// 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
}
}
})
}
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{
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.")
}
}
})
}
// scanLoop loops endlessly, scanning through ranges available via
// the range set, 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) {
stopper.RunWorker(func() {
start := time.Now()
stats := &storeStats{}
for {
if rs.ranges.EstimatedCount() == 0 {
// Just wait without processing any range.
if rs.waitAndProcess(start, clock, stopper, stats, nil) {
break
}
} else {
shouldStop := true
rs.ranges.Visit(func(rng *Range) bool {
shouldStop = rs.waitAndProcess(start, clock, stopper, stats, rng)
return !shouldStop
})
if shouldStop {
break
}
}
if !stopper.StartTask() {
// Exit the loop.
break
}
// We're done with the iteration.
// Store the most recent scan results in the scanner's stats.
atomic.StorePointer(&rs.stats, unsafe.Pointer(stats))
stats = &storeStats{}
if rs.scanFn != nil {
rs.scanFn()
}
// Increment iteration count.
rs.completedScan.L.Lock()
rs.count++
rs.total += time.Now().Sub(start)
rs.completedScan.Broadcast()
rs.completedScan.L.Unlock()
if log.V(6) {
log.Infof("reset range scan iteration")
}
// Reset iteration and start time.
start = time.Now()
stopper.FinishTask()
}
})
}
// 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
}
})
}
// 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
}
}
})
}
// 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)
}
}
})
}
// 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
}
}
})
}
// scanLoop loops endlessly, scanning through ranges available via
// the range set, 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) {
stopper.RunWorker(func() {
start := time.Now()
for {
if rs.ranges.EstimatedCount() == 0 {
// Just wait without processing any range.
if rs.waitAndProcess(start, clock, stopper, nil) {
break
}
} else {
shouldStop := true
rs.ranges.Visit(func(rng *Range) bool {
shouldStop = rs.waitAndProcess(start, clock, stopper, rng)
return !shouldStop
})
if shouldStop {
break
}
}
if !stopper.StartTask() {
// Exit the loop.
break
}
// Increment iteration count.
rs.completedScan.L.Lock()
rs.count++
rs.total += time.Now().Sub(start)
rs.completedScan.Broadcast()
rs.completedScan.L.Unlock()
if log.V(6) {
log.Infof("reset range scan iteration")
}
// Reset iteration and start time.
start = time.Now()
stopper.FinishTask()
}
})
}
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 dials the remote addr and commences gossip once connected.
// Upon exit, signals client is done by pushing it onto the done
// channel. If the client experienced an error, its err field will
// be set. This method starts client processing in a goroutine and
// returns immediately.
func (c *client) start(g *Gossip, done chan *client, context *rpc.Context, stopper *util.Stopper) {
stopper.RunWorker(func() {
c.rpcClient = rpc.NewClient(c.addr, nil, context)
select {
case <-c.rpcClient.Ready:
// Success!
case <-c.rpcClient.Closed:
c.err = util.Errorf("gossip client failed to connect")
done <- c
return
}
// Start gossipping and wait for disconnect or error.
c.lastFresh = time.Now().UnixNano()
c.err = c.gossip(g, stopper)
if c.err != nil {
c.rpcClient.Close()
}
done <- c
})
}
// 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
}
}
})
}
// 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
}
}
})
}
// NewServer creates a Server from a server.Context.
func NewServer(ctx *Context, stopper *util.Stopper) (*Server, error) {
if ctx == nil {
return nil, util.Error("ctx must not be null")
}
addr := ctx.Addr
_, err := net.ResolveTCPAddr("tcp", addr)
if err != nil {
return nil, util.Errorf("unable to resolve RPC address %q: %v", addr, err)
}
if ctx.Insecure {
log.Warning("running in insecure mode, this is strongly discouraged. See --insecure and --certs.")
}
// Try loading the TLS configs before anything else.
if _, err := ctx.GetServerTLSConfig(); err != nil {
return nil, err
}
if _, err := ctx.GetClientTLSConfig(); err != nil {
return nil, err
}
s := &Server{
ctx: ctx,
mux: http.NewServeMux(),
clock: hlc.NewClock(hlc.UnixNano),
stopper: stopper,
}
s.clock.SetMaxOffset(ctx.MaxOffset)
rpcContext := rpc.NewContext(&ctx.Context, s.clock, stopper)
stopper.RunWorker(func() {
rpcContext.RemoteClocks.MonitorRemoteOffsets(stopper)
})
s.rpc = rpc.NewServer(util.MakeUnresolvedAddr("tcp", addr), rpcContext)
s.stopper.AddCloser(s.rpc)
s.gossip = gossip.New(rpcContext, s.ctx.GossipInterval, s.ctx.GossipBootstrapResolvers)
ds := kv.NewDistSender(&kv.DistSenderContext{Clock: s.clock}, s.gossip)
sender := kv.NewTxnCoordSender(ds, s.clock, ctx.Linearizable, s.stopper)
if s.db, err = client.Open("//root@", client.SenderOpt(sender)); err != nil {
return nil, err
}
s.raftTransport, err = newRPCTransport(s.gossip, s.rpc, rpcContext)
if err != nil {
return nil, err
}
s.stopper.AddCloser(s.raftTransport)
s.kvDB = kv.NewDBServer(&s.ctx.Context, sender)
if s.ctx.ExperimentalRPCServer {
if err = s.kvDB.RegisterRPC(s.rpc); err != nil {
return nil, err
}
}
//
s.sqlServer = sqlserver.NewServer(s.db)
// TODO(bdarnell): make StoreConfig configurable.
nCtx := storage.StoreContext{
Clock: s.clock,
DB: s.db,
Gossip: s.gossip,
Transport: s.raftTransport,
ScanInterval: s.ctx.ScanInterval,
ScanMaxIdleTime: s.ctx.ScanMaxIdleTime,
EventFeed: &util.Feed{},
}
s.node = NewNode(nCtx)
s.admin = newAdminServer(s.db, s.stopper)
s.status = newStatusServer(s.db, s.gossip)
s.tsDB = ts.NewDB(s.db)
s.tsServer = ts.NewServer(s.tsDB)
s.stopper.AddCloser(nCtx.EventFeed)
return s, nil
}