What’s a good cache hit ratio to aim for?

It depends on the workload and keyspace. For a hot, read-heavy endpoint, I typically expect **70–95%** once warmed. More important: segment hit ratio by keyspace/endpoint and correlate it to backend QPS and tail latency. A global “90% hit ratio” can hide a miss storm on the one endpoint that matters.

Should we fail open (bypass cache) or fail closed when Redis is down?

Per endpoint. For user-facing reads where stale is acceptable, prefer **serve stale** (SWR) or **fail open** to origin with strict rate limiting. For authorization/session-like data, you may need **fail closed** to avoid security issues. Write the policy down and test it.

Is in-process caching safe in Kubernetes with autoscaling?

Yes, if you bound it. Use a real LRU/TTL cache with size limits and measure memory. Expect lower hit ratios during scale events and deploys. In-process caches are great for hot keys but shouldn’t be your only layer if you need stability across pods.

How do we avoid data leaks with caching in multi-tenant systems?

Treat cache keys like you treat database queries: include **tenant + auth scope** (and any personalization dimension). Add versioning, and use automated tests that confirm tenant isolation. I’ve seen ‘just cache the response’ turn into a P0 incident because the key forgot one dimension.

When should we use Redis vs a CDN?

If the response is cacheable via HTTP and doesn’t require per-user personalization, start with **CDN/edge caching**—it’s cheaper and reduces load closer to the client. Use Redis for shared, authenticated, or non-HTTP caching needs, and for coordination (locks/coalescing) when stampedes matter.

Guides · Jan 14, 2026 · 7 minute read

Your Cache Isn’t a Performance Trick — It’s a Reliability System (If You Design It Right)

A pragmatic, layered approach to caching that cuts latency *and* reduces blast radius when downstreams misbehave.

Back to all posts

Your Cache Isn’t a Performance Trick — It’s a Reliability System (If You Design It Right)

Key takeaways

Implementation checklist