[KEP] Maximum Execution Time

[trasc]

What type of PR is this?

/kind documentation

What this PR does / why we need it:

Proposed KEP fo Maximum Execution Time feature.

Which issue(s) this PR fixes:

Relates to #3125

Special notes for your reviewer:

Does this PR introduce a user-facing change?

none

[netlify]

✅ Deploy Preview for kubernetes-sigs-kueue canceled.

Name	Link
🔨 Latest commit	3fb7435
🔍 Latest deploy log	https://app.netlify.com/sites/kubernetes-sigs-kueue/deploys/67189d6abc2ded0008dcd6ed

[trasc]

/cc @tenzen-y

[mimowo]

Very initial pass to discuss the main points.

[mimowo]

/assign @mwielgus

[kannon92]

There are a few questions I have.

1. Do we want some kind of declarative API that users who don't use kueuectl can still set this?
   a) Is it okay to set the label on the Workload objects and this would be respected?

2. I am a bit unfamilar with workloads in Kueue. Will setting the active field to false cause a suspend or a deletion of all workloads?

3. What would be the suggestions for existing activeDeadlines with this feature? If a JobSet user wants an active deadline outside of Kueue, what should we implement? Same for Kubeflow, Ray, etc.

[trasc]

1. Do we want some kind of declarative API that users who don't use kueuectl can still set this?
   a) Is it okay to set the label on the Workload objects and this would be respected?

Yes, a label kueue.x-k8s.io/max-exec-time-seconds, regardless if the job type, the value will be passed into the created Workload object.

2. I am a bit unfamilar with workloads in Kueue. Will setting the active field to false cause a suspend or a deletion of all workloads?

The job is suspended.

3. What would be the suggestions for existing activeDeadlines with this feature? If a JobSet user wants an active deadline outside of Kueue, what should we implement? Same for Kubeflow, Ray, etc.

The proposal is only relevant while working with Kueue.

Other activeDeadlines (that maybe marks the job a Failed) could be used but Kueue will not be aware of them.

[kannon92]

Yes, a label kueue.x-k8s.io/max-exec-time-seconds, regardless if the job type, the value will be passed into the created Workload object.

I read that a user needs to set kueuectl -t and that would set a label in the workload. But if I do not use kueuectl at all, I would like to know can I add a label to a job and this would be propagated to the workload object?

I read that this is a bit of a hack to get a standard for representing workload time. I would like to see a future where we can add some kind of general API that works with Kueue or without Kueue to enforce a deadline for any workload.

At the moment, I worry that users would not know to use activeDeadlineSeconds (at Job or Pod) or as a person working on JobSet, I do not know what kind of API to implement in JobSet to set up a future with Kueue.

Plus we have a few feature requests for the Job API to have a smarter deadline ie kubernetes/kubernetes#111948.

It seems like you are basically saying that if Kueue label is specfiied, then we should not be using any of the workload timelimits as they will race/conflict with each other.

[trasc]

Yes, a label kueue.x-k8s.io/max-exec-time-seconds, regardless if the job type, the value will be passed into the created Workload object.

I read that a user needs to set kueuectl -t and that would set a label in the workload. But if I do not use kueuectl at all, I would like to know can I add a label to a job and this would be propagated to the workload object?

I think you are referring to kjobctl -t which is one intended use-case for this, however if the label is set on a job Kueue will perform the same regardless of the tool used to create the job.

I read that this is a bit of a hack to get a standard for representing workload time. I would like to see a future where we can add some kind of general API that works with Kueue or without Kueue to enforce a deadline for any workload.

At the moment, I worry that users would not know to use activeDeadlineSeconds (at Job or Pod) or as a person working on JobSet, I do not know what kind of API to implement in JobSet to set up a future with Kueue.

Nothing should be done for the specific jobs (JobSet, RayJob ...) the only changes needed are in Kueue and mostly in the generic implementation (jobframework).

Plus we have a few feature requests for the Job API to have a smarter deadline ie kubernetes/kubernetes#111948.

It seems like you are basically saying that if Kueue label is specfiied, then we should not be using any of the workload timelimits as they will race/conflict with each other.

Unfortunately for specific job types this could happen, however, as follow-ups, Kueue can try to detect conflicting configuration at the job specific webhooks level and alert the user when necessary.

[tenzen-y]

I will come back here after the next release.

/hold

[tenzen-y]

/cc

[trasc]

I have a draft implementation, and only need to realign it with the KEP.

[tenzen-y]

I have a draft implementation, and only need to realign it with the KEP.

Thank you for leading this feature.

[kannon92]

I still have concerns on the impact of this with existing deadline features with other workloads.

It seems that Kueue will have its own deadline syntax for all workloads. Does this mean that workloads using their own API should be discouraged from using those APIs while using Kueue?

If a pod has an active deadline will we do any validation to make sure that the Kueue timeout and this are meaningful together or would we ignore the pod one?

Same for Job with ActiveDeadlines.

And yes, any future workload can implement their own deadline but I think Kueue integration with these would be important so I am still not sure the path forward for JobSet or training operator.

[mimowo]

As mentioned in the issue we would like to have slurm-like -t option, supported across all Jobs to use for time-based scheduling reliably. We also want to be able to set the new timeout from kjobctl for any Job CRD.

Actually, take for example the ActiveDeadlineSeconds timeout from core k8s - we have it two different levels: Pod and Job, because the use-cases are different. I think it is also the case here, that the use-case of Kueue is different, and the semantics of the timeout differ.

[mimowo]

@tenzen-y @mimowo can this still be considered for v0.9?

It is still on the list of nice-to-haves, and I'm happy to move it forward, especially if @alculquicondor and @tenzen-y can carry the reviews as I'm currently focused primarly on completing TAS.

[tenzen-y]

@tenzen-y @mimowo can this still be considered for v0.9?

It is still on the list of nice-to-haves, and I'm happy to move it forward, especially if @alculquicondor and @tenzen-y can carry the reviews as I'm currently focused primarly on completing TAS.

Sure, I can lead the review for this.

[tenzen-y]

If a pod has an active deadline will we do any validation to make sure that the Kueue timeout and this are meaningful together or would we ignore the pod one?

Indeed, I asked similar questions at #3133 (comment).
But, we do not verify the relationship between Pod or Job activedeadlineSeconds and Kueue activeDeadlinSeconds since as @mimowo mentioned above, those have different use cases.

[tenzen-y]

Thank you!
/lgtm
/approve

[k8s-ci-robot]

LGTM label has been added.

Git tree hash: 776213cfe52cb0d7a5901cd6ba599a27a8a6b455

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: alculquicondor, tenzen-y, trasc

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:

• OWNERS [alculquicondor,tenzen-y]

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

[mimowo]

/hold cancel