The change was intended to fix a case where a canary alloc may fail to
be rescheduled if all the other allocs fail as well (e.g. if all allocs
happen to be placed on a node that died). However, it introduced some
unintended side-effects.
Reverting the change for now and will investigate further.
This handles the case where a job when from no-deployment to deployment
with canaries.
Consider a case where a `max_parallel=0` job is submitted as version 0,
then an update is submitted with `max_parallel=1, canary=1` as verion 1.
In this case, we will have 1 canary alloc, and all remaining allocs will
be version 0. Until the deployment is promoted, we ought to replace the
canaries with version 0 job (which isn't associated with a deployment).
This change fixes a bug where lost/failed allocations are replaced by
allocations with the latest versions, even if the version hasn't been
promoted yet.
Now, when generating a plan for lost/failed allocations, the scheduler
first checks if the current deployment is in Canary stage, and if so, it
ensures that any lost/failed allocations is replaced one with the latest
promoted version instead.
The reconcile loop sets `DeploymentState.DesiredCanaries` only on the first
pass through the loop and if the job is not paused/pending. In MRD,
deployments will make one pass though the loop while "pending", and were not
ever getting `DesiredCanaries` set. We can't set it in the initial
`DeploymentState` constructor because the first pass through setting up
canaries expects it's not there yet. However, this value is static for a given
version of a job because it's coming from the update stanza, so it's safe to
re-assign the value on subsequent passes.
The field name `Deployment.TaskGroups` contains a map of `DeploymentState`,
which makes it a little harder to follow state updates when combined with
inconsistent naming conventions, particularly when we also have the state
store or actual `TaskGroup`s in scope. This changeset changes all uses to
`dstate` so as not to be confused with actual TaskGroups.
If a job update includes a task group that has no changes, those allocations
have their version bumped in-place. The ends up triggering an eval from
`deploymentwatcher` when it verifies their health. Although this eval is a
no-op, we were only treating pending deployments the same as paused when
the deployment was a new MRD. This means that any eval after the initial one
will kick off the deployment, and that caused pending deployments to "jump
the queue" and run ahead of schedule, breaking MRD invariants and resulting in
a state with all regions blocked.
This behavior can be replicated even in the case of job updates with no
in-place updates by patching `deploymentwatcher` to inject a spurious no-op
eval. This changeset fixes the behavior by treating pending deployments the
same as paused in all cases in the reconciler.
Deployments should wait until kicked off by `Job.Register` so that we can
assert that all regions have a scheduled deployment before starting any
region. This changeset includes the OSS fixes to support the ENT work.
`IsMultiregionStarter` has no more callers in OSS, so remove it here.
* ar: support opting into binding host ports to default network IP
* fix config plumbing
* plumb node address into network resource
* struct: only handle network resource upgrade path once
Pre-0.12.0 beta, a deployment was considered "complete" if it was
successful. But with MRD we have "blocked" and "unblocking" states as well. We
did not consider the case where a concurrent alloc health status update
triggers a `Compute` call on a deployment that's moved from "blocked" to
"unblocking" (it's a small window), which caused an extra pass thru the
`nextRegion` logic in `deploymentwatcher` and triggered an error when later
transitioning to "successful".
This changeset makes sure we don't overwrite that status.
In #8209 we fixed the `max_parallel` stanza for multiregion by introducing the
`IsMultiregionStarter` check, but didn't apply it to the earlier place its
required. The result is that deployments start but don't place allocations.
If `max_parallel` is not set, all regions should begin in a `running` state
rather than a `pending` state. Otherwise the first region is set to `running`
and then all the remaining regions once it enters `blocked. That behavior is
technically correct in that we have at most `max_parallel` regions running,
but definitely not what a user expects.
The `paused` state is used as an operator safety mechanism, so that they can
debug a deployment or halt one that's causing a wider failure. By using the
`paused` state as the first state of a multiregion deployment, we risked
resuming an intentionally operator-paused deployment because of activity in a
peer region.
This changeset replaces the use of the `paused` state with a `pending` state,
and provides a `Deployment.Run` internal RPC to replace the use of the
`Deployment.Pause` (resume) RPC we were using in `deploymentwatcher`.
* `nextRegion` should take status parameter
* thread Deployment/Job RPCs thru `nextRegion`
* add `nextRegion` calls to `deploymentwatcher`
* use a better description for paused for peer
* scheduler/reconcile: set FollowupEvalID on lost stop_after_client_disconnect
* scheduler/reconcile: thread follupEvalIDs through to results.stop
* scheduler/reconcile: comment typo
* nomad/_test: correct arguments for plan.AppendStoppedAlloc
* scheduler/reconcile: avoid nil, cleanup handleDelayed(Lost|Reschedules)
* client/heartbeatstop: reversed time condition for startup grace
* scheduler/generic_sched: use `delayInstead` to avoid a loop
Without protecting the loop that creates followUpEvals, a delayed eval
is allowed to create an immediate subsequent delayed eval. For both
`stop_after_client_disconnect` and the `reschedule` block, a delayed
eval should always produce some immediate result (running or blocked)
and then only after the outcome of that eval produce a second delayed
eval.
* scheduler/reconcile: lostLater are different than delayedReschedules
Just slightly. `lostLater` allocs should be used to create batched
evaluations, but `handleDelayedReschedules` assumes that the
allocations are in the untainted set. When it creates the in-place
updates to those allocations at the end, it causes the allocation to
be treated as running over in the planner, which causes the initial
`stop_after_client_disconnect` evaluation to be retried by the worker.
* jobspec, api: add stop_after_client_disconnect
* nomad/state/state_store: error message typo
* structs: alloc methods to support stop_after_client_disconnect
1. a global AllocStates to track status changes with timestamps. We
need this to track the time at which the alloc became lost
originally.
2. ShouldClientStop() and WaitClientStop() to actually do the math
* scheduler/reconcile_util: delayByStopAfterClientDisconnect
* scheduler/reconcile: use delayByStopAfterClientDisconnect
* scheduler/util: updateNonTerminalAllocsToLost comments
This was setup to only update allocs to lost if the DesiredStatus had
already been set by the scheduler. It seems like the intention was to
update the status from any non-terminal state, and not all lost allocs
have been marked stop or evict by now
* scheduler/testing: AssertEvalStatus just use require
* scheduler/generic_sched: don't create a blocked eval if delayed
* scheduler/generic_sched_test: several scheduling cases
The BinPackIter accounted for node reservations twice when scoring nodes
which could bias scores toward nodes with reservations.
Pseudo-code for previous algorithm:
```
proposed = reservedResources + sum(allocsResources)
available = nodeResources - reservedResources
score = 1 - (proposed / available)
```
The node's reserved resources are added to the total resources used by
allocations, and then the node's reserved resources are later
substracted from the node's overall resources.
The new algorithm is:
```
proposed = sum(allocResources)
available = nodeResources - reservedResources
score = 1 - (proposed / available)
```
The node's reserved resources are no longer added to the total resources
used by allocations.
My guess as to how this bug happened is that the resource utilization
variable (`util`) is calculated and returned by the `AllocsFit` function
which needs to take reserved resources into account as a basic
feasibility check.
To avoid re-calculating alloc resource usage (because there may be a
large number of allocs), we reused `util` in the `ScoreFit` function.
`ScoreFit` properly accounts for reserved resources by subtracting them
from the node's overall resources. However since `util` _also_ took
reserved resources into account the score would be incorrect.
Prior to the fix the added test output:
```
Node: reserved Score: 1.0000
Node: reserved2 Score: 1.0000
Node: no-reserved Score: 0.9741
```
The scores being 1.0 for *both* nodes with reserved resources is a good
hint something is wrong as they should receive different scores. Upon
further inspection the double accounting of reserved resources caused
their scores to be >1.0 and clamped.
After the fix the added test outputs:
```
Node: no-reserved Score: 0.9741
Node: reserved Score: 0.9480
Node: reserved2 Score: 0.8717
```
