// sendBatch unrolls a batched command and sends each constituent
// command in parallel.
// TODO(tschottdorf): modify sendBatch so that it sends truly parallel requests
// when outside of a Transaction. This can then be used to address the TODO in
// (*TxnCoordSender).resolve().
func (tc *TxnCoordSender) sendBatch(ctx context.Context, batchArgs *proto.BatchRequest, batchReply *proto.BatchResponse) {
// Prepare the calls by unrolling the batch. If the batchReply is
// pre-initialized with replies, use those; otherwise create replies
// as needed.
// TODO(spencer): send calls in parallel.
batchReply.Txn = batchArgs.Txn
for i := range batchArgs.Requests {
args := batchArgs.Requests[i].GetValue().(proto.Request)
if err := updateForBatch(args, batchArgs.RequestHeader); err != nil {
batchReply.Header().SetGoError(err)
return
}
call := proto.Call{Args: args}
// Create a reply from the method type and add to batch response.
if i >= len(batchReply.Responses) {
call.Reply = args.CreateReply()
batchReply.Add(call.Reply)
} else {
call.Reply = batchReply.Responses[i].GetValue().(proto.Response)
}
tc.sendOne(ctx, call)
// Amalgamate transaction updates and propagate first error, if applicable.
if batchReply.Txn != nil {
batchReply.Txn.Update(call.Reply.Header().Txn)
}
if call.Reply.Header().Error != nil {
batchReply.Error = call.Reply.Header().Error
return
}
}
}
func newTestSender(pre, post func(proto.BatchRequest) (*proto.BatchResponse, *proto.Error)) SenderFunc {
txnKey := proto.Key("test-txn")
txnID := []byte(uuid.NewUUID4())
return func(_ context.Context, ba proto.BatchRequest) (*proto.BatchResponse, *proto.Error) {
ba.UserPriority = gogoproto.Int32(-1)
if ba.Txn != nil && len(ba.Txn.ID) == 0 {
ba.Txn.Key = txnKey
ba.Txn.ID = txnID
}
var br *proto.BatchResponse
var pErr *proto.Error
if pre != nil {
br, pErr = pre(ba)
} else {
br = &proto.BatchResponse{}
}
if pErr != nil {
return nil, pErr
}
var writing bool
status := proto.PENDING
if _, ok := ba.GetArg(proto.Put); ok {
br.Add(gogoproto.Clone(testPutResp).(proto.Response))
writing = true
}
if args, ok := ba.GetArg(proto.EndTransaction); ok {
et := args.(*proto.EndTransactionRequest)
writing = true
if et.Commit {
status = proto.COMMITTED
} else {
status = proto.ABORTED
}
}
br.Txn = gogoproto.Clone(ba.Txn).(*proto.Transaction)
if br.Txn != nil && pErr == nil {
br.Txn.Writing = writing
br.Txn.Status = status
}
if post != nil {
br, pErr = post(ba)
}
return br, pErr
}
}
// sendBatch unrolls a batched command and sends each constituent
// command in parallel.
func (tc *TxnCoordSender) sendBatch(batchArgs *proto.BatchRequest, batchReply *proto.BatchResponse) {
// Prepare the calls by unrolling the batch. If the batchReply is
// pre-initialized with replies, use those; otherwise create replies
// as needed.
// TODO(spencer): send calls in parallel.
batchReply.Txn = batchArgs.Txn
for i := range batchArgs.Requests {
// Initialize args header values where appropriate.
args := batchArgs.Requests[i].GetValue().(proto.Request)
method, err := proto.MethodForRequest(args)
call := &client.Call{Method: method, Args: args}
if err != nil {
batchReply.SetGoError(err)
return
}
if args.Header().User == "" {
args.Header().User = batchArgs.User
}
if args.Header().UserPriority == nil {
args.Header().UserPriority = batchArgs.UserPriority
}
args.Header().Txn = batchArgs.Txn
// Create a reply from the method type and add to batch response.
if i >= len(batchReply.Responses) {
if call.Reply, err = proto.CreateReply(method); err != nil {
batchReply.SetGoError(util.Errorf("unsupported method in batch: %s", method))
return
}
batchReply.Add(call.Reply)
} else {
call.Reply = batchReply.Responses[i].GetValue().(proto.Response)
}
tc.sendOne(call)
// Amalgamate transaction updates and propagate first error, if applicable.
if batchReply.Txn != nil {
batchReply.Txn.Update(call.Reply.Header().Txn)
}
if call.Reply.Header().Error != nil {
batchReply.Error = call.Reply.Header().Error
return
}
}
}
// TODO(tschottdorf): this method is somewhat awkward but unless we want to
// give this error back to the client, our options are limited. We'll have to
// run the whole thing for them, or any restart will still end up at the client
// which will not be prepared to be handed a Txn.
func (tc *TxnCoordSender) resendWithTxn(ba proto.BatchRequest) (*proto.BatchResponse, *proto.Error) {
// Run a one-off transaction with that single command.
if log.V(1) {
log.Infof("%s: auto-wrapping in txn and re-executing: ", ba)
}
tmpDB := client.NewDBWithPriority(tc, ba.GetUserPriority())
var br *proto.BatchResponse
err := tmpDB.Txn(func(txn *client.Txn) error {
txn.SetDebugName("auto-wrap", 0)
b := &client.Batch{}
for _, arg := range ba.Requests {
req := arg.GetInner()
b.InternalAddRequest(req)
}
var err error
br, err = txn.CommitInBatchWithResponse(b)
return err
})
if err != nil {
return nil, proto.NewError(err)
}
br.Txn = nil // hide the evidence
return br, nil
}
// updateState updates the transaction state in both the success and
// error cases, applying those updates to the corresponding txnMeta
// object when adequate. It also updates certain errors with the
// updated transaction for use by client restarts.
func (tc *TxnCoordSender) updateState(ctx context.Context, ba proto.BatchRequest, br *proto.BatchResponse, pErr *proto.Error) *proto.Error {
trace := tracer.FromCtx(ctx)
newTxn := &proto.Transaction{}
newTxn.Update(ba.GetTxn())
err := pErr.GoError()
switch t := err.(type) {
case nil:
newTxn.Update(br.GetTxn())
// Move txn timestamp forward to response timestamp if applicable.
// TODO(tschottdorf): see (*Replica).executeBatch and comments within.
// Looks like this isn't necessary any more, nor did it prevent a bug
// referenced in a TODO there.
newTxn.Timestamp.Forward(br.Timestamp)
case *proto.TransactionStatusError:
// Likely already committed or more obscure errors such as epoch or
// timestamp regressions; consider txn dead.
defer tc.cleanupTxn(trace, t.Txn)
case *proto.OpRequiresTxnError:
// TODO(tschottdorf): range-spanning autowrap currently broken.
panic("TODO(tschottdorf): disabled")
case *proto.ReadWithinUncertaintyIntervalError:
// Mark the host as certain. See the protobuf comment for
// Transaction.CertainNodes for details.
if t.NodeID == 0 {
panic("no replica set in header on uncertainty restart")
}
newTxn.CertainNodes.Add(t.NodeID)
// If the reader encountered a newer write within the uncertainty
// interval, move the timestamp forward, just past that write or
// up to MaxTimestamp, whichever comes first.
candidateTS := newTxn.MaxTimestamp
candidateTS.Backward(t.ExistingTimestamp.Add(0, 1))
newTxn.Timestamp.Forward(candidateTS)
newTxn.Restart(ba.GetUserPriority(), newTxn.Priority, newTxn.Timestamp)
t.Txn = *newTxn
case *proto.TransactionAbortedError:
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(t.Txn.Timestamp)
newTxn.Priority = t.Txn.Priority
t.Txn = *newTxn
// Clean up the freshly aborted transaction in defer(), avoiding a
// race with the state update below.
defer tc.cleanupTxn(trace, t.Txn)
case *proto.TransactionPushError:
// Increase timestamp if applicable, ensuring that we're
// just ahead of the pushee.
newTxn.Timestamp.Forward(t.PusheeTxn.Timestamp.Add(0, 1))
newTxn.Restart(ba.GetUserPriority(), t.PusheeTxn.Priority-1, newTxn.Timestamp)
t.Txn = newTxn
case *proto.TransactionRetryError:
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(t.Txn.Timestamp)
newTxn.Restart(ba.GetUserPriority(), t.Txn.Priority, newTxn.Timestamp)
t.Txn = *newTxn
case proto.TransactionRestartError:
// Assertion: The above cases should exhaust all ErrorDetails which
// carry a Transaction.
if pErr.Detail != nil {
panic(fmt.Sprintf("unhandled TransactionRestartError %T", err))
}
}
return func() *proto.Error {
if len(newTxn.ID) <= 0 {
return pErr
}
id := string(newTxn.ID)
tc.Lock()
defer tc.Unlock()
txnMeta := tc.txns[id]
// For successful transactional requests, keep the written intents and
// the updated transaction record to be sent along with the reply.
// The transaction metadata is created with the first writing operation
// TODO(tschottdorf): already computed the intents prior to sending,
// consider re-using those.
if intents := ba.GetIntents(); len(intents) > 0 && err == nil {
if txnMeta == nil {
newTxn.Writing = true
txnMeta = &txnMetadata{
txn: *newTxn,
keys: cache.NewIntervalCache(cache.Config{Policy: cache.CacheNone}),
firstUpdateNanos: tc.clock.PhysicalNow(),
lastUpdateNanos: tc.clock.PhysicalNow(),
timeoutDuration: tc.clientTimeout,
txnEnd: make(chan struct{}),
}
tc.txns[id] = txnMeta
// If the transaction is already over, there's no point in
// launching a one-off coordinator which will shut down right
// away.
if _, isEnding := ba.GetArg(proto.EndTransaction); !isEnding {
trace.Event("coordinator spawns")
if !tc.stopper.RunAsyncTask(func() {
tc.heartbeatLoop(id)
}) {
// The system is already draining and we can't start the
// heartbeat. We refuse new transactions for now because
// they're likely not going to have all intents committed.
// In principle, we can relax this as needed though.
tc.unregisterTxnLocked(id)
return proto.NewError(&proto.NodeUnavailableError{})
}
}
}
for _, intent := range intents {
txnMeta.addKeyRange(intent.Key, intent.EndKey)
}
}
// Update our record of this transaction, even on error.
if txnMeta != nil {
txnMeta.txn.Update(newTxn) // better to replace after #2300
if !txnMeta.txn.Writing {
panic("tracking a non-writing txn")
}
txnMeta.setLastUpdate(tc.clock.PhysicalNow())
}
if err == nil {
// For successful transactional requests, always send the updated txn
// record back.
if br.Txn == nil {
br.Txn = &proto.Transaction{}
}
*br.Txn = *newTxn
}
return pErr
}()
}