[twotwotwo]

> But the cache is a distraction. The performance you’re going to get will depend on your record sizes and update patterns and anyhow you probabl don’t care about the mean or median as much as the P99.

True your 99th percentile slowest requests won't hit the cache, and certainly that caching won't solve all your scaling difficulties.

However, keeping requests for commonly-needed data away from (say) a DB cluster decreases the load on it at a given level of throughput, and that can be good for P99, and (as the post notes) caching can specifically help with super-hot data which can cause problematic hotspots in some sharding strategies.

Obviously situations vary and there're limits, but a cache seems like a legit tool, not just a band-aid, for a decent number of situations.