// maybeGossipFirstRange adds the sentinel and first range metadata to gossip
// if this is the first range and a leader lease can be obtained. The Store
// calls this periodically on first range replicas.
func (r *Range) maybeGossipFirstRange() error {
if !r.IsFirstRange() {
return nil
}
ctx := r.context()
// Gossip the cluster ID from all replicas of the first range.
log.Infoc(ctx, "gossiping cluster id %s from store %d, range %d", r.rm.ClusterID(),
r.rm.StoreID(), r.Desc().RaftID)
if err := r.rm.Gossip().AddInfo(gossip.KeyClusterID, r.rm.ClusterID(), clusterIDGossipTTL); err != nil {
log.Errorc(ctx, "failed to gossip cluster ID: %s", err)
}
if ok, err := r.getLeaseForGossip(ctx); !ok || err != nil {
return err
}
log.Infoc(ctx, "gossiping sentinel from store %d, range %d", r.rm.StoreID(), r.Desc().RaftID)
if err := r.rm.Gossip().AddInfo(gossip.KeySentinel, r.rm.ClusterID(), clusterIDGossipTTL); err != nil {
log.Errorc(ctx, "failed to gossip cluster ID: %s", err)
}
log.Infoc(ctx, "gossiping first range from store %d, range %d", r.rm.StoreID(), r.Desc().RaftID)
if err := r.rm.Gossip().AddInfo(gossip.KeyFirstRangeDescriptor, *r.Desc(), configGossipTTL); err != nil {
log.Errorc(ctx, "failed to gossip first range metadata: %s", err)
}
return nil
}
// processWriteIntentError tries to push the conflicting
// transaction(s) responsible for the given WriteIntentError, and to
// resolve those intents if possible. Returns a new error to be used
// in place of the original.
//
// The returned error may be a copy of the original WriteIntentError,
// with or without the Resolved flag set, which governs the client's
// retry behavior (if the transaction is pushed, the Resolved flag is
// set to tell the client to retry immediately; otherwise it is false
// to cause the client to back off).
func (ir *intentResolver) processWriteIntentError(ctx context.Context,
wiPErr *roachpb.Error, r *Replica, args roachpb.Request, h roachpb.Header,
pushType roachpb.PushTxnType) *roachpb.Error {
wiErr, ok := wiPErr.GetDetail().(*roachpb.WriteIntentError)
if !ok {
return roachpb.NewErrorf("not a WriteIntentError: %v", wiPErr)
}
if log.V(6) {
log.Infoc(ctx, "resolving write intent %s", wiErr)
}
method := args.Method()
readOnly := roachpb.IsReadOnly(args) // TODO(tschottdorf): pass as param
resolveIntents, pushErr := ir.maybePushTransactions(ctx, wiErr.Intents, h, pushType, false)
if resErr := ir.resolveIntents(ctx, r, resolveIntents,
false /* !wait */, pushType == roachpb.PUSH_ABORT /* poison */); resErr != nil {
// When resolving without waiting, errors should not
// usually be returned here, although there are some cases
// when they may be (especially when a test cluster is in
// the process of shutting down).
log.Warningf("asynchronous resolveIntents failed: %s", resErr)
}
if pushErr != nil {
if log.V(1) {
log.Infoc(ctx, "on %s: %s", method, pushErr)
}
// For write/write conflicts within a transaction, propagate the
// push failure, not the original write intent error. The push
// failure will instruct the client to restart the transaction
// with a backoff.
if h.Txn != nil && h.Txn.ID != nil && !readOnly {
return pushErr
}
// For read/write conflicts, and non-transactional write/write
// conflicts, return the write intent error which engages
// backoff/retry (with !Resolved). We don't need to restart the
// txn, only resend the read with a backoff.
return wiPErr
}
// We pushed all transactions, so tell the client everything's
// resolved and it can retry immediately.
wiErr.Resolved = true
return wiPErr // references wiErr
}
// start starts the node by registering the storage instance for the
// RPC service "Node" and initializing stores for each specified
// engine. Launches periodic store gossiping in a goroutine.
func (n *Node) start(rpcServer *rpc.Server, addr net.Addr, engines []engine.Engine,
attrs roachpb.Attributes, stopper *stop.Stopper) error {
n.initDescriptor(addr, attrs)
const method = "Node.Batch"
if err := rpcServer.Register(method, n.executeCmd, &roachpb.BatchRequest{}); err != nil {
log.Fatalf("unable to register node service with RPC server: %s", err)
}
// Start status monitor.
n.status.StartMonitorFeed(n.ctx.EventFeed)
// Initialize stores, including bootstrapping new ones.
if err := n.initStores(engines, stopper); err != nil {
return err
}
n.startedAt = n.ctx.Clock.Now().WallTime
// Initialize publisher for Node Events. This requires the NodeID, which is
// initialized by initStores(); because of this, some Store initialization
// events will precede the StartNodeEvent on the feed.
n.feed = status.NewNodeEventFeed(n.Descriptor.NodeID, n.ctx.EventFeed)
n.feed.StartNode(n.Descriptor, n.startedAt)
n.startPublishStatuses(stopper)
n.startGossip(stopper)
log.Infoc(n.context(), "Started node with %v engine(s) and attributes %v", engines, attrs.Attrs)
return nil
}
func (r *Range) maybeGossipConfigsLocked(match func(configPrefix proto.Key) bool) {
if r.rm.Gossip() == nil || !r.isInitialized() {
return
}
ctx := r.context()
for i, cd := range configDescriptors {
if match(cd.keyPrefix) {
// Check for a bad range split. This should never happen as ranges
// cannot be split mid-config.
if !r.ContainsKey(cd.keyPrefix.PrefixEnd()) {
// If we ever implement configs that span multiple ranges,
// we must update store.startGossip accordingly. For the
// time being, it will only fire the first range.
log.Fatalc(ctx, "range splits configuration values for %s", cd.keyPrefix)
}
configMap, hash, err := loadConfigMap(r.rm.Engine(), cd.keyPrefix, cd.configI)
if err != nil {
log.Errorc(ctx, "failed loading %s config map: %s", cd.gossipKey, err)
continue
}
if r.configHashes == nil {
r.configHashes = map[int][]byte{}
}
if prevHash, ok := r.configHashes[i]; !ok || !bytes.Equal(prevHash, hash) {
r.configHashes[i] = hash
log.Infoc(ctx, "gossiping %s config from store %d, range %d", cd.gossipKey, r.rm.StoreID(), r.Desc().RaftID)
if err := r.rm.Gossip().AddInfo(cd.gossipKey, configMap, 0*time.Second); err != nil {
log.Errorc(ctx, "failed to gossip %s configMap: %s", cd.gossipKey, err)
continue
}
}
}
}
}
// start starts the node by registering the storage instance for the
// RPC service "Node" and initializing stores for each specified
// engine. Launches periodic store gossiping in a goroutine.
func (n *Node) start(rpcServer *rpc.Server, engines []engine.Engine,
attrs proto.Attributes, stopper *util.Stopper) error {
n.initDescriptor(rpcServer.Addr(), attrs)
if err := rpcServer.RegisterName("Node", (*nodeServer)(n)); err != nil {
log.Fatalf("unable to register node service with RPC server: %s", err)
}
// Start status monitor.
n.status.StartMonitorFeed(n.ctx.EventFeed)
stopper.AddCloser(n.ctx.EventFeed)
// Initialize stores, including bootstrapping new ones.
if err := n.initStores(engines, stopper); err != nil {
return err
}
// Pass NodeID to status monitor - this value is initialized in initStores,
// but the StatusMonitor must be active before initStores.
n.status.SetNodeID(n.Descriptor.NodeID)
// Initialize publisher for Node Events.
n.feed = status.NewNodeEventFeed(n.Descriptor.NodeID, n.ctx.EventFeed)
n.startedAt = n.ctx.Clock.Now().WallTime
n.startStoresScanner(stopper)
n.startPublishStatuses(stopper)
n.startGossip(stopper)
log.Infoc(n.context(), "Started node with %v engine(s) and attributes %v", engines, attrs.Attrs)
return nil
}
func (s *Store) insertRangeLogEvent(txn *client.Txn, event rangeLogEvent) error {
// Record range log event to console log.
var info string
if event.info != nil {
info = *event.info
}
log.Infoc(txn.Context, "Range Event: %q, range: %d, info: %s",
event.eventType,
event.rangeID,
info)
const insertEventTableStmt = `
INSERT INTO system.rangelog (
timestamp, rangeID, storeID, eventType, otherRangeID, info
)
VALUES(
$1, $2, $3, $4, $5, $6
)
`
args := []interface{}{
event.timestamp,
event.rangeID,
event.storeID,
event.eventType,
nil, // otherRangeID
nil, // info
}
if event.otherRangeID != nil {
args[4] = *event.otherRangeID
}
if event.info != nil {
args[5] = *event.info
}
// Update range event metrics. We do this close to the insertion of the
// corresponding range log entry to reduce potential skew between metrics and
// range log.
switch event.eventType {
case RangeEventLogSplit:
s.metrics.rangeSplits.Inc(1)
case RangeEventLogAdd:
s.metrics.rangeAdds.Inc(1)
case RangeEventLogRemove:
s.metrics.rangeRemoves.Inc(1)
}
rows, err := s.ctx.SQLExecutor.ExecuteStatementInTransaction(txn, insertEventTableStmt, args...)
if err != nil {
return err
}
if rows != 1 {
return errors.Errorf("%d rows affected by log insertion; expected exactly one row affected.", rows)
}
return nil
}
// start starts the node by registering the storage instance for the
// RPC service "Node" and initializing stores for each specified
// engine. Launches periodic store gossiping in a goroutine.
func (n *Node) start(rpcServer *rpc.Server, engines []engine.Engine,
attrs proto.Attributes, stopper *stop.Stopper) error {
n.initDescriptor(rpcServer.Addr(), attrs)
requests := []proto.Request{
&proto.BatchRequest{},
&proto.GetRequest{},
&proto.PutRequest{},
&proto.ConditionalPutRequest{},
&proto.IncrementRequest{},
&proto.DeleteRequest{},
&proto.DeleteRangeRequest{},
&proto.ScanRequest{},
&proto.ReverseScanRequest{},
&proto.EndTransactionRequest{},
&proto.AdminSplitRequest{},
&proto.AdminMergeRequest{},
&proto.HeartbeatTxnRequest{},
&proto.GCRequest{},
&proto.PushTxnRequest{},
&proto.RangeLookupRequest{},
&proto.ResolveIntentRequest{},
&proto.ResolveIntentRangeRequest{},
&proto.MergeRequest{},
&proto.TruncateLogRequest{},
&proto.LeaderLeaseRequest{},
}
for _, r := range requests {
if err := rpcServer.Register("Node."+r.Method().String(), n.executeCmd, r); err != nil {
log.Fatalf("unable to register node service with RPC server: %s", err)
}
}
// Start status monitor.
n.status.StartMonitorFeed(n.ctx.EventFeed)
// Initialize stores, including bootstrapping new ones.
if err := n.initStores(engines, stopper); err != nil {
return err
}
n.startedAt = n.ctx.Clock.Now().WallTime
// Initialize publisher for Node Events. This requires the NodeID, which is
// initialized by initStores(); because of this, some Store initialization
// events will precede the StartNodeEvent on the feed.
n.feed = status.NewNodeEventFeed(n.Descriptor.NodeID, n.ctx.EventFeed)
n.feed.StartNode(n.Descriptor, n.startedAt)
n.startPublishStatuses(stopper)
n.startGossip(stopper)
log.Infoc(n.context(), "Started node with %v engine(s) and attributes %v", engines, attrs.Attrs)
return nil
}
// start starts the node by registering the storage instance for the
// RPC service "Node" and initializing stores for each specified
// engine. Launches periodic store gossiping in a goroutine.
func (n *Node) start(rpcServer *rpc.Server, addr net.Addr, engines []engine.Engine, attrs roachpb.Attributes) error {
n.initDescriptor(addr, attrs)
// Start status monitor.
n.status.StartMonitorFeed(n.ctx.EventFeed)
// Initialize stores, including bootstrapping new ones.
if err := n.initStores(engines, n.stopper); err != nil {
if err == errNeedsBootstrap {
// This node has no initialized stores and no way to connect to
// an existing cluster, so we bootstrap it.
clusterID, err := bootstrapCluster(engines)
if err != nil {
return err
}
log.Infof("**** cluster %s has been created", clusterID)
log.Infof("**** add additional nodes by specifying --join=%s", addr)
// Make sure we add the node as a resolver.
selfResolver, err := resolver.NewResolverFromAddress(addr)
if err != nil {
return err
}
n.ctx.Gossip.SetResolvers([]resolver.Resolver{selfResolver})
// After bootstrapping, try again to initialize the stores.
if err := n.initStores(engines, n.stopper); err != nil {
return err
}
} else {
return err
}
}
n.startedAt = n.ctx.Clock.Now().WallTime
// Initialize publisher for Node Events. This requires the NodeID, which is
// initialized by initStores(); because of this, some Store initialization
// events will precede the StartNodeEvent on the feed.
n.feed = status.NewNodeEventFeed(n.Descriptor.NodeID, n.ctx.EventFeed)
n.feed.StartNode(n.Descriptor, n.startedAt)
n.startPublishStatuses(n.stopper)
n.startGossip(n.stopper)
// Register the RPC methods we support last as doing so allows RPCs to be
// received which may access state initialized above without locks.
const method = "Node.Batch"
if err := rpcServer.Register(method, n.executeCmd, &roachpb.BatchRequest{}); err != nil {
log.Fatalf("unable to register node service with RPC server: %s", err)
}
log.Infoc(n.context(), "Started node with %v engine(s) and attributes %v", engines, attrs.Attrs)
return nil
}
// start starts the node by registering the storage instance for the
// RPC service "Node" and initializing stores for each specified
// engine. Launches periodic store gossiping in a goroutine.
func (n *Node) start(addr net.Addr, engines []engine.Engine, attrs roachpb.Attributes) error {
n.initDescriptor(addr, attrs)
// Initialize stores, including bootstrapping new ones.
if err := n.initStores(engines, n.stopper); err != nil {
if err == errNeedsBootstrap {
n.initialBoot = true
// This node has no initialized stores and no way to connect to
// an existing cluster, so we bootstrap it.
clusterID, err := bootstrapCluster(engines, n.txnMetrics)
if err != nil {
return err
}
log.Infof("**** cluster %s has been created", clusterID)
log.Infof("**** add additional nodes by specifying --join=%s", addr)
// Make sure we add the node as a resolver.
selfResolver, err := resolver.NewResolverFromAddress(addr)
if err != nil {
return err
}
n.ctx.Gossip.SetResolvers([]resolver.Resolver{selfResolver})
// After bootstrapping, try again to initialize the stores.
if err := n.initStores(engines, n.stopper); err != nil {
return err
}
} else {
return err
}
}
n.startedAt = n.ctx.Clock.Now().WallTime
// Initialize the recorder with the NodeID, which is initialized by initStores().
n.recorder.NodeStarted(n.Descriptor, n.startedAt)
n.startComputePeriodicMetrics(n.stopper)
n.startGossip(n.stopper)
// Record node started event.
n.recordJoinEvent()
log.Infoc(n.context(context.TODO()), "Started node with %v engine(s) and attributes %v", engines, attrs.Attrs)
return nil
}
// InsertEventRecord inserts a single event into the event log as part of the
// provided transaction.
func (ev EventLogger) InsertEventRecord(txn *client.Txn, eventType EventLogType, targetID, reportingID int32, info interface{}) error {
// Record event record insertion in local log output.
log.Infoc(txn.Context, "Event: %q, target: %d, info: %+v",
eventType,
targetID,
info)
const insertEventTableStmt = `
INSERT INTO system.eventlog (
timestamp, eventType, targetID, reportingID, info
)
VALUES(
$1, $2, $3, $4, $5
)
`
args := []interface{}{
ev.selectEventTimestamp(txn.Proto.Timestamp),
eventType,
targetID,
reportingID,
nil, // info
}
if info != nil {
infoBytes, err := json.Marshal(info)
if err != nil {
return err
}
args[4] = string(infoBytes)
}
rows, err := ev.ExecuteStatementInTransaction(txn, insertEventTableStmt, args...)
if err != nil {
return err
}
if rows != 1 {
return errors.Errorf("%d rows affected by log insertion; expected exactly one row affected.", rows)
}
return nil
}
// applyRaftCommand applies a raft command from the replicated log to the
// underlying state machine (i.e. the engine).
// When certain critical operations fail, a replicaCorruptionError may be
// returned and must be handled by the caller.
func (r *Range) applyRaftCommand(ctx context.Context, index uint64, originNode proto.RaftNodeID, args proto.Request, reply proto.Response) (rErr error) {
if index <= 0 {
log.Fatalc(ctx, "raft command index is <= 0")
}
committed := false
// The very last thing we do before returning is move the applied index
// forward, unless that has already happened as part of a successfully
// committed batch.
defer func() {
if !committed {
// We didn't commit the batch, but advance the last applied index nonetheless.
if err := setAppliedIndex(r.rm.Engine(), r.Desc().RaftID, index); err != nil {
rErr = newReplicaCorruptionError(
util.Errorf("could not advance applied index"), err, rErr)
return
}
atomic.StoreUint64(&r.appliedIndex, index)
}
}()
if lease := r.getLease(); args.Method() != proto.InternalLeaderLease &&
(!lease.OwnedBy(originNode) || !lease.Covers(args.Header().Timestamp)) {
// Verify the leader lease is held, unless this command is trying to
// obtain it. Any other Raft command has had the leader lease held
// by the replica at proposal time, but this may no more be the case.
// Corruption aside, the most likely reason is a leadership change (the
// most recent leader assumes responsibility for all past timestamps as
// well). In that case, it's not valid to go ahead with the execution:
// Writes must be aware of the last time the mutated key was read, and
// since reads are served locally by the lease holder without going
// through Raft, a read which was not taken into account may have been
// served. Hence, we must retry at the current leader.
//
// It's crucial that we don't update the response cache for the error
// returned below since the request is going to be retried with the
// same ClientCmdID and would get the distributed sender stuck in an
// infinite loop, retrieving a stale NotLeaderError over and over
// again, even when proposing at the correct replica.
return r.newNotLeaderError(lease)
}
// Anything happening from now on needs to enter the response cache.
defer func() {
// TODO(tamird,tschottdorf): according to #1400 we intend to set the reply
// header's error as late as possible and in a central location. Range
// commands still write to the header directly, but once they don't this
// could be the authoritative location that sets the reply error for any-
// thing that makes it into Raft. Note that we must set this prior to
// signaling cmd.done below, or the waiting RPC handler might proceed
// before we've updated its reply.
//
// It is important that the error is set before the reply is saved into
// the response cache.
reply.Header().SetGoError(rErr)
if proto.IsWrite(args) {
// No matter the result, add result to the response cache if this
// is a write method. This must be done as part of the execution of
// raft commands so that every replica maintains the same responses
// to continue request idempotence, even if leadership changes.
if err := r.respCache.PutResponse(args.Header().CmdID, reply); err != nil {
rErr = newReplicaCorruptionError(
util.Errorf("could not put to response cache"), err, rErr)
return
}
}
}()
header := args.Header()
// Check the response cache to ensure idempotency.
if proto.IsWrite(args) {
if ok, err := r.respCache.GetResponse(header.CmdID, reply); ok && err == nil {
if log.V(1) {
log.Infoc(ctx, "found response cache entry for %+v", args.Header().CmdID)
}
return err
} else if ok && err != nil {
return newReplicaCorruptionError(
util.Errorf("could not read from response cache"), err)
}
}
// Create a new batch for the command to ensure all or nothing semantics.
batch := r.rm.Engine().NewBatch()
defer batch.Close()
// Create a engine.MVCCStats instance.
ms := engine.MVCCStats{}
// Execute the command; the error will also be set in the reply header.
// TODO(tschottdorf,tamird) For #1400, want to refactor executeCmd to not
// touch the reply header's error field.
intents, err := r.executeCmd(batch, &ms, args, reply)
// If the execution of the command wasn't successful, stop here.
if err != nil {
return err
}
if oldIndex := atomic.LoadUint64(&r.appliedIndex); oldIndex >= index {
return newReplicaCorruptionError(
util.Errorf("applied index moved backwards: %d >= %d", oldIndex, index))
}
// Advance the applied index atomically within the batch.
if err := setAppliedIndex(batch, r.Desc().RaftID, index); err != nil {
return newReplicaCorruptionError(
util.Errorf("could not update applied index"), err)
}
if proto.IsWrite(args) {
// On success, flush the MVCC stats to the batch and commit.
if err := r.stats.MergeMVCCStats(batch, &ms, header.Timestamp.WallTime); err != nil {
return newReplicaCorruptionError(util.Errorf("could not merge MVCC stats"), err)
}
if err := batch.Commit(); err != nil {
return newReplicaCorruptionError(util.Errorf("could not commit batch"), err)
}
committed = true
// Publish update to event feed.
r.rm.EventFeed().updateRange(r, args.Method(), &ms)
// After successful commit, update cached stats and appliedIndex value.
atomic.StoreUint64(&r.appliedIndex, index)
// If the commit succeeded, potentially add range to split queue.
r.maybeAddToSplitQueue()
// Maybe update gossip configs on a put.
switch args.(type) {
case *proto.PutRequest, *proto.DeleteRequest, *proto.DeleteRangeRequest:
if header.Key.Less(keys.SystemMax) {
// We hold the lock already.
r.maybeGossipConfigsLocked(func(configPrefix proto.Key) bool {
return bytes.HasPrefix(header.Key, configPrefix)
})
}
}
}
// On success and only on the replica on which this command originated,
// resolve skipped intents asynchronously.
if originNode == r.rm.RaftNodeID() {
r.handleSkippedIntents(args, intents)
}
return nil
}
// applyRaftCommandInBatch executes the command in a batch engine and
// returns the batch containing the results. The caller is responsible
// for committing the batch, even on error.
func (r *Range) applyRaftCommandInBatch(ctx context.Context, index uint64, originNode proto.RaftNodeID,
args proto.Request, ms *engine.MVCCStats) (engine.Engine, proto.Response, error) {
// Create a new batch for the command to ensure all or nothing semantics.
batch := r.rm.Engine().NewBatch()
if lease := r.getLease(); args.Method() != proto.InternalLeaderLease &&
(!lease.OwnedBy(originNode) || !lease.Covers(args.Header().Timestamp)) {
// Verify the leader lease is held, unless this command is trying to
// obtain it. Any other Raft command has had the leader lease held
// by the replica at proposal time, but this may no longer be the case.
// Corruption aside, the most likely reason is a leadership change (the
// most recent leader assumes responsibility for all past timestamps as
// well). In that case, it's not valid to go ahead with the execution:
// Writes must be aware of the last time the mutated key was read, and
// since reads are served locally by the lease holder without going
// through Raft, a read which was not taken into account may have been
// served. Hence, we must retry at the current leader.
//
// It's crucial that we don't update the response cache for the error
// returned below since the request is going to be retried with the
// same ClientCmdID and would get the distributed sender stuck in an
// infinite loop, retrieving a stale NotLeaderError over and over
// again, even when proposing at the correct replica.
return batch, nil, r.newNotLeaderError(lease, originNode)
}
// Check the response cache to ensure idempotency.
if proto.IsWrite(args) {
if reply, err := r.respCache.GetResponse(batch, args.Header().CmdID); err != nil {
// Any error encountered while fetching the response cache entry means corruption.
return batch, reply, newReplicaCorruptionError(util.Errorf("could not read from response cache"), err)
} else if reply != nil {
if log.V(1) {
log.Infoc(ctx, "found response cache entry for %+v", args.Header().CmdID)
}
// TODO(tamird): move this into the response cache itself
defer func() { reply.Header().Error = nil }()
// We successfully read from the response cache, so return whatever error
// was present in the cached entry (if any).
return batch, reply, reply.Header().GoError()
}
}
// Execute the command.
reply, intents, rErr := r.executeCmd(batch, ms, args)
// Regardless of error, add result to the response cache if this is
// a write method. This must be done as part of the execution of
// raft commands so that every replica maintains the same responses
// to continue request idempotence, even if leadership changes.
if proto.IsWrite(args) {
if rErr == nil {
// If command was successful, flush the MVCC stats to the batch.
if err := r.stats.MergeMVCCStats(batch, ms, args.Header().Timestamp.WallTime); err != nil {
log.Fatalc(ctx, "setting mvcc stats in a batch should never fail: %s", err)
}
} else {
// Otherwise, reset the batch to clear out partial execution and
// prepare for the failed response cache entry.
batch.Close()
batch = r.rm.Engine().NewBatch()
}
// TODO(tamird): move this into the response cache itself
if reply == nil {
reply = args.CreateReply()
}
if reply.Header().Error != nil {
panic("the world is on fire")
}
reply.Header().SetGoError(rErr)
if err := r.respCache.PutResponse(batch, args.Header().CmdID, reply); err != nil {
log.Fatalc(ctx, "putting a response cache entry in a batch should never fail: %s", err)
}
reply.Header().Error = nil
}
// If the execution of the command wasn't successful, stop here.
if rErr != nil {
return batch, reply, rErr
}
// On success and only on the replica on which this command originated,
// resolve skipped intents asynchronously.
if originNode == r.rm.RaftNodeID() {
r.handleSkippedIntents(args, intents)
}
return batch, reply, nil
}
// resolveIntents resolves the given intents. For those which are local to the
// range, we submit directly to the range-local Raft instance; the call returns
// as soon as all resolve commands have been **proposed** (not executed). This
// ensures that if a waiting client retries immediately after conflict
// resolution, it will not hit the same intents again. All non-local intents
// are resolved asynchronously in a batch.
// TODO(tschottdorf): once Txn records have a list of possibly open intents,
// resolveIntents should send an RPC to update the transaction(s) as well (for
// those intents with non-pending Txns).
func (r *Replica) resolveIntents(ctx context.Context, intents []proto.Intent) {
trace := tracer.FromCtx(ctx)
tracer.ToCtx(ctx, nil) // we're doing async stuff below; those need new traces
trace.Event("resolving intents [async]")
var wg sync.WaitGroup
bArgs := &proto.BatchRequest{}
bArgs.User = security.RootUser
for i := range intents {
intent := intents[i] // avoids a race in `i, intent := range ...`
var resolveArgs proto.Request
var local bool // whether this intent lives on this Range
{
header := proto.RequestHeader{
// Use the pushee's timestamp, which might be lower than the
// pusher's request timestamp. No need to push the intent higher
// than the pushee's txn!
Timestamp: intent.Txn.Timestamp,
Key: intent.Key,
EndKey: intent.EndKey,
User: security.RootUser,
Txn: &intent.Txn,
}
if len(intent.EndKey) == 0 {
resolveArgs = &proto.ResolveIntentRequest{RequestHeader: header}
local = r.ContainsKey(intent.Key)
} else {
resolveArgs = &proto.ResolveIntentRangeRequest{RequestHeader: header}
local = r.ContainsKeyRange(intent.Key, intent.EndKey)
}
}
// If the intent isn't (completely) local, we'll need to send an external request.
// We'll batch them all up and send at the end.
if !local {
bArgs.Add(resolveArgs)
continue
}
// If it is local, it goes directly into Raft.
// TODO(tschottdorf): this may be premature optimization. Consider just
// treating everything as an external request. This means having to
// wait for complete execution of the command (whereas now we just wait
// for proposition) and some more overhead sending things around.
wg.Add(1)
action := func() {
// Trace this under the ID of the intent owner.
ctx := tracer.ToCtx(ctx, r.rm.Tracer().NewTrace(resolveArgs.Header().Txn))
if _, err := r.addWriteCmd(ctx, resolveArgs, &wg); err != nil && log.V(1) {
log.Warningc(ctx, "resolve for key %s failed: %s", intent.Key, err)
}
}
if !r.rm.Stopper().RunAsyncTask(action) {
// Still run the task. Our caller already has a task and going async
// here again is merely for performance, but some intents need to
// be resolved because they might block other tasks. See #1684.
// Note that handleSkippedIntents has a TODO in case #1684 comes
// back.
action()
}
}
// Resolve all of the intents which aren't local to the Range. This is a
// no-op if all are local.
b := &client.Batch{}
b.InternalAddCall(proto.Call{Args: bArgs, Reply: &proto.BatchResponse{}})
action := func() {
// TODO(tschottdorf): no tracing here yet. Probably useful at some point,
// but needs a) the corresponding interface and b) facilities for tracing
// multiple tracees at the same time (batch full of possibly individual
// txns).
if err := r.rm.DB().Run(b); err != nil {
if log.V(1) {
log.Infoc(ctx, "%s", err)
}
}
}
if !r.rm.Stopper().RunAsyncTask(action) {
// As with local intents, try async to not keep the caller waiting, but
// when draining just go ahead and do it synchronously. See #1684.
action()
}
// Wait until all the local `ResolveIntent`s have been submitted to raft.
// No-op if all were external.
wg.Wait()
}