[LUCENE-5750] Speed up monotonic address access in BINARY/SORTED_SET

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.9, 6.0
• Lucene Fields: New
• Assignee: Unassigned
• Reporter: Robert Muir
• Created: 10/Jun/14 17:15
• Updated: 28/Aug/22 14:09
• Resolved: 10/Jun/14 19:34

Description

I found this while exploring LUCENE-5748, but it currently applies to both variable length BINARY and SORTED_SET, so I think its worth it to do here first.

I think its just a holdover from before MonotonicBlockPackedWriter that to access element N we currently do:

startOffset = (docID == 0 ? 0 : ordIndex.get(docID-1));
endOffset = ordIndex.get(docID);

Thats because previously we didnt have packed ints that supported > Integer.MAX_VALUE elements. But thats been fixed for a long time. If we just write a 0 first and do this:

startOffset = ordIndex.get(docID);
endOffset = ordIndex.get(docID+1);

The access is then much faster. For sorting i see around 20% improvement. We don't lose any compression because we should assume the delta from 0 .. 1 is similar to any other gap N .. N+1