Molina documentation
How to reproduce a failure you can't reproduce
A failure that happens once and never on demand is the one you cannot fix. You add a test and it passes. You change the timing and the failure disappears. The bug is still there, and so is the argument about what caused it.
Molina runs your software in small virtual computers whose disk, network and clock it controls. A replay code fixes which faults a run applies. Ask for that code again and the same faults run again.
What I measured
In one Postgres run on a disk that said saved before it was, the console, the timings and the switch log were byte for byte identical in 3 of 3 replays on the same seed. In Akka.NET, 3 pairs of a partition run and its replay matched on all 1,476 frames, with the same delivery outcomes and times. A NATS partition matched byte for byte in 3 of 3 replays.
A program with racing threads, restored from a saved starting point, gave the same result in all 90 counted worlds. The same program on an ordinary computer gave 50 different final counts in 50 runs.
Where it stops
Three pairs are too few to say how often a replay differs. The exact runs used one seed, one machine and one build, and some were made by hand. Exact mode refuses some setups today. A replay is evidence you can check, not a promise that every failure comes back.
How exact is the replay? What it cannot do
Bring one failure you cannot reproduce to the free call.