kubernetes-sigs / karpenter / 35279361127

84%

main: 83%

Build Type

Pull #3311

github

Committed by DerekFrank

Commit Message

feat: make node repair a voluntary disruption method

Retire the standalone node.health controller and its hardcoded 20% breaker;
node repair now runs as a voluntary disruption method (reason "Unhealthy"):

- Budgeted via spec.disruption.budgets (default 10%); the Unhealthy budget
  excludes still-unhealthy nodes from its own count while other reasons keep
  counting them. Retains a 20% circuit breaker as an interim correlated-failure
  backstop (to be replaced by #3031); the budget paces concurrent repairs under it.
- Ordered by rank + age/tau: score() argmaxes over a node's ELIGIBLE conditions
  (not just its highest-priority one) over a dense-ranked integer-priority set.
- matchRepairPolicy returns the highest-priority ELIGIBLE condition, governing
  eligibility and the drain-bound action; the breaker's census counts matches
  regardless of toleration.
- Pre-spins a replacement and terminates the original only once it is healthy
  (replace-then-terminate); a replacement that never comes up healthy never
  terminates the original.
- Vetoable via a do-not-repair annotation, distinct from do-not-disrupt (which
  repair ignores); repair is never stranded by a blocking pod.
- Per-candidate TerminationGracePeriod (min of matched policy and NodeClaim TGP)
  stamped as the termination deadline at actual deletion time.
- Registered only when the NodeRepair feature gate is on; NewRepair panics if the
  provider defines no RepairPolicies.
- Re-adds karpenter_nodeclaims_unhealthy_disrupted_total, emitted at termination.

Follow-ups: correlated-failure restraint / backoff (#3031, #3178); reason-aware
matching (#3263); termination contract (#3029); do-not-disrupt overlap (#2424).

Pull Request Pull Request #3311: feat: node repair as voluntary disruption

Coverage Stats

216 of 238 new or added lines in 8 files covered. (90.76%)

1 existing line in 1 file now uncovered.

17536 of 21002 relevant lines covered (83.5%)

78157.91 hits per line