I used to think: “Add a cache. Latency solved.” ⚡
We added a cache. Then I spent the next few months learning cache. 💀
1. The cache becomes a second source of truth
The same data now lives in two places. Every write has to update or evict the cache entry. Miss one eviction path and users see stale data.
2. Expiry is a traffic event
One configuration key expired. Hundreds of requests hit the database in the same millisecond. Latency went from 200 ms to 2 seconds.
Nothing else had changed. Handling cache expiry was equally important.
3. The cache is no longer optional
It started as “make it faster.” Six months later, the database couldn’t survive full traffic.
Redis goes down → the service goes down.
That’s not just a cache. That’s a dependency.
None of this means “don’t cache.” We still cache everything.
But “add a cache” is not the design decision. The design decision is what happens when it’s empty, stale, or gone.
Adding a cache is only the happy path. Production shows you the other three. ⚡