[perfloop]

Problem

Cluster-wide ListWithValues could return a terminal page (has_more=false) after an active shard could not be queried. Because callers use has_more to decide whether to request another page, they could stop before seeing keys held by that shard.

Change

Value pages now return an error instead of merging a partial response when client acquisition or a shard's local value-list operation fails. Cancellation and the no-active-node error are unchanged. Keys-only List retains its existing best-effort behavior.

Tests

The assertion is violated on the comparison and satisfied with this change: For a cluster-wide public CacheService.ListWithValues request with two active shards holding distinct matching keys and a live caller context, a page is returned only if every active shard responds: a peer client-acquisition or ListLocalWithValues RPC failure returns an error and no response; healthy shards return both sorted values in a terminal page, caller cancellation remains context.Canceled, and zero active shards retains the existing error.

Checks: 5 passed.

• make test-server, including TestCacheService_ListWithValuesRequiresEveryActiveShard
• make test-race TEST=TestCacheService_ListWithValuesRequiresEveryActiveShard
• make test-coverage TEST=TestCacheService_ListWithValuesRequiresEveryActiveShard
• make build, make lint, and make license-check

Target CI retains the full Test / build race and coverage runs, integration and end-to-end suites, Benchmark Cluster 3-Node / benchmark-cluster-3n, Benchmark Single Node / benchmark-single, DCO / dco, and Lint PR Title / lint.

---

Generated by Perfloop. Measurements and checks. Reproduce this comparison (tenant access required).

---

Note

Medium Risk
Changes cluster list semantics for value pages (stricter errors vs partial success) while keys-only list stays best-effort; incorrect rollout could surface more failures to callers that previously got incomplete pages.

Overview
Cluster-wide ListWithValues no longer returns a “complete” page when an active shard cannot be queried. fetchFromAllNodes now fails the whole value-list fan-out if any node’s client lookup or ListLocalWithValues RPC errors, instead of skipping that shard and merging a partial result that could show has_more=false while keys on the failed shard were never seen.

Keys-only cluster List behavior is unchanged: per-node failures are still logged and skipped. Context cancellation and the existing no active nodes error are unchanged.

Adds TestCacheService_ListWithValuesRequiresEveryActiveShard, a multi-node integration test covering healthy merged pages, peer RPC/client failures, all-shard failure, cancellation, and zero active nodes.

^{Reviewed by Cursor Bugbot for commit 7f1a524. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Changes how cluster-wide list operations handle node failures.

The PR appears safe to merge; no actionable regression was identified.

Summary

The PR makes cluster-wide ListWithValues reject a page when an active shard cannot be queried, while leaving keys-only List best-effort.

• Adds multi-node coverage for healthy results, shard failures, cancellation, and no active nodes.

_{Reviews (1) · Last reviewed commit: "fix: reject incomplete cluster value pag..."}