[jira] [Commented] (HIVE-12312) Excessive logging in PPD code

["Ashutosh Chauhan (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-12312?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14990523#comment-14990523 ] 

Ashutosh Chauhan commented on HIVE-12312:
-----------------------------------------

[~cartershanklin] Thanks for the patch. Patch needs to be name as per convention : https://cwiki.apache.org/confluence/display/Hive/Hive+PreCommit+Patch+Testing for Hive QA to trigger QA run on it.

> Excessive logging in PPD code
> -----------------------------
>
>                 Key: HIVE-12312
>                 URL: https://issues.apache.org/jira/browse/HIVE-12312
>             Project: Hive
>          Issue Type: Bug
>          Components: Hive
>    Affects Versions: 1.2.1
>            Reporter: Carter Shanklin
>            Priority: Minor
>         Attachments: ppd_debug.patch
>
>
> One of my very complex queries takes about 14 minutes to compile with PPD on. Profiling it I saw a lot of time spent in this stack which is called many many thousands of times.
> {code}
> java.lang.Throwable.getStackTraceElement(-2)
> java.lang.Throwable.getOurStackTrace(827)
> java.lang.Throwable.getStackTrace(816)
> sun.reflect.GeneratedMethodAccessor5.invoke(-1)
> sun.reflect.DelegatingMethodAccessorImpl.invoke(43)
> java.lang.reflect.Method.invoke(497)
> org.apache.log4j.spi.LocationInfo.<init>(139)
> org.apache.log4j.spi.LoggingEvent.getLocationInformation(253)
> org.apache.log4j.helpers.PatternParser$LocationPatternConverter.convert(500)
> org.apache.log4j.helpers.PatternConverter.format(65)
> org.apache.log4j.PatternLayout.format(506)
> org.apache.log4j.WriterAppender.subAppend(310)
> org.apache.log4j.DailyRollingFileAppender.subAppend(369)
> org.apache.log4j.WriterAppender.append(162)
> org.apache.log4j.AppenderSkeleton.doAppend(251)
> org.apache.log4j.helpers.AppenderAttachableImpl.appendLoopOnAppenders(66)
> org.apache.log4j.Category.callAppenders(206)
> org.apache.log4j.Category.forcedLog(391)
> org.apache.log4j.Category.log(856)
> org.apache.commons.logging.impl.Log4JLogger.info(176)
> org.apache.hadoop.hive.ql.ppd.OpProcFactory$DefaultPPD.logExpr(707)
> org.apache.hadoop.hive.ql.ppd.OpProcFactory$DefaultPPD.mergeWithChildrenPred(752)
> org.apache.hadoop.hive.ql.ppd.OpProcFactory$FilterPPD.process(437)
> {code}
> logExpr is set to log at INFO level, but I think DEBUG is more appropriate. When I set log level to debug I see > 20% speedup in compile time:
> Before:
> {code}
> real    14m47.972s
> user    15m25.609s
> sys    0m20.282s
> {code}
> After:
> {code}
> real    11m30.946s
> user    12m10.870s
> sys    0m7.320s
> {code}
> It looks like there's a lot of stuff in the PPD code that could be optimized, when I turn PPD off the query compiles in 2m 30s. But this seems like an easy and low risk win.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)