Using the facet.heatmap feature will internally constructs a Bits instance (to be passed down into Lucene) from the Solr-provided DocSet. The current procedure is sometimes fine but it can be optimized:
- if docSet is equal to SolrIndexSearcher.getLiveDocs then use a null Bits, which signals to the Lucene code that it can skip reading the postings and simply use the doc freq when there are no deleted docs either. The difference is huge.
- if docSet has size 0 then use Bits.MatchNoBits. After
LUCENE-7787is committed, this is a fast do-nothing optimization that could otherwise take a lot of time.
- if the docSet is not a BitDocSet then build a FixedBitSet. Presently the code wraps a Bits around the DocSet, likely a SortedIntDocSet, which is a huge performance hit for large shards when this happens – log(N) over possibly thousands of doc IDs for each doc accumulation
(This development was funded by the Harvard Center for Geographic Analysis as part of the HHypermap project)