[
https://issues.apache.org/jira/browse/LUCENE-7828?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16012600#comment-16012600
]
Alan Woodward commented on LUCENE-7828:
---------------------------------------
Yes, this looks like it only affects range fields, so the various shortcuts can
be implemented in RangeFieldQuery's IntersectVisitor rather than changing the
core API. This change gives me a 35% speedup on this particular dataset:
{code}
@@ -136,6 +136,11 @@ public void visit(int docID, byte[] leaf) throws
IOException {
}
@Override
public Relation compare(byte[] minPackedValue, byte[]
maxPackedValue) {
+ if (Arrays.equals(minPackedValue, maxPackedValue)) {
+ if (queryType == QueryType.CONTAINS &&
comparator.isWithin(minPackedValue)) {
+ return Relation.CELL_INSIDE_QUERY;
+ }
+ }
byte[] node = getInternalRange(minPackedValue,
maxPackedValue);
// compute range relation for BKD traversal
if (comparator.isDisjoint(node)) {
{code}
> Improve PointValues visitor calls when all docs in a leaf share a value
> -----------------------------------------------------------------------
>
> Key: LUCENE-7828
> URL: https://issues.apache.org/jira/browse/LUCENE-7828
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Alan Woodward
> Assignee: Nicholas Knize
>
> When all the docs in a leaf node have the same value, range queries can waste
> a lot of processing if the node itself returns CELL_CROSSES_QUERY when
> compare() is called, in effect performing the same calculation in visit(int,
> byte[]) over and over again. In the case I'm looking at (very low
> cardinality indexed LongRange fields), this causes something of a perfect
> storm for performance. PointValues can detect up front if a given node has a
> single value (because it's min value and max value will be equal), so this
> case should be fairly simple to identify and shortcut.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
