fix(resilience-ranking): return warmed scores from memory, skip lossy re-read

[koala73]

Summary

Fixes the silent ranking publish failure that leaves resilience:ranking:v9 absent while the seeder keeps writing a "fresh" seed-meta:resilience:ranking, producing resilienceRanking: { status: EMPTY_ON_DEMAND, records: 0, seedAgeMin: 277 } in /api/health even when the cron is firing normally.

Root cause

server/worldmonitor/resilience/v1/get-resilience-ranking.ts warmed ~222 country scores, then immediately re-read them via getCachedResilienceScores (a /pipeline GET) to compute coverage. Upstash REST has documented write→re-read visibility lag within a single Vercel invocation (PR #3057, feedback_upstash_write_reread_race_in_handler.md). The pipeline GET could return 0 results despite every SET succeeding — coverage collapsed to 0% < 75% threshold, handler logged [resilience] ranking not cached … and skipped both the ranking SET and the meta SET. Seeder then wrote its own meta anyway (based on its own pipeline GET, which by then had propagated), leaving the aggregate permanently absent.

Fix

• warmMissingResilienceScores now returns Map<countryCode, GetResilienceScoreResponse> with every successfully computed score.
• Handler merges warm results into cachedScores directly and drops the post-warm re-read — coverage reflects what was actually computed in-memory in this invocation.

No API changes, no schema changes, ~20 lines across 2 files.

Test plan

• npx tsx --test tests/resilience-ranking.test.mts — 9/9 pass (8 existing + 1 new race regression).
• Full resilience suite: 369/369 pass.
• New test publishes ranking via in-memory warm results even when Upstash pipeline-GET lags after /set writes (race regression) simulates the exact Upstash lag: pipeline-GET of score keys returns null after /set writes. Fails against old code, passes with the fix.
• After deploy, probe Redis: TTL resilience:ranking:v9 should be positive; /api/health should show resilienceRanking: { status: OK, records: ~150+, seedAgeMin: <360 }.
• Seeder's [resilience] ranking not cached — coverage X/222 below 75% threshold warnings should stop appearing.

Follow-up (not in this PR)

Slice B: make seed-resilience-scores.mjs write seed-meta:resilience:ranking only when the ranking key is actually present (currently it lies about freshness based on per-country score count).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
worldmonitor	Ignored	Preview	Apr 16, 2026 5:59am

[greptile-apps]

Greptile Summary

Fixes a silent ranking-publish failure caused by an Upstash REST write→read visibility lag: after warming missing scores, the handler now merges in-memory results directly into cachedScores instead of issuing a second pipeline GET that could return zero results and collapse coverage below the 75% gate. The change is minimal (~20 lines), well-commented, and accompanied by a regression test that reproduces the exact race condition.

Confidence Score: 5/5

Safe to merge — targeted fix with no API or schema changes and a direct regression test.

All changes are narrowly scoped to the race condition path. The return-type widening of warmMissingResilienceScores is backward-compatible, const instead of let is correct since the map is mutated not reassigned, country-code normalisation is idempotent so key alignment is guaranteed, and the scoreInterval field on warm results is not consumed by buildRankingItem (it reads from the separate fetchIntervals map), so no data inconsistency is introduced. No P0/P1 findings.

No files require special attention.

Important Files Changed

Filename	Overview
server/worldmonitor/resilience/v1/_shared.ts	Return type of warmMissingResilienceScores changed from void to Map<string, GetResilienceScoreResponse>; fulfilled results are now collected and returned to the caller for in-memory merge. Logic and normalization are correct.
server/worldmonitor/resilience/v1/get-resilience-ranking.ts	Drops the post-warm getCachedResilienceScores re-read; merges the in-memory warmed map into cachedScores directly. Coverage gate now reflects in-memory truth rather than a Upstash-lagged pipeline GET.
tests/resilience-ranking.test.mts	Adds a targeted race-regression test that intercepts all pipeline-GET calls for score keys and returns null, directly reproducing the Upstash lag; verifies ranking and seed-meta are published despite the simulated lag.

Sequence Diagram

sequenceDiagram
    participant H as Handler
    participant R as Redis (Upstash)
    participant W as warmMissing...

    Note over H,R: Before fix — pipeline-GET lag causes silent ranking drop
    H->>R: getCachedResilienceScores (pipeline GET)
    R-->>H: [] (lag: no scores yet)
    H->>W: warmMissingResilienceScores(missing)
    W->>R: ensureResilienceScoreCached × N (SET writes)
    R-->>W: OK
    W-->>H: void
    H->>R: getCachedResilienceScores (pipeline GET) ← race
    R-->>H: [] (Upstash lag: writes not yet visible)
    Note over H: coverage = 0/222 < 75% → ranking NOT published ❌

    Note over H,R: After fix — in-memory merge bypasses re-read lag
    H->>R: getCachedResilienceScores (pipeline GET)
    R-->>H: [] (lag: no scores yet)
    H->>W: warmMissingResilienceScores(missing)
    W->>R: ensureResilienceScoreCached × N (SET writes)
    R-->>W: OK
    W-->>H: Map<countryCode, score> (in-memory)
    H->>H: merge warmed into cachedScores
    Note over H: coverage = 222/222 = 100% ≥ 75% → ranking published ✅
    H->>R: SET resilience:ranking:v9 + seed-meta

Loading

_{Reviews (1): Last reviewed commit: "fix(resilience-ranking): return warmed s..." | Re-trigger Greptile}