When replaying customer queries against the next version, how do you compare results for queries using now() or approximate aggregates? Curious how you separate expected differences from actual regressions.