After https://gerrit.wikimedia.org/r/#/c/119041/ we will be running the suggester with faster, less accurate settings. They'll be good enough to find lots of suggestions, but not all.
Proposal: if the search didn't find many results then rerun the suggester with more aggressive settings. Stuff like:
Also, we might want to only do this if the suggest run with the query didn't find anything. Not sure. We could also check the score on the result and use some cutoff. Dunno.
Note: prefix_length and max_errors can cause much more CPU usage. I'm suggesting just a single notch more on each setting but that still would be quite a bit.