[anarazel]

Entirely depends on the the type of workload / application architecture. In some cases the rollback ratios will be massive (I've seen 70%). In other cases adding such a retry loop is unattractive because the latency jitter. Or retaining all involved data for a retry is unattrictive.

Yet Another problem is that you need to enforce all sessions potentially involved in a data race need to use serializable; that can be easy or hard, depending on the scenario.