Performance engineering
Performance is a measurement discipline. Change the bottleneck you can show, not the one that is fashionable.
Personal experience
TRIKO’s backend stack included Redis, Memcached, and Supervisor — tools that exist to keep hot data close and workers alive. Cache hit ratios and latency budgets are not published (TODO_METRIC).
Technical perspective
Profile before you cache. A missing index, an N+1 query, and a JSON encode in a loop will not be saved by Redis. Caching a wrong query is how you serve wrong data faster.
Caches need a TTL, a stampede strategy, and a story for invalidation. Memcached and Redis fail differently; treating them as interchangeable “speed layers” is how you debug the wrong layer.
Supervisor (and its descendants) is performance in the operational sense: a dead worker is a throughput graph that goes to zero. Process management is part of the performance design, not an afterthought for ops.