kubernetes-sigs / karpenter / 25582899070

Pull #3008

github

feat(scheduling): add opt-in NodeClaim optimization pass

Adds an opt-in post-scheduling optimization pass behind the
NodeClaimOptimization feature gate that detects when a NodeClaim's
greedy instance-type growth left earlier pods on a machine larger
than they need, and splits the claim at the transition point.
Displaced pods are re-queued and re-scheduled onto a fresh,
right-sized NodeClaim.

The full design, alternatives, and correctness argument live in
designs/nodeclaim-optimization.md.

Off by default. Enable via FEATURE_GATES=NodeClaimOptimization=true.
The disruption controller's simulated scheduler is intentionally
not wired to the gate — consolidation rejects multi-NodeClaim
replacements, so running the splitting pass inside a consolidation
simulation would erase valid consolidation opportunities.

Signed-off-by: Michael Manders <mmanders@amazon.com>

287 of 359 new or added lines in 8 files covered. (79.94%)

12461 of 15524 relevant lines covered (80.27%)

118437.38 hits per line