If only it always worked that way....
Differences in the test bed environment that are hard to reproduce can play a big part, for example latency and DB load.
(ie I have difficulty simulating the temps and high vibration my devices experience onboard a DASH / EVO heavy haul locomotives, in which normal grade PC parts only last hours before failing.)