I'll repeat what I wrote on the Phabricator task:
It probably goes without saying that AbuseFilter is indispensable. Especially on larger wikis, it automates a great deal of work for us, and I hear horror stories of what the patrollers went through before AbuseFilter was a thing. Ongoing maintenance and support of this extension should be paramount.
By design there's always potential for performance implications, but I would argue that under the current or a similar system, the primary responsibility of keeping it efficient lies with the filter author. The issue here is that often the author is unaware they are slowing anything down. T177017 and T161059 helped us on enwiki, but the this profiling is still absent on many other wikis. Ideas for further improvements can be found at T176895, T102903, T179604, and T175954.
I would like to also stress that the extension seems to be slowly degrading, or has had regressions introduced over time. On enwiki at least, we've found malfunctions are more common these days, and nearly something we've come to expect. See T175933 for a parent task.