Currently while sending pivot facet requests to each shard, the facet.pivot.mincount is set to 0 if the facet is sorted by count with a specified limit > 0. However with a mincount of 0, the pivot facet will use exponentially more wasted memory for every pivot field added. This is because there will be a total of limit^(# of pivots) pivot values created in memory, even though the vast majority of them will have counts of 0, and are therefore useless.
Imagine the scenario of a pivot facet with 3 levels, and facet.limit=1000. There will be a billion pivot values created, and there will almost definitely be nowhere near a billion pivot values with counts > 0.
This likely due to the reasoning mentioned in this comment in the original distributed pivot facet ticket. Basically it was thought that the refinement code would need to know that a count was 0 for a shard so that a refinement request wasn't sent to that shard. However this is checked in the code, in this part of the refinement candidate checking. Therefore if the pivot.mincount was set to 1, the non-existent values would either:
- Not be known, because the facet.limit was smaller than the number of facet values with positive counts. This isn't an issue, because they wouldn't have been returned with pivot.mincount set to 0.
- Would be known, because the facet.limit would be larger than the number of facet values returned. therefore this conditional would return false (since we are only talking about pivot facets sorted by count).
The solution, is to use the same pivot mincount as would be used if no limit was specified.
This also relates to a similar problem in field faceting that was "fixed" in SOLR-8988. The solution was to add a flag, facet.distrib.mco, which would enable not choosing a mincount of 0 when unnessesary. Since this flag can only increase performance, and doesn't break any queries I have removed it as an option and replaced the code to use the feature always.
There was one code change necessary to fix the MCO option, since the refinement candidate selection logic had a bug. The bug only occured with a minCount > 0 and limit > 0 specified. When a shard replied with less than the limit requested, it would assume the next maximum count on that shard was the mincount, where it would actually be the mincount-1 (because a facet value with a count of mincount would have been returned). Therefore the MCO didn't cause any errors, but with a mincount of 1 the refinement logic always assumed that the shard had more values with a count of 1.