[ https://issues.apache.org/jira/browse/LUCENE-10233?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Feng Guo updated LUCENE-10233: ------------------------------ Description: In low cardinality points cases, id blocks will usually store doc ids that have the same point value, and {{intersect}} will get into {{addAll}} logic. If we store ids as bitset, and give the IntersectVisitor bulk visiting ability, we can speed up addAll because we can just execute the 'or' logic between the result and the block ids. Optimization will be triggered when the following conditions are met at the same time: # leafCardinality = 1 # max(docId) - min(docId) <= 16 * pointCount (in order to avoid expanding too much storage) # no duplicate doc id I mocked a field that has 10,000,000 docs per value and search it with a 1 term PointInSetQuery, the build scorer time decreased from 71ms to 8ms. (WIP, Just post this first to see whether you think this optimization makes sense) was: In low cardinality points cases, id blocks will usually store doc ids that have the same point value, and intersect will get into addAll logic. If we store ids as bitset when the leafCadinality = 1, and give the IntersectVisitor bulk visit ability, we can speed up addAll because we can just execute the 'or' logic between the result and the block ids. I mocked a field that has 10,000,000 docs per value and search it with a PointInSetQuery with 1 term, the build scorer time decreased from 71ms to 8ms. Concerns: 1. Bitset could occupy more disk space.(Maybe we can force this optimization only works when block's (max-min) <= n * count?) 2. MergeReader will become a bit slower because it needs to iterate docIds one by one. > Store docIds as bitset when leafCardinality = 1 to speed up addAll > ------------------------------------------------------------------ > > Key: LUCENE-10233 > URL: https://issues.apache.org/jira/browse/LUCENE-10233 > Project: Lucene - Core > Issue Type: Improvement > Components: core/codecs > Reporter: Feng Guo > Priority: Major > Time Spent: 10m > Remaining Estimate: 0h > > In low cardinality points cases, id blocks will usually store doc ids that > have the same point value, and {{intersect}} will get into {{addAll}} logic. > If we store ids as bitset, and give the IntersectVisitor bulk visiting > ability, we can speed up addAll because we can just execute the 'or' logic > between the result and the block ids. > Optimization will be triggered when the following conditions are met at the > same time: > # leafCardinality = 1 > # max(docId) - min(docId) <= 16 * pointCount (in order to avoid expanding > too much storage) > # no duplicate doc id > I mocked a field that has 10,000,000 docs per value and search it with a 1 > term PointInSetQuery, the build scorer time decreased from 71ms to 8ms. > (WIP, Just post this first to see whether you think this optimization makes > sense) -- This message was sent by Atlassian Jira (v8.20.1#820001) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org