Support Stateful JobSet

[tenzen-y]

What would you like to be added:
I would like to support features to create a single PVC and mount the PV to some replicatedJobs like this:

apiVersion: jobset.x-k8s.io/v1alpha2
kind: JobSet
metadata:
  name: volume-sample
spec:
  volumePolicy:
    replicatedJobs:
     - name: A
     - name: B
    volumeClaimTemplates:
    - metadata:
        name: pretrained-model
      spec:
        accessModes: [ "ReadWriteMany" ]
        storageClassName: "my-storage-class"
        resources:
          requests:
            storage: 1000Gi
  replicatedJobs:
  - name: workers
[...]

In this example, JobSet creates a PVC, "pretrained-model" and then the created PV is mounted to replicatedJobs specified in the .spec.volumePolicy.replicatedJobs

This feature is similar to kubernetes/kubernetes#115066

Why is this needed:
In large distributed training, we often store the base model, and then we want to share the pre-trained model with all workers so that we can avoid downloading the pre-trained model many times.

[tenzen-y]

cc: @ahg-g @andreyvelich

[tenzen-y]

/kind feature

[danielvegamyhre]

Is this shared volume containing the pretrained model not possible to do using a regular PV, and specifying a PVC in the JobTemplate? Or is the purpose of this just to provide some automation and lifecycle management for the PV so the user doesn't need to specify the PV manifest separately?

[andreyvelich]

Or is the purpose of this just to provide some automation and lifecycle management for the PV so the user doesn't need to specify the PV manifest separately?

That's right, we want to manage lifecycle of the storage on the controller side, not on the client side.
This should simplify the ability to fine-tune models where each worker should share pre-trained model and dataset that we download in the master pod.

[ahg-g]

I like this feature, we should track it under 0.7 release.

[danielvegamyhre]

@tenzen-y I'm tentatively labeling the issue to be marked as part of the v0.7.0 release, under the assumption you plan to work on this - let me know if that's not the case.

[andreyvelich]

@danielvegamyhre I can try to take this feature in the mid Q4 - Q1.
Do we know when are you planning to cut the v0.7.0 for JobSet ?

[danielvegamyhre]

@danielvegamyhre I can try to take this feature in the mid Q4 - Q1. Do we know when are you planning to cut the v0.7.0 for JobSet ?

Sounds good, we have release cycle of roughly every 3 months, with 0.6 planned to release any day now. So 0.7 will be around October 1st, and 0.8 we can plan for around January 1st.

For now I've removed the 0.7 label for this issue and we can tentatively plan on including it in 0.8, I'll follow up on this once we get closer to that time of year.

[tenzen-y]

Sorry for the delay.
Actually, I'm okay with Andray taking this issue.

Maybe, from kubeflow v2 perspective, we need to order the priority for the additional JobSet features.
But I guess that is irrelevant to the JobSet community.

[k8s-triage-robot]

The Kubernetes project currently lacks enough contributors to adequately respond to all issues.

This bot triages un-triaged issues according to the following rules:

• After 90d of inactivity, lifecycle/stale is applied
• After 30d of inactivity since lifecycle/stale was applied, lifecycle/rotten is applied
• After 30d of inactivity since lifecycle/rotten was applied, the issue is closed

You can:

• Mark this issue as fresh with /remove-lifecycle stale
• Close this issue with /close
• Offer to help out with Issue Triage

Please send feedback to sig-contributor-experience at kubernetes/community.

/lifecycle stale

[tenzen-y]

/remove-lifecycle stale