1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
15 "git.arvados.org/arvados.git/lib/cloud"
16 "git.arvados.org/arvados.git/sdk/go/arvados"
17 "git.arvados.org/arvados.git/sdk/go/stats"
18 "github.com/sirupsen/logrus"
23 maxPingFailTime = 10 * time.Minute
26 // State indicates whether a worker is available to do work, and (if
27 // not) whether/when it is expected to become ready.
31 StateUnknown State = iota // might be running a container already
32 StateBooting // instance is booting
33 StateIdle // instance booted, no containers are running
34 StateRunning // instance is running one or more containers
35 StateShutdown // worker has stopped monitoring the instance
38 var stateString = map[State]string{
39 StateUnknown: "unknown",
40 StateBooting: "booting",
42 StateRunning: "running",
43 StateShutdown: "shutdown",
46 // String implements fmt.Stringer.
47 func (s State) String() string {
51 // MarshalText implements encoding.TextMarshaler so a JSON encoding of
52 // map[State]anything uses the state's string representation.
53 func (s State) MarshalText() ([]byte, error) {
54 return []byte(stateString[s]), nil
57 // IdleBehavior indicates the behavior desired when a node becomes idle.
58 type IdleBehavior string
61 IdleBehaviorRun IdleBehavior = "run" // run containers, or shutdown on idle timeout
62 IdleBehaviorHold IdleBehavior = "hold" // don't shutdown or run more containers
63 IdleBehaviorDrain IdleBehavior = "drain" // shutdown immediately when idle
66 var validIdleBehavior = map[IdleBehavior]bool{
67 IdleBehaviorRun: true,
68 IdleBehaviorHold: true,
69 IdleBehaviorDrain: true,
73 logger logrus.FieldLogger
77 mtx sync.Locker // must be wp's Locker.
79 idleBehavior IdleBehavior
80 instance cloud.Instance
81 instType arvados.InstanceType
90 running map[string]*remoteRunner // remember to update state idle<->running when this changes
91 starting map[string]*remoteRunner // remember to update state idle<->running when this changes
95 func (wkr *worker) onUnkillable(uuid string) {
97 defer wkr.mtx.Unlock()
98 logger := wkr.logger.WithField("ContainerUUID", uuid)
99 if wkr.idleBehavior == IdleBehaviorHold {
100 logger.Warn("unkillable container, but worker has IdleBehavior=Hold")
103 logger.Warn("unkillable container, draining worker")
104 wkr.setIdleBehavior(IdleBehaviorDrain)
107 func (wkr *worker) onKilled(uuid string) {
109 defer wkr.mtx.Unlock()
110 wkr.closeRunner(uuid)
114 // caller must have lock.
115 func (wkr *worker) setIdleBehavior(idleBehavior IdleBehavior) {
116 wkr.logger.WithField("IdleBehavior", idleBehavior).Info("set idle behavior")
117 wkr.idleBehavior = idleBehavior
122 // caller must have lock.
123 func (wkr *worker) startContainer(ctr arvados.Container) {
124 logger := wkr.logger.WithFields(logrus.Fields{
125 "ContainerUUID": ctr.UUID,
126 "Priority": ctr.Priority,
128 logger.Debug("starting container")
129 rr := newRemoteRunner(ctr.UUID, wkr)
130 wkr.starting[ctr.UUID] = rr
131 if wkr.state != StateRunning {
132 wkr.state = StateRunning
138 defer wkr.mtx.Unlock()
142 delete(wkr.starting, ctr.UUID)
143 wkr.running[ctr.UUID] = rr
144 wkr.lastUUID = ctr.UUID
148 // ProbeAndUpdate conducts appropriate boot/running probes (if any)
149 // for the worker's curent state. If a previous probe is still
150 // running, it does nothing.
152 // It should be called in a new goroutine.
153 func (wkr *worker) ProbeAndUpdate() {
155 case wkr.probing <- struct{}{}:
159 wkr.logger.Debug("still waiting for last probe to finish")
163 // probeAndUpdate calls probeBooted and/or probeRunning if needed, and
164 // updates state accordingly.
166 // In StateUnknown: Call both probeBooted and probeRunning.
167 // In StateBooting: Call probeBooted; if successful, call probeRunning.
168 // In StateRunning: Call probeRunning.
169 // In StateIdle: Call probeRunning.
170 // In StateShutdown: Do nothing.
172 // If both probes succeed, wkr.state changes to
173 // StateIdle/StateRunning.
175 // If probeRunning succeeds, wkr.running is updated. (This means
176 // wkr.running might be non-empty even in StateUnknown, if the boot
179 // probeAndUpdate should be called in a new goroutine.
180 func (wkr *worker) probeAndUpdate() {
182 updated := wkr.updated
183 initialState := wkr.state
190 stderr []byte // from probeBooted
193 switch initialState {
196 case StateIdle, StateRunning:
198 case StateUnknown, StateBooting:
200 panic(fmt.Sprintf("unknown state %s", initialState))
203 probeStart := time.Now()
204 logger := wkr.logger.WithField("ProbeStart", probeStart)
207 booted, stderr = wkr.probeBooted()
209 // Pretend this probe succeeded if another
210 // concurrent attempt succeeded.
212 booted = wkr.state == StateRunning || wkr.state == StateIdle
216 logger.Info("instance booted; will try probeRunning")
219 reportedBroken := false
220 if booted || wkr.state == StateUnknown {
221 ctrUUIDs, reportedBroken, ok = wkr.probeRunning()
224 defer wkr.mtx.Unlock()
225 if reportedBroken && wkr.idleBehavior == IdleBehaviorRun {
226 logger.Info("probe reported broken instance")
227 // FIXME add prometheus tick
228 wkr.setIdleBehavior(IdleBehaviorDrain)
230 if !ok || (!booted && len(ctrUUIDs) == 0 && len(wkr.running) == 0) {
231 if wkr.state == StateShutdown && wkr.updated.After(updated) {
232 // Skip the logging noise if shutdown was
233 // initiated during probe.
236 // Using the start time of the probe as the timeout
237 // threshold ensures we always initiate at least one
238 // probe attempt after the boot/probe timeout expires
239 // (otherwise, a slow probe failure could cause us to
240 // shutdown an instance even though it did in fact
241 // boot/recover before the timeout expired).
242 dur := probeStart.Sub(wkr.probed)
243 if wkr.shutdownIfBroken(dur) {
244 // stderr from failed run-probes will have
245 // been logged already, but boot-probe
246 // failures are normal so they are logged only
247 // at Debug level. This is our chance to log
248 // some evidence about why the node never
249 // booted, even in non-debug mode.
251 logger.WithFields(logrus.Fields{
253 "stderr": string(stderr),
254 }).Info("boot failed")
260 updateTime := time.Now()
261 wkr.probed = updateTime
263 if updated != wkr.updated {
264 // Worker was updated after the probe began, so
265 // wkr.running might have a container UUID that was
266 // not yet running when ctrUUIDs was generated. Leave
267 // wkr.running alone and wait for the next probe to
268 // catch up on any changes.
272 if len(ctrUUIDs) > 0 {
273 wkr.busy = updateTime
274 wkr.lastUUID = ctrUUIDs[0]
275 } else if len(wkr.running) > 0 {
276 // Actual last-busy time was sometime between wkr.busy
277 // and now. Now is the earliest opportunity to take
278 // advantage of the non-busy state, though.
279 wkr.busy = updateTime
282 changed := wkr.updateRunning(ctrUUIDs)
284 // Update state if this was the first successful boot-probe.
285 if booted && (wkr.state == StateUnknown || wkr.state == StateBooting) {
286 // Note: this will change again below if
287 // len(wkr.starting)+len(wkr.running) > 0.
288 wkr.state = StateIdle
292 // If wkr.state and wkr.running aren't changing then there's
293 // no need to log anything, notify the scheduler, move state
294 // back and forth between idle/running, etc.
299 // Log whenever a run-probe reveals crunch-run processes
300 // appearing/disappearing before boot-probe succeeds.
301 if wkr.state == StateUnknown && changed {
302 logger.WithFields(logrus.Fields{
303 "RunningContainers": len(wkr.running),
305 }).Info("crunch-run probe succeeded, but boot probe is still failing")
308 if wkr.state == StateIdle && len(wkr.starting)+len(wkr.running) > 0 {
309 wkr.state = StateRunning
310 } else if wkr.state == StateRunning && len(wkr.starting)+len(wkr.running) == 0 {
311 wkr.state = StateIdle
313 wkr.updated = updateTime
314 if booted && (initialState == StateUnknown || initialState == StateBooting) {
315 logger.WithFields(logrus.Fields{
316 "RunningContainers": len(wkr.running),
318 }).Info("probes succeeded, instance is in service")
323 func (wkr *worker) probeRunning() (running []string, reportsBroken, ok bool) {
324 cmd := wkr.wp.runnerCmd + " --list"
325 if u := wkr.instance.RemoteUser(); u != "root" {
328 stdout, stderr, err := wkr.executor.Execute(nil, cmd, nil)
330 wkr.logger.WithFields(logrus.Fields{
332 "stdout": string(stdout),
333 "stderr": string(stderr),
334 }).WithError(err).Warn("probe failed")
338 for _, s := range strings.Split(string(stdout), "\n") {
342 running = append(running, s)
348 func (wkr *worker) probeBooted() (ok bool, stderr []byte) {
349 cmd := wkr.wp.bootProbeCommand
353 stdout, stderr, err := wkr.executor.Execute(nil, cmd, nil)
354 logger := wkr.logger.WithFields(logrus.Fields{
356 "stdout": string(stdout),
357 "stderr": string(stderr),
360 logger.WithError(err).Debug("boot probe failed")
363 logger.Info("boot probe succeeded")
364 if err = wkr.wp.loadRunnerData(); err != nil {
365 wkr.logger.WithError(err).Warn("cannot boot worker: error loading runner binary")
367 } else if len(wkr.wp.runnerData) == 0 {
368 // Assume crunch-run is already installed
369 } else if _, stderr2, err := wkr.copyRunnerData(); err != nil {
370 wkr.logger.WithError(err).WithField("stderr", string(stderr2)).Warn("error copying runner binary")
371 return false, stderr2
373 stderr = append(stderr, stderr2...)
378 func (wkr *worker) copyRunnerData() (stdout, stderr []byte, err error) {
379 hash := fmt.Sprintf("%x", wkr.wp.runnerMD5)
380 dstdir, _ := filepath.Split(wkr.wp.runnerCmd)
381 logger := wkr.logger.WithFields(logrus.Fields{
383 "path": wkr.wp.runnerCmd,
386 stdout, stderr, err = wkr.executor.Execute(nil, `md5sum `+wkr.wp.runnerCmd, nil)
387 if err == nil && len(stderr) == 0 && bytes.Equal(stdout, []byte(hash+" "+wkr.wp.runnerCmd+"\n")) {
388 logger.Info("runner binary already exists on worker, with correct hash")
392 // Note touch+chmod come before writing data, to avoid the
393 // possibility of md5 being correct while file mode is
395 cmd := `set -e; dstdir="` + dstdir + `"; dstfile="` + wkr.wp.runnerCmd + `"; mkdir -p "$dstdir"; touch "$dstfile"; chmod 0755 "$dstdir" "$dstfile"; cat >"$dstfile"`
396 if wkr.instance.RemoteUser() != "root" {
397 cmd = `sudo sh -c '` + strings.Replace(cmd, "'", "'\\''", -1) + `'`
399 logger.WithField("cmd", cmd).Info("installing runner binary on worker")
400 stdout, stderr, err = wkr.executor.Execute(nil, cmd, bytes.NewReader(wkr.wp.runnerData))
404 // caller must have lock.
405 func (wkr *worker) shutdownIfBroken(dur time.Duration) bool {
406 if wkr.idleBehavior == IdleBehaviorHold {
410 label, threshold := "", wkr.wp.timeoutProbe
411 if wkr.state == StateUnknown || wkr.state == StateBooting {
412 label, threshold = "new ", wkr.wp.timeoutBooting
417 wkr.logger.WithFields(logrus.Fields{
421 }).Warnf("%sinstance unresponsive, shutting down", label)
426 // Returns true if the instance is eligible for shutdown: either it's
427 // been idle too long, or idleBehavior=Drain and nothing is running.
429 // caller must have lock.
430 func (wkr *worker) eligibleForShutdown() bool {
431 if wkr.idleBehavior == IdleBehaviorHold {
434 draining := wkr.idleBehavior == IdleBehaviorDrain
439 return draining || time.Since(wkr.busy) >= wkr.wp.timeoutIdle
444 for _, rr := range wkr.running {
449 for _, rr := range wkr.starting {
454 // draining, and all remaining runners are just trying
455 // to force-kill their crunch-run procs
462 // caller must have lock.
463 func (wkr *worker) shutdownIfIdle() bool {
464 if !wkr.eligibleForShutdown() {
467 wkr.logger.WithFields(logrus.Fields{
469 "IdleDuration": stats.Duration(time.Since(wkr.busy)),
470 "IdleBehavior": wkr.idleBehavior,
471 }).Info("shutdown worker")
476 // caller must have lock.
477 func (wkr *worker) shutdown() {
481 wkr.state = StateShutdown
484 err := wkr.instance.Destroy()
486 wkr.logger.WithError(err).Warn("shutdown failed")
492 // Save worker tags to cloud provider metadata, if they don't already
493 // match. Caller must have lock.
494 func (wkr *worker) saveTags() {
495 instance := wkr.instance
496 tags := instance.Tags()
497 update := cloud.InstanceTags{
498 wkr.wp.tagKeyPrefix + tagKeyInstanceType: wkr.instType.Name,
499 wkr.wp.tagKeyPrefix + tagKeyIdleBehavior: string(wkr.idleBehavior),
502 for k, v := range update {
510 err := instance.SetTags(tags)
512 wkr.wp.logger.WithField("Instance", instance.ID()).WithError(err).Warnf("error updating tags")
518 func (wkr *worker) Close() {
519 // This might take time, so do it after unlocking mtx.
520 defer wkr.executor.Close()
523 defer wkr.mtx.Unlock()
524 for uuid, rr := range wkr.running {
525 wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process abandoned")
528 for uuid, rr := range wkr.starting {
529 wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process abandoned")
534 // Add/remove entries in wkr.running to match ctrUUIDs returned by a
535 // probe. Returns true if anything was added or removed.
537 // Caller must have lock.
538 func (wkr *worker) updateRunning(ctrUUIDs []string) (changed bool) {
539 alive := map[string]bool{}
540 for _, uuid := range ctrUUIDs {
542 if _, ok := wkr.running[uuid]; ok {
544 } else if rr, ok := wkr.starting[uuid]; ok {
545 wkr.running[uuid] = rr
546 delete(wkr.starting, uuid)
549 // We didn't start it -- it must have been
550 // started by a previous dispatcher process.
551 wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process detected")
552 wkr.running[uuid] = newRemoteRunner(uuid, wkr)
556 for uuid := range wkr.running {
558 wkr.closeRunner(uuid)
565 // caller must have lock.
566 func (wkr *worker) closeRunner(uuid string) {
567 rr := wkr.running[uuid]
571 wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process ended")
572 delete(wkr.running, uuid)
577 wkr.wp.exited[uuid] = now
578 if wkr.state == StateRunning && len(wkr.running)+len(wkr.starting) == 0 {
579 wkr.state = StateIdle