// handleNodeStatus handles GET requests for a single node's status.
func (s *statusServer) Node(
ctx context.Context, req *serverpb.NodeRequest,
) (*status.NodeStatus, error) {
ctx = s.AnnotateCtx(ctx)
nodeID, _, err := s.parseNodeID(req.NodeId)
if err != nil {
return nil, grpc.Errorf(codes.InvalidArgument, err.Error())
}
key := keys.NodeStatusKey(nodeID)
b := &client.Batch{}
b.Get(key)
if err := s.db.Run(ctx, b); err != nil {
log.Error(ctx, err)
return nil, grpc.Errorf(codes.Internal, err.Error())
}
var nodeStatus status.NodeStatus
if err := b.Results[0].Rows[0].ValueProto(&nodeStatus); err != nil {
err = errors.Errorf("could not unmarshal NodeStatus from %s: %s", key, err)
log.Error(ctx, err)
return nil, grpc.Errorf(codes.Internal, err.Error())
}
return &nodeStatus, nil
}
// Nodes returns all node statuses.
func (s *statusServer) Nodes(
ctx context.Context, req *serverpb.NodesRequest,
) (*serverpb.NodesResponse, error) {
ctx = s.AnnotateCtx(ctx)
startKey := keys.StatusNodePrefix
endKey := startKey.PrefixEnd()
b := &client.Batch{}
b.Scan(startKey, endKey)
if err := s.db.Run(ctx, b); err != nil {
log.Error(ctx, err)
return nil, grpc.Errorf(codes.Internal, err.Error())
}
rows := b.Results[0].Rows
resp := serverpb.NodesResponse{
Nodes: make([]status.NodeStatus, len(rows)),
}
for i, row := range rows {
if err := row.ValueProto(&resp.Nodes[i]); err != nil {
log.Error(ctx, err)
return nil, grpc.Errorf(codes.Internal, err.Error())
}
}
return &resp, nil
}
func (ds *ServerImpl) setupFlow(
ctx context.Context, req *SetupFlowRequest, simpleFlowConsumer RowReceiver,
) (*Flow, error) {
sp, err := tracing.JoinOrNew(ds.AmbientContext.Tracer, req.TraceContext, "flow")
if err != nil {
return nil, err
}
ctx = opentracing.ContextWithSpan(ctx, sp)
txn := ds.setupTxn(ctx, &req.Txn)
flowCtx := FlowCtx{
Context: ctx,
id: req.Flow.FlowID,
evalCtx: &ds.evalCtx,
rpcCtx: ds.RPCContext,
txn: txn,
}
f := newFlow(flowCtx, ds.flowRegistry, simpleFlowConsumer)
if err := f.setupFlow(&req.Flow); err != nil {
log.Error(ctx, err)
sp.Finish()
return nil, err
}
return f, nil
}
func (c *v3Conn) finish(ctx context.Context) {
// This is better than always flushing on error.
if err := c.wr.Flush(); err != nil {
log.Error(ctx, err)
}
_ = c.conn.Close()
}
func (ds *ServerImpl) setupFlow(
ctx context.Context, req *SetupFlowRequest, syncFlowConsumer RowReceiver,
) (*Flow, error) {
sp, err := tracing.JoinOrNew(ds.AmbientContext.Tracer, req.TraceContext, "flow")
if err != nil {
return nil, err
}
ctx = opentracing.ContextWithSpan(ctx, sp)
// TODO(radu): we should sanity check some of these fields (especially
// txnProto).
flowCtx := FlowCtx{
Context: ctx,
id: req.Flow.FlowID,
evalCtx: &ds.evalCtx,
rpcCtx: ds.RPCContext,
txnProto: &req.Txn,
clientDB: ds.DB,
}
f := newFlow(flowCtx, ds.flowRegistry, syncFlowConsumer)
if err := f.setupFlow(&req.Flow); err != nil {
log.Error(ctx, err)
sp.Finish()
return nil, err
}
return f, nil
}
func (a *allocSim) maybeLogError(err error) {
if localcluster.IsUnavailableError(err) {
return
}
log.Error(context.Background(), err)
atomic.AddUint64(&a.stats.errors, 1)
}
func (z *zeroSum) maybeLogError(err error) {
if localcluster.IsUnavailableError(err) || strings.Contains(err.Error(), "range is frozen") {
return
}
log.Error(context.Background(), err)
atomic.AddUint64(&z.stats.errors, 1)
}
func (s *statusServer) handleVars(w http.ResponseWriter, r *http.Request) {
w.Header().Set(httputil.ContentTypeHeader, httputil.PlaintextContentType)
err := s.metricSource.PrintAsText(w)
if err != nil {
log.Error(r.Context(), err)
http.Error(w, err.Error(), http.StatusInternalServerError)
}
}
// FlowStream is part of the DistSQLServer interface.
func (ds *ServerImpl) FlowStream(stream DistSQL_FlowStreamServer) error {
ctx := ds.AnnotateCtx(context.TODO())
err := ds.flowStreamInt(stream)
if err != nil {
log.Error(ctx, err)
}
return err
}
// SetStorage provides an instance of the Storage interface
// for reading and writing gossip bootstrap data from persistent
// storage. This should be invoked as early in the lifecycle of a
// gossip instance as possible, but can be called at any time.
func (g *Gossip) SetStorage(storage Storage) error {
ctx := g.AnnotateCtx(context.TODO())
// Maintain lock ordering.
var storedBI BootstrapInfo
if err := storage.ReadBootstrapInfo(&storedBI); err != nil {
log.Warningf(ctx, "failed to read gossip bootstrap info: %s", err)
}
g.mu.Lock()
defer g.mu.Unlock()
g.storage = storage
// Merge the stored bootstrap info addresses with any we've become
// aware of through gossip.
existing := map[string]struct{}{}
makeKey := func(a util.UnresolvedAddr) string { return fmt.Sprintf("%s,%s", a.Network(), a.String()) }
for _, addr := range g.bootstrapInfo.Addresses {
existing[makeKey(addr)] = struct{}{}
}
for _, addr := range storedBI.Addresses {
// If the address is new, and isn't our own address, add it.
if _, ok := existing[makeKey(addr)]; !ok && addr != g.mu.is.NodeAddr {
g.maybeAddBootstrapAddress(addr)
}
}
// Persist merged addresses.
if numAddrs := len(g.bootstrapInfo.Addresses); numAddrs > len(storedBI.Addresses) {
if err := g.storage.WriteBootstrapInfo(&g.bootstrapInfo); err != nil {
log.Error(ctx, err)
}
}
// Cycle through all persisted bootstrap hosts and add resolvers for
// any which haven't already been added.
newResolverFound := false
for _, addr := range g.bootstrapInfo.Addresses {
if !g.maybeAddResolver(addr) {
continue
}
// If we find a new resolver, reset the resolver index so that the
// next resolver we try is the first of the new resolvers.
if !newResolverFound {
newResolverFound = true
g.resolverIdx = len(g.resolvers) - 1
}
}
// If a new resolver was found, immediately signal bootstrap.
if newResolverFound {
if log.V(1) {
log.Infof(ctx, "found new resolvers from storage; signalling bootstrap")
}
g.signalStalledLocked()
}
return nil
}
// DrainQueue locks the queue and processes the remaining queued replicas. It
// processes the replicas in the order they're queued in, one at a time.
// Exposed for testing only.
//
// TODO(bdarnell): this method may race with the call to bq.pop() in
// the main loop, in which case it does not guarantee that all
// replicas have been processed by the time it returns. This is most
// noticeable with ForceReplicaGCScanAndProcess, since the replica GC
// queue has many event-driven triggers. This should synchronize
// somehow with processLoop so we wait for anything being handled
// there to finish too. When that's done, the SucceedsSoon at the end
// of TestRemoveRangeWithoutGC (and perhaps others) can be replaced
// with a one-time check.
func (bq *baseQueue) DrainQueue(clock *hlc.Clock) {
ctx := bq.AnnotateCtx(context.TODO())
for repl := bq.pop(); repl != nil; repl = bq.pop() {
annotatedCtx := repl.AnnotateCtx(ctx)
if err := bq.processReplica(annotatedCtx, repl, clock); err != nil {
bq.failures.Inc(1)
log.Error(annotatedCtx, err)
}
}
}
// updateNodeAddress is a gossip callback which fires with each
// update to the node address. This allows us to compute the
// total size of the gossip network (for determining max peers
// each gossip node is allowed to have), as well as to create
// new resolvers for each encountered host and to write the
// set of gossip node addresses to persistent storage when it
// changes.
func (g *Gossip) updateNodeAddress(_ string, content roachpb.Value) {
ctx := g.AnnotateCtx(context.TODO())
var desc roachpb.NodeDescriptor
if err := content.GetProto(&desc); err != nil {
log.Error(ctx, err)
return
}
g.mu.Lock()
defer g.mu.Unlock()
// Skip if the node has already been seen.
if _, ok := g.nodeDescs[desc.NodeID]; ok {
return
}
g.nodeDescs[desc.NodeID] = &desc
// Recompute max peers based on size of network and set the max
// sizes for incoming and outgoing node sets.
maxPeers := g.maxPeers(len(g.nodeDescs))
g.mu.incoming.setMaxSize(maxPeers)
g.outgoing.setMaxSize(maxPeers)
// Skip if it's our own address.
if desc.Address == g.mu.is.NodeAddr {
return
}
// Add this new node address (if it's not already there) to our list
// of resolvers so we can keep connecting to gossip if the original
// resolvers go offline.
g.maybeAddResolver(desc.Address)
// Add new address (if it's not already there) to bootstrap info and
// persist if possible.
if g.storage != nil && g.maybeAddBootstrapAddress(desc.Address) {
if err := g.storage.WriteBootstrapInfo(&g.bootstrapInfo); err != nil {
log.Error(ctx, err)
}
}
}
// process() is called on every range for which this node is a lease holder.
func (q *consistencyQueue) process(ctx context.Context, repl *Replica, _ config.SystemConfig) error {
req := roachpb.CheckConsistencyRequest{}
if _, pErr := repl.CheckConsistency(ctx, req); pErr != nil {
log.Error(ctx, pErr.GoError())
}
// Update the last processed time for this queue.
if err := repl.setQueueLastProcessed(ctx, q.name, repl.store.Clock().Now()); err != nil {
log.ErrEventf(ctx, "failed to update last processed time: %v", err)
}
return nil
}
// Addr returns the TCP address to connect to.
func (c *Container) Addr(port nat.Port) *net.TCPAddr {
containerInfo, err := c.Inspect()
if err != nil {
log.Error(context.TODO(), err)
return nil
}
bindings, ok := containerInfo.NetworkSettings.Ports[port]
if !ok || len(bindings) == 0 {
return nil
}
portNum, err := strconv.Atoi(bindings[0].HostPort)
if err != nil {
log.Error(context.TODO(), err)
return nil
}
return &net.TCPAddr{
IP: dockerIP(),
Port: portNum,
}
}
// GRPCDial calls grpc.Dial with the options appropriate for the context.
func (ctx *Context) GRPCDial(target string, opts ...grpc.DialOption) (*grpc.ClientConn, error) {
ctx.conns.Lock()
meta, ok := ctx.conns.cache[target]
if !ok {
meta = &connMeta{}
ctx.conns.cache[target] = meta
}
ctx.conns.Unlock()
meta.Do(func() {
var dialOpt grpc.DialOption
if ctx.Insecure {
dialOpt = grpc.WithInsecure()
} else {
tlsConfig, err := ctx.GetClientTLSConfig()
if err != nil {
meta.err = err
return
}
dialOpt = grpc.WithTransportCredentials(credentials.NewTLS(tlsConfig))
}
dialOpts := make([]grpc.DialOption, 0, 2+len(opts))
dialOpts = append(dialOpts, dialOpt)
dialOpts = append(dialOpts, grpc.WithBackoffMaxDelay(maxBackoff))
dialOpts = append(dialOpts, opts...)
if log.V(1) {
log.Infof(ctx.masterCtx, "dialing %s", target)
}
meta.conn, meta.err = grpc.DialContext(ctx.masterCtx, target, dialOpts...)
if meta.err == nil {
if err := ctx.Stopper.RunTask(func() {
ctx.Stopper.RunWorker(func() {
err := ctx.runHeartbeat(meta.conn, target)
if err != nil && !grpcutil.IsClosedConnection(err) {
log.Error(ctx.masterCtx, err)
}
ctx.removeConn(target, meta)
})
}); err != nil {
meta.err = err
// removeConn and ctx's cleanup worker both lock ctx.conns. However,
// to avoid racing with meta's initialization, the cleanup worker
// blocks on meta.Do while holding ctx.conns. Invoke removeConn
// asynchronously to avoid deadlock.
go ctx.removeConn(target, meta)
}
}
})
return meta.conn, meta.err
}
func (rq *replicateQueue) shouldQueue(
ctx context.Context, now hlc.Timestamp, repl *Replica, sysCfg config.SystemConfig,
) (shouldQ bool, priority float64) {
if !repl.store.splitQueue.Disabled() && repl.needsSplitBySize() {
// If the range exceeds the split threshold, let that finish first.
// Ranges must fit in memory on both sender and receiver nodes while
// being replicated. This supplements the check provided by
// acceptsUnsplitRanges, which looks at zone config boundaries rather
// than data size.
//
// This check is ignored if the split queue is disabled, since in that
// case, the split will never come.
return
}
// Find the zone config for this range.
desc := repl.Desc()
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(ctx, err)
return
}
action, priority := rq.allocator.ComputeAction(zone, desc)
if action != AllocatorNoop {
if log.V(2) {
log.Infof(ctx, "%s repair needed (%s), enqueuing", repl, action)
}
return true, priority
}
// See if there is a rebalancing opportunity present.
leaseStoreID := repl.store.StoreID()
if lease, _ := repl.getLease(); lease != nil {
leaseStoreID = lease.Replica.StoreID
}
target := rq.allocator.RebalanceTarget(
zone.Constraints,
desc.Replicas,
leaseStoreID,
desc.RangeID,
)
if log.V(2) {
if target != nil {
log.Infof(ctx, "%s rebalance target found, enqueuing", repl)
} else {
log.Infof(ctx, "%s no rebalance target found, not enqueuing", repl)
}
}
return target != nil, 0
}
// SendNext invokes the specified RPC on the supplied client when the
// client is ready. On success, the reply is sent on the channel;
// otherwise an error is sent.
func (gt *grpcTransport) SendNext(done chan<- BatchCall) {
client := gt.orderedClients[gt.clientIndex]
gt.clientIndex++
gt.setPending(client.args.Replica, true)
addr := client.remoteAddr
if log.V(2) {
log.Infof(gt.opts.ctx, "sending request to %s: %+v", addr, client.args)
}
if localServer := gt.rpcContext.GetLocalInternalServerForAddr(addr); enableLocalCalls && localServer != nil {
// Clone the request. At the time of writing, Replica may mutate it
// during command execution which can lead to data races.
//
// TODO(tamird): we should clone all of client.args.Header, but the
// assertions in protoutil.Clone fire and there seems to be no
// reasonable workaround.
origTxn := client.args.Txn
if origTxn != nil {
clonedTxn := origTxn.Clone()
client.args.Txn = &clonedTxn
}
reply, err := localServer.Batch(gt.opts.ctx, &client.args)
gt.setPending(client.args.Replica, false)
done <- BatchCall{Reply: reply, Err: err}
return
}
go func() {
// HACK: GRPC leaks if client calls are made with a context which
// is cancelable but doesn't actually get canceled. Insulate this
// call from our outer context, which may last for the lifetime of
// a client session.
// TODO(bdarnell): remove after https://github.com/grpc/grpc-go/issues/888
// is fixed.
ctx, cancel := context.WithCancel(gt.opts.ctx)
defer cancel()
reply, err := client.client.Batch(ctx, &client.args)
if reply != nil {
for i := range reply.Responses {
if err := reply.Responses[i].GetInner().Verify(client.args.Requests[i].GetInner()); err != nil {
log.Error(gt.opts.ctx, err)
}
}
}
gt.setPending(client.args.Replica, false)
done <- BatchCall{Reply: reply, Err: err}
}()
}
// maybeCleanupBootstrapAddresses cleans up the stored bootstrap addresses to
// include only those currently available via gossip. The gossip mutex must
// be held by the caller.
func (g *Gossip) maybeCleanupBootstrapAddressesLocked() {
if g.storage == nil || g.hasCleanedBS {
return
}
defer func() { g.hasCleanedBS = true }()
ctx := g.AnnotateCtx(context.TODO())
log.Event(ctx, "cleaning up bootstrap addresses")
g.resolvers = g.resolvers[:0]
g.resolverIdx = 0
g.bootstrapInfo.Addresses = g.bootstrapInfo.Addresses[:0]
g.bootstrapAddrs = map[util.UnresolvedAddr]struct{}{}
g.resolverAddrs = map[util.UnresolvedAddr]resolver.Resolver{}
g.resolversTried = map[int]struct{}{}
var desc roachpb.NodeDescriptor
if err := g.mu.is.visitInfos(func(key string, i *Info) error {
if strings.HasPrefix(key, KeyNodeIDPrefix) {
if err := i.Value.GetProto(&desc); err != nil {
return err
}
if desc.Address == g.mu.is.NodeAddr {
return nil
}
g.maybeAddResolver(desc.Address)
g.maybeAddBootstrapAddress(desc.Address)
}
return nil
}); err != nil {
log.Error(ctx, err)
return
}
if err := g.storage.WriteBootstrapInfo(&g.bootstrapInfo); err != nil {
log.Error(ctx, err)
}
}
// storeGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) storeGossipUpdate(_ string, content roachpb.Value) {
var storeDesc roachpb.StoreDescriptor
if err := content.GetProto(&storeDesc); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
// Does this storeDetail exist yet?
detail := sp.getStoreDetailLocked(storeDesc.StoreID)
detail.markAlive(sp.clock.Now(), &storeDesc)
sp.mu.queue.enqueue(detail)
}
// storeGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) storeGossipUpdate(_ string, content roachpb.Value) {
var storeDesc roachpb.StoreDescriptor
if err := content.GetProto(&storeDesc); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
detail := sp.getStoreDetailLocked(storeDesc.StoreID)
detail.desc = &storeDesc
detail.lastUpdatedTime = sp.clock.PhysicalTime()
sp.mu.nodeLocalities[storeDesc.Node.NodeID] = storeDesc.Node.Locality
}
func (r *Replica) computeChecksumPostApply(
ctx context.Context, args roachpb.ComputeChecksumRequest,
) {
stopper := r.store.Stopper()
id := args.ChecksumID
now := timeutil.Now()
r.mu.Lock()
var notify chan struct{}
if c, ok := r.mu.checksums[id]; !ok {
// There is no record of this ID. Make a new notification.
notify = make(chan struct{})
} else if !c.started {
// A CollectChecksumRequest is waiting on the existing notification.
notify = c.notify
} else {
// A previous attempt was made to compute the checksum.
r.mu.Unlock()
return
}
r.gcOldChecksumEntriesLocked(now)
// Create an entry with checksum == nil and gcTimestamp unset.
r.mu.checksums[id] = replicaChecksum{started: true, notify: notify}
desc := *r.mu.state.Desc
r.mu.Unlock()
snap := r.store.NewSnapshot()
// Compute SHA asynchronously and store it in a map by UUID.
if err := stopper.RunAsyncTask(ctx, func(ctx context.Context) {
defer snap.Close()
var snapshot *roachpb.RaftSnapshotData
if args.Snapshot {
snapshot = &roachpb.RaftSnapshotData{}
}
sha, err := r.sha512(desc, snap, snapshot)
if err != nil {
log.Errorf(ctx, "%v", err)
sha = nil
}
r.computeChecksumDone(ctx, id, sha, snapshot)
}); err != nil {
defer snap.Close()
log.Error(ctx, errors.Wrapf(err, "could not run async checksum computation (ID = %s)", id))
// Set checksum to nil.
r.computeChecksumDone(ctx, id, nil, nil)
}
}
// livenessGossipUpdate is the gossip callback used to keep the
// in-memory liveness info up to date.
func (nl *NodeLiveness) livenessGossipUpdate(key string, content roachpb.Value) {
var liveness Liveness
if err := content.GetProto(&liveness); err != nil {
log.Error(context.TODO(), err)
return
}
// If there's an existing liveness record, only update the received
// timestamp if this is our first receipt of this node's liveness
// or if the expiration or epoch was advanced.
nl.mu.Lock()
defer nl.mu.Unlock()
exLiveness, ok := nl.mu.nodes[liveness.NodeID]
if !ok || exLiveness.Expiration.Less(liveness.Expiration) || exLiveness.Epoch < liveness.Epoch {
nl.mu.nodes[liveness.NodeID] = liveness
}
}
// SendNext invokes the specified RPC on the supplied client when the
// client is ready. On success, the reply is sent on the channel;
// otherwise an error is sent.
func (gt *grpcTransport) SendNext(ctx context.Context, done chan<- BatchCall) {
client := gt.orderedClients[gt.clientIndex]
gt.clientIndex++
gt.setPending(client.args.Replica, true)
batchFn := func(ctx context.Context, args *roachpb.BatchRequest) (*roachpb.BatchResponse, error) {
reply, err := client.client.Batch(ctx, args)
if reply != nil {
for i := range reply.Responses {
if err := reply.Responses[i].GetInner().Verify(client.args.Requests[i].GetInner()); err != nil {
log.Error(ctx, err)
}
}
}
return reply, err
}
addr := client.remoteAddr
if localServer := gt.rpcContext.GetLocalInternalServerForAddr(addr); enableLocalCalls && localServer != nil {
batchFn = func(ctx context.Context, args *roachpb.BatchRequest) (*roachpb.BatchResponse, error) {
gt.opts.metrics.LocalSentCount.Inc(1)
return localServer.Batch(ctx, args)
}
// Clone the request. At the time of writing, Replica may mutate it
// during command execution which can lead to data races.
//
// TODO(tamird): we should clone all of client.args.Header, but the
// assertions in protoutil.Clone fire and there seems to be no
// reasonable workaround.
origTxn := client.args.Txn
if origTxn != nil {
clonedTxn := origTxn.Clone()
client.args.Txn = &clonedTxn
}
}
go func() {
gt.opts.metrics.SentCount.Inc(1)
if log.V(2) {
log.Infof(ctx, "sending request to %s: %+v", addr, client.args)
}
reply, err := batchFn(ctx, &client.args)
gt.setPending(client.args.Replica, false)
done <- BatchCall{Reply: reply, Err: err}
}()
}
// SendNext invokes the specified RPC on the supplied client when the
// client is ready. On success, the reply is sent on the channel;
// otherwise an error is sent.
func (gt *grpcTransport) SendNext(ctx context.Context, done chan<- BatchCall) {
client := gt.orderedClients[gt.clientIndex]
gt.clientIndex++
gt.setPending(client.args.Replica, true)
// Fork the original context as this async send may outlast the
// caller's context.
ctx, sp := tracing.ForkCtxSpan(ctx, "grpcTransport SendNext")
go func() {
defer tracing.FinishSpan(sp)
gt.opts.metrics.SentCount.Inc(1)
reply, err := func() (*roachpb.BatchResponse, error) {
if enableLocalCalls {
if localServer := gt.rpcContext.GetLocalInternalServerForAddr(client.remoteAddr); localServer != nil {
// Clone the request. At the time of writing, Replica may mutate it
// during command execution which can lead to data races.
//
// TODO(tamird): we should clone all of client.args.Header, but the
// assertions in protoutil.Clone fire and there seems to be no
// reasonable workaround.
origTxn := client.args.Txn
if origTxn != nil {
clonedTxn := origTxn.Clone()
client.args.Txn = &clonedTxn
}
gt.opts.metrics.LocalSentCount.Inc(1)
log.VEvent(ctx, 2, "sending request to local server")
return localServer.Batch(ctx, &client.args)
}
}
log.VEventf(ctx, 2, "sending request to %s", client.remoteAddr)
reply, err := client.client.Batch(ctx, &client.args)
if reply != nil {
for i := range reply.Responses {
if err := reply.Responses[i].GetInner().Verify(client.args.Requests[i].GetInner()); err != nil {
log.Error(ctx, err)
}
}
}
return reply, err
}()
gt.setPending(client.args.Replica, false)
done <- BatchCall{Reply: reply, Err: err}
}()
}
// deadReplicasGossipUpdate is the gossip callback used to keep the StorePool up to date.
func (sp *StorePool) deadReplicasGossipUpdate(_ string, content roachpb.Value) {
var replicas roachpb.StoreDeadReplicas
if err := content.GetProto(&replicas); err != nil {
ctx := sp.AnnotateCtx(context.TODO())
log.Error(ctx, err)
return
}
sp.mu.Lock()
defer sp.mu.Unlock()
detail := sp.getStoreDetailLocked(replicas.StoreID)
deadReplicas := make(map[roachpb.RangeID][]roachpb.ReplicaDescriptor)
for _, r := range replicas.Replicas {
deadReplicas[r.RangeID] = append(deadReplicas[r.RangeID], r.Replica)
}
detail.deadReplicas = deadReplicas
}
// RunSyncFlow is part of the DistSQLServer interface.
func (ds *ServerImpl) RunSyncFlow(req *SetupFlowRequest, stream DistSQL_RunSyncFlowServer) error {
// Set up the outgoing mailbox for the stream.
mbox := newOutboxSyncFlowStream(stream)
ctx := ds.AnnotateCtx(stream.Context())
f, err := ds.SetupSyncFlow(ctx, req, mbox)
if err != nil {
log.Error(ctx, err)
return err
}
mbox.setFlowCtx(&f.FlowCtx)
if err := ds.Stopper.RunTask(func() {
f.waitGroup.Add(1)
mbox.start(&f.waitGroup)
f.Start(func() {})
f.Wait()
f.Cleanup()
}); err != nil {
return err
}
return mbox.err
}
// shouldQueue determines whether a range should be queued for
// splitting. This is true if the range is intersected by a zone config
// prefix or if the range's size in bytes exceeds the limit for the zone.
func (sq *splitQueue) shouldQueue(
ctx context.Context, now hlc.Timestamp, repl *Replica, sysCfg config.SystemConfig,
) (shouldQ bool, priority float64) {
desc := repl.Desc()
if len(sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)) > 0 {
// Set priority to 1 in the event the range is split by zone configs.
priority = 1
shouldQ = true
}
// Add priority based on the size of range compared to the max
// size for the zone it's in.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(ctx, err)
return
}
if ratio := float64(repl.GetMVCCStats().Total()) / float64(zone.RangeMaxBytes); ratio > 1 {
priority += ratio
shouldQ = true
}
return
}
// updateNodeAddress is a gossip callback which fires with each
// update to the node address. This allows us to compute the
// total size of the gossip network (for determining max peers
// each gossip node is allowed to have), as well as to create
// new resolvers for each encountered host and to write the
// set of gossip node addresses to persistent storage when it
// changes.
func (g *Gossip) updateNodeAddress(key string, content roachpb.Value) {
ctx := g.AnnotateCtx(context.TODO())
var desc roachpb.NodeDescriptor
if err := content.GetProto(&desc); err != nil {
log.Error(ctx, err)
return
}
g.mu.Lock()
defer g.mu.Unlock()
// If desc is the empty descriptor, that indicates that the node has been
// removed from the cluster. If that's the case, remove it from our map of
// nodes to prevent other parts of the system from trying to talk to it.
// We can't directly compare the node against the empty descriptor because
// the proto has a repeated field and thus isn't comparable.
if desc.NodeID == 0 && desc.Address.IsEmpty() {
nodeID, err := NodeIDFromKey(key)
if err != nil {
log.Errorf(ctx, "unable to update node address for removed node: %s", err)
return
}
log.Infof(ctx, "removed node %d from gossip", nodeID)
delete(g.nodeDescs, nodeID)
return
}
// Skip if the node has already been seen.
if _, ok := g.nodeDescs[desc.NodeID]; ok {
return
}
g.nodeDescs[desc.NodeID] = &desc
// Recompute max peers based on size of network and set the max
// sizes for incoming and outgoing node sets.
maxPeers := g.maxPeers(len(g.nodeDescs))
g.mu.incoming.setMaxSize(maxPeers)
g.outgoing.setMaxSize(maxPeers)
// Skip if it's our own address.
if desc.Address == g.mu.is.NodeAddr {
return
}
// Add this new node address (if it's not already there) to our list
// of resolvers so we can keep connecting to gossip if the original
// resolvers go offline.
g.maybeAddResolver(desc.Address)
// We ignore empty addresses for the sake of not breaking the many tests
// that don't bother specifying addresses.
if desc.Address.IsEmpty() {
return
}
// If the new node's address conflicts with another node's address, then it
// must be the case that the new node has replaced the previous one. Remove
// it from our set of tracked descriptors to ensure we don't attempt to
// connect to its previous identity (as came up in issue #10266).
oldNodeID, ok := g.bootstrapAddrs[desc.Address]
if ok && oldNodeID != unknownNodeID && oldNodeID != desc.NodeID {
log.Infof(ctx, "removing node %d which was at same address (%s) as new node %v",
oldNodeID, desc.Address, desc)
delete(g.nodeDescs, oldNodeID)
// Deleting the local copy isn't enough to remove the node from the gossip
// network. We also have to clear it out in the infoStore by overwriting
// it with an empty descriptor, which can be represented as just an empty
// byte array due to how protocol buffers are serialied.
// Calling addInfoLocked here is somewhat recursive since
// updateNodeAddress is typically called in response to the infoStore
// being updated but won't lead to deadlock because it's called
// asynchronously.
key := MakeNodeIDKey(oldNodeID)
var emptyProto []byte
if err := g.addInfoLocked(key, emptyProto, ttlNodeDescriptorGossip); err != nil {
log.Errorf(ctx, "failed to empty node descriptor for node %d: %s", oldNodeID, err)
}
}
// Add new address (if it's not already there) to bootstrap info and
// persist if possible.
added := g.maybeAddBootstrapAddress(desc.Address, desc.NodeID)
if added && g.storage != nil {
if err := g.storage.WriteBootstrapInfo(&g.bootstrapInfo); err != nil {
log.Error(ctx, err)
}
}
}
// Start starts the server on the specified port, starts gossip and initializes
// the node using the engines from the server's context.
//
// The passed context can be used to trace the server startup. The context
// should represent the general startup operation.
func (s *Server) Start(ctx context.Context) error {
ctx = s.AnnotateCtx(ctx)
startTime := timeutil.Now()
tlsConfig, err := s.cfg.GetServerTLSConfig()
if err != nil {
return err
}
httpServer := netutil.MakeServer(s.stopper, tlsConfig, s)
plainRedirectServer := netutil.MakeServer(s.stopper, tlsConfig, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
http.Redirect(w, r, "https://"+r.Host+r.RequestURI, http.StatusPermanentRedirect)
}))
// The following code is a specialization of util/net.go's ListenAndServe
// which adds pgwire support. A single port is used to serve all protocols
// (pg, http, h2) via the following construction:
//
// non-TLS case:
// net.Listen -> cmux.New
// |
// - -> pgwire.Match -> pgwire.Server.ServeConn
// - -> cmux.Any -> grpc.(*Server).Serve
//
// TLS case:
// net.Listen -> cmux.New
// |
// - -> pgwire.Match -> pgwire.Server.ServeConn
// - -> cmux.Any -> grpc.(*Server).Serve
//
// Note that the difference between the TLS and non-TLS cases exists due to
// Go's lack of an h2c (HTTP2 Clear Text) implementation. See inline comments
// in util.ListenAndServe for an explanation of how h2c is implemented there
// and here.
ln, err := net.Listen("tcp", s.cfg.Addr)
if err != nil {
return err
}
log.Eventf(ctx, "listening on port %s", s.cfg.Addr)
unresolvedListenAddr, err := officialAddr(s.cfg.Addr, ln.Addr())
if err != nil {
return err
}
s.cfg.Addr = unresolvedListenAddr.String()
unresolvedAdvertAddr, err := officialAddr(s.cfg.AdvertiseAddr, ln.Addr())
if err != nil {
return err
}
s.cfg.AdvertiseAddr = unresolvedAdvertAddr.String()
s.rpcContext.SetLocalInternalServer(s.node)
m := cmux.New(ln)
pgL := m.Match(pgwire.Match)
anyL := m.Match(cmux.Any())
httpLn, err := net.Listen("tcp", s.cfg.HTTPAddr)
if err != nil {
return err
}
unresolvedHTTPAddr, err := officialAddr(s.cfg.HTTPAddr, httpLn.Addr())
if err != nil {
return err
}
s.cfg.HTTPAddr = unresolvedHTTPAddr.String()
workersCtx := s.AnnotateCtx(context.Background())
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
if err := httpLn.Close(); err != nil {
log.Fatal(workersCtx, err)
}
})
if tlsConfig != nil {
httpMux := cmux.New(httpLn)
clearL := httpMux.Match(cmux.HTTP1())
tlsL := httpMux.Match(cmux.Any())
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(httpMux.Serve())
})
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(plainRedirectServer.Serve(clearL))
})
httpLn = tls.NewListener(tlsL, tlsConfig)
}
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(httpServer.Serve(httpLn))
})
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
netutil.FatalIfUnexpected(anyL.Close())
<-s.stopper.ShouldStop()
s.grpc.Stop()
})
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(s.grpc.Serve(anyL))
})
s.stopper.RunWorker(func() {
pgCtx := s.pgServer.AmbientCtx.AnnotateCtx(context.Background())
netutil.FatalIfUnexpected(httpServer.ServeWith(s.stopper, pgL, func(conn net.Conn) {
connCtx := log.WithLogTagStr(pgCtx, "client", conn.RemoteAddr().String())
if err := s.pgServer.ServeConn(connCtx, conn); err != nil && !netutil.IsClosedConnection(err) {
// Report the error on this connection's context, so that we
// know which remote client caused the error when looking at
// the logs.
log.Error(connCtx, err)
}
}))
})
if len(s.cfg.SocketFile) != 0 {
// Unix socket enabled: postgres protocol only.
unixLn, err := net.Listen("unix", s.cfg.SocketFile)
if err != nil {
return err
}
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
if err := unixLn.Close(); err != nil {
log.Fatal(workersCtx, err)
}
})
s.stopper.RunWorker(func() {
pgCtx := s.pgServer.AmbientCtx.AnnotateCtx(context.Background())
netutil.FatalIfUnexpected(httpServer.ServeWith(s.stopper, unixLn, func(conn net.Conn) {
connCtx := log.WithLogTagStr(pgCtx, "client", conn.RemoteAddr().String())
if err := s.pgServer.ServeConn(connCtx, conn); err != nil && !netutil.IsClosedConnection(err) {
// Report the error on this connection's context, so that we
// know which remote client caused the error when looking at
// the logs.
log.Error(connCtx, err)
}
}))
})
}
// Enable the debug endpoints first to provide an earlier window
// into what's going on with the node in advance of exporting node
// functionality.
// TODO(marc): when cookie-based authentication exists,
// apply it for all web endpoints.
s.mux.HandleFunc(debugEndpoint, http.HandlerFunc(handleDebug))
s.gossip.Start(unresolvedAdvertAddr)
log.Event(ctx, "started gossip")
s.engines, err = s.cfg.CreateEngines()
if err != nil {
return errors.Wrap(err, "failed to create engines")
}
s.stopper.AddCloser(&s.engines)
// We might have to sleep a bit to protect against this node producing non-
// monotonic timestamps. Before restarting, its clock might have been driven
// by other nodes' fast clocks, but when we restarted, we lost all this
// information. For example, a client might have written a value at a
// timestamp that's in the future of the restarted node's clock, and if we
// don't do something, the same client's read would not return the written
// value. So, we wait up to MaxOffset; we couldn't have served timestamps more
// than MaxOffset in the future (assuming that MaxOffset was not changed, see
// #9733).
//
// As an optimization for tests, we don't sleep if all the stores are brand
// new. In this case, the node will not serve anything anyway until it
// synchronizes with other nodes.
{
anyStoreBootstrapped := false
for _, e := range s.engines {
if _, err := storage.ReadStoreIdent(ctx, e); err != nil {
// NotBootstrappedError is expected.
if _, ok := err.(*storage.NotBootstrappedError); !ok {
return err
}
} else {
anyStoreBootstrapped = true
break
}
}
if anyStoreBootstrapped {
sleepDuration := s.clock.MaxOffset() - timeutil.Since(startTime)
if sleepDuration > 0 {
log.Infof(ctx, "sleeping for %s to guarantee HLC monotonicity", sleepDuration)
time.Sleep(sleepDuration)
}
}
}
// Now that we have a monotonic HLC wrt previous incarnations of the process,
// init all the replicas.
err = s.node.start(
ctx,
unresolvedAdvertAddr,
s.engines,
s.cfg.NodeAttributes,
s.cfg.Locality,
)
if err != nil {
return err
}
log.Event(ctx, "started node")
s.nodeLiveness.StartHeartbeat(ctx, s.stopper)
// We can now add the node registry.
s.recorder.AddNode(s.registry, s.node.Descriptor, s.node.startedAt)
// Begin recording runtime statistics.
s.startSampleEnvironment(s.cfg.MetricsSampleInterval)
// Begin recording time series data collected by the status monitor.
s.tsDB.PollSource(
s.cfg.AmbientCtx, s.recorder, s.cfg.MetricsSampleInterval, ts.Resolution10s, s.stopper,
)
// Begin recording status summaries.
s.node.startWriteSummaries(s.cfg.MetricsSampleInterval)
// Create and start the schema change manager only after a NodeID
// has been assigned.
testingKnobs := &sql.SchemaChangerTestingKnobs{}
if s.cfg.TestingKnobs.SQLSchemaChanger != nil {
testingKnobs = s.cfg.TestingKnobs.SQLSchemaChanger.(*sql.SchemaChangerTestingKnobs)
}
sql.NewSchemaChangeManager(testingKnobs, *s.db, s.gossip, s.leaseMgr).Start(s.stopper)
s.distSQLServer.Start()
log.Infof(ctx, "starting %s server at %s", s.cfg.HTTPRequestScheme(), unresolvedHTTPAddr)
log.Infof(ctx, "starting grpc/postgres server at %s", unresolvedListenAddr)
log.Infof(ctx, "advertising CockroachDB node at %s", unresolvedAdvertAddr)
if len(s.cfg.SocketFile) != 0 {
log.Infof(ctx, "starting postgres server at unix:%s", s.cfg.SocketFile)
}
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(m.Serve())
})
log.Event(ctx, "accepting connections")
// Initialize grpc-gateway mux and context.
jsonpb := &protoutil.JSONPb{
EnumsAsInts: true,
EmitDefaults: true,
Indent: " ",
}
protopb := new(protoutil.ProtoPb)
gwMux := gwruntime.NewServeMux(
gwruntime.WithMarshalerOption(gwruntime.MIMEWildcard, jsonpb),
gwruntime.WithMarshalerOption(httputil.JSONContentType, jsonpb),
gwruntime.WithMarshalerOption(httputil.AltJSONContentType, jsonpb),
gwruntime.WithMarshalerOption(httputil.ProtoContentType, protopb),
gwruntime.WithMarshalerOption(httputil.AltProtoContentType, protopb),
)
gwCtx, gwCancel := context.WithCancel(s.AnnotateCtx(context.Background()))
s.stopper.AddCloser(stop.CloserFn(gwCancel))
// Setup HTTP<->gRPC handlers.
conn, err := s.rpcContext.GRPCDial(s.cfg.Addr)
if err != nil {
return errors.Errorf("error constructing grpc-gateway: %s; are your certificates valid?", err)
}
for _, gw := range []grpcGatewayServer{s.admin, s.status, &s.tsServer} {
if err := gw.RegisterGateway(gwCtx, gwMux, conn); err != nil {
return err
}
}
var uiFileSystem http.FileSystem
uiDebug := envutil.EnvOrDefaultBool("COCKROACH_DEBUG_UI", false)
if uiDebug {
uiFileSystem = http.Dir("pkg/ui")
} else {
uiFileSystem = &assetfs.AssetFS{
Asset: ui.Asset,
AssetDir: ui.AssetDir,
AssetInfo: ui.AssetInfo,
}
}
uiFileServer := http.FileServer(uiFileSystem)
s.mux.HandleFunc("/", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.URL.Path == "/" {
if uiDebug {
r.URL.Path = "debug.html"
} else {
r.URL.Path = "release.html"
}
}
uiFileServer.ServeHTTP(w, r)
}))
// TODO(marc): when cookie-based authentication exists,
// apply it for all web endpoints.
s.mux.Handle(adminPrefix, gwMux)
s.mux.Handle(ts.URLPrefix, gwMux)
s.mux.Handle(statusPrefix, gwMux)
s.mux.Handle("/health", gwMux)
s.mux.Handle(statusVars, http.HandlerFunc(s.status.handleVars))
log.Event(ctx, "added http endpoints")
if err := sdnotify.Ready(); err != nil {
log.Errorf(ctx, "failed to signal readiness using systemd protocol: %s", err)
}
log.Event(ctx, "server ready")
return nil
}
// maybeAddToPurgatory possibly adds the specified replica to the
// purgatory queue, which holds replicas which have failed
// processing. To be added, the failing error must implement
// purgatoryError and the queue implementation must have its own
// mechanism for signaling re-processing of replicas held in
// purgatory.
func (bq *baseQueue) maybeAddToPurgatory(
ctx context.Context, repl *Replica, triggeringErr error, clock *hlc.Clock, stopper *stop.Stopper,
) {
// Increment failures metric here to capture all error returns from
// process().
bq.failures.Inc(1)
// Check whether the failure is a purgatory error and whether the queue supports it.
if _, ok := triggeringErr.(purgatoryError); !ok || bq.impl.purgatoryChan() == nil {
log.Error(ctx, triggeringErr)
return
}
bq.mu.Lock()
defer bq.mu.Unlock()
// First, check whether the replica has already been re-added to queue.
if _, ok := bq.mu.replicas[repl.RangeID]; ok {
return
}
log.Error(ctx, errors.Wrap(triggeringErr, "purgatory"))
item := &replicaItem{value: repl.RangeID}
bq.mu.replicas[repl.RangeID] = item
defer func() {
bq.purgatory.Update(int64(len(bq.mu.purgatory)))
}()
// If purgatory already exists, just add to the map and we're done.
if bq.mu.purgatory != nil {
bq.mu.purgatory[repl.RangeID] = triggeringErr
return
}
// Otherwise, create purgatory and start processing.
bq.mu.purgatory = map[roachpb.RangeID]error{
repl.RangeID: triggeringErr,
}
stopper.RunWorker(func() {
ctx := bq.AnnotateCtx(context.Background())
ticker := time.NewTicker(purgatoryReportInterval)
for {
select {
case <-bq.impl.purgatoryChan():
// Remove all items from purgatory into a copied slice.
bq.mu.Lock()
ranges := make([]roachpb.RangeID, 0, len(bq.mu.purgatory))
for rangeID := range bq.mu.purgatory {
item := bq.mu.replicas[rangeID]
ranges = append(ranges, item.value)
bq.remove(item)
}
bq.mu.Unlock()
for _, id := range ranges {
repl, err := bq.store.GetReplica(id)
if err != nil {
log.Errorf(ctx, "range %s no longer exists on store: %s", id, err)
return
}
if stopper.RunTask(func() {
annotatedCtx := repl.AnnotateCtx(ctx)
if err := bq.processReplica(annotatedCtx, repl, clock); err != nil {
bq.maybeAddToPurgatory(annotatedCtx, repl, err, clock, stopper)
}
}) != nil {
return
}
}
bq.mu.Lock()
if len(bq.mu.purgatory) == 0 {
log.Infof(ctx, "purgatory is now empty")
bq.mu.purgatory = nil
bq.mu.Unlock()
return
}
bq.mu.Unlock()
case <-ticker.C:
// Report purgatory status.
bq.mu.Lock()
errMap := map[string]int{}
for _, err := range bq.mu.purgatory {
errMap[err.Error()]++
}
bq.mu.Unlock()
for errStr, count := range errMap {
log.Errorf(ctx, "%d replicas failing with %q", count, errStr)
}
case <-stopper.ShouldStop():
return
}
}
})
}