I've faced that problem before: Someone claimed their autoscaling algorithm had to be working correctly, as he had written a proof and had a working simulator b

I've faced that problem before: Someone claimed their autoscaling algorithm had to be working correctly, as he had written a proof and had a working simulator behaving perfectly. And yes, the code matched the algorithm as written, and the proof was correct... as long as latency was zero. Once I added lags to the right places in the simulator, we got the exact same problems than in production.