timeseries_predicates.js takes a few minutes to run, because it tests many combinations: many ways to bucket a handful of events.

It compares results against a normal collection. I think we can avoid dropping and recreating the normal collection in many cases.