Priority/preemption

[davidopp]

Feature Description

• One-line feature description (can be used as a release note): A mechanism for imposing a total order on pods, that determines which pods run and which go pending when the cluster is overcomitted, implemented through a priority scheme, preemption of running pods by the default scheduler, having kubelet eviction order take the priority into account, and modifications to the quota mechanism to take priority into account.
• Primary contact (assignee): @bsalamat
• Responsible SIGs: @kubernetes/sig-scheduling-feature-requests, @kubernetes/sig-node-feature-requests
• Design proposal link (community repo): high-level architecture doc (currently under review/discussion/iteration) here; detailed design doc for first chunk is in Design proposal for adding priority to Kubernetes API community#604
• Reviewer(s) - (for LGTM) recommend having 2+ reviewers (at least one from code-area OWNERS file) agreed to review. Reviewers from multiple companies preferred: @davidopp, @derekwaynecarr, @erictune, @k82cn, @timothysc, @wojtek-t
• Approver (likely from SIG/area to which feature belongs): @davidopp
• Initial target stage (alpha/beta/stable) and release (x.y): for 1.7 we anticipate having the detailed design, API (alpha), and implementation of kubelet eviction ordering (alpha); some implementation work for the other pieces will hopefully have begun but will not be in 1.7

ref/ kubernetes/kubernetes#22212
cc/ @dchen1107

Note: this feature is intimately related to #269.

[davidopp]

The minimum we want for 1.7 is an alpha API fo expressing pod priority, and hopefully incorporating that into kubelet pod eviction ordering when out of resource. Most of this feature is not expected to ship in 1.7, especially since the design is still under discussion.

[0xmichalis]

when the cell is overcomitted

This is Borg terminology, right? Many people coming to k8s will not be familiar with Borg so I would suggest switching "cell" to "node".

[davidopp]

You caught me! I meant cluster, not cell. Fixed now.

Glad to know someone is reading this stuff. :)

[obeattie]

Is the document linked in the description (and that of #269) intended to be private? This is a feature which is very important for our deployment. I'm mostly just curious to follow its development, so if this is private for a reason that's fine – I just figure it may be an oversight.

[davidopp]

You can view (and commnt in) the document by joining the kubernetes-dev mailing list
https://groups.google.com/forum/#!forum/kubernetes-dev

Any feedback on the proposal is welcome. (It's still in the proposal phase.)

[obeattie]

Aha, thank you 👍