import (
"sort"
+ "time"
- "git.curoverse.com/arvados.git/lib/cloud"
"git.curoverse.com/arvados.git/lib/dispatchcloud/container"
"git.curoverse.com/arvados.git/sdk/go/arvados"
- "github.com/Sirupsen/logrus"
+ "github.com/sirupsen/logrus"
)
func (sch *Scheduler) runQueue() {
overquota = sorted[i:]
break tryrun
}
- sch.bgLock(logger, ctr.UUID)
+ go sch.lockContainer(logger, ctr.UUID)
unalloc[it]--
case arvados.ContainerStateLocked:
- if unalloc[it] < 1 {
- if sch.pool.AtQuota() {
- logger.Debug("not starting: AtQuota and no unalloc workers")
- overquota = sorted[i:]
- break tryrun
- }
+ if unalloc[it] > 0 {
+ unalloc[it]--
+ } else if sch.pool.AtQuota() {
+ logger.Debug("not starting: AtQuota and no unalloc workers")
+ overquota = sorted[i:]
+ break tryrun
+ } else {
logger.Info("creating new instance")
- err := sch.pool.Create(it)
- if err != nil {
- if _, ok := err.(cloud.QuotaError); !ok {
- logger.WithError(err).Warn("error creating worker")
- }
+ if !sch.pool.Create(it) {
+ // (Note pool.Create works
+ // asynchronously and logs its
+ // own failures, so we don't
+ // need to log this as a
+ // failure.)
+
sch.queue.Unlock(ctr.UUID)
// Don't let lower-priority
// containers starve this one
overquota = sorted[i:]
break tryrun
}
- unalloc[it]++
}
if dontstart[it] {
// same instance type. Don't let this
// one sneak in ahead of it.
} else if sch.pool.StartContainer(it, ctr) {
- unalloc[it]--
+ // Success.
} else {
dontstart[it] = true
}
}
}
-// Start an API call to lock the given container, and return
-// immediately while waiting for the response in a new goroutine. Do
-// nothing if a lock request is already in progress for this
-// container.
-func (sch *Scheduler) bgLock(logger logrus.FieldLogger, uuid string) {
- logger.Debug("locking")
+// Lock the given container. Should be called in a new goroutine.
+func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
+ if !sch.uuidLock(uuid, "lock") {
+ return
+ }
+ defer sch.uuidUnlock(uuid)
+ if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
+ // This happens if the container has been cancelled or
+ // locked since runQueue called sch.queue.Entries(),
+ // possibly by a lockContainer() call from a previous
+ // runQueue iteration. In any case, we will respond
+ // appropriately on the next runQueue iteration, which
+ // will have already been triggered by the queue
+ // update.
+ logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
+ return
+ }
+ err := sch.queue.Lock(uuid)
+ if err != nil {
+ logger.WithError(err).Warn("error locking container")
+ return
+ }
+ logger.Debug("lock succeeded")
+ ctr, ok := sch.queue.Get(uuid)
+ if !ok {
+ logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
+ } else if ctr.State != arvados.ContainerStateLocked {
+ logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
+ }
+}
+
+// Acquire a non-blocking lock for specified UUID, returning true if
+// successful. The op argument is used only for debug logs.
+//
+// If the lock is not available, uuidLock arranges to wake up the
+// scheduler after a short delay, so it can retry whatever operation
+// is trying to get the lock (if that operation is still worth doing).
+//
+// This mechanism helps avoid spamming the controller/database with
+// concurrent updates for any single container, even when the
+// scheduler loop is running frequently.
+func (sch *Scheduler) uuidLock(uuid, op string) bool {
sch.mtx.Lock()
defer sch.mtx.Unlock()
- if sch.locking[uuid] {
- logger.Debug("locking in progress, doing nothing")
- return
+ logger := sch.logger.WithFields(logrus.Fields{
+ "ContainerUUID": uuid,
+ "Op": op,
+ })
+ if op, locked := sch.uuidOp[uuid]; locked {
+ logger.Debugf("uuidLock not available, Op=%s in progress", op)
+ // Make sure the scheduler loop wakes up to retry.
+ sch.wakeup.Reset(time.Second / 4)
+ return false
}
- sch.locking[uuid] = true
- go func() {
- defer func() {
- sch.mtx.Lock()
- defer sch.mtx.Unlock()
- delete(sch.locking, uuid)
- }()
- err := sch.queue.Lock(uuid)
- if err != nil {
- logger.WithError(err).Warn("error locking container")
- return
- }
- ctr, ok := sch.queue.Get(uuid)
- if !ok {
- logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
- } else if ctr.State != arvados.ContainerStateLocked {
- logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
- }
- }()
+ logger.Debug("uuidLock acquired")
+ sch.uuidOp[uuid] = op
+ return true
+}
+
+func (sch *Scheduler) uuidUnlock(uuid string) {
+ sch.mtx.Lock()
+ defer sch.mtx.Unlock()
+ delete(sch.uuidOp, uuid)
}