Summary

Solves #408

Six optimizations in the PCS / PCSG / PodClique reconcile paths. Measured on the 500-replica / 1000-pod scale test (k3d + 100 KWOK nodes, operator concurrency=20).

An earlier revision of this PR included a per-PodClique / per-PCSG spec-hash short-circuit (annotation grove.io/spec-hash). It was dropped after CR feedback flagged the hash as fragile (hand-picked field selection, MNNVL exclusion that depends on a webhook invariant, drift risk if buildResource changes silently). The remaining changes capture the same wall-clock improvement on this benchmark; steady-state CPU regresses ~10 ms/sec versus the short-circuited variant but stays ~50% under baseline.

Wall-clock

CPU — deploy phase (full pprof, 500 PodCliques + 500 PodGangs being created)

CPU — steady-state phase (30s window, single no-op reconcile triggered by annotation patch)

The common production case: a watch event fires but nothing actually changed.

A note on wall-clock vs. CPU

The wall-clock reduction is somewhat smaller than the CPU reduction because this test is not CPU-bound: the operator spends most of its wall-clock time waiting on kube-apiserver latency, informer cache syncs, and KWOK's own pod-scheduling delays. Those don't shrink just because our client-side code got faster. Where the CPU savings pay off in production:

• Operator headroom: lower baseline CPU means less resource pressure under bursts and tighter container limits
• Multi-tenant density: at N concurrent PCSes the per-reconcile savings scale with N
• Event storms: steady-state −46% on a no-op reconcile is exactly the common case where an unrelated watch event wakes the controller
• Larger clusters: at 10k+ PodCliques the CPU-per-reconcile would start gating wall-clock

All unit tests pass. No semantic changes to reconcile behavior (behavior is asserted via equality.Semantic.DeepEqual before skipping status writes, ownership still validated by metav1.IsControlledBy, etc.).

What's in the PR

Parallel component sync — three dependency-ordered groups in PCS reconcilespec.go:

1. RBAC + infra (ServiceAccount, Role, RoleBinding, SATokenSecret, HeadlessService, HPA, PCSReplica, ComputeDomain, ResourceClaim)
2. PodClique
3. PCSG + PodGang

Instead of 12 sequential Syncs. Mostly benefits the delete phase (delete wall-clock −24%) where cleanup tasks overlap.

GetPCLQPods narrowed selector — MatchingLabels{LabelPodClique} only. That label is unique cluster-wide, so the parent managed-by/part-of labels only added per-pod labels.Set.Lookup work for no filtering benefit. Ownership is still validated by metav1.IsControlledBy.

ctx-scoped memoization — WithPCLQPodsCache and WithPodCliqueSetCache in internal/controller/common/component/utils/. PodClique's reconcileSpec and reconcileStatus share one pod list; GetPodCliqueSet drops from 4 calls per PodClique reconcile (and 3 per PCSG reconcile) to 1.

No-op status skip — in podclique/pcs/pcsg reconcileStatus, snapshot status before mutations, skip the API round-trip when equality.Semantic.DeepEqual says nothing changed.

Test additions

To make the steady-state CPU measurable in a scale test:

• operator/e2e/measurement/condition/timer.go — TimerCondition holds the measurement window open for a fixed duration after a trigger phase fires.
• WorkloadManager.TriggerPCSReconcile in operator/e2e/grove/workload/workload.go — bumps grove.io/reconcile-trigger annotation to force a no-op reconcile cycle without touching the spec.
• New steady-state-reconcile phase in operator/e2e/tests/scale/scale_test.go — patches the annotation, waits 30s, and pprof captured during that window isolates the cache-hit reconcile cost.

🤖 Generated with Claude Code