On Thu, Apr 28, 2016 at 11:50 AM, Nick Vasilyev <nick.vasily...@gmail.com> wrote: > mmfr_exact is a string field. key_phrases is a multivalued string field.
One guess is that top-level field caches (and UnInvertedField use) were removed in https://issues.apache.org/jira/browse/LUCENE-5666 While this is better for NRT (a quickly changing index), it's worse in CPU, and can be worse in memory overhead for very static indexes. Multi-valued string faceting was hit hardest: https://issues.apache.org/jira/browse/SOLR-8096 Although I only measured the CPU impact, and not memory. The 4.x method of faceting was restored as part of https://issues.apache.org/jira/browse/SOLR-8466 If this is the issue, you can: - try reindexing with docValues... that should solve memory issues at the expense of some speed - upgrade to a more recent Solr version and use facet.method=uif for your multi-valued string fields -Yonik