// 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 roachpb.BatchRequest,
) (*roachpb.BatchResponse, *roachpb.Error) {
ctx := tc.AnnotateCtx(context.TODO())
// Run a one-off transaction with that single command.
if log.V(1) {
log.Infof(ctx, "%s: auto-wrapping in txn and re-executing: ", ba)
}
// TODO(bdarnell): need to be able to pass other parts of DBContext
// through here.
dbCtx := client.DefaultDBContext()
dbCtx.UserPriority = ba.UserPriority
tmpDB := client.NewDBWithContext(tc, dbCtx)
var br *roachpb.BatchResponse
err := tmpDB.Txn(ctx, func(txn *client.Txn) error {
txn.SetDebugName("auto-wrap", 0)
b := txn.NewBatch()
b.Header = ba.Header
for _, arg := range ba.Requests {
req := arg.GetInner()
b.AddRawRequest(req)
}
err := txn.CommitInBatch(b)
br = b.RawResponse()
return err
})
if err != nil {
return nil, roachpb.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,
startNS int64,
ba roachpb.BatchRequest,
br *roachpb.BatchResponse,
pErr *roachpb.Error,
) *roachpb.Error {
tc.Lock()
defer tc.Unlock()
if ba.Txn == nil {
// Not a transactional request.
return pErr
}
var newTxn roachpb.Transaction
newTxn.Update(ba.Txn)
if pErr == nil {
newTxn.Update(br.Txn)
} else if errTxn := pErr.GetTxn(); errTxn != nil {
newTxn.Update(errTxn)
}
switch t := pErr.GetDetail().(type) {
case *roachpb.OpRequiresTxnError:
panic("OpRequiresTxnError must not happen at this level")
case *roachpb.ReadWithinUncertaintyIntervalError:
// If the reader encountered a newer write within the uncertainty
// interval, we advance the txn's timestamp just past the last observed
// timestamp from the node.
restartTS, ok := newTxn.GetObservedTimestamp(pErr.OriginNode)
if !ok {
pErr = roachpb.NewError(errors.Errorf("no observed timestamp for node %d found on uncertainty restart", pErr.OriginNode))
} else {
newTxn.Timestamp.Forward(restartTS)
newTxn.Restart(ba.UserPriority, newTxn.Priority, newTxn.Timestamp)
}
case *roachpb.TransactionAbortedError:
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(pErr.GetTxn().Timestamp)
newTxn.Priority = pErr.GetTxn().Priority
// Clean up the freshly aborted transaction in defer(), avoiding a
// race with the state update below.
defer tc.cleanupTxnLocked(ctx, newTxn)
case *roachpb.TransactionPushError:
// Increase timestamp if applicable, ensuring that we're
// just ahead of the pushee.
newTxn.Timestamp.Forward(t.PusheeTxn.Timestamp)
newTxn.Restart(ba.UserPriority, t.PusheeTxn.Priority-1, newTxn.Timestamp)
case *roachpb.TransactionRetryError:
// Increase timestamp so on restart, we're ahead of any timestamp
// cache entries or newer versions which caused the restart.
newTxn.Restart(ba.UserPriority, pErr.GetTxn().Priority, newTxn.Timestamp)
case *roachpb.WriteTooOldError:
newTxn.Restart(ba.UserPriority, newTxn.Priority, t.ActualTimestamp)
case nil:
// Nothing to do here, avoid the default case.
default:
// Do not clean up the transaction since we're leaving cancellation of
// the transaction up to the client. For example, on seeing an error,
// like TransactionStatusError or ConditionFailedError, the client
// will call Txn.CleanupOnError() which will cleanup the transaction
// and its intents. Therefore leave the transaction in the PENDING
// state and do not call cleanTxnLocked().
}
txnID := *newTxn.ID
txnMeta := tc.txns[txnID]
// 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.
// A tricky edge case is that of a transaction which "fails" on the
// first writing request, but actually manages to write some intents
// (for example, due to being multi-range). In this case, there will
// be an error, but the transaction will be marked as Writing and the
// coordinator must track the state, for the client's retry will be
// performed with a Writing transaction which the coordinator rejects
// unless it is tracking it (on top of it making sense to track it;
// after all, it **has** laid down intents and only the coordinator
// can augment a potential EndTransaction call). See #3303.
if txnMeta != nil || pErr == nil || newTxn.Writing {
// Adding the intents even on error reduces the likelihood of dangling
// intents blocking concurrent writers for extended periods of time.
// See #3346.
var keys []roachpb.Span
if txnMeta != nil {
keys = txnMeta.keys
}
ba.IntentSpanIterate(br, func(key, endKey roachpb.Key) {
keys = append(keys, roachpb.Span{
Key: key,
EndKey: endKey,
})
})
if txnMeta != nil {
txnMeta.keys = keys
} else if len(keys) > 0 {
if !newTxn.Writing {
panic("txn with intents marked as non-writing")
}
// If the transaction is already over, there's no point in
// launching a one-off coordinator which will shut down right
// away. If we ended up here with an error, we'll always start
// the coordinator - the transaction has laid down intents, so
// we expect it to be committed/aborted at some point in the
// future.
if _, isEnding := ba.GetArg(roachpb.EndTransaction); pErr != nil || !isEnding {
log.Event(ctx, "coordinator spawns")
txnMeta = &txnMetadata{
txn: newTxn,
keys: keys,
firstUpdateNanos: startNS,
lastUpdateNanos: tc.clock.PhysicalNow(),
timeoutDuration: tc.clientTimeout,
txnEnd: make(chan struct{}),
}
tc.txns[txnID] = txnMeta
if err := tc.stopper.RunAsyncTask(ctx, func(ctx context.Context) {
tc.heartbeatLoop(ctx, txnID)
}); err != nil {
// 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(txnID)
return roachpb.NewError(err)
}
} else {
// If this was a successful one phase commit, update stats
// directly as they won't otherwise be updated on heartbeat
// loop shutdown.
etArgs, ok := br.Responses[len(br.Responses)-1].GetInner().(*roachpb.EndTransactionResponse)
tc.updateStats(tc.clock.PhysicalNow()-startNS, 0, newTxn.Status, ok && etArgs.OnePhaseCommit)
}
}
}
// Update our record of this transaction, even on error.
if txnMeta != nil {
txnMeta.txn.Update(&newTxn)
if !txnMeta.txn.Writing {
panic("tracking a non-writing txn")
}
txnMeta.setLastUpdate(tc.clock.PhysicalNow())
}
if pErr == nil {
// For successful transactional requests, always send the updated txn
// record back. Note that we make sure not to share data with newTxn
// (which may have made it into txnMeta).
if br.Txn != nil {
br.Txn.Update(&newTxn)
} else {
clonedTxn := newTxn.Clone()
br.Txn = &clonedTxn
}
} else if pErr.GetTxn() != nil {
// Avoid changing existing errors because sometimes they escape into
// goroutines and data races can occur.
pErrShallow := *pErr
pErrShallow.SetTxn(&newTxn) // SetTxn clones newTxn
pErr = &pErrShallow
}
return pErr
}
// TestSender mocks out some of the txn coordinator sender's
// functionality. It responds to PutRequests using testPutResp.
func newTestSender(
pre, post func(roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error),
) SenderFunc {
txnID := uuid.MakeV4()
return func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if ba.UserPriority == 0 {
ba.UserPriority = 1
}
if ba.Txn != nil && ba.Txn.ID == nil {
ba.Txn.Key = txnKey
ba.Txn.ID = &txnID
}
var br *roachpb.BatchResponse
var pErr *roachpb.Error
if pre != nil {
br, pErr = pre(ba)
} else {
br = ba.CreateReply()
}
if pErr != nil {
return nil, pErr
}
var writing bool
status := roachpb.PENDING
for i, req := range ba.Requests {
args := req.GetInner()
if _, ok := args.(*roachpb.PutRequest); ok {
testPutRespCopy := testPutResp
union := &br.Responses[i] // avoid operating on copy
union.MustSetInner(&testPutRespCopy)
}
if roachpb.IsTransactionWrite(args) {
writing = true
}
}
if args, ok := ba.GetArg(roachpb.EndTransaction); ok {
et := args.(*roachpb.EndTransactionRequest)
writing = true
if et.Commit {
status = roachpb.COMMITTED
} else {
status = roachpb.ABORTED
}
}
if ba.Txn != nil {
txnClone := ba.Txn.Clone()
br.Txn = &txnClone
if pErr == nil {
br.Txn.Writing = writing
br.Txn.Status = status
}
}
if post != nil {
br, pErr = post(ba)
}
return br, pErr
}
}