How do I choose TTLs without breaking freshness?

Classify data. For content that changes infrequently (marketing pages), 5–30 minutes at edge is safe with surrogate-key purges on publish. For semi-dynamic data (catalog, non-personalized pricing), 1–10 minutes edge, 10–30 minutes service cache, with SWR 10–60 minutes. For highly dynamic or personalized data (cart, balances), avoid shared edge caching; use short in-process/Redis TTLs and ETags for conditional requests.

What’s the fastest way to see ROI from caching?

Start at the edge: standardize `Cache-Control`, strip noisy cookies, enable SWR, and add surrogate keys. We routinely see 25–40% cost reductions and 2–3x p95 improvements in under 2 weeks. Then move to service-level cache-aside for your most expensive aggregations.

How do I avoid cache stampedes during traffic spikes?

Enable request coalescing at the proxy (Varnish/Envoy/Nginx), add Redis locks (`SET NX PX`) around refresh, keep a stale copy for 30–60 minutes, and add TTL jitter. Canary TTL/policy changes and monitor miss penalty. Pre-warm hot keys on deploy.

Is caching safe with authenticated users and GDPR/PII?

Yes, with boundaries. Don’t put PII or personalized responses in shared caches. Use `Cache-Control: private` for personalized content, segment keys by tenant/org, and encrypt at rest for Redis. Audit keys/values for sensitive fields and set short TTLs for any token-bearing entries. Document a data retention policy.

Can GraphQL be cached effectively?

Edge caching of GraphQL is hard because most clients POST and responses mix fields. Cache at the resolver/service layer: hot entities and lists keyed by args. Support GET for idempotent queries with persisted query hashes; normalize arguments; then cache GETs at edge/proxy with short TTLs and SWR.

Performance-optimization · Nov 9, 2025 · 9 minute read

Stop Recomputing the Same Bytes: Caching Architectures That Cut p95 In Half and Your Cloud Bill by a Third

Design a cache stack that hits user-facing SLOs and trims your infra bill—without gambling on consistency.

Back to all posts

Stop Recomputing the Same Bytes: Caching Architectures That Cut p95 In Half and Your Cloud Bill by a Third

Key takeaways

Implementation checklist