A distributed job scheduler, part 3: measuring before scaling
The spec said add a Redis hot index for speed. So I built it, measured it against the Postgres claim it would replace, and it came out 21% slower — because the thing it optimizes was never the bottleneck.