9c9CPf8tAxs=
 github.com/shurcooL/sanitized_anchor_name v1.0.0/go.mod h1:1NzhyTcUVG4SuEtjjoZeVRXNmyL/1OwPU0+IJeTBvfc=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=

internal/util/error.go

@@ -20,8 +20,9 @@ import (
 // `ErrorCode` newtype, but that requires a more invasive change to everything
 // that uses error codes.
 const (
-	LockFailed              int = 107
-	SampleTooManyDuplicates int = 28799
+	LockFailed              = 107
+	SampleTooManyDuplicates = 28799
+	CursorKilled            = 237
 )
 
 //

internal/verifier/change_stream.go

@@ -6,8 +6,11 @@ import (
 	"time"
 
 	"github.com/10gen/migration-verifier/internal/keystring"
+	"github.com/10gen/migration-verifier/internal/retry"
+	"github.com/10gen/migration-verifier/internal/util"
 	"github.com/pkg/errors"
 	"github.com/rs/zerolog"
+	"github.com/samber/mo"
 	"go.mongodb.org/mongo-driver/bson"
 	"go.mongodb.org/mongo-driver/bson/primitive"
 	"go.mongodb.org/mongo-driver/mongo"
@@ -63,7 +66,7 @@ func (verifier *Verifier) HandleChangeStreamEvents(ctx context.Context, batch []
 	for i, changeEvent := range batch {
 		if changeEvent.ClusterTime != nil &&
 			(verifier.lastChangeEventTime == nil ||
-				verifier.lastChangeEventTime.Compare(*changeEvent.ClusterTime) < 0) {
+				verifier.lastChangeEventTime.Before(*changeEvent.ClusterTime)) {
 			verifier.lastChangeEventTime = changeEvent.ClusterTime
 		}
 		switch changeEvent.OpType {
@@ -136,6 +139,7 @@ func (verifier *Verifier) GetChangeStreamFilter() []bson.D {
 // shouldn’t really happen anyway by definition.
 func (verifier *Verifier) readAndHandleOneChangeEventBatch(
 	ctx context.Context,
+	ri *retry.Info,
 	cs *mongo.ChangeStream,
 ) error {
 	eventsRead := 0
@@ -163,6 +167,8 @@ func (verifier *Verifier) readAndHandleOneChangeEventBatch(
 		eventsRead++
 	}
 
+	ri.IterationSuccess()
+
 	if eventsRead == 0 {
 		return nil
 	}
@@ -175,9 +181,11 @@ func (verifier *Verifier) readAndHandleOneChangeEventBatch(
 	return nil
 }
 
-func (verifier *Verifier) iterateChangeStream(ctx context.Context, cs *mongo.ChangeStream) {
-	defer cs.Close(ctx)
-
+func (verifier *Verifier) iterateChangeStream(
+	ctx context.Context,
+	ri *retry.Info,
+	cs *mongo.ChangeStream,
+) error {
 	var lastPersistedTime time.Time
 
 	persistResumeTokenIfNeeded := func() error {
@@ -201,10 +209,7 @@ func (verifier *Verifier) iterateChangeStream(ctx context.Context, cs *mongo.Cha
 
 		// If the context is canceled, return immmediately.
 		case <-ctx.Done():
-			verifier.logger.Debug().
-				Err(ctx.Err()).
-				Msg("Change stream quitting.")
-			return
+			return ctx.Err()
 
 		// If the changeStreamEnderChan has a message, the user has indicated that
 		// source writes are ended. This means we should exit rather than continue
@@ -222,10 +227,11 @@ func (verifier *Verifier) iterateChangeStream(ctx context.Context, cs *mongo.Cha
 				var curTs primitive.Timestamp
 				curTs, err = extractTimestampFromResumeToken(cs.ResumeToken())
 				if err != nil {
-					err = errors.Wrap(err, "failed to extract timestamp from change stream's resume token")
-					break
+					return errors.Wrap(err, "failed to extract timestamp from change stream's resume token")
 				}
 
+				// writesOffTs never refers to a real event,
+				// so we can stop once curTs >= writesOffTs.
 				if !curTs.Before(writesOffTs) {
 					verifier.logger.Debug().
 						Interface("currentTimestamp", curTs).
@@ -235,30 +241,22 @@ func (verifier *Verifier) iterateChangeStream(ctx context.Context, cs *mongo.Cha
 					break
 				}
 
-				err = verifier.readAndHandleOneChangeEventBatch(ctx, cs)
+				err = verifier.readAndHandleOneChangeEventBatch(ctx, ri, cs)
 
 				if err != nil {
-					break
+					return err
 				}
 			}
 
 		default:
-			err = verifier.readAndHandleOneChangeEventBatch(ctx, cs)
+			err = verifier.readAndHandleOneChangeEventBatch(ctx, ri, cs)
 
 			if err == nil {
 				err = persistResumeTokenIfNeeded()
 			}
-		}
 
-		if err != nil && !errors.Is(err, context.Canceled) {
-			verifier.logger.Debug().
-				Err(err).
-				Msg("Sending change stream error.")
-
-			verifier.changeStreamErrChan <- err
-
-			if !gotwritesOffTimestamp {
-				break
+			if err != nil {
+				return err
 			}
 		}
 
@@ -284,10 +282,13 @@ func (verifier *Verifier) iterateChangeStream(ctx context.Context, cs *mongo.Cha
 	}
 
 	infoLog.Msg("Change stream is done.")
+
+	return nil
 }
 
-// StartChangeStream starts the change stream.
-func (verifier *Verifier) StartChangeStream(ctx context.Context) error {
+func (verifier *Verifier) createChangeStream(
+	ctx context.Context,
+) (*mongo.ChangeStream, primitive.Timestamp, error) {
 	pipeline := verifier.GetChangeStreamFilter()
 	opts := options.ChangeStream().
 		SetMaxAwaitTime(1 * time.Second).
@@ -299,7 +300,7 @@ func (verifier *Verifier) StartChangeStream(ctx context.Context) error {
 
 	savedResumeToken, err := verifier.loadChangeStreamResumeToken(ctx)
 	if err != nil {
-		return errors.Wrap(err, "failed to load persisted change stream resume token")
+		return nil, primitive.Timestamp{}, errors.Wrap(err, "failed to load persisted change stream resume token")
 	}
 
 	csStartLogEvent := verifier.logger.Info()
@@ -326,40 +327,100 @@ func (verifier *Verifier) StartChangeStream(ctx context.Context) error {
 
 	sess, err := verifier.srcClient.StartSession()
 	if err != nil {
-		return errors.Wrap(err, "failed to start session")
+		return nil, primitive.Timestamp{}, errors.Wrap(err, "failed to start session")
 	}
 	sctx := mongo.NewSessionContext(ctx, sess)
 	srcChangeStream, err := verifier.srcClient.Watch(sctx, pipeline, opts)
 	if err != nil {
-		return errors.Wrap(err, "failed to open change stream")
+		return nil, primitive.Timestamp{}, errors.Wrap(err, "failed to open change stream")
 	}
 
 	err = verifier.persistChangeStreamResumeToken(ctx, srcChangeStream)
 	if err != nil {
-		return err
+		return nil, primitive.Timestamp{}, err
 	}
 
-	csTimestamp, err := extractTimestampFromResumeToken(srcChangeStream.ResumeToken())
+	startTs, err := extractTimestampFromResumeToken(srcChangeStream.ResumeToken())
 	if err != nil {
-		return errors.Wrap(err, "failed to extract timestamp from change stream's resume token")
+		return nil, primitive.Timestamp{}, errors.Wrap(err, "failed to extract timestamp from change stream's resume token")
 	}
 
+	// With sharded clusters the resume token might lead the cluster time
+	// by 1 increment. In that case we need the actual cluster time;
+	// otherwise we will get errors.
 	clusterTime, err := getClusterTimeFromSession(sess)
 	if err != nil {
-		return errors.Wrap(err, "failed to read cluster time from session")
+		return nil, primitive.Timestamp{}, errors.Wrap(err, "failed to read cluster time from session")
+	}
+
+	if startTs.After(clusterTime) {
+		startTs = clusterTime
 	}
 
-	verifier.srcStartAtTs = &csTimestamp
-	if csTimestamp.After(clusterTime) {
-		verifier.srcStartAtTs = &clusterTime
+	return srcChangeStream, startTs, nil
+}
+
+// StartChangeStream starts the change stream.
+func (verifier *Verifier) StartChangeStream(ctx context.Context) error {
+	// This channel holds the first change stream creation's result, whether
+	// success or failure. Rather than using a Result we could make separate
+	// Timestamp and error channels, but the single channel is cleaner since
+	// there's no chance of "nonsense" like both channels returning a payload.
+	initialCreateResultChan := make(chan mo.Result[primitive.Timestamp])
+
+	go func() {
+		retryer := retry.New(retry.DefaultDurationLimit)
+		retryer = retryer.WithErrorCodes(util.CursorKilled)
+
+		parentThreadWaiting := true
+
+		err := retryer.RunForTransientErrorsOnly(
+			ctx,
+			verifier.logger,
+			func(ri *retry.Info) error {
+				srcChangeStream, startTs, err := verifier.createChangeStream(ctx)
+				if err != nil {
+					if parentThreadWaiting {
+						initialCreateResultChan <- mo.Err[primitive.Timestamp](err)
+						return nil
+					}
+
+					return err
+				}
+
+				defer srcChangeStream.Close(ctx)
+
+				if parentThreadWaiting {
+					initialCreateResultChan <- mo.Ok(startTs)
+					close(initialCreateResultChan)
+					parentThreadWaiting = false
+				}
+
+				return verifier.iterateChangeStream(ctx, ri, srcChangeStream)
+			},
+		)
+
+		if err != nil {
+			// NB: This failure always happens after the initial change stream
+			// creation.
+			verifier.changeStreamErrChan <- err
+			close(verifier.changeStreamErrChan)
+		}
+	}()
+
+	result := <-initialCreateResultChan
+
+	startTs, err := result.Get()
+	if err != nil {
+		return err
 	}
 
+	verifier.srcStartAtTs = &startTs
+
 	verifier.mux.Lock()
 	verifier.changeStreamRunning = true
 	verifier.mux.Unlock()
 
-	go verifier.iterateChangeStream(ctx, srcChangeStream)
-
 	return nil
 }
 

internal/verifier/change_stream_test.go

@@ -13,6 +13,8 @@ import (
 	"go.mongodb.org/mongo-driver/bson"
 	"go.mongodb.org/mongo-driver/bson/primitive"
 	"go.mongodb.org/mongo-driver/mongo"
+	"go.mongodb.org/mongo-driver/mongo/options"
+	"go.mongodb.org/mongo-driver/mongo/readconcern"
 )
 
 func TestChangeStreamFilter(t *testing.T) {
@@ -263,6 +265,92 @@ func (suite *IntegrationTestSuite) TestWithChangeEventsBatching() {
 
 }
 
+func (suite *IntegrationTestSuite) TestCursorKilledResilience() {
+	ctx := suite.Context()
+
+	verifier := suite.BuildVerifier()
+
+	db := suite.srcMongoClient.Database(suite.DBNameForTest())
+	coll := db.Collection("mycoll")
+	suite.Require().NoError(
+		db.CreateCollection(ctx, coll.Name()),
+	)
+
+	// start verifier
+	verifierRunner := RunVerifierCheck(suite.Context(), suite.T(), verifier)
+
+	// wait for generation 0 to end
+	verifierRunner.AwaitGenerationEnd()
+
+	const mvName = "Migration Verifier"
+
+	// Kill verifier’s change stream.
+	cursor, err := suite.srcMongoClient.Database(
+		"admin",
+		options.Database().SetReadConcern(readconcern.Local()),
+	).Aggregate(
+		ctx,
+		mongo.Pipeline{
+			{
+				{"$currentOp", bson.D{
+					{"idleCursors", true},
+				}},
+			},
+			{
+				{"$match", bson.D{
+					{"clientMetadata.application.name", mvName},
+					{"command.collection", "$cmd.aggregate"},
+					{"cursor.originatingCommand.pipeline.0.$_internalChangeStreamOplogMatch",
+						bson.D{{"$type", "object"}},
+					},
+				}},
+			},
+		},
+	)
+	suite.Require().NoError(err)
+
+	var ops []bson.Raw
+	suite.Require().NoError(cursor.All(ctx, &ops))
+
+	for _, cursorRaw := range ops {
+		opId, err := cursorRaw.LookupErr("opid")
+		suite.Require().NoError(err, "should get opid from op")
+
+		suite.T().Logf("Killing change stream op %+v", opId)
+
+		suite.Require().NoError(
+			suite.srcMongoClient.Database("admin").RunCommand(
+				ctx,
+				bson.D{
+					{"killOp", 1},
+					{"op", opId},
+				},
+			).Err(),
+		)
+	}
+
+	_, err = coll.InsertOne(
+		ctx,
+		bson.D{{"_id", "after kill"}},
+	)
+	suite.Require().NoError(err)
+
+	suite.Require().NoError(verifier.WritesOff(ctx))
+
+	suite.Require().NoError(verifierRunner.Await())
+
+	failedTasks, incompleteTasks, err := FetchFailedAndIncompleteTasks(
+		ctx,
+		verifier.verificationTaskCollection(),
+		verificationTaskVerifyDocuments,
+		verifier.generation,
+	)
+	suite.Require().NoError(err)
+
+	suite.Assert().Zero(incompleteTasks, "no incomplete tasks")
+	suite.Require().Len(failedTasks, 1, "expect one failed task")
+}
+
 func (suite *IntegrationTestSuite) TestManyInsertsBeforeWritesOff() {
 	suite.testInsertsBeforeWritesOff(10_000)
 }

internal/verifier/check.go

@@ -87,7 +87,7 @@ func (verifier *Verifier) CheckWorker(ctx context.Context) error {
 		select {
 		case err := <-verifier.changeStreamErrChan:
 			cancel()
-			return err
+			return errors.Wrap(err, "change stream failed")
 		case <-ctx.Done():
 			cancel()
 			return nil