mirror of
https://github.com/stack-auth/stack.git
synced 2026-06-04 21:04:37 +08:00
91b8e4caa4
Fixes Sentry [STACK-BACKEND-16H](https://stackframe-pw.sentry.io/issues/STACK-BACKEND-16H) — the `/api/v1/internal/metrics` endpoint was triggering the cluster's 10.8 GiB OvercommitTracker kill on tenants with months of `$token-refresh` history. ## Root cause Three queries in `loadAnalyticsOverview` plus `loadUsersByCountry` did `GROUP BY user_id` over the events table with **no lower `event_at` bound**, so their hash table working set scaled with cumulative-distinct-users-ever-seen instead of the 30-day metrics window. ## Changes - Add 30-day `event_at` lower bound to `loadUsersByCountry` and to the `analyticsUserJoin` inner subquery (used by `dailyEvents`, `totalVisitors`, `topReferrers`). - New `getClickhouseAdminClientForMetrics()` factory in `lib/clickhouse.tsx` with connection-level safety net: per-query + per-user memory caps, external GROUP BY spill, and `join_algorithm: 'grace_hash,parallel_hash,hash'` (grace_hash measured to give 48% memory reduction at zero latency cost — see benchmark notes in the file). - Inline comment + concrete next steps for the long-term fix (option C: stamp `is_anonymous` at ingest on page-view/click events, then drop the join entirely). - Extend `scripts/benchmark-internal-metrics.ts` with the historical-seed knob and three new modes (`BENCH_BACKFILL_COMPARE`, `BENCH_JOIN_ALGO_COMPARE`, plus the existing `BENCH_ROUTE_QUERIES` updated) used to validate the choices above. ## Benchmark — pre-PR vs post-PR Synthetic seed: 300k users × 9 events spread over 365 days (~2.7M events). | | pre-PR | post-PR | delta | |---|---:|---:|---:| | Sum peak memory | 2.18 GiB | 515 MiB | **4.3× less** | | Max query duration | 1293 ms | 101 ms | **12.8× faster** | | Sum CPU duration | 5119 ms | 394 ms | 13× less work | | Sum bytes read | 3.87 GiB | 929 MiB | 4.3× less I/O | Per-query at 300k users: - `analyticsOverview:dailyEvents` 561 → 44 MiB (12.8× less) - `analyticsOverview:totalVisitors` 560 → 50 MiB (11.2× less) - `analyticsOverview:topReferrers` 546 → 50 MiB (10.9× less) - `loadUsersByCountry` 388 → 44 MiB (8.9× less) ## Caveats - `loadDailyActiveSplitFromClickhouse` still scans all-history on its `min(event_at)` subquery. It can't be naively bounded — `first_date` is used to classify entities as new vs reactivated, and a 30d bound would silently mislabel old-but-active entities as "new." The new SETTINGS cap+spill it; the proper fix is option C (documented inline). - A user with a page-view but no `$token-refresh` in the last 30 days now falls through to `coalesce(NULL, 0)` and is classified non-anonymous. Token-refresh fires every few minutes per active session, so this is rare but not impossible (embedded SDKs that poll less frequently, sessions straddling the 30d boundary). - `max_memory_usage_for_user: 9 GB` trades "cluster-wide OvercommitTracker kill of a random query" for "clean per-user memory error attributed to the specific query." After our 30d bounds, no query is anywhere near 9 GB. ## Test plan - [x] `pnpm typecheck` passes - [x] `pnpm lint` passes - [x] `pnpm test run apps/e2e/tests/backend/endpoints/api/v1/internal-metrics.test.ts` — 9/10 pass; the 1 failure (`risk_scores` snapshot drift) reproduces on clean `dev` and is unrelated - [x] `pnpm test run apps/e2e/tests/backend/endpoints/api/v1/analytics-{events,events-batch,query}.test.ts apps/e2e/tests/backend/endpoints/api/v1/token-refresh-events.test.ts apps/e2e/tests/backend/performance/metrics.test.ts` — all passing tests pass; 10 pre-existing `PRODUCT_DOES_NOT_EXIST` setup failures reproduce on clean `dev` - [x] Benchmark `BENCH_ROUTE_QUERIES=1` at 300k users shows the deltas above <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Chores** * Improved internal metrics collection to use metrics-specific DB settings for more reliable, safer analytical reads. * Added guardrails to metrics queries to enforce time-window bounds and avoid unbounded scans. * Expanded benchmark modes (backfill and join-algo comparisons), extended perf seeding, and improved logging/retry behavior to capture more complete stats and reduce missing log rows. <!-- review_stack_entry_start --> [](https://app.coderabbit.ai/change-stack/hexclave/stack-auth/pull/1457?utm_source=github_walkthrough&utm_medium=github&utm_campaign=change_stack) <!-- review_stack_entry_end --> <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|---|---|---|
| .. | ||
| prisma | ||
| scripts | ||
| src | ||
| .env | ||
| .env.development | ||
| .eslintrc.cjs | ||
| .gitignore | ||
| instrumentation-client.ts | ||
| LICENSE | ||
| next.config.mjs | ||
| package.json | ||
| prisma.config.ts | ||
| tsconfig.json | ||
| vercel.json | ||
| vitest.config.ts | ||
| vitest.setup.ts | ||
