Right now AbuseFilter provides per filter stats on mean run time and number of conditions utilized (or it would if it wasn't horribly broken per T53294). I would like to propose adding stats on max run time / conditions to help identify filters that are sometimes very slow but only rarely so. Since most filters have a series of fast early selections criteria (e.g. checking article namespace or not autoconfirmed user) it isn't always obvious how fast or slow the guts of the filter may be. Recording the longest run time in an observed sample would help to identify filters that occasionally perform badly. The present system tracks the mean properties of the last 10,000 actions, and also adding the max values to the statistics would not be difficult.