gistfile1.txt

Hello!

I would like to ask for an exception that allows
https://github.com/kubernetes/kubernetes/pull/111023 to be merged into
1.26 after the code freeze.

Enhancement name: dynamic resource allocation
Enhancement status: alpha
SIG: Node, with Scheduling as participating SIG
k/enhancements repo issue #: #3063
PR #’s: #111023
Additional time needed (in days): 3 (= till Friday this week)

   This is needed to give various code owners time to add their final
   approval. Key stakeholders (Aldo for scheduling, Tim for architecture
   and API) were basically ready for that already yesterday before the
   code freeze, but some smaller recent changes still needed to be
   checked again and other reviewers need more time.

Reason this enhancement is critical for this milestone:

   There is a lot of interest and momentum behind this feature right
   now. Hardware vendors are ready to showcase it to customers, but
   that will be harder when using the feature depends on building
   a fork of Kubernetes from source. Not merging it now risks loosing
   this momentum.

   Merging the feature also won't be easier for 1.27. It's now
   fresh in the minds of reviewers, delaying until then would
   imply that they need to make themselves familiar with it anew.

Risks from adding code late:

   All the new code is behind a feature gate, so the risk for stability
   of other features when not enabled is low.

   The core API gets changed. The new fields are also feature gated, so
   without the feature gate they won't be visible and not impact clients.
   This in particular has been heavily scrutinized and changes were
   made to support future extensions (for example, using a struct
   instead of a plain string in one place), so the risk of making some
   undesirable change is low to medium.

   All code has unit tests that are passing. Test coverage (despite
   being alpha) is getting close to that of comparable code that is GA.
   For example, the GA volumebinding scheduler plugin has 82%
   statement coverage while the new plugin has 70%.
   E2E tests also exist and are passing in a new, optional Prow
   pre-merge job that runs for PRs touching the code.
   The risk for testing stability is low.

Risks from cutting enhancement:

   The biggest risk is that if don't merge, we will loose the momentum
   and then also won't get it merged in the future. This affects various
   customer use cases where the currently possible workarounds are
   not fully solving the problems that customers are having.

   For example, Kevin Klues said in
   https://github.com/kubernetes/kubernetes/pull/111023#discussion_r1017545708:
   "the ability to share resources is one of the main features that
   NVIDIA is excited about finally being able to support with DRA. We
   already have an out-of-tree DRA driver running with this
   functionality and have been communicating to customers that this is
   the (long-awaited) path towards finally doing GPU-sharing "the right
   way" in Kubernetes."

-- 
Best Regards

Patrick Ohly
Cloud Software Architect