Unfortunately, the Java code is much less reliable than the C++ code at reporting UMAs, so filtering by FailureReasons==0 eliminates all users. We need a metric that we can filter by in both control and experiment, so we can reduce selection bias.

Suggestion: Make PrecacheManager::OnGetUnfinishedWorkDone report an UMA (e.g. a new Precache.Events value) if IsInControlGroup(), just before OnDone(). This would be comparable, then, to Precache.Fetch.PercentCompleted on the experiment side.