Ejemplo n.º 1
0
// TestTxnCoordSenderErrorWithIntent validates that if a transactional request
// returns an error but also indicates a Writing transaction, the coordinator
// tracks it just like a successful request.
func TestTxnCoordSenderErrorWithIntent(t *testing.T) {
	defer leaktest.AfterTest(t)
	stopper := stop.NewStopper()
	manual := hlc.NewManualClock(0)
	clock := hlc.NewClock(manual.UnixNano)
	clock.SetMaxOffset(20)

	ts := NewTxnCoordSender(senderFn(func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
		txn := ba.Txn.Clone()
		txn.Writing = true
		pErr := roachpb.NewError(roachpb.NewTransactionRetryError())
		pErr.SetTxn(txn)
		return nil, pErr
	}), clock, false, nil, stopper)
	defer stopper.Stop()

	var ba roachpb.BatchRequest
	key := roachpb.Key("test")
	ba.Add(&roachpb.BeginTransactionRequest{Span: roachpb.Span{Key: key}})
	ba.Add(&roachpb.PutRequest{Span: roachpb.Span{Key: key}})
	ba.Add(&roachpb.EndTransactionRequest{})
	ba.Txn = &roachpb.Transaction{Name: "test"}
	if _, pErr := ts.Send(context.Background(), ba); !testutils.IsPError(pErr, "retry txn") {
		t.Fatalf("unexpected error: %v", pErr)
	}

	defer teardownHeartbeats(ts)
	ts.Lock()
	defer ts.Unlock()
	if len(ts.txns) != 1 {
		t.Fatalf("expected transaction to be tracked")
	}
}
Ejemplo n.º 2
0
// TestTxnCoordSenderSingleRoundtripTxn checks that a batch which completely
// holds the writing portion of a Txn (including EndTransaction) does not
// launch a heartbeat goroutine at all.
func TestTxnCoordSenderSingleRoundtripTxn(t *testing.T) {
	defer leaktest.AfterTest(t)
	stopper := stop.NewStopper()
	manual := hlc.NewManualClock(0)
	clock := hlc.NewClock(manual.UnixNano)
	clock.SetMaxOffset(20)

	ts := NewTxnCoordSender(senderFn(func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
		br := ba.CreateReply()
		br.Txn = ba.Txn.Clone()
		br.Txn.Writing = true
		return br, nil
	}), clock, false, nil, stopper)

	// Stop the stopper manually, prior to trying the transaction. This has the
	// effect of returning a NodeUnavailableError for any attempts at launching
	// a heartbeat goroutine.
	stopper.Stop()

	var ba roachpb.BatchRequest
	key := roachpb.Key("test")
	ba.Add(&roachpb.BeginTransactionRequest{Span: roachpb.Span{Key: key}})
	ba.Add(&roachpb.PutRequest{Span: roachpb.Span{Key: key}})
	ba.Add(&roachpb.EndTransactionRequest{})
	ba.Txn = &roachpb.Transaction{Name: "test"}
	_, pErr := ts.Send(context.Background(), ba)
	if pErr != nil {
		t.Fatal(pErr)
	}
}
Ejemplo n.º 3
0
Archivo: txn.go Proyecto: l2x/cockroach
func (ts *txnSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	// Send call through wrapped sender.
	ba.Txn = &ts.Proto
	ba.SetNewRequest()
	br, pErr := ts.wrapped.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(ts.wrapped, br))
	}

	// Only successful requests can carry an updated Txn in their response
	// header. Any error (e.g. a restart) can have a Txn attached to them as
	// well; those update our local state in the same way for the next attempt.
	// The exception is if our transaction was aborted and needs to restart
	// from scratch, in which case we do just that.
	if pErr == nil {
		ts.Proto.Update(br.Txn)
		return br, nil
	} else if _, ok := pErr.GoError().(*roachpb.TransactionAbortedError); ok {
		// On Abort, reset the transaction so we start anew on restart.
		ts.Proto = roachpb.Transaction{
			Name:      ts.Proto.Name,
			Isolation: ts.Proto.Isolation,
		}
		// Acts as a minimum priority on restart.
		if pErr.GetTxn() != nil {
			ts.Proto.Priority = pErr.GetTxn().Priority
		}
	} else if pErr.TransactionRestart != roachpb.TransactionRestart_ABORT {
		ts.Proto.Update(pErr.GetTxn())
	}
	return nil, pErr
}
Ejemplo n.º 4
0
func (ts *txnSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	// Send call through wrapped sender.
	ba.Txn = &ts.Proto
	br, pErr := ts.wrapped.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(ts.wrapped, br))
	}

	// TODO(tschottdorf): see about using only the top-level *roachpb.Error
	// information for this restart logic (includes adding the Txn).
	err := pErr.GoError()
	// Only successful requests can carry an updated Txn in their response
	// header. Any error (e.g. a restart) can have a Txn attached to them as
	// well; those update our local state in the same way for the next attempt.
	// The exception is if our transaction was aborted and needs to restart
	// from scratch, in which case we do just that.
	if err == nil {
		ts.Proto.Update(br.Txn)
		return br, nil
	} else if abrtErr, ok := err.(*roachpb.TransactionAbortedError); ok {
		// On Abort, reset the transaction so we start anew on restart.
		ts.Proto = roachpb.Transaction{
			Name:      ts.Proto.Name,
			Isolation: ts.Proto.Isolation,
		}
		if abrtTxn := abrtErr.Transaction(); abrtTxn != nil {
			// Acts as a minimum priority on restart.
			ts.Proto.Priority = abrtTxn.Priority
		}
	} else if txnErr, ok := err.(roachpb.TransactionRestartError); ok {
		ts.Proto.Update(txnErr.Transaction())
	}
	return nil, pErr
}
Ejemplo n.º 5
0
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	var store *Store
	var err error

	// If we aren't given a Replica, then a little bending over
	// backwards here. This case applies exclusively to unittests.
	if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
		var repl *roachpb.ReplicaDescriptor
		var rangeID roachpb.RangeID
		rs := keys.Range(ba)
		rangeID, repl, err = ls.lookupReplica(rs.Key, rs.EndKey)
		if err == nil {
			ba.RangeID = rangeID
			ba.Replica = *repl
		}
	}

	ctx = log.Add(ctx,
		log.RangeID, ba.RangeID)

	if err == nil {
		store, err = ls.GetStore(ba.Replica.StoreID)
	}

	if err != nil {
		return nil, roachpb.NewError(err)
	}

	sp, cleanupSp := tracing.SpanFromContext(opStores, store.Tracer(), ctx)
	defer cleanupSp()
	if ba.Txn != nil {
		// For calls that read data within a txn, we keep track of timestamps
		// observed from the various participating nodes' HLC clocks. If we have
		// a timestamp on file for this Node which is smaller than MaxTimestamp,
		// we can lower MaxTimestamp accordingly. If MaxTimestamp drops below
		// OrigTimestamp, we effectively can't see uncertainty restarts any
		// more.
		// Note that it's not an issue if MaxTimestamp propagates back out to
		// the client via a returned Transaction update - when updating a Txn
		// from another, the larger MaxTimestamp wins.
		if maxTS, ok := ba.Txn.GetObservedTimestamp(ba.Replica.NodeID); ok && maxTS.Less(ba.Txn.MaxTimestamp) {
			// Copy-on-write to protect others we might be sharing the Txn with.
			shallowTxn := *ba.Txn
			// The uncertainty window is [OrigTimestamp, maxTS), so if that window
			// is empty, there won't be any uncertainty restarts.
			if !ba.Txn.OrigTimestamp.Less(maxTS) {
				sp.LogEvent("read has no clock uncertainty")
			}
			shallowTxn.MaxTimestamp.Backward(maxTS)
			ba.Txn = &shallowTxn
		}
	}
	br, pErr := store.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(store, br))
	}
	return br, pErr
}
Ejemplo n.º 6
0
func (tc *TxnCoordSender) heartbeat(id string, trace *tracer.Trace, ctx context.Context) bool {
	tc.Lock()
	proceed := true
	txnMeta := tc.txns[id]
	// Before we send a heartbeat, determine whether this transaction
	// should be considered abandoned. If so, exit heartbeat.
	if txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow()) {
		// TODO(tschottdorf): should we be more proactive here?
		// The client might be continuing the transaction
		// through another coordinator, but in the most likely
		// case it's just gone and the open transaction record
		// could block concurrent operations.
		if log.V(1) {
			log.Infof("transaction %s abandoned; stopping heartbeat",
				txnMeta.txn)
		}
		proceed = false
	}
	// txnMeta.txn is possibly replaced concurrently,
	// so grab a copy before unlocking.
	txn := txnMeta.txn
	tc.Unlock()
	if !proceed {
		return false
	}

	hb := &roachpb.HeartbeatTxnRequest{}
	hb.Key = txn.Key
	ba := roachpb.BatchRequest{}
	ba.Timestamp = tc.clock.Now()
	ba.CmdID = ba.GetOrCreateCmdID(ba.Timestamp.WallTime)
	ba.Txn = txn.Clone()
	ba.Add(hb)

	epochEnds := trace.Epoch("heartbeat")
	_, err := tc.wrapped.Send(ctx, ba)
	epochEnds()
	// If the transaction is not in pending state, then we can stop
	// the heartbeat. It's either aborted or committed, and we resolve
	// write intents accordingly.
	if err != nil {
		log.Warningf("heartbeat to %s failed: %s", txn, err)
	}
	// TODO(bdarnell): once we have gotten a heartbeat response with
	// Status != PENDING, future heartbeats are useless. However, we
	// need to continue the heartbeatLoop until the client either
	// commits or abandons the transaction. We could save a little
	// pointless work by restructuring this loop to stop sending
	// heartbeats between the time that the transaction is aborted and
	// the client finds out. Furthermore, we could use this information
	// to send TransactionAbortedErrors to the client so it can restart
	// immediately instead of running until its EndTransaction.
	return true
}
Ejemplo n.º 7
0
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	sp := tracing.SpanFromContext(ctx)
	var store *Store
	var pErr *roachpb.Error

	// If we aren't given a Replica, then a little bending over
	// backwards here. This case applies exclusively to unittests.
	if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
		var repl *roachpb.ReplicaDescriptor
		var rangeID roachpb.RangeID
		rs := keys.Range(ba)
		rangeID, repl, pErr = ls.lookupReplica(rs.Key, rs.EndKey)
		if pErr == nil {
			ba.RangeID = rangeID
			ba.Replica = *repl
		}
	}

	ctx = log.Add(ctx,
		log.RangeID, ba.RangeID)

	if pErr == nil {
		store, pErr = ls.GetStore(ba.Replica.StoreID)
	}

	var br *roachpb.BatchResponse
	if pErr != nil {
		return nil, pErr
	}
	// For calls that read data within a txn, we can avoid uncertainty
	// related retries in certain situations. If the node is in
	// "CertainNodes", we need not worry about uncertain reads any
	// more. Setting MaxTimestamp=OrigTimestamp for the operation
	// accomplishes that. See roachpb.Transaction.CertainNodes for details.
	if ba.Txn != nil && ba.Txn.CertainNodes.Contains(ba.Replica.NodeID) {
		// MaxTimestamp = Timestamp corresponds to no clock uncertainty.
		sp.LogEvent("read has no clock uncertainty")
		// Copy-on-write to protect others we might be sharing the Txn with.
		shallowTxn := *ba.Txn
		// We set to OrigTimestamp because that works for both SNAPSHOT and
		// SERIALIZABLE: If we used Timestamp instead, we could run into
		// unnecessary retries at SNAPSHOT. For example, a SNAPSHOT txn at
		// OrigTimestamp = 1000.0, Timestamp = 2000.0, MaxTimestamp = 3000.0
		// will always read at 1000, so a MaxTimestamp of 2000 will still let
		// it restart with uncertainty when it finds a value in (1000, 2000).
		shallowTxn.MaxTimestamp = ba.Txn.OrigTimestamp
		ba.Txn = &shallowTxn
	}
	br, pErr = store.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(store, br))
	}
	return br, pErr
}
Ejemplo n.º 8
0
// TestTxnCoordSenderErrorWithIntent validates that if a transactional request
// returns an error but also indicates a Writing transaction, the coordinator
// tracks it just like a successful request.
func TestTxnCoordSenderErrorWithIntent(t *testing.T) {
	defer leaktest.AfterTest(t)()
	stopper := stop.NewStopper()
	defer stopper.Stop()
	manual := hlc.NewManualClock(0)
	clock := hlc.NewClock(manual.UnixNano)
	clock.SetMaxOffset(20)

	testCases := []struct {
		roachpb.Error
		errMsg string
	}{
		{*roachpb.NewError(roachpb.NewTransactionRetryError()), "retry txn"},
		{*roachpb.NewError(roachpb.NewTransactionPushError(roachpb.Transaction{
			TxnMeta: enginepb.TxnMeta{
				ID: uuid.NewV4(),
			}})), "failed to push"},
		{*roachpb.NewErrorf("testError"), "testError"},
	}
	for i, test := range testCases {
		func() {
			senderFunc := func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
				txn := ba.Txn.Clone()
				txn.Writing = true
				pErr := &roachpb.Error{}
				*pErr = test.Error
				pErr.SetTxn(&txn)
				return nil, pErr
			}
			ctx := tracing.WithTracer(context.Background(), tracing.NewTracer())
			ts := NewTxnCoordSender(ctx, senderFn(senderFunc), clock, false, stopper, MakeTxnMetrics())

			var ba roachpb.BatchRequest
			key := roachpb.Key("test")
			ba.Add(&roachpb.BeginTransactionRequest{Span: roachpb.Span{Key: key}})
			ba.Add(&roachpb.PutRequest{Span: roachpb.Span{Key: key}})
			ba.Add(&roachpb.EndTransactionRequest{})
			ba.Txn = &roachpb.Transaction{Name: "test"}
			_, pErr := ts.Send(context.Background(), ba)
			if !testutils.IsPError(pErr, test.errMsg) {
				t.Errorf("%d: error did not match %s: %v", i, test.errMsg, pErr)
			}

			defer teardownHeartbeats(ts)
			ts.Lock()
			defer ts.Unlock()
			if len(ts.txns) != 1 {
				t.Errorf("%d: expected transaction to be tracked", i)
			}
		}()
	}
}
Ejemplo n.º 9
0
func (ts *txnSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	// Send call through wrapped sender.
	ba.Txn = &ts.Proto
	if ts.UserPriority > 0 {
		ba.UserPriority = ts.UserPriority
	}

	ctx = opentracing.ContextWithSpan(ctx, ts.Trace)

	ba.SetNewRequest()
	br, pErr := ts.wrapped.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(ts.wrapped, br))
	}

	if br != nil {
		for _, encSp := range br.CollectedSpans {
			var newSp basictracer.RawSpan
			if err := tracing.DecodeRawSpan(encSp, &newSp); err != nil {
				return nil, roachpb.NewError(err)
			}
			ts.CollectedSpans = append(ts.CollectedSpans, newSp)
		}
	}
	// Only successful requests can carry an updated Txn in their response
	// header. Any error (e.g. a restart) can have a Txn attached to them as
	// well; those update our local state in the same way for the next attempt.
	// The exception is if our transaction was aborted and needs to restart
	// from scratch, in which case we do just that.
	if pErr == nil {
		ts.Proto.Update(br.Txn)
		return br, nil
	} else if _, ok := pErr.GetDetail().(*roachpb.TransactionAbortedError); ok {
		// On Abort, reset the transaction so we start anew on restart.
		ts.Proto = roachpb.Transaction{
			TxnMeta: roachpb.TxnMeta{
				Isolation: ts.Proto.Isolation,
			},
			Name: ts.Proto.Name,
		}
		// Acts as a minimum priority on restart.
		if pErr.GetTxn() != nil {
			ts.Proto.Priority = pErr.GetTxn().Priority
		}
	} else if pErr.TransactionRestart != roachpb.TransactionRestart_ABORT {
		ts.Proto.Update(pErr.GetTxn())
	}
	return nil, pErr
}
Ejemplo n.º 10
0
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	sp := tracing.SpanFromContext(ctx)
	var store *Store
	var pErr *roachpb.Error

	// If we aren't given a Replica, then a little bending over
	// backwards here. This case applies exclusively to unittests.
	if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
		var repl *roachpb.ReplicaDescriptor
		var rangeID roachpb.RangeID
		rs := keys.Range(ba)
		rangeID, repl, pErr = ls.lookupReplica(rs.Key, rs.EndKey)
		if pErr == nil {
			ba.RangeID = rangeID
			ba.Replica = *repl
		}
	}

	ctx = log.Add(ctx,
		log.RangeID, ba.RangeID)

	if pErr == nil {
		store, pErr = ls.GetStore(ba.Replica.StoreID)
	}

	var br *roachpb.BatchResponse
	if pErr != nil {
		return nil, pErr
	}
	// For calls that read data within a txn, we can avoid uncertainty
	// related retries in certain situations. If the node is in
	// "CertainNodes", we need not worry about uncertain reads any
	// more. Setting MaxTimestamp=Timestamp for the operation
	// accomplishes that. See roachpb.Transaction.CertainNodes for details.
	if ba.Txn != nil && ba.Txn.CertainNodes.Contains(ba.Replica.NodeID) {
		// MaxTimestamp = Timestamp corresponds to no clock uncertainty.
		sp.LogEvent("read has no clock uncertainty")
		// Copy-on-write to protect others we might be sharing the Txn with.
		shallowTxn := *ba.Txn
		shallowTxn.MaxTimestamp = ba.Txn.Timestamp
		ba.Txn = &shallowTxn
	}
	br, pErr = store.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(store, br))
	}
	return br, pErr
}
Ejemplo n.º 11
0
// maybeBeginTxn begins a new transaction if a txn has been specified
// in the request but has a nil ID. The new transaction is initialized
// using the name and isolation in the otherwise uninitialized txn.
// The Priority, if non-zero is used as a minimum.
//
// No transactional writes are allowed unless preceded by a begin
// transaction request within the same batch. The exception is if the
// transaction is already in state txn.Writing=true.
func (tc *TxnCoordSender) maybeBeginTxn(ba *roachpb.BatchRequest) error {
	if ba.Txn == nil {
		return nil
	}
	if len(ba.Requests) == 0 {
		return util.Errorf("empty batch with txn")
	}
	if ba.Txn.ID == nil {
		// Create transaction without a key. The key is set when a begin
		// transaction request is received.

		// The initial timestamp may be communicated by a higher layer.
		// If so, use that. Otherwise make up a new one.
		timestamp := ba.Txn.OrigTimestamp
		if timestamp == roachpb.ZeroTimestamp {
			timestamp = tc.clock.Now()
		}
		newTxn := roachpb.NewTransaction(ba.Txn.Name, nil, ba.UserPriority,
			ba.Txn.Isolation, timestamp, tc.clock.MaxOffset().Nanoseconds())
		// Use existing priority as a minimum. This is used on transaction
		// aborts to ratchet priority when creating successor transaction.
		if newTxn.Priority < ba.Txn.Priority {
			newTxn.Priority = ba.Txn.Priority
		}
		ba.Txn = newTxn
	}

	// Check for a begin transaction to set txn key based on the key of
	// the first transactional write. Also enforce that no transactional
	// writes occur before a begin transaction.
	var haveBeginTxn bool
	for _, req := range ba.Requests {
		args := req.GetInner()
		if bt, ok := args.(*roachpb.BeginTransactionRequest); ok {
			if haveBeginTxn || ba.Txn.Writing {
				return util.Errorf("begin transaction requested twice in the same transaction: %s", ba.Txn)
			}
			haveBeginTxn = true
			if ba.Txn.Key == nil {
				ba.Txn.Key = bt.Key
			}
		}
		if roachpb.IsTransactionWrite(args) && !haveBeginTxn && !ba.Txn.Writing {
			return util.Errorf("transactional write before begin transaction")
		}
	}
	return nil
}
Ejemplo n.º 12
0
func (tc *TxnCoordSender) clientHasAbandoned(txnID uuid.UUID) {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	var intentSpans []roachpb.Span

	// TODO(tschottdorf): should we be more proactive here?
	// The client might be continuing the transaction
	// through another coordinator, but in the most likely
	// case it's just gone and the open transaction record
	// could block concurrent operations.
	if log.V(1) {
		log.Infof("transaction %s abandoned; stopping heartbeat", txnMeta.txn)
	}
	// Grab the intents here to avoid potential race.
	intentSpans = collectIntentSpans(txnMeta.keys)
	txnMeta.keys.Clear()

	// txnMeta.txn is possibly replaced concurrently,
	// so grab a copy before unlocking.
	txn := txnMeta.txn.Clone()
	tc.Unlock()

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	// Actively abort the transaction and its intents since we assume it's abandoned.
	et := &roachpb.EndTransactionRequest{
		Span: roachpb.Span{
			Key: txn.Key,
		},
		Commit:      false,
		IntentSpans: intentSpans,
	}
	ba.Add(et)
	tc.stopper.RunAsyncTask(func() {
		// Use the wrapped sender since the normal Sender
		// does not allow clients to specify intents.
		// TODO(tschottdorf): not using the existing context here since that
		// leads to use-after-finish of the contained trace. Should fork off
		// before the goroutine.
		if _, pErr := tc.wrapped.Send(context.Background(), ba); pErr != nil {
			if log.V(1) {
				log.Warningf("abort due to inactivity failed for %s: %s ", txn, pErr)
			}
		}
	})
}
Ejemplo n.º 13
0
// tryAsyncAbort (synchronously) grabs a copy of the txn proto and the intents
// (which it then clears from txnMeta), and asynchronously tries to abort the
// transaction.
func (tc *TxnCoordSender) tryAsyncAbort(txnID uuid.UUID) {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	// Clone the intents and the txn to avoid data races.
	txnMeta.keys = append([]roachpb.Span(nil), txnMeta.keys...)
	roachpb.MergeSpans(&txnMeta.keys)
	intentSpans := txnMeta.keys
	txnMeta.keys = nil
	txn := txnMeta.txn.Clone()
	tc.Unlock()

	// Since we don't hold the lock continuously, it's possible that two aborts
	// raced here. That's fine (and probably better than the alternative, which
	// is missing new intents sometimes).
	if txn.Status != roachpb.PENDING {
		return
	}

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	et := &roachpb.EndTransactionRequest{
		Span: roachpb.Span{
			Key: txn.Key,
		},
		Commit:      false,
		IntentSpans: intentSpans,
	}
	ba.Add(et)
	if err := tc.stopper.RunAsyncTask(func() {
		// Use the wrapped sender since the normal Sender does not allow
		// clients to specify intents.
		// TODO(tschottdorf): not using the existing context here since that
		// leads to use-after-finish of the contained trace. Should fork off
		// before the goroutine.
		if _, pErr := tc.wrapped.Send(context.Background(), ba); pErr != nil {
			if log.V(1) {
				log.Warningf("abort due to inactivity failed for %s: %s ", txn, pErr)
			}
		}
	}); err != nil {
		log.Warning(err)
	}
}
Ejemplo n.º 14
0
func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	txn := txnMeta.txn.Clone()
	tc.Unlock()

	// Before we send a heartbeat, determine whether this transaction should be
	// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
	// it is a cancellable Context and we skip this check and use the ctx lifetime
	// instead of a timeout.
	if ctx.Done() == nil && txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow()) {
		tc.clientHasAbandoned(txnID)
		return false
	}

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	hb := &roachpb.HeartbeatTxnRequest{
		Now: tc.clock.Now(),
	}
	hb.Key = txn.Key
	ba.Add(hb)

	log.Trace(ctx, "heartbeat")
	_, err := tc.wrapped.Send(ctx, ba)
	// If the transaction is not in pending state, then we can stop
	// the heartbeat. It's either aborted or committed, and we resolve
	// write intents accordingly.
	if err != nil {
		log.Warningf("heartbeat to %s failed: %s", txn, err)
	}
	// TODO(bdarnell): once we have gotten a heartbeat response with
	// Status != PENDING, future heartbeats are useless. However, we
	// need to continue the heartbeatLoop until the client either
	// commits or abandons the transaction. We could save a little
	// pointless work by restructuring this loop to stop sending
	// heartbeats between the time that the transaction is aborted and
	// the client finds out. Furthermore, we could use this information
	// to send TransactionAbortedErrors to the client so it can restart
	// immediately instead of running until its EndTransaction.
	return true
}
Ejemplo n.º 15
0
// maybeBeginTxn begins a new transaction if a txn has been specified
// in the request but has a nil ID. The new transaction is initialized
// using the name and isolation in the otherwise uninitialized txn.
// The Priority, if non-zero is used as a minimum.
func (tc *TxnCoordSender) maybeBeginTxn(ba *roachpb.BatchRequest) {
	if ba.Txn == nil {
		return
	}
	if len(ba.Requests) == 0 {
		panic("empty batch with txn")
	}
	if len(ba.Txn.ID) == 0 {
		// TODO(tschottdorf): should really choose the first txn write here.
		firstKey := ba.Requests[0].GetInner().Header().Key
		newTxn := roachpb.NewTransaction(ba.Txn.Name, keys.KeyAddress(firstKey), ba.GetUserPriority(),
			ba.Txn.Isolation, tc.clock.Now(), tc.clock.MaxOffset().Nanoseconds())
		// Use existing priority as a minimum. This is used on transaction
		// aborts to ratchet priority when creating successor transaction.
		if newTxn.Priority < ba.Txn.Priority {
			newTxn.Priority = ba.Txn.Priority
		}
		ba.Txn = newTxn
	}
}
Ejemplo n.º 16
0
// TestTruncateWithSpanAndDescriptor verifies that a batch request is truncated with a
// range span and the range of a descriptor found in cache.
func TestTruncateWithSpanAndDescriptor(t *testing.T) {
	defer leaktest.AfterTest(t)
	g, s := makeTestGossip(t)
	defer s()

	g.SetNodeID(1)
	if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{NodeID: 1}); err != nil {
		t.Fatal(err)
	}
	nd := &roachpb.NodeDescriptor{
		NodeID:  roachpb.NodeID(1),
		Address: util.MakeUnresolvedAddr(testAddress.Network(), testAddress.String()),
	}
	if err := g.AddInfoProto(gossip.MakeNodeIDKey(roachpb.NodeID(1)), nd, time.Hour); err != nil {
		t.Fatal(err)
	}

	// Fill mockRangeDescriptorDB with two descriptors. When a
	// range descriptor is looked up by key "b", return the second
	// descriptor whose range is ["a", "c") and partially overlaps
	// with the first descriptor's range.
	var descriptor1 = roachpb.RangeDescriptor{
		RangeID:  1,
		StartKey: roachpb.RKeyMin,
		EndKey:   roachpb.RKey("b"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	var descriptor2 = roachpb.RangeDescriptor{
		RangeID:  2,
		StartKey: roachpb.RKey("a"),
		EndKey:   roachpb.RKey("c"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	descDB := mockRangeDescriptorDB(func(key roachpb.RKey, _, _ bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
		desc := descriptor1
		if key.Equal(roachpb.RKey("b")) {
			desc = descriptor2
		}
		return []roachpb.RangeDescriptor{desc}, nil
	})

	// Define our rpcSend stub which checks the span of the batch
	// requests. The first request should be the point request on
	// "a". The second request should be on "b".
	first := true
	var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) proto.Message, getReply func() proto.Message, _ *rpc.Context) ([]proto.Message, error) {
		if method != "Node.Batch" {
			return nil, util.Errorf("unexpected method %v", method)
		}

		ba := getArgs(testAddress).(*roachpb.BatchRequest)
		rs := keys.Range(*ba)
		if first {
			if !(rs.Key.Equal(roachpb.RKey("a")) && rs.EndKey.Equal(roachpb.RKey("a").Next())) {
				t.Errorf("Unexpected span [%s,%s)", rs.Key, rs.EndKey)
			}
			first = false
		} else {
			if !(rs.Key.Equal(roachpb.RKey("b")) && rs.EndKey.Equal(roachpb.RKey("b").Next())) {
				t.Errorf("Unexpected span [%s,%s)", rs.Key, rs.EndKey)
			}
		}

		batchReply := getReply().(*roachpb.BatchResponse)
		reply := &roachpb.PutResponse{}
		batchReply.Add(reply)
		return []proto.Message{batchReply}, nil
	}

	ctx := &DistSenderContext{
		RPCSend:           testFn,
		RangeDescriptorDB: descDB,
	}
	ds := NewDistSender(ctx, g)

	// Send a batch request contains two puts. In the first
	// attempt, the range of the descriptor found in the cache is
	// ["a", "b"). The request is truncated to contain only the put
	// on "a".
	//
	// In the second attempt, The range of the descriptor found in
	// the cache is ["a", c"), but the put on "a" will not be
	// resent. The request is truncated to contain only the put on "b".
	ba := roachpb.BatchRequest{}
	ba.Txn = &roachpb.Transaction{Name: "test"}
	val := roachpb.MakeValueFromString("val")
	ba.Add(roachpb.NewPut(keys.RangeTreeNodeKey(roachpb.RKey("a")), val).(*roachpb.PutRequest))
	ba.Add(roachpb.NewPut(keys.RangeTreeNodeKey(roachpb.RKey("b")), val).(*roachpb.PutRequest))

	_, pErr := ds.Send(context.Background(), ba)
	if err := pErr.GoError(); err != nil {
		t.Fatal(err)
	}
}
Ejemplo n.º 17
0
// TestMultiRangeSplitEndTransaction verifies that when a chunk of batch looks
// like it's going to be dispatched to more than one range, it will be split
// up if it it contains EndTransaction.
func TestMultiRangeSplitEndTransaction(t *testing.T) {
	defer leaktest.AfterTest(t)
	g, s := makeTestGossip(t)
	defer s()

	testCases := []struct {
		put1, put2, et roachpb.Key
		exp            [][]roachpb.Method
	}{
		{
			// Everything hits the first range, so we get a 1PC txn.
			roachpb.Key("a1"), roachpb.Key("a2"), roachpb.Key("a3"),
			[][]roachpb.Method{{roachpb.Put, roachpb.Put, roachpb.EndTransaction}},
		},
		{
			// Only EndTransaction hits the second range.
			roachpb.Key("a1"), roachpb.Key("a2"), roachpb.Key("b"),
			[][]roachpb.Method{{roachpb.Put, roachpb.Put}, {roachpb.EndTransaction}},
		},
		{
			// One write hits the second range, so EndTransaction has to be split off.
			// In this case, going in the usual order without splitting off
			// would actually be fine, but it doesn't seem worth optimizing at
			// this point.
			roachpb.Key("a1"), roachpb.Key("b1"), roachpb.Key("a1"),
			[][]roachpb.Method{{roachpb.Put, roachpb.Noop}, {roachpb.Noop, roachpb.Put}, {roachpb.EndTransaction}},
		},
		{
			// Both writes go to the second range, but not EndTransaction.
			roachpb.Key("b1"), roachpb.Key("b2"), roachpb.Key("a1"),
			[][]roachpb.Method{{roachpb.Put, roachpb.Put}, {roachpb.EndTransaction}},
		},
	}

	if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{NodeID: 1}); err != nil {
		t.Fatal(err)
	}
	nd := &roachpb.NodeDescriptor{
		NodeID:  roachpb.NodeID(1),
		Address: util.MakeUnresolvedAddr(testAddress.Network(), testAddress.String()),
	}
	if err := g.AddInfoProto(gossip.MakeNodeIDKey(roachpb.NodeID(1)), nd, time.Hour); err != nil {
		t.Fatal(err)

	}

	// Fill mockRangeDescriptorDB with two descriptors.
	var descriptor1 = roachpb.RangeDescriptor{
		RangeID:  1,
		StartKey: roachpb.RKeyMin,
		EndKey:   roachpb.RKey("b"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	var descriptor2 = roachpb.RangeDescriptor{
		RangeID:  2,
		StartKey: roachpb.RKey("b"),
		EndKey:   roachpb.RKeyMax,
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	descDB := mockRangeDescriptorDB(func(key roachpb.RKey, _, _ bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
		desc := descriptor1
		if !key.Less(roachpb.RKey("b")) {
			desc = descriptor2
		}
		return []roachpb.RangeDescriptor{desc}, nil
	})

	for _, test := range testCases {
		var act [][]roachpb.Method
		var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, ga func(addr net.Addr) proto.Message, _ func() proto.Message, _ *rpc.Context) ([]proto.Message, error) {
			ba := ga(testAddress).(*roachpb.BatchRequest)
			var cur []roachpb.Method
			for _, union := range ba.Requests {
				cur = append(cur, union.GetInner().Method())
			}
			act = append(act, cur)
			return []proto.Message{ba.CreateReply()}, nil
		}

		ctx := &DistSenderContext{
			RPCSend:           testFn,
			RangeDescriptorDB: descDB,
		}
		ds := NewDistSender(ctx, g)

		// Send a batch request containing two puts.
		var ba roachpb.BatchRequest
		ba.Txn = &roachpb.Transaction{Name: "test"}
		val := roachpb.MakeValueFromString("val")
		ba.Add(roachpb.NewPut(roachpb.Key(test.put1), val).(*roachpb.PutRequest))
		ba.Add(roachpb.NewPut(roachpb.Key(test.put2), val).(*roachpb.PutRequest))
		ba.Add(&roachpb.EndTransactionRequest{Span: roachpb.Span{Key: test.et}})

		_, pErr := ds.Send(context.Background(), ba)
		if err := pErr.GoError(); err != nil {
			t.Fatal(err)
		}

		if !reflect.DeepEqual(test.exp, act) {
			t.Fatalf("expected %v, got %v", test.exp, act)
		}
	}
}
Ejemplo n.º 18
0
// TestTruncateWithLocalSpanAndDescriptor verifies that a batch request with local keys
// is truncated with a range span and the range of a descriptor found in cache.
func TestTruncateWithLocalSpanAndDescriptor(t *testing.T) {
	defer leaktest.AfterTest(t)()
	g, s := makeTestGossip(t)
	defer s()

	if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{NodeID: 1}); err != nil {
		t.Fatal(err)
	}
	nd := &roachpb.NodeDescriptor{
		NodeID:  roachpb.NodeID(1),
		Address: util.MakeUnresolvedAddr(testAddress.Network(), testAddress.String()),
	}
	if err := g.AddInfoProto(gossip.MakeNodeIDKey(roachpb.NodeID(1)), nd, time.Hour); err != nil {
		t.Fatal(err)
	}

	// Fill mockRangeDescriptorDB with two descriptors.
	var descriptor1 = roachpb.RangeDescriptor{
		RangeID:  1,
		StartKey: roachpb.RKeyMin,
		EndKey:   roachpb.RKey("b"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	var descriptor2 = roachpb.RangeDescriptor{
		RangeID:  2,
		StartKey: roachpb.RKey("b"),
		EndKey:   roachpb.RKey("c"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	var descriptor3 = roachpb.RangeDescriptor{
		RangeID:  3,
		StartKey: roachpb.RKey("c"),
		EndKey:   roachpb.RKeyMax,
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}

	descDB := mockRangeDescriptorDB(func(key roachpb.RKey, _, _ bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
		switch {
		case !key.Less(roachpb.RKey("c")):
			return []roachpb.RangeDescriptor{descriptor3}, nil
		case !key.Less(roachpb.RKey("b")):
			return []roachpb.RangeDescriptor{descriptor2}, nil
		default:
			return []roachpb.RangeDescriptor{descriptor1}, nil
		}
	})

	// Define our rpcSend stub which checks the span of the batch
	// requests.
	requests := 0
	sendStub := func(_ SendOptions, _ ReplicaSlice, ba roachpb.BatchRequest, _ *rpc.Context) (*roachpb.BatchResponse, error) {
		h := ba.Requests[0].GetInner().Header()
		switch requests {
		case 0:
			wantStart := keys.RangeDescriptorKey(roachpb.RKey("a"))
			wantEnd := keys.MakeRangeKeyPrefix(roachpb.RKey("b"))
			if !(h.Key.Equal(wantStart) && h.EndKey.Equal(wantEnd)) {
				t.Errorf("Unexpected span [%s,%s), want [%s,%s)", h.Key, h.EndKey, wantStart, wantEnd)
			}
		case 1:
			wantStart := keys.MakeRangeKeyPrefix(roachpb.RKey("b"))
			wantEnd := keys.MakeRangeKeyPrefix(roachpb.RKey("c"))
			if !(h.Key.Equal(wantStart) && h.EndKey.Equal(wantEnd)) {
				t.Errorf("Unexpected span [%s,%s), want [%s,%s)", h.Key, h.EndKey, wantStart, wantEnd)
			}
		case 2:
			wantStart := keys.MakeRangeKeyPrefix(roachpb.RKey("c"))
			wantEnd := keys.RangeDescriptorKey(roachpb.RKey("c"))
			if !(h.Key.Equal(wantStart) && h.EndKey.Equal(wantEnd)) {
				t.Errorf("Unexpected span [%s,%s), want [%s,%s)", h.Key, h.EndKey, wantStart, wantEnd)
			}
		}
		requests++

		batchReply := &roachpb.BatchResponse{}
		reply := &roachpb.ScanResponse{}
		batchReply.Add(reply)
		return batchReply, nil
	}

	ctx := &DistSenderContext{
		RPCSend:           sendStub,
		RangeDescriptorDB: descDB,
	}
	ds := NewDistSender(ctx, g)

	// Send a batch request contains two scans. In the first
	// attempt, the range of the descriptor found in the cache is
	// ["", "b"). The request is truncated to contain only the scan
	// on local keys that address up to "b".
	//
	// In the second attempt, The range of the descriptor found in
	// the cache is ["b", "d"), The request is truncated to contain
	// only the scan on local keys that address from "b" to "d".
	ba := roachpb.BatchRequest{}
	ba.Txn = &roachpb.Transaction{Name: "test"}
	ba.Add(roachpb.NewScan(keys.RangeDescriptorKey(roachpb.RKey("a")), keys.RangeDescriptorKey(roachpb.RKey("c")), 0))

	if _, pErr := ds.Send(context.Background(), ba); pErr != nil {
		t.Fatal(pErr)
	}
	if want := 3; requests != want {
		t.Errorf("expected request to be split into %d parts, found %d", want, requests)
	}
}
Ejemplo n.º 19
0
// Send implements the batch.Sender interface. It subdivides
// the Batch into batches admissible for sending (preventing certain
// illegal mixtures of requests), executes each individual part
// (which may span multiple ranges), and recombines the response.
// When the request spans ranges, it is split up and the corresponding
// ranges queried serially, in ascending order.
// In particular, the first write in a transaction may not be part of the first
// request sent. This is relevant since the first write is a BeginTransaction
// request, thus opening up a window of time during which there may be intents
// of a transaction, but no entry. Pushing such a transaction will succeed, and
// may lead to the transaction being aborted early.
func (ds *DistSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	tracing.AnnotateTrace()

	// In the event that timestamp isn't set and read consistency isn't
	// required, set the timestamp using the local clock.
	if ba.ReadConsistency == roachpb.INCONSISTENT && ba.Timestamp.Equal(hlc.ZeroTimestamp) {
		ba.Timestamp = ds.clock.Now()
	}

	if ba.Txn != nil {
		// Make a copy here since the code below modifies it in different places.
		// TODO(tschottdorf): be smarter about this - no need to do it for
		// requests that don't get split.
		txnClone := ba.Txn.Clone()
		ba.Txn = &txnClone

		if len(ba.Txn.ObservedTimestamps) == 0 {
			// Ensure the local NodeID is marked as free from clock offset;
			// the transaction's timestamp was taken off the local clock.
			if nDesc := ds.getNodeDescriptor(); nDesc != nil {
				// TODO(tschottdorf): future refactoring should move this to txn
				// creation in TxnCoordSender, which is currently unaware of the
				// NodeID (and wraps *DistSender through client.Sender since it
				// also needs test compatibility with *LocalSender).
				//
				// Taking care below to not modify any memory referenced from
				// our BatchRequest which may be shared with others.
				//
				// We already have a clone of our txn (see above), so we can
				// modify it freely.
				//
				// Zero the existing data. That makes sure that if we had
				// something of size zero but with capacity, we don't re-use the
				// existing space (which others may also use). This is just to
				// satisfy paranoia/OCD and not expected to matter in practice.
				ba.Txn.ResetObservedTimestamps()
				// OrigTimestamp is the HLC timestamp at which the Txn started, so
				// this effectively means no more uncertainty on this node.
				ba.Txn.UpdateObservedTimestamp(nDesc.NodeID, ba.Txn.OrigTimestamp)
			}
		}
	}

	if len(ba.Requests) < 1 {
		panic("empty batch")
	}

	if ba.MaxSpanRequestKeys != 0 {
		// Verify that the batch contains only specific range requests or the
		// Begin/EndTransactionRequest. Verify that a batch with a ReverseScan
		// only contains ReverseScan range requests.
		isReverse := ba.IsReverse()
		for _, req := range ba.Requests {
			inner := req.GetInner()
			switch inner.(type) {
			case *roachpb.ScanRequest, *roachpb.DeleteRangeRequest:
				// Accepted range requests. All other range requests are still
				// not supported.
				// TODO(vivek): don't enumerate all range requests.
				if isReverse {
					return nil, roachpb.NewErrorf("batch with limit contains both forward and reverse scans")
				}

			case *roachpb.BeginTransactionRequest, *roachpb.EndTransactionRequest, *roachpb.ReverseScanRequest:
				continue

			default:
				return nil, roachpb.NewErrorf("batch with limit contains %T request", inner)
			}
		}
	}

	var rplChunks []*roachpb.BatchResponse
	parts := ba.Split(false /* don't split ET */)
	if len(parts) > 1 && ba.MaxSpanRequestKeys != 0 {
		// We already verified above that the batch contains only scan requests of the same type.
		// Such a batch should never need splitting.
		panic("batch with MaxSpanRequestKeys needs splitting")
	}
	for len(parts) > 0 {
		part := parts[0]
		ba.Requests = part
		rpl, pErr, shouldSplitET := ds.sendChunk(ctx, ba)
		if shouldSplitET {
			// If we tried to send a single round-trip EndTransaction but
			// it looks like it's going to hit multiple ranges, split it
			// here and try again.
			if len(parts) != 1 {
				panic("EndTransaction not in last chunk of batch")
			}
			parts = ba.Split(true /* split ET */)
			if len(parts) != 2 {
				panic("split of final EndTransaction chunk resulted in != 2 parts")
			}
			continue
		}
		if pErr != nil {
			return nil, pErr
		}
		// Propagate transaction from last reply to next request. The final
		// update is taken and put into the response's main header.
		ba.UpdateTxn(rpl.Txn)
		rplChunks = append(rplChunks, rpl)
		parts = parts[1:]
	}

	reply := rplChunks[0]
	for _, rpl := range rplChunks[1:] {
		reply.Responses = append(reply.Responses, rpl.Responses...)
		reply.CollectedSpans = append(reply.CollectedSpans, rpl.CollectedSpans...)
	}
	reply.BatchResponse_Header = rplChunks[len(rplChunks)-1].BatchResponse_Header
	return reply, nil
}
Ejemplo n.º 20
0
// Send implements the batch.Sender interface. It subdivides
// the Batch into batches admissible for sending (preventing certain
// illegal mixtures of requests), executes each individual part
// (which may span multiple ranges), and recombines the response.
// When the request spans ranges, it is split up and the corresponding
// ranges queried serially, in ascending order.
// In particular, the first write in a transaction may not be part of the first
// request sent. This is relevant since the first write is a BeginTransaction
// request, thus opening up a window of time during which there may be intents
// of a transaction, but no entry. Pushing such a transaction will succeed, and
// may lead to the transaction being aborted early.
func (ds *DistSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	tracing.AnnotateTrace()

	// In the event that timestamp isn't set and read consistency isn't
	// required, set the timestamp using the local clock.
	if ba.ReadConsistency == roachpb.INCONSISTENT && ba.Timestamp.Equal(roachpb.ZeroTimestamp) {
		ba.Timestamp = ds.clock.Now()
	}

	if ba.Txn != nil && len(ba.Txn.CertainNodes.Nodes) == 0 {
		// Ensure the local NodeID is marked as free from clock offset;
		// the transaction's timestamp was taken off the local clock.
		if nDesc := ds.getNodeDescriptor(); nDesc != nil {
			// TODO(tschottdorf): future refactoring should move this to txn
			// creation in TxnCoordSender, which is currently unaware of the
			// NodeID (and wraps *DistSender through client.Sender since it
			// also needs test compatibility with *LocalSender).
			//
			// Taking care below to not modify any memory referenced from
			// our BatchRequest which may be shared with others.
			// First, get a shallow clone of our txn (since that holds the
			// NodeList struct).
			txnShallow := *ba.Txn
			// Next, zero out the NodeList pointer. That makes sure that
			// if we had something of size zero but with capacity, we don't
			// re-use the existing space (which others may also use).
			txnShallow.CertainNodes.Nodes = nil
			txnShallow.CertainNodes.Add(nDesc.NodeID)
			ba.Txn = &txnShallow
		}
	}

	if len(ba.Requests) < 1 {
		panic("empty batch")
	}

	var rplChunks []*roachpb.BatchResponse
	parts := ba.Split(false /* don't split ET */)
	for len(parts) > 0 {
		part := parts[0]
		ba.Requests = part
		rpl, pErr, shouldSplitET := ds.sendChunk(ctx, ba)
		if shouldSplitET {
			// If we tried to send a single round-trip EndTransaction but
			// it looks like it's going to hit multiple ranges, split it
			// here and try again.
			if len(parts) != 1 {
				panic("EndTransaction not in last chunk of batch")
			}
			parts = ba.Split(true /* split ET */)
			if len(parts) != 2 {
				panic("split of final EndTransaction chunk resulted in != 2 parts")
			}
			continue
		}
		if pErr != nil {
			return nil, pErr
		}
		// Propagate transaction from last reply to next request. The final
		// update is taken and put into the response's main header.
		ba.Txn.Update(rpl.Header().Txn)
		rplChunks = append(rplChunks, rpl)
		parts = parts[1:]
	}

	reply := rplChunks[0]
	for _, rpl := range rplChunks[1:] {
		reply.Responses = append(reply.Responses, rpl.Responses...)
	}
	*reply.Header() = rplChunks[len(rplChunks)-1].BatchResponse_Header
	return reply, nil
}
Ejemplo n.º 21
0
// TestSequenceUpdate verifies txn sequence number is incremented
// on successive commands.
func TestSequenceUpdate(t *testing.T) {
	defer leaktest.AfterTest(t)()
	g, s := makeTestGossip(t)
	defer s()

	if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{NodeID: 1}); err != nil {
		t.Fatal(err)
	}
	nd := &roachpb.NodeDescriptor{
		NodeID:  roachpb.NodeID(1),
		Address: util.MakeUnresolvedAddr(testAddress.Network(), testAddress.String()),
	}
	if err := g.AddInfoProto(gossip.MakeNodeIDKey(roachpb.NodeID(1)), nd, time.Hour); err != nil {
		t.Fatal(err)

	}

	descDB := mockRangeDescriptorDB(func(key roachpb.RKey, _, _ bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
		return []roachpb.RangeDescriptor{
			{
				RangeID:  1,
				StartKey: roachpb.RKeyMin,
				EndKey:   roachpb.RKeyMax,
				Replicas: []roachpb.ReplicaDescriptor{
					{
						NodeID:  1,
						StoreID: 1,
					},
				},
			},
		}, nil
	})

	var expSequence uint32
	var testFn rpcSendFn = func(_ SendOptions, _ ReplicaSlice, ba roachpb.BatchRequest, _ *rpc.Context) (*roachpb.BatchResponse, error) {
		expSequence++
		if expSequence != ba.Txn.Sequence {
			t.Errorf("expected sequence %d; got %d", expSequence, ba.Txn.Sequence)
		}
		br := ba.CreateReply()
		br.Txn = ba.Txn
		return br, nil
	}

	ctx := &DistSenderContext{
		RPCSend:           testFn,
		RangeDescriptorDB: descDB,
	}
	ds := NewDistSender(ctx, g)

	// Send 5 puts and verify sequence number increase.
	txn := &roachpb.Transaction{Name: "test"}
	for i := 0; i < 5; i++ {
		var ba roachpb.BatchRequest
		ba.Txn = txn
		ba.Add(roachpb.NewPut(roachpb.Key("a"), roachpb.MakeValueFromString("foo")).(*roachpb.PutRequest))
		br, pErr := ds.Send(context.Background(), ba)
		if pErr != nil {
			t.Fatal(pErr)
		}
		txn = br.Txn
	}
}
Ejemplo n.º 22
0
// TestTxnCoordSenderHeartbeat verifies periodic heartbeat of the
// transaction record.
func TestTxnCoordSenderHeartbeat(t *testing.T) {
	defer leaktest.AfterTest(t)()
	s, sender := createTestDB(t)
	defer s.Stop()
	defer teardownHeartbeats(sender)

	// Set heartbeat interval to 1ms for testing.
	sender.heartbeatInterval = 1 * time.Millisecond

	initialTxn := client.NewTxn(context.Background(), *s.DB)
	if err := initialTxn.Put(roachpb.Key("a"), []byte("value")); err != nil {
		t.Fatal(err)
	}

	// Verify 3 heartbeats.
	var heartbeatTS roachpb.Timestamp
	for i := 0; i < 3; i++ {
		util.SucceedsSoon(t, func() error {
			txn, pErr := getTxn(sender, &initialTxn.Proto)
			if pErr != nil {
				t.Fatal(pErr)
			}
			// Advance clock by 1ns.
			// Locking the TxnCoordSender to prevent a data race.
			sender.Lock()
			s.Manual.Increment(1)
			sender.Unlock()
			if txn.LastHeartbeat != nil && heartbeatTS.Less(*txn.LastHeartbeat) {
				heartbeatTS = *txn.LastHeartbeat
				return nil
			}
			return util.Errorf("expected heartbeat")
		})
	}

	// Sneakily send an ABORT right to DistSender (bypassing TxnCoordSender).
	{
		var ba roachpb.BatchRequest
		ba.Add(&roachpb.EndTransactionRequest{
			Commit: false,
			Span:   roachpb.Span{Key: initialTxn.Proto.Key},
		})
		ba.Txn = &initialTxn.Proto
		if _, pErr := sender.wrapped.Send(context.Background(), ba); pErr != nil {
			t.Fatal(pErr)
		}
	}

	util.SucceedsSoon(t, func() error {
		sender.Lock()
		defer sender.Unlock()
		if txnMeta, ok := sender.txns[*initialTxn.Proto.ID]; !ok {
			t.Fatal("transaction unregistered prematurely")
		} else if txnMeta.txn.Status != roachpb.ABORTED {
			return fmt.Errorf("transaction is not aborted")
		}
		return nil
	})

	// Trying to do something else should give us a TransactionAbortedError.
	_, err := initialTxn.Get("a")
	assertTransactionAbortedError(t, err)
}
Ejemplo n.º 23
0
// TestSequenceUpdateOnMultiRangeQueryLoop reproduces #3206 and
// verifies that the sequence is updated in the DistSender
// multi-range-query loop.
//
// More specifically, the issue was that DistSender might send
// multiple batch requests to the same replica when it finds a
// post-split range descriptor in the cache while the split has not
// yet been fully completed. By giving a higher sequence to the second
// request, we can avoid an infinite txn restart error (otherwise
// caused by hitting the sequence cache).
func TestSequenceUpdateOnMultiRangeQueryLoop(t *testing.T) {
	defer leaktest.AfterTest(t)
	g, s := makeTestGossip(t)
	defer s()

	if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{NodeID: 1}); err != nil {
		t.Fatal(err)
	}
	nd := &roachpb.NodeDescriptor{
		NodeID:  roachpb.NodeID(1),
		Address: util.MakeUnresolvedAddr(testAddress.Network(), testAddress.String()),
	}
	if err := g.AddInfoProto(gossip.MakeNodeIDKey(roachpb.NodeID(1)), nd, time.Hour); err != nil {
		t.Fatal(err)

	}

	// Fill mockRangeDescriptorDB with two descriptors.
	var descriptor1 = roachpb.RangeDescriptor{
		RangeID:  1,
		StartKey: roachpb.RKeyMin,
		EndKey:   roachpb.RKey("b"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	var descriptor2 = roachpb.RangeDescriptor{
		RangeID:  2,
		StartKey: roachpb.RKey("b"),
		EndKey:   roachpb.RKey("c"),
		Replicas: []roachpb.ReplicaDescriptor{
			{
				NodeID:  1,
				StoreID: 1,
			},
		},
	}
	descDB := mockRangeDescriptorDB(func(key roachpb.RKey, _, _ bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
		desc := descriptor1
		if key.Equal(roachpb.RKey("b")) {
			desc = descriptor2
		}
		return []roachpb.RangeDescriptor{desc}, nil
	})

	// Define our rpcSend stub which checks the span of the batch
	// requests. The first request should be the point request on
	// "a". The second request should be on "b". The sequence of the
	// second request will be incremented by one from that of the
	// first request.
	first := true
	var firstSequence uint32
	var testFn rpcSendFn = func(_ rpc.Options, method string, addrs []net.Addr, getArgs func(addr net.Addr) proto.Message, getReply func() proto.Message, _ *rpc.Context) ([]proto.Message, error) {
		if method != "Node.Batch" {
			return nil, util.Errorf("unexpected method %v", method)
		}

		ba := getArgs(testAddress).(*roachpb.BatchRequest)
		rs := keys.Range(*ba)
		if first {
			if !(rs.Key.Equal(roachpb.RKey("a")) && rs.EndKey.Equal(roachpb.RKey("a").Next())) {
				t.Errorf("unexpected span [%s,%s)", rs.Key, rs.EndKey)
			}
			first = false
			firstSequence = ba.Txn.Sequence
		} else {
			if !(rs.Key.Equal(roachpb.RKey("b")) && rs.EndKey.Equal(roachpb.RKey("b").Next())) {
				t.Errorf("unexpected span [%s,%s)", rs.Key, rs.EndKey)
			}
			if ba.Txn.Sequence != firstSequence+1 {
				t.Errorf("unexpected sequence; expected %d, but got %d", firstSequence+1, ba.Txn.Sequence)
			}
		}
		return []proto.Message{ba.CreateReply()}, nil
	}

	ctx := &DistSenderContext{
		RPCSend:           testFn,
		RangeDescriptorDB: descDB,
	}
	ds := NewDistSender(ctx, g)

	// Send a batch request containing two puts.
	var ba roachpb.BatchRequest
	ba.Txn = &roachpb.Transaction{Name: "test"}
	val := roachpb.MakeValueFromString("val")
	ba.Add(roachpb.NewPut(roachpb.Key("a"), val).(*roachpb.PutRequest))
	ba.Add(roachpb.NewPut(roachpb.Key("b"), val).(*roachpb.PutRequest))

	_, pErr := ds.Send(context.Background(), ba)
	if err := pErr.GoError(); err != nil {
		t.Fatal(err)
	}
}
Ejemplo n.º 24
0
func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	txn := txnMeta.txn.Clone()
	hasAbandoned := txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow())
	tc.Unlock()

	if txn.Status != roachpb.PENDING {
		// A previous iteration has already determined that the transaction is
		// already finalized, so we wait for the client to realize that and
		// want to keep our state for the time being (to dish out the right
		// error once it returns).
		return true
	}

	// Before we send a heartbeat, determine whether this transaction should be
	// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
	// it is a cancellable Context and we skip this check and use the ctx lifetime
	// instead of a timeout.
	if ctx.Done() == nil && hasAbandoned {
		if log.V(1) {
			log.Infof(ctx, "transaction %s abandoned; stopping heartbeat", txnMeta.txn)
		}
		tc.tryAsyncAbort(txnID)
		return false
	}

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	hb := &roachpb.HeartbeatTxnRequest{
		Now: tc.clock.Now(),
	}
	hb.Key = txn.Key
	ba.Add(hb)

	log.Trace(ctx, "heartbeat")
	br, pErr := tc.wrapped.Send(ctx, ba)

	// Correctness mandates that when we can't heartbeat the transaction, we
	// make sure the client doesn't keep going. This is particularly relevant
	// in the case of an ABORTED transaction, but if we can't reach the
	// transaction record at all, we're going to have to assume we're aborted
	// as well.
	if pErr != nil {
		log.Warningf(ctx, "heartbeat to %s failed: %s", txn, pErr)
		// We're not going to let the client carry out additional requests, so
		// try to clean up.
		tc.tryAsyncAbort(*txn.ID)
		txn.Status = roachpb.ABORTED
	} else {
		txn.Update(br.Responses[0].GetInner().(*roachpb.HeartbeatTxnResponse).Txn)
	}

	// Give the news to the txn in the txns map. This will update long-running
	// transactions (which may find out that they have to restart in that way),
	// but in particular makes sure that they notice when they've been aborted
	// (in which case we'll give them an error on their next request).
	tc.Lock()
	tc.txns[txnID].txn.Update(&txn)
	tc.Unlock()

	return true
}
Ejemplo n.º 25
0
func (tc *TxnCoordSender) heartbeat(txnID uuid.UUID, trace opentracing.Span, ctx context.Context) bool {
	tc.Lock()
	proceed := true
	txnMeta := tc.txns[txnID]
	var intentSpans []roachpb.Span
	// Before we send a heartbeat, determine whether this transaction
	// should be considered abandoned. If so, exit heartbeat.
	if txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow()) {
		// TODO(tschottdorf): should we be more proactive here?
		// The client might be continuing the transaction
		// through another coordinator, but in the most likely
		// case it's just gone and the open transaction record
		// could block concurrent operations.
		if log.V(1) {
			log.Infof("transaction %s abandoned; stopping heartbeat",
				txnMeta.txn)
		}
		proceed = false
		// Grab the intents here to avoid potential race.
		intentSpans = collectIntentSpans(txnMeta.keys)
		txnMeta.keys.Clear()
	}
	// txnMeta.txn is possibly replaced concurrently,
	// so grab a copy before unlocking.
	txn := txnMeta.txn.Clone()
	tc.Unlock()

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	if !proceed {
		// Actively abort the transaction and its intents since we assume it's abandoned.
		et := &roachpb.EndTransactionRequest{
			Span: roachpb.Span{
				Key: txn.Key,
			},
			Commit:      false,
			IntentSpans: intentSpans,
		}
		ba.Add(et)
		tc.stopper.RunAsyncTask(func() {
			// Use the wrapped sender since the normal Sender
			// does not allow clients to specify intents.
			// TODO(tschottdorf): not using the existing context here since that
			// leads to use-after-finish of the contained trace. Should fork off
			// before the goroutine.
			if _, pErr := tc.wrapped.Send(context.Background(), ba); pErr != nil {
				if log.V(1) {
					log.Warningf("abort due to inactivity failed for %s: %s ", txn, pErr)
				}
			}
		})
		return false
	}

	hb := &roachpb.HeartbeatTxnRequest{
		Now: tc.clock.Now(),
	}
	hb.Key = txn.Key
	ba.Add(hb)

	trace.LogEvent("heartbeat")
	_, err := tc.wrapped.Send(ctx, ba)
	// If the transaction is not in pending state, then we can stop
	// the heartbeat. It's either aborted or committed, and we resolve
	// write intents accordingly.
	if err != nil {
		log.Warningf("heartbeat to %s failed: %s", txn, err)
	}
	// TODO(bdarnell): once we have gotten a heartbeat response with
	// Status != PENDING, future heartbeats are useless. However, we
	// need to continue the heartbeatLoop until the client either
	// commits or abandons the transaction. We could save a little
	// pointless work by restructuring this loop to stop sending
	// heartbeats between the time that the transaction is aborted and
	// the client finds out. Furthermore, we could use this information
	// to send TransactionAbortedErrors to the client so it can restart
	// immediately instead of running until its EndTransaction.
	return true
}