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
{{ message }}
Repository navigation
perf(observability): roll up log histogram counts by scoped severity #276
Cache per-minute counts by namespace, service, and severity for requests without free-text search. Merge exact partial-minute boundaries and uncached/late batches. Free-text searches match redacted bodies and must retain their raw predicate.
Validation: entries, bucket counts, and matched totals remain consistent for unfiltered/severity/service/namespace/search cases, nonaligned bounds, mixed severity spelling, late data, restart and publication changes. Measure scan rows/CPU as retained data grows.
Implementation is authorized in PR #272; this issue tracks the fix and its validation.
Implemented scoped severity minute counts plus exact boundary-time counts and uncached active files. Duplicate timestamps retain their multiplicity. Free-text searches continue counting matches on redacted bodies. Tests compare raw/production entries and histogram data across namespaces, services, severities, searches, clipped windows, cold/warm caches, and late publication. Broad/narrow log medians improve 60.60 to 18.13ms and 38.56 to 15.25ms on identical files.
Final parity hardening is in cd916d5: caches retain the original severity text, so case-insensitive filters and uppercase histogram groups reproduce raw SQL even for Unicode labels. Tests cover i/İ, duplicates, searches, scopes, and exact boundaries; query/observability race tests pass. The repeated identical-file benchmark gives broad endpoint/trace/log medians 32.38/20.04/18.21ms, with unchanged qualitative limits. The report and JSON retain both trials.
64e3ce9 also clarifies that skip-to-latest applies to service/edge rollups; completed-batch caches still rebuild retained history. Both commits passed the full normal pre-push gate.
Activity
vishr commentedon Oct 1, 2026
Implemented scoped severity minute counts plus exact boundary-time counts and uncached active files. Duplicate timestamps retain their multiplicity. Free-text searches continue counting matches on redacted bodies. Tests compare raw/production entries and histogram data across namespaces, services, severities, searches, clipped windows, cold/warm caches, and late publication. Broad/narrow log medians improve 60.60 to 18.13ms and 38.56 to 15.25ms on identical files.
Implemented in ea13675; tracked in #276. 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.vishr commentedon Oct 1, 2026
Final parity hardening is in cd916d5: caches retain the original severity text, so case-insensitive filters and uppercase histogram groups reproduce raw SQL even for Unicode labels. Tests cover i/İ, duplicates, searches, scopes, and exact boundaries; query/observability race tests pass. The repeated identical-file benchmark gives broad endpoint/trace/log medians 32.38/20.04/18.21ms, with unchanged qualitative limits. The report and JSON retain both trials.
64e3ce9 also clarifies that skip-to-latest applies to service/edge rollups; completed-batch caches still rebuild retained history. Both commits passed the full normal pre-push gate.
cd916d5c
64e3ce9d