We could also record a sparse histogram of specifically which user actions were dropped.  This might more easily allow us to identify candidate actions that should not be recorded as "user actions".

+1

We can use the same hashing as UserActions.Counts server-side so that the two are easy to compare.

We should probably record two histograms, one for which ones are dropped and one for which are kept/total, since I think initial metrics logs will contain histograms but not user actions, so the data might get harder to compare with UserActions.Counts

For the record, this problem started with M-63.
https://uma.googleplex.com/p/chrome/timeline_v2/?sid=f247e1cc2dfc70bb88ac6f5b1bba5479

Also, jwd@ added logging for what actions are truncated.
https://uma.googleplex.com/histograms?endDate=20180204&dayCount=1&histograms=UMA.TruncatedEvents.UserAction.Type&fixupData=true&showMax=true&filters=platform%2Ceq%2CW%2Cchannel%2Ceq%2C1%2Csimple_version%2Cone_of%2C66.0.3338.0%7C66.0.3339.0%2Cisofficial%2Ceq%2CTrue&implicitFilters=isofficial

Based on the data, and ss I said on an e-mail thread, I think the likely problem is InProductHelp.*, as that feature started launching at around this time (in M-63?).  I think many of those actions are spammy.