Re: [Exception request] KEP-6268: Stall and resume slow watch streams (v1.38 KEP readiness deadline and enhancements freeze)

[Joe Betz]

As SIG API Machinery TL, I support this exception. I've been tracking this KEP and it's in good shape to merge by enhancement freeze.

On Fri, Sep 25, 2026 at 7:15 PM 'Nami Nasserazad' via kubernetes-sig-release <kubernetes-...@googlegroups.com> wrote:

Hello release team,

I am requesting an exception for KEP-6268 for the v1.38 KEP readiness deadline and, because the two dates are close, for the enhancements freeze in the same request.

- Enhancement name: Stall and resume slow watch streams instead of terminating them (WatchCacheStallResume)
- Enhancement status (alpha/beta/stable): alpha
- SIG: sig-api-machinery
- k/enhancements repo issue #: https://www.google.com/url?q=https://github.com/kubernetes/enhancements/issues/6268&source=gmail&ust=1790459392825000&sa=E
- PR #'s: KEP https://www.google.com/url?q=https://github.com/kubernetes/enhancements/pull/6269&source=gmail&ust=1790459392825000&sa=E (updated with the PRR file and marked ready for review); code: https://www.google.com/url?q=https://github.com/kubernetes/kubernetes/pull/141450&source=gmail&ust=1790459392825000&sa=E (feature gate and mechanism), https://www.google.com/url?q=https://github.com/kubernetes/kubernetes/pull/141451&source=gmail&ust=1790459392825000&sa=E (metrics gauge and cull), https://www.google.com/url?q=https://github.com/kubernetes/kubernetes/pull/141452&source=gmail&ust=1790459392825000&sa=E (integration test); already merged: https://www.google.com/url?q=https://github.com/kubernetes/kubernetes/pull/141449&source=gmail&ust=1790459392825000&sa=E and https://www.google.com/url?q=https://github.com/kubernetes/kubernetes/pull/141475&source=gmail&ust=1790459392825000&sa=E
- Additional time needed (in calendar days, due end of day AoE): 4 days, through Tuesday 29 September 2026 AoE (the enhancements freeze), to get the KEP PR reviewed and merged. The PRR reviewer (@wojtek-t) and the lead (@jpbetz) have opted in.
- Reason this enhancement is critical for this milestone: large clusters lose watch clients (kubelets, controllers, proxies) whenever one slow client fills its buffer, because the apiserver blocks the shared dispatch path and then terminates the client. The code is written, benchmarked and validated with the watch correctness harness; two of the five PRs are merged and the mechanism PR is ready for review. The KEP missed the readiness deadline only because the design was revised at the reviewers' request (server-driven batched resync, the etcd model) and the PRR file was added late.
- Risks from adding code late: low. The change is behind a new alpha feature gate that is off by default; with the gate off the control flow is unchanged. It touches one package (the apiserver watch cache). It has unit tests under the race detector, an integration test, a benchmark, and was validated against the model-based watch correctness harness with slow readers.
- Risks from cutting enhancement: the fix slips one release (about four months) while the slow-watcher terminations keep happening on large clusters; no partial implementation lands if it is cut, because the whole series is behind the gate.

Thank you,
Nami Nasserazad (@nami-ant)

--
You received this message because you are subscribed to the Google Groups "kubernetes-sig-release" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kubernetes-sig-re...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/kubernetes-sig-release/CAFQn_hZT6usthbRU2BTEP8Wef%3DRwhutFhOxPmpUY5SN98zizfw%40mail.gmail.com.