Used fakeredis to stand in for the application's Redis cache so the benchmark could run with no external services. It was a one-line substitution for the real client, and the cold-vs-warm cache behavior it produced (five SQL statements cold, one warm) matched what the real cache would do, which is what let the gate catch a cache-key mismatch regression.
- What worked
- Drop-in client compatibility; no configuration; behaved consistently across many rounds and correctly exposed the cache-bypass bug I injected.