lib/dispatchcloud/scheduler/sync.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package scheduler
   6
   7 import (
   8         "fmt"
   9
  10         "git.curoverse.com/arvados.git/lib/dispatchcloud/container"
  11         "git.curoverse.com/arvados.git/sdk/go/arvados"
  12         "github.com/sirupsen/logrus"
  13 )
  14
  15 // sync resolves discrepancies between the queue and the pool:
  16 //
  17 // Lingering crunch-run processes for finalized and unlocked/requeued
  18 // containers are killed.
  19 //
  20 // Locked containers whose crunch-run processes have exited are
  21 // requeued.
  22 //
  23 // Running containers whose crunch-run processes have exited are
  24 // cancelled.
  25 func (sch *Scheduler) sync() {
  26         running := sch.pool.Running()
  27         qEntries, qUpdated := sch.queue.Entries()
  28         for uuid, ent := range qEntries {
  29                 exited, running := running[uuid]
  30                 switch ent.Container.State {
  31                 case arvados.ContainerStateRunning:
  32                         if !running {
  33                                 go sch.cancel(uuid, "not running on any worker")
  34                         } else if !exited.IsZero() && qUpdated.After(exited) {
  35                                 go sch.cancel(uuid, "state=Running after crunch-run exited")
  36                         } else if ent.Container.Priority == 0 {
  37                                 go sch.kill(uuid, "priority=0")
  38                         }
  39                 case arvados.ContainerStateComplete, arvados.ContainerStateCancelled:
  40                         if running {
  41                                 // Kill crunch-run in case it's stuck;
  42                                 // nothing it does now will matter
  43                                 // anyway. If crunch-run has already
  44                                 // exited and we just haven't found
  45                                 // out about it yet, the only effect
  46                                 // of kill() will be to make the
  47                                 // worker available for the next
  48                                 // container.
  49                                 go sch.kill(uuid, fmt.Sprintf("state=%s", ent.Container.State))
  50                         } else {
  51                                 sch.logger.WithFields(logrus.Fields{
  52                                         "ContainerUUID": uuid,
  53                                         "State":         ent.Container.State,
  54                                 }).Info("container finished")
  55                                 sch.queue.Forget(uuid)
  56                         }
  57                 case arvados.ContainerStateQueued:
  58                         if running {
  59                                 // Can happen if a worker returns from
  60                                 // a network outage and is still
  61                                 // preparing to run a container that
  62                                 // has already been unlocked/requeued.
  63                                 go sch.kill(uuid, fmt.Sprintf("state=%s", ent.Container.State))
  64                         }
  65                 case arvados.ContainerStateLocked:
  66                         if running && !exited.IsZero() && qUpdated.After(exited) {
  67                                 go sch.requeue(ent, "crunch-run exited")
  68                         } else if running && exited.IsZero() && ent.Container.Priority == 0 {
  69                                 go sch.kill(uuid, "priority=0")
  70                         } else if !running && ent.Container.Priority == 0 {
  71                                 go sch.requeue(ent, "priority=0")
  72                         }
  73                 default:
  74                         sch.logger.WithFields(logrus.Fields{
  75                                 "ContainerUUID": uuid,
  76                                 "State":         ent.Container.State,
  77                         }).Error("BUG: unexpected state")
  78                 }
  79         }
  80         for uuid := range running {
  81                 if _, known := qEntries[uuid]; !known {
  82                         go sch.kill(uuid, "not in queue")
  83                 }
  84         }
  85 }
  86
  87 func (sch *Scheduler) cancel(uuid string, reason string) {
  88         if !sch.uuidLock(uuid, "cancel") {
  89                 return
  90         }
  91         defer sch.uuidUnlock(uuid)
  92         logger := sch.logger.WithField("ContainerUUID", uuid)
  93         logger.Infof("cancelling container because %s", reason)
  94         err := sch.queue.Cancel(uuid)
  95         if err != nil {
  96                 logger.WithError(err).Print("error cancelling container")
  97         }
  98 }
  99
 100 func (sch *Scheduler) kill(uuid string, reason string) {
 101         sch.pool.KillContainer(uuid, reason)
 102 }
 103
 104 func (sch *Scheduler) requeue(ent container.QueueEnt, reason string) {
 105         uuid := ent.Container.UUID
 106         if !sch.uuidLock(uuid, "cancel") {
 107                 return
 108         }
 109         defer sch.uuidUnlock(uuid)
 110         logger := sch.logger.WithFields(logrus.Fields{
 111                 "ContainerUUID": uuid,
 112                 "State":         ent.Container.State,
 113                 "Priority":      ent.Container.Priority,
 114         })
 115         logger.Infof("requeueing locked container because %s", reason)
 116         err := sch.queue.Unlock(uuid)
 117         if err != nil {
 118                 logger.WithError(err).Error("error requeueing container")
 119         }
 120 }