// Batch implements the roachpb.KVServer interface.
func (s *DBServer) Batch(
ctx context.Context, args *roachpb.BatchRequest,
) (br *roachpb.BatchResponse, err error) {
// TODO(marc,bdarnell): this code is duplicated in server/node.go,
// which should be fixed.
defer func() {
// We always return errors via BatchResponse.Error so structure is
// preserved; plain errors are presumed to be from the RPC
// framework and not from cockroach.
if err != nil {
if br == nil {
br = &roachpb.BatchResponse{}
}
if br.Error != nil {
panic(fmt.Sprintf(
"attempting to return both a plain error (%s) and roachpb.Error (%s)", err, br.Error))
}
br.Error = roachpb.NewError(err)
err = nil
}
}()
// TODO(marc): grpc's authentication model (which gives credential access in
// the request handler) doesn't really fit with the current design of the
// security package (which assumes that TLS state is only given at connection
// time) - that should be fixed.
if peer, ok := peer.FromContext(ctx); ok {
if tlsInfo, ok := peer.AuthInfo.(credentials.TLSInfo); ok {
certUser, err := security.GetCertificateUser(&tlsInfo.State)
if err != nil {
return nil, err
}
if certUser != security.NodeUser {
return nil, errors.Errorf("user %s is not allowed", certUser)
}
}
}
if err = verifyRequest(args); err != nil {
return br, err
}
err = s.stopper.RunTask(func() {
var pErr *roachpb.Error
// TODO(wiz): This is required to be a different context from the one
// provided by grpc since it has to last for the entire transaction and not
// just this one RPC call. See comment for (*TxnCoordSender).hearbeatLoop.
br, pErr = s.sender.Send(context.TODO(), *args)
if pErr != nil {
br = &roachpb.BatchResponse{}
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(s.sender, br))
}
br.Error = pErr
})
return br, err
}
// 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) {
// 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 {
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err)
}
rangeID, repDesc, err := ls.LookupReplica(rs.Key, rs.EndKey)
if err != nil {
return nil, roachpb.NewError(err)
}
ba.RangeID = rangeID
ba.Replica = repDesc
}
store, err := ls.GetStore(ba.Replica.StoreID)
if err != nil {
return nil, roachpb.NewError(err)
}
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) {
log.Event(ctx, "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
}
func (ss *notifyingSender) Send(
ctx context.Context, ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, *roachpb.Error) {
br, pErr := ss.wrapped.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(ss.wrapped, br))
}
select {
case ss.notify <- struct{}{}:
default:
}
return br, pErr
}
func (s *senderTransport) SendNext(done chan<- BatchCall) {
if s.called {
panic("called an exhausted transport")
}
s.called = true
sp := s.tracer.StartSpan("node")
defer sp.Finish()
ctx := opentracing.ContextWithSpan(context.TODO(), sp)
log.Event(ctx, s.args.String())
br, pErr := s.sender.Send(ctx, s.args)
if br == nil {
br = &roachpb.BatchResponse{}
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(s.sender, br))
}
br.Error = pErr
if pErr != nil {
log.Event(ctx, "error: "+pErr.String())
}
done <- BatchCall{Reply: br}
}
// sendInternal sends the batch and updates the transaction on error. Depending
// on the error type, the transaction might be replaced by a new one.
func (txn *Txn) sendInternal(ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(ba.Requests) == 0 {
return nil, nil
}
if pErr := txn.db.prepareToSend(&ba); pErr != nil {
return nil, pErr
}
// Send call through the DB's sender.
ba.Txn = &txn.Proto
// For testing purposes, txn.UserPriority can be a negative value (see
// MakePriority).
if txn.UserPriority != 0 {
ba.UserPriority = txn.UserPriority
}
// TODO(radu): when db.send supports a context, we can just use that (and
// remove the prepareToSend call above).
br, pErr := txn.db.sender.Send(txn.Context, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(txn.db.sender, 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)
}
txn.CollectedSpans = append(txn.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 {
txn.Proto.Update(br.Txn)
return br, nil
}
if log.V(1) {
log.Infof(txn.Context, "failed batch: %s", pErr)
}
if _, ok := pErr.GetDetail().(*roachpb.TransactionAbortedError); ok {
// On Abort, reset the transaction so we start anew on restart.
txn.Proto = roachpb.Transaction{
TxnMeta: enginepb.TxnMeta{
Isolation: txn.Proto.Isolation,
},
Name: txn.Proto.Name,
}
// Acts as a minimum priority on restart.
if pErr.GetTxn() != nil {
txn.Proto.Priority = pErr.GetTxn().Priority
}
} else if pErr.TransactionRestart != roachpb.TransactionRestart_NONE {
txn.Proto.Update(pErr.GetTxn())
}
return nil, pErr
}
func (n *Node) batchInternal(
ctx context.Context, args *roachpb.BatchRequest,
) (*roachpb.BatchResponse, error) {
// TODO(marc): grpc's authentication model (which gives credential access in
// the request handler) doesn't really fit with the current design of the
// security package (which assumes that TLS state is only given at connection
// time) - that should be fixed.
if peer, ok := peer.FromContext(ctx); ok {
if tlsInfo, ok := peer.AuthInfo.(credentials.TLSInfo); ok {
certUser, err := security.GetCertificateUser(&tlsInfo.State)
if err != nil {
return nil, err
}
if certUser != security.NodeUser {
return nil, errors.Errorf("user %s is not allowed", certUser)
}
}
}
var br *roachpb.BatchResponse
type snowballInfo struct {
syncutil.Mutex
collectedSpans [][]byte
done bool
}
var snowball *snowballInfo
if err := n.stopper.RunTaskWithErr(func() error {
const opName = "node.Batch"
sp, err := tracing.JoinOrNew(n.storeCfg.AmbientCtx.Tracer, args.TraceContext, opName)
if err != nil {
return err
}
// If this is a snowball span, it gets special treatment: It skips the
// regular tracing machinery, and we instead send the collected spans
// back with the response. This is more expensive, but then again,
// those are individual requests traced by users, so they can be.
if sp.BaggageItem(tracing.Snowball) != "" {
sp.LogEvent("delegating to snowball tracing")
sp.Finish()
snowball = new(snowballInfo)
recorder := func(rawSpan basictracer.RawSpan) {
snowball.Lock()
defer snowball.Unlock()
if snowball.done {
// This is a late span that we must discard because the request was
// already completed.
return
}
encSp, err := tracing.EncodeRawSpan(&rawSpan, nil)
if err != nil {
log.Warning(ctx, err)
}
snowball.collectedSpans = append(snowball.collectedSpans, encSp)
}
if sp, err = tracing.JoinOrNewSnowball(opName, args.TraceContext, recorder); err != nil {
return err
}
}
defer sp.Finish()
traceCtx := opentracing.ContextWithSpan(ctx, sp)
log.Event(traceCtx, args.Summary())
tStart := timeutil.Now()
var pErr *roachpb.Error
br, pErr = n.stores.Send(traceCtx, *args)
if pErr != nil {
br = &roachpb.BatchResponse{}
log.ErrEventf(traceCtx, "%T", pErr.GetDetail())
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(n.stores, br))
}
n.metrics.callComplete(timeutil.Since(tStart), pErr)
br.Error = pErr
return nil
}); err != nil {
return nil, err
}
if snowball != nil {
snowball.Lock()
br.CollectedSpans = snowball.collectedSpans
snowball.done = true
snowball.Unlock()
}
return br, nil
}