What you propose would catch many changes by looking at things like cpu and latency, but the space of potential hidden resource constraints is vast enough that without empirical verifician, we cannot have high confidence that performance regressions have not occurred (eg: lock contention can be not a problem at all until it is a huge problem...)
as the new version gets deployed and starts to take more traffic it doesnt keep up and is scaled up accordingly which establishes a new performance curve for the failover prediction system.
