X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/a5fe34a8cd05c4e55deaec599347f65e6d662d22..be8ed479042df4fdefe1fd18c1e2e984e1c99bc0:/lib/dispatchcloud/worker/pool.go diff --git a/lib/dispatchcloud/worker/pool.go b/lib/dispatchcloud/worker/pool.go index 1e759e38f3..1665a1e43d 100644 --- a/lib/dispatchcloud/worker/pool.go +++ b/lib/dispatchcloud/worker/pool.go @@ -5,7 +5,7 @@ package worker import ( - "bytes" + "errors" "io" "sort" "strings" @@ -14,30 +14,31 @@ import ( "git.curoverse.com/arvados.git/lib/cloud" "git.curoverse.com/arvados.git/sdk/go/arvados" - "github.com/Sirupsen/logrus" "github.com/prometheus/client_golang/prometheus" + "github.com/sirupsen/logrus" ) const ( tagKeyInstanceType = "InstanceType" - tagKeyHold = "Hold" + tagKeyIdleBehavior = "IdleBehavior" ) // An InstanceView shows a worker's current state and recent activity. type InstanceView struct { - Instance string - Price float64 - ArvadosInstanceType string - ProviderInstanceType string - LastContainerUUID string - Unallocated time.Time - WorkerState string + Instance cloud.InstanceID `json:"instance"` + Price float64 `json:"price"` + ArvadosInstanceType string `json:"arvados_instance_type"` + ProviderInstanceType string `json:"provider_instance_type"` + LastContainerUUID string `json:"last_container_uuid"` + LastBusy time.Time `json:"last_busy"` + WorkerState string `json:"worker_state"` + IdleBehavior IdleBehavior `json:"idle_behavior"` } // An Executor executes shell commands on a remote host. type Executor interface { // Run cmd on the current target. - Execute(cmd string, stdin io.Reader) (stdout, stderr []byte, err error) + Execute(env map[string]string, cmd string, stdin io.Reader) (stdout, stderr []byte, err error) // Use the given target for subsequent operations. The new // target is the same host as the previous target, but it @@ -61,6 +62,14 @@ const ( defaultTimeoutIdle = time.Minute defaultTimeoutBooting = time.Minute * 10 defaultTimeoutProbe = time.Minute * 10 + defaultTimeoutShutdown = time.Second * 10 + + // Time after a quota error to try again anyway, even if no + // instances have been shutdown. + quotaErrorTTL = time.Minute + + // Time between "X failed because rate limiting" messages + logRateLimitErrorInterval = time.Second * 10 ) func duration(conf arvados.Duration, def time.Duration) time.Duration { @@ -75,10 +84,11 @@ func duration(conf arvados.Duration, def time.Duration) time.Duration { // // New instances are configured and set up according to the given // cluster configuration. -func NewPool(logger logrus.FieldLogger, reg *prometheus.Registry, instanceSet cloud.InstanceSet, newExecutor func(cloud.Instance) Executor, cluster *arvados.Cluster) *Pool { +func NewPool(logger logrus.FieldLogger, arvClient *arvados.Client, reg *prometheus.Registry, instanceSet cloud.InstanceSet, newExecutor func(cloud.Instance) Executor, cluster *arvados.Cluster) *Pool { wp := &Pool{ logger: logger, - instanceSet: instanceSet, + arvClient: arvClient, + instanceSet: &throttledInstanceSet{InstanceSet: instanceSet}, newExecutor: newExecutor, bootProbeCommand: cluster.CloudVMs.BootProbeCommand, imageID: cloud.ImageID(cluster.CloudVMs.ImageID), @@ -89,6 +99,8 @@ func NewPool(logger logrus.FieldLogger, reg *prometheus.Registry, instanceSet cl timeoutIdle: duration(cluster.CloudVMs.TimeoutIdle, defaultTimeoutIdle), timeoutBooting: duration(cluster.CloudVMs.TimeoutBooting, defaultTimeoutBooting), timeoutProbe: duration(cluster.CloudVMs.TimeoutProbe, defaultTimeoutProbe), + timeoutShutdown: duration(cluster.CloudVMs.TimeoutShutdown, defaultTimeoutShutdown), + stop: make(chan bool), } wp.registerMetrics(reg) go func() { @@ -105,7 +117,8 @@ func NewPool(logger logrus.FieldLogger, reg *prometheus.Registry, instanceSet cl type Pool struct { // configuration logger logrus.FieldLogger - instanceSet cloud.InstanceSet + arvClient *arvados.Client + instanceSet *throttledInstanceSet newExecutor func(cloud.Instance) Executor bootProbeCommand string imageID cloud.ImageID @@ -116,19 +129,25 @@ type Pool struct { timeoutIdle time.Duration timeoutBooting time.Duration timeoutProbe time.Duration + timeoutShutdown time.Duration // private state subscribers map[<-chan struct{}]chan<- struct{} - creating map[arvados.InstanceType]int // goroutines waiting for (InstanceSet)Create to return + creating map[arvados.InstanceType][]time.Time // start times of unfinished (InstanceSet)Create calls workers map[cloud.InstanceID]*worker loaded bool // loaded list of instances from InstanceSet at least once exited map[string]time.Time // containers whose crunch-run proc has exited, but KillContainer has not been called atQuotaUntil time.Time + atQuotaErr cloud.QuotaError stop chan bool mtx sync.RWMutex setupOnce sync.Once + throttleCreate throttle + throttleInstances throttle + mInstances prometheus.Gauge + mInstancesPrice prometheus.Gauge mContainersRunning prometheus.Gauge mVCPUs prometheus.Gauge mVCPUsInuse prometheus.Gauge @@ -136,33 +155,21 @@ type Pool struct { mMemoryInuse prometheus.Gauge } -type worker struct { - state State - instance cloud.Instance - executor Executor - instType arvados.InstanceType - vcpus int64 - memory int64 - booted bool - probed time.Time - updated time.Time - busy time.Time - unallocated time.Time - lastUUID string - running map[string]struct{} - starting map[string]struct{} - probing chan struct{} -} - -// Subscribe returns a channel that becomes ready whenever a worker's -// state changes. +// Subscribe returns a buffered channel that becomes ready after any +// change to the pool's state that could have scheduling implications: +// a worker's state changes, a new worker appears, the cloud +// provider's API rate limiting period ends, etc. +// +// Additional events that occur while the channel is already ready +// will be dropped, so it is OK if the caller services the channel +// slowly. // // Example: // // ch := wp.Subscribe() // defer wp.Unsubscribe(ch) // for range ch { -// // ...try scheduling some work... +// tryScheduling(wp) // if done { // break // } @@ -185,49 +192,93 @@ func (wp *Pool) Unsubscribe(ch <-chan struct{}) { } // Unallocated returns the number of unallocated (creating + booting + -// idle + unknown) workers for each instance type. +// idle + unknown) workers for each instance type. Workers in +// hold/drain mode are not included. func (wp *Pool) Unallocated() map[arvados.InstanceType]int { wp.setupOnce.Do(wp.setup) wp.mtx.RLock() defer wp.mtx.RUnlock() - u := map[arvados.InstanceType]int{} - for it, c := range wp.creating { - u[it] = c + unalloc := map[arvados.InstanceType]int{} + creating := map[arvados.InstanceType]int{} + for it, times := range wp.creating { + creating[it] = len(times) } for _, wkr := range wp.workers { - if len(wkr.running)+len(wkr.starting) == 0 && (wkr.state == StateRunning || wkr.state == StateBooting || wkr.state == StateUnknown) { - u[wkr.instType]++ + if !(wkr.state == StateIdle || wkr.state == StateBooting || wkr.state == StateUnknown) || wkr.idleBehavior != IdleBehaviorRun { + continue } + it := wkr.instType + unalloc[it]++ + if wkr.state == StateUnknown && creating[it] > 0 && wkr.appeared.After(wp.creating[it][0]) { + // If up to N new workers appear in + // Instances() while we are waiting for N + // Create() calls to complete, we assume we're + // just seeing a race between Instances() and + // Create() responses. + // + // The other common reason why nodes have + // state==Unknown is that they appeared at + // startup, before any Create calls. They + // don't match the above timing condition, so + // we never mistakenly attribute them to + // pending Create calls. + creating[it]-- + } + } + for it, c := range creating { + unalloc[it] += c } - return u + return unalloc } // Create a new instance with the given type, and add it to the worker // pool. The worker is added immediately; instance creation runs in // the background. -func (wp *Pool) Create(it arvados.InstanceType) error { +// +// Create returns false if a pre-existing error state prevents it from +// even attempting to create a new instance. Those errors are logged +// by the Pool, so the caller does not need to log anything in such +// cases. +func (wp *Pool) Create(it arvados.InstanceType) bool { logger := wp.logger.WithField("InstanceType", it.Name) wp.setupOnce.Do(wp.setup) wp.mtx.Lock() defer wp.mtx.Unlock() - tags := cloud.InstanceTags{tagKeyInstanceType: it.Name} - wp.creating[it]++ + if time.Now().Before(wp.atQuotaUntil) || wp.throttleCreate.Error() != nil { + return false + } + tags := cloud.InstanceTags{ + tagKeyInstanceType: it.Name, + tagKeyIdleBehavior: string(IdleBehaviorRun), + } + now := time.Now() + wp.creating[it] = append(wp.creating[it], now) go func() { defer wp.notify() inst, err := wp.instanceSet.Create(it, wp.imageID, tags, nil) wp.mtx.Lock() defer wp.mtx.Unlock() - wp.creating[it]-- - if err, ok := err.(cloud.QuotaError); ok && err.IsQuotaError() { - wp.atQuotaUntil = time.Now().Add(time.Minute) + // Remove our timestamp marker from wp.creating + for i, t := range wp.creating[it] { + if t == now { + copy(wp.creating[it][i:], wp.creating[it][i+1:]) + wp.creating[it] = wp.creating[it][:len(wp.creating[it])-1] + break + } } if err != nil { + if err, ok := err.(cloud.QuotaError); ok && err.IsQuotaError() { + wp.atQuotaErr = err + wp.atQuotaUntil = time.Now().Add(quotaErrorTTL) + time.AfterFunc(quotaErrorTTL, wp.notify) + } logger.WithError(err).Error("create failed") + wp.instanceSet.throttleCreate.CheckRateLimitError(err, wp.logger, "create instance", wp.notify) return } wp.updateWorker(inst, it, StateBooting) }() - return nil + return true } // AtQuota returns true if Create is not expected to work at the @@ -238,44 +289,84 @@ func (wp *Pool) AtQuota() bool { return time.Now().Before(wp.atQuotaUntil) } +// SetIdleBehavior determines how the indicated instance will behave +// when it has no containers running. +func (wp *Pool) SetIdleBehavior(id cloud.InstanceID, idleBehavior IdleBehavior) error { + wp.mtx.Lock() + defer wp.mtx.Unlock() + wkr, ok := wp.workers[id] + if !ok { + return errors.New("requested instance does not exist") + } + wkr.idleBehavior = idleBehavior + wkr.saveTags() + wkr.shutdownIfIdle() + return nil +} + // Add or update worker attached to the given instance. Use -// initialState if a new worker is created. Caller must have lock. +// initialState if a new worker is created. // -// Returns true when a new worker is created. -func (wp *Pool) updateWorker(inst cloud.Instance, it arvados.InstanceType, initialState State) bool { +// The second return value is true if a new worker is created. +// +// Caller must have lock. +func (wp *Pool) updateWorker(inst cloud.Instance, it arvados.InstanceType, initialState State) (*worker, bool) { id := inst.ID() - if wp.workers[id] != nil { - wp.workers[id].executor.SetTarget(inst) - wp.workers[id].instance = inst - wp.workers[id].updated = time.Now() - if initialState == StateBooting && wp.workers[id].state == StateUnknown { - wp.workers[id].state = StateBooting + if wkr := wp.workers[id]; wkr != nil { + wkr.executor.SetTarget(inst) + wkr.instance = inst + wkr.updated = time.Now() + if initialState == StateBooting && wkr.state == StateUnknown { + wkr.state = StateBooting } - return false + wkr.saveTags() + return wkr, false } - if initialState == StateUnknown && inst.Tags()[tagKeyHold] != "" { - initialState = StateHold + + // If an instance has a valid IdleBehavior tag when it first + // appears, initialize the new worker accordingly (this is how + // we restore IdleBehavior that was set by a prior dispatch + // process); otherwise, default to "run". After this, + // wkr.idleBehavior is the source of truth, and will only be + // changed via SetIdleBehavior(). + idleBehavior := IdleBehavior(inst.Tags()[tagKeyIdleBehavior]) + if !validIdleBehavior[idleBehavior] { + idleBehavior = IdleBehaviorRun } - wp.logger.WithFields(logrus.Fields{ + + logger := wp.logger.WithFields(logrus.Fields{ "InstanceType": it.Name, "Instance": inst, + }) + logger.WithFields(logrus.Fields{ "State": initialState, + "IdleBehavior": idleBehavior, }).Infof("instance appeared in cloud") now := time.Now() - wp.workers[id] = &worker{ - executor: wp.newExecutor(inst), - state: initialState, - instance: inst, - instType: it, - probed: now, - busy: now, - updated: now, - unallocated: now, - running: make(map[string]struct{}), - starting: make(map[string]struct{}), - probing: make(chan struct{}, 1), - } - return true + wkr := &worker{ + mtx: &wp.mtx, + wp: wp, + logger: logger, + executor: wp.newExecutor(inst), + state: initialState, + idleBehavior: idleBehavior, + instance: inst, + instType: it, + appeared: now, + probed: now, + busy: now, + updated: now, + running: make(map[string]struct{}), + starting: make(map[string]struct{}), + probing: make(chan struct{}, 1), + } + wp.workers[id] = wkr + return wkr, true +} + +// caller must have lock. +func (wp *Pool) notifyExited(uuid string, t time.Time) { + wp.exited[uuid] = t } // Shutdown shuts down a worker with the given type, or returns false @@ -286,45 +377,22 @@ func (wp *Pool) Shutdown(it arvados.InstanceType) bool { defer wp.mtx.Unlock() logger := wp.logger.WithField("InstanceType", it.Name) logger.Info("shutdown requested") - for _, tryState := range []State{StateBooting, StateRunning} { + for _, tryState := range []State{StateBooting, StateIdle} { // TODO: shutdown the worker with the longest idle - // time (Running) or the earliest create time - // (Booting) + // time (Idle) or the earliest create time (Booting) for _, wkr := range wp.workers { - if wkr.state != tryState || len(wkr.running)+len(wkr.starting) > 0 { - continue - } - if wkr.instType != it { - continue + if wkr.idleBehavior != IdleBehaviorHold && wkr.state == tryState && wkr.instType == it { + logger.WithField("Instance", wkr.instance).Info("shutting down") + wkr.shutdown() + return true } - logger = logger.WithField("Instance", wkr.instance) - logger.Info("shutting down") - wp.shutdown(wkr, logger) - return true } } return false } -// caller must have lock -func (wp *Pool) shutdown(wkr *worker, logger logrus.FieldLogger) { - wkr.updated = time.Now() - wkr.state = StateShutdown - go func() { - err := wkr.instance.Destroy() - if err != nil { - logger.WithError(err).WithField("Instance", wkr.instance).Warn("shutdown failed") - return - } - wp.mtx.Lock() - wp.atQuotaUntil = time.Now() - wp.mtx.Unlock() - wp.notify() - }() -} - -// Workers returns the current number of workers in each state. -func (wp *Pool) Workers() map[State]int { +// CountWorkers returns the current number of workers in each state. +func (wp *Pool) CountWorkers() map[State]int { wp.setupOnce.Do(wp.setup) wp.mtx.Lock() defer wp.mtx.Unlock() @@ -336,6 +404,12 @@ func (wp *Pool) Workers() map[State]int { } // Running returns the container UUIDs being prepared/run on workers. +// +// In the returned map, the time value indicates when the Pool +// observed that the container process had exited. A container that +// has not yet exited has a zero time value. The caller should use +// KillContainer() to garbage-collect the entries for exited +// containers. func (wp *Pool) Running() map[string]time.Time { wp.setupOnce.Do(wp.setup) wp.mtx.Lock() @@ -358,17 +432,12 @@ func (wp *Pool) Running() map[string]time.Time { // StartContainer starts a container on an idle worker immediately if // possible, otherwise returns false. func (wp *Pool) StartContainer(it arvados.InstanceType, ctr arvados.Container) bool { - logger := wp.logger.WithFields(logrus.Fields{ - "InstanceType": it.Name, - "ContainerUUID": ctr.UUID, - "Priority": ctr.Priority, - }) wp.setupOnce.Do(wp.setup) wp.mtx.Lock() defer wp.mtx.Unlock() var wkr *worker for _, w := range wp.workers { - if w.instType == it && w.state == StateRunning && len(w.running)+len(w.starting) == 0 { + if w.instType == it && w.state == StateIdle { if wkr == nil || w.busy.After(wkr.busy) { wkr = w } @@ -377,33 +446,7 @@ func (wp *Pool) StartContainer(it arvados.InstanceType, ctr arvados.Container) b if wkr == nil { return false } - logger = logger.WithField("Instance", wkr.instance) - logger.Debug("starting container") - wkr.starting[ctr.UUID] = struct{}{} - go func() { - stdout, stderr, err := wkr.executor.Execute("crunch-run --detach '"+ctr.UUID+"'", nil) - wp.mtx.Lock() - defer wp.mtx.Unlock() - now := time.Now() - wkr.updated = now - wkr.busy = now - delete(wkr.starting, ctr.UUID) - wkr.running[ctr.UUID] = struct{}{} - wkr.lastUUID = ctr.UUID - if err != nil { - logger.WithField("stdout", string(stdout)). - WithField("stderr", string(stderr)). - WithError(err). - Error("error starting crunch-run process") - // Leave uuid in wkr.running, though: it's - // possible the error was just a communication - // failure and the process was in fact - // started. Wait for next probe to find out. - return - } - logger.Info("crunch-run process started") - wkr.lastUUID = ctr.UUID - }() + wkr.startContainer(ctr) return true } @@ -435,7 +478,7 @@ func (wp *Pool) kill(wkr *worker, uuid string) { "Instance": wkr.instance, }) logger.Debug("killing process") - stdout, stderr, err := wkr.executor.Execute("crunch-run --kill "+uuid, nil) + stdout, stderr, err := wkr.executor.Execute(nil, "crunch-run --kill 15 "+uuid, nil) if err != nil { logger.WithFields(logrus.Fields{ "stderr": string(stderr), @@ -449,6 +492,9 @@ func (wp *Pool) kill(wkr *worker, uuid string) { defer wp.mtx.Unlock() if _, ok := wkr.running[uuid]; ok { delete(wkr.running, uuid) + if wkr.state == StateRunning && len(wkr.running)+len(wkr.starting) == 0 { + wkr.state = StateIdle + } wkr.updated = time.Now() go wp.notify() } @@ -465,6 +511,13 @@ func (wp *Pool) registerMetrics(reg *prometheus.Registry) { Help: "Number of cloud VMs including pending, booting, running, held, and shutting down.", }) reg.MustRegister(wp.mInstances) + wp.mInstancesPrice = prometheus.NewGauge(prometheus.GaugeOpts{ + Namespace: "arvados", + Subsystem: "dispatchcloud", + Name: "instances_price_total", + Help: "Sum of prices of all cloud VMs including pending, booting, running, held, and shutting down.", + }) + reg.MustRegister(wp.mInstancesPrice) wp.mContainersRunning = prometheus.NewGauge(prometheus.GaugeOpts{ Namespace: "arvados", Subsystem: "dispatchcloud", @@ -515,8 +568,10 @@ func (wp *Pool) updateMetrics() { wp.mtx.RLock() defer wp.mtx.RUnlock() + var price float64 var alloc, cpu, cpuInuse, mem, memInuse int64 for _, wkr := range wp.workers { + price += wkr.instType.Price cpu += int64(wkr.instType.VCPUs) mem += int64(wkr.instType.RAM) if len(wkr.running)+len(wkr.starting) == 0 { @@ -527,6 +582,7 @@ func (wp *Pool) updateMetrics() { memInuse += int64(wkr.instType.RAM) } wp.mInstances.Set(float64(len(wp.workers))) + wp.mInstancesPrice.Set(price) wp.mContainersRunning.Set(float64(alloc)) wp.mVCPUs.Set(float64(cpu)) wp.mMemory.Set(float64(mem)) @@ -550,7 +606,7 @@ func (wp *Pool) runProbes() { workers = workers[:0] wp.mtx.Lock() for id, wkr := range wp.workers { - if wkr.state == StateShutdown || wp.shutdownIfIdle(wkr) { + if wkr.state == StateShutdown || wkr.shutdownIfIdle() { continue } workers = append(workers, id) @@ -561,20 +617,12 @@ func (wp *Pool) runProbes() { wp.mtx.Lock() wkr, ok := wp.workers[id] wp.mtx.Unlock() - if !ok || wkr.state == StateShutdown { - // Deleted/shutdown while we - // were probing others + if !ok { + // Deleted while we were probing + // others continue } - select { - case wkr.probing <- struct{}{}: - go func() { - wp.probeAndUpdate(wkr) - <-wkr.probing - }() - default: - wp.logger.WithField("Instance", wkr.instance).Debug("still waiting for last probe to finish") - } + go wkr.ProbeAndUpdate() select { case <-wp.stop: return @@ -603,45 +651,6 @@ func (wp *Pool) runSync() { } } -// caller must have lock. -func (wp *Pool) shutdownIfBroken(wkr *worker, dur time.Duration) { - if wkr.state == StateHold { - return - } - label, threshold := "", wp.timeoutProbe - if wkr.state == StateBooting { - label, threshold = "new ", wp.timeoutBooting - } - if dur < threshold { - return - } - wp.logger.WithFields(logrus.Fields{ - "Instance": wkr.instance, - "Duration": dur, - "Since": wkr.probed, - "State": wkr.state, - }).Warnf("%sinstance unresponsive, shutting down", label) - wp.shutdown(wkr, wp.logger) -} - -// caller must have lock. -func (wp *Pool) shutdownIfIdle(wkr *worker) bool { - if len(wkr.running)+len(wkr.starting) > 0 || wkr.state != StateRunning { - return false - } - age := time.Since(wkr.unallocated) - if age < wp.timeoutIdle { - return false - } - logger := wp.logger.WithFields(logrus.Fields{ - "Age": age, - "Instance": wkr.instance, - }) - logger.Info("shutdown idle worker") - wp.shutdown(wkr, logger) - return true -} - // Stop synchronizing with the InstanceSet. func (wp *Pool) Stop() { wp.setupOnce.Do(wp.setup) @@ -656,24 +665,25 @@ func (wp *Pool) Instances() []InstanceView { wp.mtx.Lock() for _, w := range wp.workers { r = append(r, InstanceView{ - Instance: w.instance.String(), + Instance: w.instance.ID(), Price: w.instType.Price, ArvadosInstanceType: w.instType.Name, ProviderInstanceType: w.instType.ProviderType, LastContainerUUID: w.lastUUID, - Unallocated: w.unallocated, + LastBusy: w.busy, WorkerState: w.state.String(), + IdleBehavior: w.idleBehavior, }) } wp.mtx.Unlock() sort.Slice(r, func(i, j int) bool { - return strings.Compare(r[i].Instance, r[j].Instance) < 0 + return strings.Compare(string(r[i].Instance), string(r[j].Instance)) < 0 }) return r } func (wp *Pool) setup() { - wp.creating = map[arvados.InstanceType]int{} + wp.creating = map[arvados.InstanceType][]time.Time{} wp.exited = map[string]time.Time{} wp.workers = map[cloud.InstanceID]*worker{} wp.subscribers = map[<-chan struct{}]chan<- struct{}{} @@ -692,10 +702,14 @@ func (wp *Pool) notify() { func (wp *Pool) getInstancesAndSync() error { wp.setupOnce.Do(wp.setup) + if err := wp.instanceSet.throttleInstances.Error(); err != nil { + return err + } wp.logger.Debug("getting instance list") threshold := time.Now() instances, err := wp.instanceSet.Instances(cloud.InstanceTags{}) if err != nil { + wp.instanceSet.throttleInstances.CheckRateLimitError(err, wp.logger, "list instances", wp.notify) return err } wp.sync(threshold, instances) @@ -719,8 +733,11 @@ func (wp *Pool) sync(threshold time.Time, instances []cloud.Instance) { wp.logger.WithField("Instance", inst).Errorf("unknown InstanceType tag %q --- ignoring", itTag) continue } - if wp.updateWorker(inst, it, StateUnknown) { + if wkr, isNew := wp.updateWorker(inst, it, StateUnknown); isNew { notify = true + } else if wkr.state == StateShutdown && time.Since(wkr.destroyed) > wp.timeoutShutdown { + wp.logger.WithField("Instance", inst).Info("worker still listed after shutdown; retrying") + wkr.shutdown() } } @@ -747,139 +764,3 @@ func (wp *Pool) sync(threshold time.Time, instances []cloud.Instance) { go wp.notify() } } - -// should be called in a new goroutine -func (wp *Pool) probeAndUpdate(wkr *worker) { - logger := wp.logger.WithField("Instance", wkr.instance) - wp.mtx.Lock() - updated := wkr.updated - booted := wkr.booted - wp.mtx.Unlock() - - var ( - ctrUUIDs []string - ok bool - stderr []byte - ) - if !booted { - booted, stderr = wp.probeBooted(wkr) - wp.mtx.Lock() - if booted && !wkr.booted { - wkr.booted = booted - logger.Info("instance booted") - } else { - booted = wkr.booted - } - wp.mtx.Unlock() - } - if booted { - ctrUUIDs, ok, stderr = wp.probeRunning(wkr) - } - logger = logger.WithField("stderr", string(stderr)) - wp.mtx.Lock() - defer wp.mtx.Unlock() - if !ok { - if wkr.state == StateShutdown { - return - } - dur := time.Since(wkr.probed) - logger := logger.WithFields(logrus.Fields{ - "Duration": dur, - "State": wkr.state, - }) - if wkr.state == StateBooting { - logger.Debug("new instance not responding") - } else { - logger.Info("instance not responding") - } - wp.shutdownIfBroken(wkr, dur) - return - } - - updateTime := time.Now() - wkr.probed = updateTime - if wkr.state == StateShutdown || wkr.state == StateHold { - } else if booted { - if wkr.state != StateRunning { - wkr.state = StateRunning - go wp.notify() - } - } else { - wkr.state = StateBooting - } - - if updated != wkr.updated { - // Worker was updated after the probe began, so - // wkr.running might have a container UUID that was - // not yet running when ctrUUIDs was generated. Leave - // wkr.running alone and wait for the next probe to - // catch up on any changes. - return - } - - if len(ctrUUIDs) > 0 { - wkr.busy = updateTime - wkr.lastUUID = ctrUUIDs[0] - } else if len(wkr.running) > 0 { - wkr.unallocated = updateTime - } - running := map[string]struct{}{} - changed := false - for _, uuid := range ctrUUIDs { - running[uuid] = struct{}{} - if _, ok := wkr.running[uuid]; !ok { - changed = true - } - } - for uuid := range wkr.running { - if _, ok := running[uuid]; !ok { - logger.WithField("ContainerUUID", uuid).Info("crunch-run process ended") - wp.exited[uuid] = updateTime - changed = true - } - } - if changed { - wkr.running = running - wkr.updated = updateTime - go wp.notify() - } -} - -func (wp *Pool) probeRunning(wkr *worker) (running []string, ok bool, stderr []byte) { - cmd := "crunch-run --list" - stdout, stderr, err := wkr.executor.Execute(cmd, nil) - if err != nil { - wp.logger.WithFields(logrus.Fields{ - "Instance": wkr.instance, - "Command": cmd, - "stdout": string(stdout), - "stderr": string(stderr), - }).WithError(err).Warn("probe failed") - return nil, false, stderr - } - stdout = bytes.TrimRight(stdout, "\n") - if len(stdout) == 0 { - return nil, true, stderr - } - return strings.Split(string(stdout), "\n"), true, stderr -} - -func (wp *Pool) probeBooted(wkr *worker) (ok bool, stderr []byte) { - cmd := wp.bootProbeCommand - if cmd == "" { - cmd = "true" - } - stdout, stderr, err := wkr.executor.Execute(cmd, nil) - logger := wp.logger.WithFields(logrus.Fields{ - "Instance": wkr.instance, - "Command": cmd, - "stdout": string(stdout), - "stderr": string(stderr), - }) - if err != nil { - logger.WithError(err).Debug("boot probe failed") - return false, stderr - } - logger.Info("boot probe succeeded") - return true, stderr -}