Weight service overview error rates by sample rate

[Makisuo]

Summary

• Add estimatedErrorCount to service overview data from the query engine through mobile and web consumers.
• Compute overview error rates from weighted error spans when reliable sampling weights are available, with a fallback to raw retained-span rates.
• Update sampling documentation to describe the new error-rate behavior.
• Expand tests to cover weighted, unsampled, and stale-response cases.

Testing

• Added unit tests for the query engine SQL shape.
• Added unit tests for service overview aggregation in the web and mobile clients.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.

[pullfrog]

✅ No new issues found.

Reviewed changes — makes the service-overview error rate sampling-aware so aggressively-retained error spans no longer over-represent the error rate under mixed per-span sampling.

• Weighted error aggregate in SQL — adds estimatedErrorCount: sumIf(SampleRate, StatusCode = 'Error') in serviceOverviewQuery, mirroring the existing estimatedSpanCount: sum(SampleRate).
• Domain type — estimatedErrorCount: number added (required) to ServiceOverviewOutput in both @maple/query-engine and @maple/domain; the HTTP ServiceOverviewResponse.data is a passthrough Record<string, unknown> so the field reaches clients without a schema change.
• Client aggregation — apps/web/src/api/warehouse/services.ts and apps/mobile/lib/api.ts compute errorRate from weighted errors/spans when every row in a group carries a finite estimatedErrorCount, falling back to the raw retained-span rate otherwise (tolerates stale cached responses that predate the field).
• Docs + tests — docs/sampling-throughput.md updated to describe the weighting; tests cover the weighted (1/101), unsampled (weighted == raw), and stale-response fallback cases.

The change is correct and well-tested. estimatedErrorCount/estimatedSpanCount derive from sum(SampleRate) (Float64 → JSON number), so they sidestep the BYO-ClickHouse UInt64→string decode issue and the clients' Number() coercion covers both backends. The unsampled path is provably equivalent to the previous errorCount/spanCount behavior (all SampleRate = 1.0), the hasEstimatedErrors guard is conservatively all-or-nothing per group, and the totalEstimated > 0 fallback prevents divide-by-zero. Web and mobile implementations are identical. Nothing to change.

  ^{｜ ⚠️ this action is pinned to a commit SHA, which freezes the cleanup step — switch to @v0 or keep the SHA fresh with Dependabot ｜ View workflow run ｜ Using Claude Opus ｜ 𝕏}