[ https://issues.apache.org/jira/browse/LUCENE-7882?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16597756#comment-16597756 ]
Tony Xu commented on LUCENE-7882: --------------------------------- A simple idea here can be caching the compiled Expression instances. One way of doing it is to have a class that wraps the *JavascriptCompiler* and it makes sure one unique expression text only get compiled once and cached. Note that errors can be cached, too, because same invalid expression should fail at exactly the same point (e.g the ParseException)! Future compilation of the same expression should return the cached compilation result. This new class can be a static so we cache compilation in JVM scope or it can be instantiable so the caller decides the scope. > Maybe expression compiler should cache recently compiled expressions? > --------------------------------------------------------------------- > > Key: LUCENE-7882 > URL: https://issues.apache.org/jira/browse/LUCENE-7882 > Project: Lucene - Core > Issue Type: Improvement > Components: modules/expressions > Reporter: Michael McCandless > Priority: Major > > I've been running search performance tests using a simple expression > ({{_score + ln(1000+unit_sales)}}) for sorting and hit this odd bottleneck: > {noformat} > "pool-1-thread-30" #70 prio=5 os_prio=0 tid=0x00007eea7000a000 nid=0x1ea8a > waiting for monitor entry [0x00007eea867dd000] > java.lang.Thread.State: BLOCKED (on object monitor) > at > org.apache.lucene.expressions.js.JavascriptCompiler$CompiledExpression.evaluate(_score > + ln(1000+unit_sales)) > at > org.apache.lucene.expressions.ExpressionFunctionValues.doubleValue(ExpressionFunctionValues.java:49) > at > com.amazon.lucene.OrderedVELeafCollector.collectInternal(OrderedVELeafCollector.java:123) > at > com.amazon.lucene.OrderedVELeafCollector.collect(OrderedVELeafCollector.java:108) > at > org.apache.lucene.search.MultiCollectorManager$Collectors$LeafCollectors.collect(MultiCollectorManager.java:102) > at > org.apache.lucene.search.Weight$DefaultBulkScorer.scoreAll(Weight.java:241) > at > org.apache.lucene.search.Weight$DefaultBulkScorer.score(Weight.java:184) > at org.apache.lucene.search.BulkScorer.score(BulkScorer.java:39) > at org.apache.lucene.search.IndexSearcher.search(IndexSearcher.java:658) > at org.apache.lucene.search.IndexSearcher$5.call(IndexSearcher.java:600) > at org.apache.lucene.search.IndexSearcher$5.call(IndexSearcher.java:597) > at java.util.concurrent.FutureTask.run(FutureTask.java:266) > at > java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142) > at > java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) > at java.lang.Thread.run(Thread.java:745) > {noformat} > I couldn't see any {{synchronized}} in the sources here, so I'm not sure > which object monitor it's blocked on. > I was accidentally compiling a new expression for every query, and that > bottleneck would cause overall QPS to slow down drastically (~4X slower after > ~1 hour of redline tests), as if the JVM is getting slower and slower to > evaluate each expression the more expressions I had compiled. > I tested JDK 9-ea and it also kept slowing down over time as the performance > test ran. > Maybe we should put a small cache in front of the expressions compiler to > make it less trappy? Or maybe we can get to the root cause of why the JVM > slows down more and more, the more expressions you compile? > I won't have time to work on this in the near future so if anyone else feels > the itch, please scratch it! -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org