Publicaciones

APQ Case Study

Period: Nov 2025 —

Shipped

The challenge

Large GraphQL schemas pay full request-body cost on every operation even when the operation is a repeat. The cost shows up as bandwidth, CPU on the server, and p95 latency on the wire.

My role

I published a case study on Automatic Persisted Queries at scale: 90.9% cache hit rate over a 135,504-line schema, 2,089 query types, 707 Prisma models, with a cgroup v2-isolated compile runner and a five-phase auto-warmup.

What I did

01
Two-layer cache (server-side APQ + client-side persist)
Why: Server-side APQ cuts the request body to a hash; client-side IndexedDB + Service Worker cache cuts the round trip entirely for repeat queries. The two layers cover different cases — same query, different transport.
Trade-off: Two cache surfaces to invalidate on schema change. The schema-hash invalidation pattern handles the bulk of the problem; the edge cases need manual review.
02
Auto-warmup pre-caches the hot 2,130 queries
Why: Cold-start latency is the cost a new replica pays until the cache fills. The five-phase auto-warmup pre-caches the queries most clients will hit in the first 30 seconds, so a replica warms up before the first real client arrives.
Trade-off: Five-minute startup ceremony per replica. Worth it for a 30-second cold-start improvement on the live traffic path.
03
cgroup v2 runner isolation for the compile step
Why: Schema compilation can OOM a shared node. The compile runner runs in a cgroup v2 slice with MemoryMax=16G so a compile spike cannot cascade into the production workers.
Trade-off: Compile runs slower in the slice than it would on bare metal. The latency cost is bounded; the safety win is total.

What changed

Cache hit rate in production
Before: 0%
After: 90.9%
Evidence: case-study-apq-performance telemetry, 2026-Q3
p95 latency reduction
Before: 25 ms without APQ
After: 12 ms with APQ
Payload reduction on cached queries
Before: 100% body
After: 25% body (75% reduction)
External SaaS spend for the APQ stack
Before: $0 — not yet implemented
After: $0/month — self-hosted Redis + Lua EVAL

Trade-offs

The cgroup v2 isolation adds operational surface. We chose it over a hosted compile service because the audit chain had to be self-hostable end to end.

What I learned

Two-layer caches pay back only when invalidation is honest. Schema-hash invalidation is the honest version; TTL-based invalidation is a lie that compounds over time.

Stack

GraphQL
Apollo Server v4
Redis 7
IndexedDB