// RangeLookup dispatches an RangeLookup request for the given
// metadata key to the replicas of the given range. Note that we allow
// inconsistent reads when doing range lookups for efficiency. Getting
// stale data is not a correctness problem but instead may
// infrequently result in additional latency as additional range
// lookups may be required. Note also that rangeLookup bypasses the
// DistSender's Send() method, so there is no error inspection and
// retry logic here; this is not an issue since the lookup performs a
// single inconsistent read only.
func (ds *DistSender) RangeLookup(key roachpb.RKey, desc *roachpb.RangeDescriptor, considerIntents, useReverseScan bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
ba := roachpb.BatchRequest{}
ba.ReadConsistency = roachpb.INCONSISTENT
ba.Add(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
// We can interpret the RKey as a Key here since it's a metadata
// lookup; those are never local.
Key: key.AsRawKey(),
},
MaxRanges: ds.rangeLookupMaxRanges,
ConsiderIntents: considerIntents,
Reverse: useReverseScan,
})
replicas := newReplicaSlice(ds.gossip, desc)
// TODO(tschottdorf) consider a Trace here, potentially that of the request
// that had the cache miss and waits for the result.
br, err := ds.sendRPC(tracing.NilSpan(), desc.RangeID, replicas, orderRandom, ba)
if err != nil {
return nil, err
}
if br.Error != nil {
return nil, br.Error
}
return br.Responses[0].GetInner().(*roachpb.RangeLookupResponse).Ranges, nil
}
// processReplica processes a single replica. This should not be
// called externally to the queue. bq.mu.Lock should not be held
// while calling this method.
func (bq *baseQueue) processReplica(repl *Replica, clock *hlc.Clock) error {
// Load the system config.
cfg := bq.gossip.GetSystemConfig()
if cfg == nil {
bq.eventLog.Infof(log.V(1), "no system config available. skipping")
return nil
}
desc := repl.Desc()
if !bq.impl.acceptsUnsplitRanges() && cfg.NeedsSplit(desc.StartKey, desc.EndKey) {
// Range needs to be split due to zone configs, but queue does
// not accept unsplit ranges.
bq.eventLog.Infof(log.V(3), "%s: split needed; skipping", repl)
return nil
}
// If the queue requires a replica to have the range leader lease in
// order to be processed, check whether this replica has leader lease
// and renew or acquire if necessary.
if bq.impl.needsLeaderLease() {
// Create a "fake" get request in order to invoke redirectOnOrAcquireLease.
if err := repl.redirectOnOrAcquireLeaderLease(tracing.NilSpan()); err != nil {
bq.eventLog.Infof(log.V(3), "%s: could not acquire leader lease; skipping", repl)
return nil
}
}
bq.eventLog.Infof(log.V(3), "%s: processing", repl)
start := time.Now()
if err := bq.impl.process(clock.Now(), repl, cfg); err != nil {
return err
}
bq.eventLog.Infof(log.V(2), "%s: done: %s", repl, time.Since(start))
return nil
}
// Send sends one or more RPCs to clients specified by the slice of
// replicas. On success, Send returns the first successful reply. Otherwise,
// Send returns an error if and as soon as the number of failed RPCs exceeds
// the available endpoints less the number of required replies.
//
// TODO(pmattis): Get rid of the getArgs function which requires the caller to
// maintain a map from address to replica. Instead, pass in the list of
// replicas instead of a list of addresses and use that to populate the
// requests.
func send(opts SendOptions, replicas ReplicaSlice,
args roachpb.BatchRequest, context *rpc.Context) (proto.Message, error) {
sp := opts.Trace
if sp == nil {
sp = tracing.NilSpan()
}
if len(replicas) < 1 {
return nil, roachpb.NewSendError(
fmt.Sprintf("insufficient replicas (%d) to satisfy send request of %d",
len(replicas), 1), false)
}
done := make(chan *netrpc.Call, len(replicas))
clients := make([]batchClient, 0, len(replicas))
for i, replica := range replicas {
clients = append(clients, batchClient{
Client: rpc.NewClient(&replica.NodeDesc.Address, context),
replica: &replicas[i],
args: args,
})
}
var orderedClients []batchClient
switch opts.Ordering {
case orderStable:
orderedClients = clients
case orderRandom:
// Randomly permute order, but keep known-unhealthy clients last.
var nHealthy int
for i, client := range clients {
select {
case <-client.Healthy():
clients[i], clients[nHealthy] = clients[nHealthy], clients[i]
nHealthy++
default:
}
}
shuffleClients(clients[:nHealthy])
shuffleClients(clients[nHealthy:])
orderedClients = clients
}
// TODO(spencer): going to need to also sort by affinity; closest
// ping time should win. Makes sense to have the rpc client/server
// heartbeat measure ping times. With a bit of seasoning, each
// node will be able to order the healthy replicas based on latency.
// Send the first request.
sendOneFn(&orderedClients[0], opts.Timeout, context, sp, done)
orderedClients = orderedClients[1:]
var errors, retryableErrors int
// Wait for completions.
var sendNextTimer util.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 len(orderedClients) > 0 {
sp.LogEvent("timeout, trying next peer")
sendOneFn(&orderedClients[0], opts.Timeout, context, sp, done)
orderedClients = orderedClients[1:]
}
case call := <-done:
if call.Error == nil {
// Verify response data integrity if this is a proto response.
if req, reqOk := call.Args.(roachpb.Request); reqOk {
if resp, respOk := call.Reply.(roachpb.Response); respOk {
if err := resp.Verify(req); err != nil {
call.Error = err
}
} else {
call.Error = util.Errorf("response to proto request must be a proto")
}
}
}
err := call.Error
if err == nil {
if log.V(2) {
log.Infof("successful reply: %+v", call.Reply)
}
return call.Reply.(proto.Message), nil
}
// Error handling.
if log.V(1) {
log.Warningf("error reply: %s", err)
}
errors++
// Since we have a reconnecting client here, disconnect errors are retryable.
disconnected := err == netrpc.ErrShutdown || err == io.ErrUnexpectedEOF
if retryErr, ok := err.(retry.Retryable); disconnected || (ok && retryErr.CanRetry()) {
retryableErrors++
}
if remainingNonErrorRPCs := len(replicas) - errors; remainingNonErrorRPCs < 1 {
return nil, roachpb.NewSendError(
fmt.Sprintf("too many errors encountered (%d of %d total): %v",
errors, len(clients), err), remainingNonErrorRPCs+retryableErrors >= 1)
}
// Send to additional replicas if available.
if len(orderedClients) > 0 {
sp.LogEvent("error, trying next peer")
sendOneFn(&orderedClients[0], opts.Timeout, context, sp, done)
orderedClients = orderedClients[1:]
}
}
}
}