func (e *eventDemux) start(stopper *util.Stopper) {
stopper.RunWorker(func() {
for {
select {
case event := <-e.events:
switch event := event.(type) {
case *EventLeaderElection:
e.LeaderElection <- event
case *EventCommandCommitted:
e.CommandCommitted <- event
case *EventMembershipChangeCommitted:
e.MembershipChangeCommitted <- event
default:
panic(fmt.Sprintf("got unknown event type %T", event))
}
case <-stopper.ShouldStop():
close(e.CommandCommitted)
close(e.MembershipChangeCommitted)
close(e.LeaderElection)
return
}
}
})
}
// maybeWarnAboutInit looks for signs indicating a cluster which
// hasn't been initialized and warns. There's no absolutely sure way
// to determine whether the current node is simply waiting to be
// bootstrapped to an existing cluster vs. the operator having failed
// to initialize the cluster via the "cockroach init" command, so
// we can only warn.
//
// This method checks whether all gossip bootstrap hosts are
// connected, and whether the node itself is a bootstrap host, but
// there is still no sentinel gossip.
func (g *Gossip) maybeWarnAboutInit(stopper *util.Stopper) {
stopper.RunWorker(func() {
// Wait 5s before first check.
select {
case <-stopper.ShouldStop():
return
case <-time.After(5 * time.Second):
}
retryOptions := retry.Options{
Tag: "check cluster initialization",
Backoff: 5 * time.Second, // first backoff at 5s
MaxBackoff: 60 * time.Second, // max backoff is 60s
Constant: 2, // doubles
MaxAttempts: 0, // indefinite retries
Stopper: stopper, // stop no matter what on stopper
}
// will never error because infinite retries
_ = retry.WithBackoff(retryOptions, func() (retry.Status, error) {
g.mu.Lock()
hasSentinel := g.is.getInfo(KeySentinel) != nil
g.mu.Unlock()
// If we have the sentinel, exit the retry loop.
if hasSentinel {
return retry.Break, nil
}
// Otherwise, if all bootstrap hosts are connected, warn.
if g.triedAll {
log.Warningf("connected to gossip but missing sentinel. Has the cluster been initialized? " +
"Use \"cockroach init\" to initialize.")
}
return retry.Continue, nil
})
})
}
// bootstrap connects the node to the gossip network. Bootstrapping
// commences in the event there are no connected clients or the
// sentinel gossip info is not available. After a successful bootstrap
// connection, this method will block on the stalled condvar, which
// receives notifications that gossip network connectivity has been
// lost and requires re-bootstrapping.
func (g *Gossip) bootstrap(stopper *util.Stopper) {
stopper.RunWorker(func() {
for {
g.mu.Lock()
if g.closed {
g.mu.Unlock()
return
}
// Check whether or not we need bootstrap.
haveClients := g.outgoing.len() > 0
haveSentinel := g.is.getInfo(KeySentinel) != nil
if !haveClients || !haveSentinel {
// Try to get another bootstrap address from the resolvers.
if addr := g.getNextBootstrapAddress(); addr != nil {
g.startClient(addr, g.bsRPCContext, stopper)
}
}
g.mu.Unlock()
// Block until we need bootstrapping again.
select {
case <-g.stalled:
// continue
case <-stopper.ShouldStop():
return
}
}
})
}
// maybeWarnAboutInit looks for signs indicating a cluster which
// hasn't been initialized and warns. There's no absolutely sure way
// to determine whether the current node is simply waiting to be
// bootstrapped to an existing cluster vs. the operator having failed
// to initialize the cluster via the "cockroach init" command, so
// we can only warn.
//
// This method checks whether all gossip bootstrap hosts are
// connected, and whether the node itself is a bootstrap host, but
// there is still no sentinel gossip.
func (g *Gossip) maybeWarnAboutInit(stopper *util.Stopper) {
stopper.RunWorker(func() {
// Wait 5s before first check.
select {
case <-stopper.ShouldStop():
return
case <-time.After(5 * time.Second):
}
retryOptions := retry.Options{
InitialBackoff: 5 * time.Second, // first backoff at 5s
MaxBackoff: 60 * time.Second, // max backoff is 60s
Multiplier: 2, // doubles
Stopper: stopper, // stop no matter what on stopper
}
// will never error because infinite retries
for r := retry.Start(retryOptions); r.Next(); {
g.mu.Lock()
hasSentinel := g.is.getInfo(KeySentinel) != nil
g.mu.Unlock()
// If we have the sentinel, exit the retry loop.
if hasSentinel {
break
}
// Otherwise, if all bootstrap hosts are connected, warn.
if g.triedAll {
log.Warningf("connected to gossip but missing sentinel. Has the cluster been initialized? " +
"Use \"cockroach init\" to initialize.")
}
}
})
}
// waitAndProcess waits for the pace interval and processes the range
// if rng is not nil. The method returns true when the scanner needs
// to be stopped. The method also removes a range from queues when it
// is signaled via the removed channel.
func (rs *rangeScanner) waitAndProcess(start time.Time, clock *hlc.Clock, stopper *util.Stopper,
rng *Range) bool {
waitInterval := rs.paceInterval(start, time.Now())
nextTime := time.After(waitInterval)
if log.V(6) {
log.Infof("Wait time interval set to %s", waitInterval)
}
for {
select {
case <-nextTime:
if rng == nil {
return false
}
if !stopper.StartTask() {
return true
}
// Try adding range to all queues.
for _, q := range rs.queues {
q.MaybeAdd(rng, clock.Now())
}
stopper.FinishTask()
return false
case rng := <-rs.removed:
// Remove range from all queues as applicable.
for _, q := range rs.queues {
q.MaybeRemove(rng)
}
if log.V(6) {
log.Infof("removed range %s", rng)
}
case <-stopper.ShouldStop():
return true
}
}
}
// scanLoop loops endlessly, scanning through ranges available via
// the range iterator, or until the scanner is stopped. The iteration
// is paced to complete a full scan in approximately the scan interval.
func (rs *rangeScanner) scanLoop(clock *hlc.Clock, stopper *util.Stopper) {
start := time.Now()
stats := &storeStats{}
for {
elapsed := time.Now().Sub(start)
remainingNanos := rs.interval.Nanoseconds() - elapsed.Nanoseconds()
if remainingNanos < 0 {
remainingNanos = 0
}
nextIteration := time.Duration(remainingNanos)
if count := rs.iter.EstimatedCount(); count > 0 {
nextIteration = time.Duration(remainingNanos / int64(count))
}
log.V(6).Infof("next range scan iteration in %s", nextIteration)
select {
case <-time.After(nextIteration):
rng := rs.iter.Next()
if rng != nil {
// Try adding range to all queues.
for _, q := range rs.queues {
q.MaybeAdd(rng, clock.Now())
}
stats.RangeCount++
stats.MVCC.Accumulate(rng.stats.GetMVCC())
} else {
// Otherwise, we're done with the iteration. Reset iteration and start time.
rs.iter.Reset()
start = time.Now()
// Increment iteration counter.
atomic.AddInt64(&rs.count, 1)
// Store the most recent scan results in the scanner's stats.
atomic.StorePointer(&rs.stats, unsafe.Pointer(stats))
stats = &storeStats{}
log.V(6).Infof("reset range scan iteration")
}
case rng := <-rs.removed:
// Remove range from all queues as applicable.
for _, q := range rs.queues {
q.MaybeRemove(rng)
}
log.V(6).Infof("removed range %s", rng)
case <-stopper.ShouldStop():
// Exit the loop.
stopper.SetStopped()
return
}
}
}
// start runs the storage loop in a goroutine.
func (w *writeTask) start(stopper *util.Stopper) {
stopper.RunWorker(func() {
for {
var request *writeRequest
select {
case <-w.ready:
continue
case <-stopper.ShouldStop():
return
case request = <-w.in:
}
if log.V(6) {
log.Infof("writeTask got request %#v", *request)
}
response := &writeResponse{make(map[proto.RaftID]*groupWriteResponse)}
for groupID, groupReq := range request.groups {
group := w.storage.GroupStorage(groupID)
if group == nil {
if log.V(4) {
log.Infof("dropping write to group %v", groupID)
}
continue
}
groupResp := &groupWriteResponse{raftpb.HardState{}, -1, -1, groupReq.entries}
response.groups[groupID] = groupResp
if !raft.IsEmptyHardState(groupReq.state) {
err := group.SetHardState(groupReq.state)
if err != nil {
panic(err) // TODO(bdarnell): mark this node dead on storage errors
}
groupResp.state = groupReq.state
}
if !raft.IsEmptySnap(groupReq.snapshot) {
err := group.ApplySnapshot(groupReq.snapshot)
if err != nil {
panic(err) // TODO(bdarnell)
}
}
if len(groupReq.entries) > 0 {
err := group.Append(groupReq.entries)
if err != nil {
panic(err) // TODO(bdarnell)
}
}
}
w.out <- response
}
})
}
// manage manages outgoing clients. Periodically, the infostore is
// scanned for infos with hop count exceeding maxToleratedHops()
// threshold. If the number of outgoing clients doesn't exceed
// MaxPeers, a new gossip client is connected to a randomly selected
// peer beyond maxToleratedHops threshold. Otherwise, the least useful
// peer node is cut off to make room for a replacement. Disconnected
// clients are processed via the disconnected channel and taken out of
// the outgoing address set. If there are no longer any outgoing
// connections or the sentinel gossip is unavailable, the bootstrapper
// is notified via the stalled conditional variable.
func (g *Gossip) manage(stopper *util.Stopper) {
stopper.RunWorker(func() {
// Loop until closed and there are no remaining outgoing connections.
for {
select {
case <-stopper.ShouldStop():
return
case c := <-g.disconnected:
g.doDisconnected(stopper, c)
case <-time.After(g.jitteredGossipInterval()):
g.doCheckTimeout(stopper)
}
}
})
}
// startGossip loops on a periodic ticker to gossip node-related
// information. Starts a goroutine to loop until the node is closed.
func (n *Node) startGossip(stopper *util.Stopper) {
stopper.RunWorker(func() {
ticker := time.NewTicker(gossipInterval)
defer ticker.Stop()
n.gossipCapacities() // one-off run before going to sleep
for {
select {
case <-ticker.C:
n.gossipCapacities()
case <-stopper.ShouldStop():
return
}
}
})
}
// processEventsUntil reads and acknowledges messages from the given channel
// until either the given conditional returns true, the channel is closed or a
// read on the channel times out.
func processEventsUntil(ch <-chan *interceptMessage, stopper *util.Stopper, f func(*RaftMessageRequest) bool) {
for {
select {
case e, ok := <-ch:
if !ok {
return
}
e.ack <- struct{}{}
if f(e.args.(*RaftMessageRequest)) {
return
}
case <-stopper.ShouldStop():
return
}
}
}
// runHeartbeat sends periodic heartbeats to client. Closes the
// connection on error. Heartbeats are sent in an infinite loop until
// an error is encountered.
func (c *Client) runHeartbeat(stopper *util.Stopper) {
if log.V(2) {
log.Infof("client %s starting heartbeat", c.Addr())
}
for {
select {
case <-stopper.ShouldStop():
return
case <-time.After(heartbeatInterval):
if err := c.heartbeat(); err != nil {
log.Infof("client %s heartbeat failed: %v; recycling...", c.Addr(), err)
return
}
}
}
}
// startPublishStatuses starts a loop which periodically instructs each store to
// publish its current status to the event feed.
func (n *Node) startPublishStatuses(stopper *util.Stopper) {
stopper.RunWorker(func() {
// Publish status at the same frequency as metrics are collected.
ticker := time.NewTicker(publishStatusInterval)
defer ticker.Stop()
for {
select {
case <-ticker.C:
err := n.publishStoreStatuses()
if err != nil {
log.Error(err)
}
case <-stopper.ShouldStop():
return
}
}
})
}
func (tq *testQueue) Start(clock *hlc.Clock, stopper *util.Stopper) {
stopper.RunWorker(func() {
for {
select {
case <-time.After(1 * time.Millisecond):
tq.Lock()
if !tq.disabled && len(tq.ranges) > 0 {
tq.ranges = tq.ranges[1:]
tq.processed++
}
tq.Unlock()
case <-stopper.ShouldStop():
tq.Lock()
tq.done = true
tq.Unlock()
return
}
}
})
}
// start initializes the infostore with the rpc server address and
// then begins processing connecting clients in an infinite select
// loop via goroutine. Periodically, clients connected and awaiting
// the next round of gossip are awoken via the conditional variable.
func (s *server) start(rpcServer *rpc.Server, stopper *util.Stopper) {
s.is.NodeAddr = rpcServer.Addr()
if err := rpcServer.RegisterName("Gossip", s); err != nil {
log.Fatalf("unable to register gossip service with RPC server: %s", err)
}
rpcServer.AddCloseCallback(s.onClose)
stopper.RunWorker(func() {
// Periodically wakeup blocked client gossip requests.
for {
select {
case <-time.After(s.jitteredGossipInterval()):
// Wakeup all blocked gossip requests.
s.ready.Broadcast()
case <-stopper.ShouldStop():
s.stop()
return
}
}
})
}
// MonitorRemoteOffsets periodically checks that the offset of this server's
// clock from the true cluster time is within MaxOffset. If the offset exceeds
// MaxOffset, then this method will trigger a fatal error, causing the node to
// suicide.
func (r *RemoteClockMonitor) MonitorRemoteOffsets(stopper *util.Stopper) {
if log.V(1) {
log.Infof("monitoring cluster offset")
}
for {
select {
case <-stopper.ShouldStop():
return
case <-time.After(monitorInterval):
offsetInterval, err := r.findOffsetInterval()
// By the contract of the hlc, if the value is 0, then safety checking
// of the max offset is disabled. However we may still want to
// propagate the information to a status node.
// TODO(embark): once there is a framework for collecting timeseries
// data about the db, propagate the offset status to that.
// Don't forget to protect r.offsets through the Mutex if those
// Fatalf's below ever turn into something less destructive.
if r.lClock.MaxOffset() != 0 {
if err != nil {
log.Fatalf("clock offset from the cluster time "+
"for remote clocks %v could not be determined: %s",
r.offsets, err)
}
if !isHealthyOffsetInterval(offsetInterval, r.lClock.MaxOffset()) {
log.Fatalf("clock offset from the cluster time "+
"for remote clocks: %v is in interval: %s, which "+
"indicates that the true offset is greater than %s",
r.offsets, offsetInterval, time.Duration(r.lClock.MaxOffset()))
}
if log.V(1) {
log.Infof("healthy cluster offset: %s", offsetInterval)
}
}
r.mu.Lock()
r.lastMonitoredAt = r.lClock.PhysicalNow()
r.mu.Unlock()
}
}
}
func (tq *testQueue) Start(clock *hlc.Clock, stopper *util.Stopper) {
stopper.Add(1)
go func() {
for {
select {
case <-time.After(1 * time.Millisecond):
tq.Lock()
if len(tq.ranges) > 0 {
tq.ranges = tq.ranges[1:]
tq.processed++
}
tq.Unlock()
case <-stopper.ShouldStop():
tq.Lock()
tq.done = true
tq.Unlock()
stopper.SetStopped()
return
}
}
}()
}
// processLoop processes the entries in the queue until the provided
// stopper signals exit.
//
// TODO(spencer): current load should factor into range processing timer.
func (bq *baseQueue) processLoop(clock *hlc.Clock, stopper *util.Stopper) {
stopper.RunWorker(func() {
// nextTime is initially nil; we don't start any timers until the queue
// becomes non-empty.
var nextTime <-chan time.Time
for {
select {
// Incoming signal sets the next time to process if there were previously
// no ranges in the queue.
case <-bq.incoming:
if nextTime == nil {
// When the first range is added, wake up immediately. This is
// mainly to facilitate testing without unnecessary sleeps.
nextTime = time.After(0 * time.Millisecond)
}
// Process ranges as the timer expires.
case <-nextTime:
bq.processOne(clock, stopper)
if bq.Length() == 0 {
nextTime = nil
} else {
nextTime = time.After(bq.impl.timer())
}
// Exit on stopper.
case <-stopper.ShouldStop():
bq.Lock()
bq.ranges = map[proto.RaftID]*rangeItem{}
bq.priorityQ = nil
bq.Unlock()
return
}
}
})
}
// startStoresScanner will walk through all the stores in the node every
// ctx.ScanInterval and store the status in the db.
func (n *Node) startStoresScanner(stopper *util.Stopper) {
stopper.RunWorker(func() {
// Pick the smaller of the two intervals.
var minScanInterval time.Duration
if n.ctx.ScanInterval <= n.ctx.ScanMaxIdleTime || n.ctx.ScanMaxIdleTime == 0 {
minScanInterval = n.ctx.ScanInterval
} else {
minScanInterval = n.ctx.ScanMaxIdleTime
}
// TODO(bram): The number of stores is small. The node status should be
// updated whenever a store status is updated.
for interval := time.Duration(0); true; interval = minScanInterval {
select {
case <-time.After(interval):
if !stopper.StartTask() {
continue
}
// Walk through all the stores on this node.
var rangeCount, leaderRangeCount, replicatedRangeCount, availableRangeCount int32
stats := &engine.MVCCStats{}
accessedStoreIDs := []proto.StoreID{}
// will never error because `return nil` below
_ = n.lSender.VisitStores(func(store *storage.Store) error {
storeStatus, err := store.GetStatus()
if err != nil {
log.Error(err)
return nil
}
if storeStatus == nil {
// The store scanner hasn't run on this node yet.
return nil
}
accessedStoreIDs = append(accessedStoreIDs, store.Ident.StoreID)
rangeCount += storeStatus.RangeCount
leaderRangeCount += storeStatus.LeaderRangeCount
replicatedRangeCount += storeStatus.ReplicatedRangeCount
availableRangeCount += storeStatus.AvailableRangeCount
stats.Add(&storeStatus.Stats)
return nil
})
// Store the combined stats in the db.
now := n.ctx.Clock.Now().WallTime
status := &NodeStatus{
Desc: n.Descriptor,
StoreIDs: accessedStoreIDs,
UpdatedAt: now,
StartedAt: n.startedAt,
RangeCount: rangeCount,
Stats: *stats,
LeaderRangeCount: leaderRangeCount,
ReplicatedRangeCount: replicatedRangeCount,
AvailableRangeCount: availableRangeCount,
}
key := keys.NodeStatusKey(int32(n.Descriptor.NodeID))
if err := n.ctx.DB.Put(key, status); err != nil {
log.Error(err)
}
// Increment iteration count.
n.completedScan.L.Lock()
n.scanCount++
n.completedScan.Broadcast()
n.completedScan.L.Unlock()
if log.V(6) {
log.Infof("store scan iteration completed")
}
stopper.FinishTask()
case <-stopper.ShouldStop():
// Exit the loop.
return
}
}
})
}
// gossip loops, sending deltas of the infostore and receiving deltas
// in turn. If an alternate is proposed on response, the client addr
// is modified and method returns for forwarding by caller.
func (c *client) gossip(g *Gossip, stopper *util.Stopper) error {
localMaxSeq := int64(0)
remoteMaxSeq := int64(-1)
for {
// Compute the delta of local node's infostore to send with request.
g.mu.Lock()
delta := g.is.delta(c.peerID, localMaxSeq)
nodeID := g.is.NodeID // needs to be accessed with the lock held
g.mu.Unlock()
var deltaBytes []byte
if delta != nil {
localMaxSeq = delta.MaxSeq
var buf bytes.Buffer
if err := gob.NewEncoder(&buf).Encode(delta); err != nil {
return util.Errorf("infostore could not be encoded: %s", err)
}
deltaBytes = buf.Bytes()
}
// Send gossip with timeout.
args := &proto.GossipRequest{
NodeID: nodeID,
Addr: *proto.FromNetAddr(g.is.NodeAddr),
LAddr: *proto.FromNetAddr(c.rpcClient.LocalAddr()),
MaxSeq: remoteMaxSeq,
Delta: deltaBytes,
}
reply := &proto.GossipResponse{}
gossipCall := c.rpcClient.Go("Gossip.Gossip", args, reply, nil)
select {
case <-gossipCall.Done:
if gossipCall.Error != nil {
return gossipCall.Error
}
case <-c.rpcClient.Closed:
return util.Error("client closed")
case <-c.closer:
return nil
case <-stopper.ShouldStop():
return nil
case <-time.After(g.interval * 10):
return util.Errorf("timeout after: %s", g.interval*10)
}
// Handle remote forwarding.
if reply.Alternate != nil {
var err error
if c.forwardAddr, err = reply.Alternate.NetAddr(); err != nil {
return util.Errorf("unable to resolve alternate address: %s: %s", reply.Alternate, err)
}
return util.Errorf("received forward from %s to %s", c.addr, reply.Alternate)
}
// Combine remote node's infostore delta with ours.
now := time.Now().UnixNano()
if reply.Delta != nil {
delta := &infoStore{}
if err := gob.NewDecoder(bytes.NewBuffer(reply.Delta)).Decode(delta); err != nil {
return util.Errorf("infostore could not be decoded: %s", err)
}
if delta.infoCount() > 0 {
if log.V(1) {
log.Infof("gossip: received %s", delta)
} else {
log.Infof("gossip: received %d info(s) from %s", delta.infoCount(), c.addr)
}
}
g.mu.Lock()
c.peerID = delta.NodeID
g.outgoing.addNode(c.peerID)
freshCount := g.is.combine(delta)
if freshCount > 0 {
c.lastFresh = now
}
remoteMaxSeq = delta.MaxSeq
// If we have the sentinel gossip, we're considered connected.
g.checkHasConnected()
g.mu.Unlock()
}
// Check whether this outgoing client is duplicating work already
// being done by an incoming client. To avoid mutual shutdown, we
// only shutdown our client if our node ID is less than the peer's.
if g.hasIncoming(c.peerID) && nodeID < c.peerID {
return util.Errorf("stopping outgoing client %d @ %s; already have incoming", c.peerID, c.addr)
}
// Check whether peer node is too boring--disconnect if yes.
if nodeID != c.peerID && (now-c.lastFresh) > int64(maxWaitForNewGossip) {
return util.Errorf("peer is too boring")
}
}
}
