Fix deployment watchers to avoid creating unnecessary deployment watcher goroutines and blocking queries. `deploymentWatcher.getAllocsCh` creates a new goroutine that makes a blocking query to fetch updates of deployment allocs.
## Background
When operators submit a new or updated service job, Nomad create a new deployment by default. The deployment object controls how fast to place the allocations through [`max_parallel`](https://www.nomadproject.io/docs/job-specification/update#max_parallel) and health checks configurations.
The `scheduler` and `deploymentwatcher` package collaborate to achieve deployment logic: The scheduler only places the canaries and `max_parallel` allocations for a new deployment; the `deploymentwatcher` monitors for alloc progress and then enqueues a new evaluation whenever the scheduler should reprocess a job and places the next `max_parallel` round of allocations.
The `deploymentwatcher` package makes blocking queries against the state store, to fetch all deployments and the relevant allocs for each running deployments. If `deploymentwatcher` fails or is hindered from fetching the state, the deployments fail to make progress.
`Deploymentwatcher` logic only runs on the leader.
## Why unnecessary deployment watchers can halt cluster progress
Previously, `getAllocsCh` is called on every for loop iteration in `deploymentWatcher.watch()` function. However, the for-loop may iterate many times before the allocs get updated. In fact, whenever a new deployment is created/updated/deleted, *all* `deploymentWatcher`s get notified through `w.deploymentUpdateCh`. The `getAllocsCh` goroutines and blocking queries spike significantly and grow quadratically with respect to the number of running deployments. The growth leads to two adverse outcomes:
1. it spikes the CPU/Memory usage resulting potentially leading to OOM or very slow processing
2. it activates the [query rate limiter](abaa9c5c5b/nomad/deploymentwatcher/deployment_watcher.go (L896-L898)), so later the watcher fails to get updates and consequently fails to make progress towards placing new allocations for the deployment!
So the cluster fails to catch up and fails to make progress in almost all deployments. The cluster recovers after a leader transition: the deposed leader stops all watchers and free up goroutines and blocking queries; the new leader recreates the watchers without the quadratic growth and remaining under the rate limiter. Well, until a spike of deployments are created triggering the condition again.
### Relevant Code References
Path for deployment monitoring:
* [`Watcher.watchDeployments`](abaa9c5c5b/nomad/deploymentwatcher/deployments_watcher.go (L164-L192)) loops waiting for deployment updates.
* On every deployment update, [`w.getDeploys`](abaa9c5c5b/nomad/deploymentwatcher/deployments_watcher.go (L194-L229)) returns all deployments in the system
* `watchDeployments` calls `w.add(d)` on every active deployment
* which in turns, [updates existing watcher if one is found](abaa9c5c5b/nomad/deploymentwatcher/deployments_watcher.go (L251-L255)).
* The deployment watcher [updates local local deployment field and trigger `deploymentUpdateCh` channel]( abaa9c5c5b/nomad/deploymentwatcher/deployment_watcher.go (L136-L147))
* The [deployment watcher `deploymentUpdateCh` selector is activated](abaa9c5c5b/nomad/deploymentwatcher/deployment_watcher.go (L455-L489)). Most of the time the selector clause is a no-op, because the flow was triggered due to another deployment update
* The `watch` for-loop iterates again and in the previous code we create yet another goroutine and blocking call that risks being rate limited.
Co-authored-by: Tim Gross <tgross@hashicorp.com>
This PR edits the computed agent version that is returned upon hitting
the agent self request endpoint. The reason is because we believe that
the Agent Member Tag property sometimes returns null because we may have
cases where there are only clients and no servers and only servers are
included in the Serf Gossip Protocol. There may be other cases where we
do in fact have servers but the node is erased for some reason. We are
unsure how to replicate that issue, however.
edit mirage config
This commit updates the Mirage Config because our acceptance tests
depend on the Mirage Config, while we rely on Mirage Factories to
populate fixture data for us to use when to run the Nomad UI locally
Revert "update the open-button disability functionality depending on a job's state"
This reverts commit 5190b308a51d55a7b0617854164c155d36d7e513.
The `QuotaIterator` is used as the source of nodes passed into feasibility
checking for constraints. Every node that passes the quota check counts the
allocation resources agains the quota, and as a result we count nodes which
will be later filtered out by constraints. Therefore for jobs with
constraints, nodes that are feasibility checked but fail have been counted
against quotas. This failure mode is order dependent; if all the unfiltered
nodes happen to be quota checked first, everything works as expected.
This changeset moves the `QuotaIterator` to happen last among all feasibility
checkers (but before ranking). The `QuotaIterator` will never receive filtered
nodes so it will calculate quotas correctly.
The test fails reliably locally on my machine. The test uses non-dev mode
where Raft actions get committed to disk, causing operations to exceed
the 50ms tight Raft deadlines.
So, here we ensure that non-dev servers use default Raft config
files with longer timeouts.
Also, noticed that the test queries a server, that may a follower with a
stale state.
I've updated the test to ensure we query the leader for its state. The
Barrier call ensures that the leader is a "stable" leader with committed
entries. Protects against a window where a new leader reports the
previous term before it commits a raft log entry.
Ensure that all servers are joined to each other before test proceed,
instead of just joining them to the first server and relying on
background serf propagation.
Relying on backgorund serf propagation is a cause of flakiness,
specially for tests with multiple regions. The server receiving the RPC
may not be aware of the region and fail to forward RPC accordingly.
For example, consider `TestMonitor_Monitor_RemoteServer` failure in https://app.circleci.com/pipelines/github/hashicorp/nomad/16402/workflows/7f327235-7d0c-40ba-9757-600522afca51/jobs/158045 you can observe:
* `nomad-117` is joined to `nomad-118` and `nomad-119`
* `nomad-119` is the foreign region
* `nomad-117` gains leadership in the default region, `nomad-118` is the non-leader
* search logs for `nomad: adding server` and notice that `nomad-118`
only added `nomad-118` and `nomad-118`, but not `nomad-119`!
* so the query to the non-leader in the test fails to be forwarded to
the appopriate region.
This updates `client.Ready()` so it returns once the client node got
registered at the servers. Previously, it returns when the
fingerprinters first batch completes, wtihout ensuring that the node is
stored in the Raft data. The tests may fail later when it with unknown
node errors later.
`client.Reedy()` seem to be only called in CSI and some client stats
now.
This class of bug, assuming client is registered without checking, is a
source of flakiness elsewhere. Other tests use other mechanisms for
checking node readiness, though not consistently.
Glint pulled in an updated version of mitchellh/go-testing-interface
which broke some existing tests because the update added a Parallel()
method to testing.T. This switches to the standard library testing.TB
which doesn't have a Parallel() method.
Adding '-verbose' will print out the allocation information for the
deployment. This also changes the job run command so that it now blocks
until deployment is complete and adds timestamps to the output so that
it's more in line with the output of node drain.
This uses glint to print in place in running in a tty. Because glint
doesn't yet support cmd/powershell, Windows workflows use a different
library to print in place, which results in slightly different
formatting: 1) different margins, and 2) no spinner indicating
deployment in progress.
(cherry-pick ent back to oss)
This PR moves a lot of Consul ACL token validation tests into ent files,
so that we can verify correct behavior difference between OSS and ENT
Nomad versions.
Old description of `{plan,worker}.wait_for_index` described the metric
in terms of waiting for a snapshot which has two problems:
1. "Snapshot" is an overloaded term in Nomad and operators can't be
expected to know which use we're referring to here.
2. The most important thing about the metric is what we're waiting *on*
before taking a snapshot: the raft index of the object to be
processed (plan or eval).
The new description tries to cram all of that context into the tiny
space provided.
See #5791 for details about the `wait_for_index` mechanism in general.
This PR fixes the Nomad Object Namespace <-> Consul ACL Token relationship
check when using Consul OSS (or Consul ENT without namespace support).
Nomad v1.1.0 introduced a regression where Nomad would fail the validation
when submitting Connect jobs and allow_unauthenticated set to true, with
Consul OSS - because it would do the namespace check against the Consul ACL
token assuming the "default" namespace, which does not work because Consul OSS
does not have namespaces.
Instead of making the bad assumption, expand the namespace check to handle
each special case explicitly.
Fixes#10718
