Add --concurrent-job-syncs flag to kube-controller-manager

[tosi3k]

What type of PR is this?

/kind feature

What this PR does / why we need it:

Number of job controller workers now defaults to 5 but cannot be overridden through CLI flags like one can already do for many other controllers in KCM.

Which issue(s) this PR fixes:

Fixes #80397.

Special notes for your reviewer:

Does this PR introduce a user-facing change?

Add '--concurrent-job-syncs' flag for kube-controller-manager to set the number of job controller workers

Additional documentation e.g., KEPs (Kubernetes Enhancement Proposals), usage docs, etc.:

[k8s-ci-robot]

Please note that we're already in Test Freeze for the release-1.27 branch. This means every merged PR will be automatically fast-forwarded via the periodic ci-fast-forward job to the release branch of the upcoming v1.27.0 release.

Fast forwards are scheduled to happen every 6 hours, whereas the most recent run was: Thu Apr 6 07:59:36 UTC 2023.

[tosi3k]

/test pull-kubernetes-integration

[tosi3k]

/assign @wojtek-t

[wojtek-t]

/assign @soltysh

I know there were some similar discussion about it recently (IIRC in the context of daemonset?)

[cici37]

/triage accepted

[cici37]

/sig cli

[soltysh]

This proposal is similar to #110433 and #111800 which both were closed. Especially the last comments in the latter propose increasing qps for controllers which would help in some scenarios, but there are still some issues holding us back. For now I'm going to close this as won't do it in the near term.

/close

[k8s-ci-robot]

@soltysh: Closed this PR.

In response to this:

This proposal is similar to #110433 and #111800 which both were closed. Especially the last comments in the latter propose increasing qps for controllers which would help in some scenarios, but there are still some issues holding us back. For now I'm going to close this as won't do it in the near term.

/close

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[mborsz]

Thanks @soltysh for review!

The main difference between daemonsets (mentioned by you #111800) and jobs is that it is a valid use case to have a lot of tiny jobs (while daemonsets will usually be larger), so the problem is that even with the current qps controller limits there are use cases when we fail to saturate the job's controller qps limit due to insufficient concurrency (think of e.g. 1000 jobs, each with 1 pod to be created, if we have 5 workers, we would have max 5 concurrent pod creations).

With pod creation latency around 10ms (based on performance tests perf-dash, but technically p99 slo is 1s for such mutating calls), we will get ~50 qps which is far from client-side qps limit of 100 qps.

The reason for adding the flag by @tosi3k is to support such use cases.

I agree that adding such flag will require us to support it in the future, but it's actually hard to imagine an implementation where we won't have a concept of worker (which is now used by nearly all controllers in kube-controller-manager)

You mentioned that we need community support to prove value of the flag and I think we already have quite a lot of use cases:

• this PR
• Add flag for job and cronjob concurrent syncs to kube-controller-manager #110433 attempts to add similar flag
• @millerick in Allow ConcurrentJobSyncs to be set for kube-controller-manager #80397)

Alternative approach would be to increase default worker count -- in most cases we should be throttled by qps anyway so no difference, cost of additional few goroutines is low, but in some corner cases the default value of 5 may not be sufficient and this is why people are asking for the flag.

[mborsz]

/reopen

[k8s-ci-robot]

@mborsz: Reopened this PR.

In response to this:

/reopen

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[tosi3k]

Per offline discussion - fixed the description and added validation for the flag value.

@soltysh PTAL.

[soltysh]

Based on the provided reasoning behind this change, and based on an offline discussion
/lgtm
/approve

[k8s-ci-robot]

LGTM label has been added.

Git tree hash: 0d882346c4427f2859ca312ef08ba639f031a840

[soltysh]

/triage accepted
/priority important-longterm

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: soltysh, tosi3k

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Needs approval from an approver in each of these files:

• cmd/kube-controller-manager/OWNERS [soltysh]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment