Owen, we're seeing a millions of those log entries. There are three, one for each package listed below in the revised hive-log4j.settings., one full example provided below. Seems to repeat fewer than per-row (that would be billions). Perhaps repeats for each and every partition in a table (1000's to 10'000s).
To reproduce, create a y/m/d partitioned table and do select * from <table> where year=2015 and month=4 and day = 6 limit 1; - Douglas From: Owen O'Malley <[email protected]<mailto:[email protected]>> Reply-To: <[email protected]<mailto:[email protected]>> Date: Mon, 6 Apr 2015 11:13:28 -0700 To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Subject: Re: Over-logging by ORC packages Sorry for the excessive logging. The pushdown logging should only be at the start, is there a particular message that was being repeated per a row? Thanks, Owen On Mon, Apr 6, 2015 at 9:15 AM, Moore, Douglas <[email protected]<mailto:[email protected]>> wrote: On a cluster recently upgraded to Hive 0.14 (HDP 2.2) we found that Gigabytes and millions more INFO level hive.log entries from ORC packages were being logged. I feel these log entries should be at the DEBUG level. Is there an existing bug in Hive or ORC? Here is one example: 2015-04-06 15:12:43,212 INFO orc.OrcInputFormat (OrcInputFormat.java:setSearchArgument(298)) - ORC pushdown predicate: leaf-0 = (EQUALS company XYZ) leaf-1 = (EQUALS site DEF) leaf-2 = (EQUALS table ABC) expr = (and leaf-0 leaf-1 leaf-2) To get an acceptable amount of logging that did not fill /tmp we had to add these entries to /etc/hive/conf/hive-log4j.settings: log4j.logger.org.apache.hadoop.hive.ql.io.orc.OrcRawRecordMerger=WARN,DRFA log4j.logger.org.apache.hadoop.hive.ql.io.orc.ReaderImpl=WARN,DRFA log4j.logger.org.apache.hadoop.hive.ql.io.orc.OrcInputFormat=WARN,DRFA While I'm on the subject, to operationally harden Hive, I think Hive should use a more aggressive rolling file appender by default, one that can roll hourly or max size, compress the rolled logs⦠- Douglas
