You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Replace repeated high-cardinality GROUP BY for the no-trace-ID selection path with maintained trace summaries/candidates. Keep error priority, elapsed bounds across selected spans, service/namespace scope, and late/incomplete traces. Do not substitute a root-span heuristic or shorten the requested window.
Validation: compare selected trace against the current raw query across complete and partial windows, multiple batches/services/namespaces, cross-boundary traces, late updates, retention/compaction, and cancellation. Benchmark candidate selection on millions of traces; avoid rebuilding every trace group per request.
Implementation is authorized in PR #272; this issue tracks the fix and its validation.
Implemented an exact candidate index for all namespace/service scope combinations, including traces spanning batches. Candidate reads merge uncached contributions and recompute partial-window traces; they preserve error priority and duration ranking, with a deterministic trace-ID tie. Retired markers disable the index until rebuilt. Parity tests compare production and raw selection across scopes, clipped windows, late spans, compaction, and retention. Broad trace median improves 60.09 to 19.48ms on identical files; narrow latency is approximately flat.
Activity
vishr commentedon Oct 1, 2026
Implemented an exact candidate index for all namespace/service scope combinations, including traces spanning batches. Candidate reads merge uncached contributions and recompute partial-window traces; they preserve error priority and duration ranking, with a deterministic trace-ID tie. Retired markers disable the index until rebuilt. Parity tests compare production and raw selection across scopes, clipped windows, late spans, compaction, and retention. Broad trace median improves 60.09 to 19.48ms on identical files; narrow latency is approximately flat.
Implemented in ea13675; tracked in #275. Measurement report and reproduction.
just check, the full Go race suite, and native Linux ARM64 full Go tests pass. Issue closure is linked to merge of #272.