New, Patch Available
Most implementations of the LeafReader cache an instance of FieldInfos which is returned in the LeafReader.getFieldInfos() method. There are a few places that currently do not and this can cause performance problems.
The most notable example is the lack of caching in Solr's SlowCompositeReaderWrapper which caused unexpected performance slowdowns when trying to use Solr's JSON Facets compared to the legacy facets.
This proposed change is mostly relevant to Solr but touches a few Lucene classes. Specifically:
1. Adds a check to TestUtil.checkReader to verify that LeafReader.getFieldInfos() returns the same instance:
I'm not entirely sure this is wanted or needed but adding it uncovered most of the other LeafReader implementations that were not caching FieldInfos. I'm happy to remove this part of the patch though.
2. Adds a FieldInfos.EMPTY that can be used in a handful of places
There are several places in the Lucene/Solr tests that were creating empty instances of FieldInfos which were causing the check in #1 to fail. This fixes those failures and cleans up the code a bit.
3. Fixes a few LeafReader implementations that were not caching FieldInfos
- MemoryIndex.MemoryIndexReader - The constructor was already looping over the fields so it seemed natural to just create the FieldInfos at that time
- SlowCompositeReaderWrapper - This was the one causing me trouble. I've moved the caching of FieldInfos from SolrIndexSearcher to SlowCompositeReaderWrapper.
- CollapsingQParserPlugin.ReaderWrapper - getFieldInfos() is immediately called twice after this is constructed
- ExpandComponent.ReaderWrapper - getFieldInfos() is immediately called twice after this is constructed
4. Minor Solr tweak to avoid calling SolrIndexSearcher.getSlowAtomicReader in FacetFieldProcessorByHashDV. This change is now optional since SlowCompositeReaderWrapper caches FieldInfos.