// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold. The transaction and abort cache records are also
// scanned and old entries evicted. During normal operation, both of these
// records are cleaned up when their respective transaction finishes, so the
// amount of work done here is expected to be small.
//
// Some care needs to be taken to avoid cyclic recreation of entries during GC:
// * a Push initiated due to an intent may recreate a transaction entry
// * resolving an intent may write a new abort cache entry
// * obtaining the transaction for a abort cache entry requires a Push
//
// The following order is taken below:
// 1) collect all intents with sufficiently old txn record
// 2) collect these intents' transactions
// 3) scan the transaction table, collecting abandoned or completed txns
// 4) push all of these transactions (possibly recreating entries)
// 5) resolve all intents (unless the txn is still PENDING), which will recreate
// abort cache entries (but with the txn timestamp; i.e. likely gc'able)
// 6) scan the abort cache table for old entries
// 7) push these transactions (again, recreating txn entries).
// 8) send a GCRequest.
func (gcq *gcQueue) process(
ctx context.Context, now hlc.Timestamp, repl *Replica, sysCfg config.SystemConfig,
) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return errors.Errorf("could not find zone config for range %s: %s", repl, err)
}
gcKeys, info, err := RunGC(ctx, desc, snap, now, zone.GC,
func(now hlc.Timestamp, txn *roachpb.Transaction, typ roachpb.PushTxnType) {
pushTxn(ctx, gcq.store.DB(), now, txn, typ)
},
func(intents []roachpb.Intent, poison bool, wait bool) error {
return repl.store.intentResolver.resolveIntents(ctx, intents, poison, wait)
})
if err != nil {
return err
}
log.VEventf(ctx, 1, "completed with stats %+v", info)
info.updateMetrics(gcq.store.metrics)
var ba roachpb.BatchRequest
var gcArgs roachpb.GCRequest
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
gcArgs.Keys = gcKeys
gcArgs.Threshold = info.Threshold
gcArgs.TxnSpanGCThreshold = info.TxnSpanGCThreshold
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(&gcArgs)
if _, pErr := repl.Send(ctx, ba); pErr != nil {
log.ErrEvent(ctx, pErr.String())
return pErr.GoError()
}
return nil
}
// sendToReplicas sends one or more RPCs to clients specified by the
// slice of replicas. On success, Send returns the first successful
// reply. If an error occurs which is not specific to a single
// replica, it's returned immediately. Otherwise, when all replicas
// have been tried and failed, returns a send error.
func (ds *DistSender) sendToReplicas(
opts SendOptions,
rangeID roachpb.RangeID,
replicas ReplicaSlice,
args roachpb.BatchRequest,
rpcContext *rpc.Context,
) (*roachpb.BatchResponse, error) {
if len(replicas) < 1 {
return nil, roachpb.NewSendError(
fmt.Sprintf("insufficient replicas (%d) to satisfy send request of %d",
len(replicas), 1))
}
var ambiguousResult bool
var haveCommit bool
// We only check for committed txns, not aborts because aborts may
// be retried without any risk of inconsistencies.
if etArg, ok := args.GetArg(roachpb.EndTransaction); ok &&
etArg.(*roachpb.EndTransactionRequest).Commit {
haveCommit = true
}
done := make(chan BatchCall, len(replicas))
transportFactory := opts.transportFactory
if transportFactory == nil {
transportFactory = grpcTransportFactory
}
transport, err := transportFactory(opts, rpcContext, replicas, args)
if err != nil {
return nil, err
}
defer transport.Close()
if transport.IsExhausted() {
return nil, roachpb.NewSendError(
fmt.Sprintf("sending to all %d replicas failed", len(replicas)))
}
// Send the first request.
pending := 1
log.VEventf(opts.ctx, 2, "sending RPC for batch: %s", args.Summary())
transport.SendNext(done)
// Wait for completions. This loop will retry operations that fail
// with errors that reflect per-replica state and may succeed on
// other replicas.
var sendNextTimer timeutil.Timer
defer sendNextTimer.Stop()
for {
sendNextTimer.Reset(opts.SendNextTimeout)
select {
case <-sendNextTimer.C:
sendNextTimer.Read = true
// On successive RPC timeouts, send to additional replicas if available.
if !transport.IsExhausted() {
log.VEventf(opts.ctx, 2, "timeout, trying next peer")
pending++
transport.SendNext(done)
}
case call := <-done:
pending--
err := call.Err
if err == nil {
if log.V(2) {
log.Infof(opts.ctx, "RPC reply: %s", call.Reply)
} else if log.V(1) && call.Reply.Error != nil {
log.Infof(opts.ctx, "application error: %s", call.Reply.Error)
}
if call.Reply.Error == nil {
return call.Reply, nil
} else if !ds.handlePerReplicaError(opts.ctx, transport, rangeID, call.Reply.Error) {
// The error received is not specific to this replica, so we
// should return it instead of trying other replicas. However,
// if we're trying to commit a transaction and there are
// still other RPCs outstanding or an ambiguous RPC error
// was already received, we must return an ambiguous commit
// error instead of returned error.
if haveCommit && (pending > 0 || ambiguousResult) {
return nil, roachpb.NewAmbiguousResultError()
}
return call.Reply, nil
}
// Extract the detail so it can be included in the error
// message if this is our last replica.
//
// TODO(bdarnell): The last error is not necessarily the best
// one to return; we may want to remember the "best" error
// we've seen (for example, a NotLeaseHolderError conveys more
// information than a RangeNotFound).
err = call.Reply.Error.GoError()
} else {
if log.V(1) {
log.Warningf(opts.ctx, "RPC error: %s", err)
}
// All connection errors except for an unavailable node (this
// is GRPC's fail-fast error), may mean that the request
// succeeded on the remote server, but we were unable to
// receive the reply. Set the ambiguous commit flag.
//
// We retry ambiguous commit batches to avoid returning the
// unrecoverable AmbiguousResultError. This is safe because
// repeating an already-successfully applied batch is
// guaranteed to return either a TransactionReplayError (in
// case the replay happens at the original leader), or a
// TransactionRetryError (in case the replay happens at a new
// leader). If the original attempt merely timed out or was
// lost, then the batch will succeed and we can be assured the
// commit was applied just once.
//
// The Unavailable code is used by GRPC to indicate that a
// request fails fast and is not sent, so we can be sure there
// is no ambiguity on these errors. Note that these are common
// if a node is down.
// See https://github.com/grpc/grpc-go/blob/52f6504dc290bd928a8139ba94e3ab32ed9a6273/call.go#L182
// See https://github.com/grpc/grpc-go/blob/52f6504dc290bd928a8139ba94e3ab32ed9a6273/stream.go#L158
if haveCommit && grpc.Code(err) != codes.Unavailable {
ambiguousResult = true
}
}
// Send to additional replicas if available.
if !transport.IsExhausted() {
log.VEventf(opts.ctx, 2, "error, trying next peer: %s", err)
pending++
transport.SendNext(done)
}
if pending == 0 {
if ambiguousResult {
err = roachpb.NewAmbiguousResultError()
} else {
err = roachpb.NewSendError(
fmt.Sprintf("sending to all %d replicas failed; last error: %v", len(replicas), err),
)
}
if log.V(2) {
log.ErrEvent(opts.ctx, err.Error())
}
return nil, err
}
}
}
}