Coscheduilng.

[k82cn]

Signed-off-by: Da K. Ma klaus1982.cn@gmail.com

Moved kubernetes/community#2337 to k/enhancements.

/cc @bsalamat, @vishh, @mindprince, @jlewi

/cc @kubernetes/sig-scheduling-feature-requests

[k82cn]

xref #583

[k82cn]

/assign bsalamat

[yastij]

/cc

adding this one to my backlog

[bsalamat]

I think this proposal is a good starting point for introducing gang scheduling. Given that our current plan is to build an early version in kube-batch (which is an incubator, not a core component), this looks good to me. We should expand the proposal in the future with more in-depth information about life-cycle and quota management of gangs.

[MaciekPytel]

I would like a section on how this will interact with cluster autoscaling added to the doc. I don't expect it to work with CA in initial version, but we should make sure the design is not fundamentally conflicting with how CA works if we want to promote it to core kubernetes later on.
Cluster Autoscaler works by importing and internally running scheduler predicates to simulate how scheduler would behave if nodes were added to cluster or removed from it. Ideally any scheduling logic should be in form of predicates that we can inform. Otherwise we should make sure it's something that can be duplicated in CA.

[k82cn]

Ideally any scheduling logic should be in form of predicates that we can inform.

Coscheduling is not a predicate in scheduler; if necessary, CA need to check PodGroup.Status and pending pods to get how many additional resource are required.

Cluster Autoscaler works by importing and internally running scheduler predicates ...

How CA handle scheduler extender ?

[k82cn]

@MaciekPytel

When discussing CA + kube-batch at kubernetes-retired/kube-batch#526 (comment) , I'm thinking to add more info in PodGroupStatus for CA to scale out.

[MaciekPytel]

I replied on kubernetes-retired/kube-batch#533 (review). Let's continue the discussion in a single thread.

[k82cn]

I replied on kubernetes-retired/kube-batch#533 (review). Let's continue the discussion in a single thread.

Thanks for your reply. I marked Cluster-Autoscaler as out-of-scope in this version; I'll open other PRs when we have alternative solution for it.

[MaciekPytel]

Thanks for your reply. I marked Cluster-Autoscaler as out-of-scope in this version; I'll open other PRs when we have alternative solution for it.

I'm perfectly fine with that as long as the feature lives in kube-batch (which is incompatible with CA anyway). We need to figure out how to make it work with CA if/when we want to move the feature to default scheduler.

[mattfarina]

Any chance you could format this following the template?

For example, what would the graduation criteria for this look like?

[k82cn]

Any chance you could format this following the template?

Done :)

[bsalamat]

/lgtm

More improvements to this design is needed in the area of life-cycle management of pod groups, but those can be added later, given that this design is going to an incubator project for now and won't be integrated into standard components.

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: bsalamat, k82cn

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:

• keps/sig-scheduling/OWNERS [bsalamat,k82cn]

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