[ https://issues.apache.org/jira/browse/DRILL-4278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15104083#comment-15104083 ]
jean-claude commented on DRILL-4278: ------------------------------------ I did a dump of the JVM memory using jmap -dump:format=b,file=drill2.hprof 29739 I then loaded it into Eclipse Memory Analyzer Tool. It showed that the Netty io.netty.util.internal.InternalThreadLocalMap has 4411 References to a Recycler class. Note that's about the number of queries I executed before the heap was full. java.lang.Thread @ 0x78041ba38 BitServer-4 Thread 4,411 120 176,440 6,704,056 \ threadLocals java.lang.ThreadLocal$ThreadLocalMap @ 0x781101230 4,411 24 176,440 6,701,608 .\ table java.lang.ThreadLocal$ThreadLocalMap$Entry[32] @ 0x782f552e0 4,411 144 176,440 6,701,584 ..\ [11] java.lang.ThreadLocal$ThreadLocalMap$Entry @ 0x7811013c0 4,411 32 176,440 6,681,728 ...\ value io.netty.util.internal.InternalThreadLocalMap @ 0x7811013e0 4,411 128 176,440 6,681,696 ....\ indexedVariables java.lang.Object[8192] @ 0x79ca14de0 4,411 32,784 176,440 6,679,848 .....+ [4829] io.netty.util.Recycler$Stack @ 0x7a11fe670 » 1 48 40 1,088 .....+ [615] io.netty.util.Recycler$Stack @ 0x7869ee3c0 » 1 48 40 1,088 .....+ I then found this issue in Netty which apparently has been resolved by adding an API (FastThreadLocal.removeAll()) to release this thread local map. However I'm not sure where and when in Drill's code this API could be called to free up the thread local ... https://github.com/netty/netty/pull/2578 > Memory leak when using LIMIT > ---------------------------- > > Key: DRILL-4278 > URL: https://issues.apache.org/jira/browse/DRILL-4278 > Project: Apache Drill > Issue Type: Bug > Components: Execution - RPC > Affects Versions: 1.4.0, 1.5.0 > Environment: OS X > 0: jdbc:drill:zk=local> select * from sys.version; > +----------+-------------------------------------------+-----------------------------------------------------+----------------------------+----------------------------+----------------------------+ > | version | commit_id | > commit_message | commit_time | > build_email | build_time | > +----------+-------------------------------------------+-----------------------------------------------------+----------------------------+----------------------------+----------------------------+ > | 1.4.0 | 32b871b24c7b69f59a1d2e70f444eed6e599e825 | > [maven-release-plugin] prepare release drill-1.4.0 | 08.12.2015 @ 00:24:59 > PST | venki.koruka...@gmail.com | 08.12.2015 @ 01:14:39 PST | > +----------+-------------------------------------------+-----------------------------------------------------+----------------------------+----------------------------+----------------------------+ > 0: jdbc:drill:zk=local> select * from sys.options where status <> 'DEFAULT'; > +-----------------------------+-------+---------+----------+----------+-------------+-----------+------------+ > | name | kind | type | status | num_val | > string_val | bool_val | float_val | > +-----------------------------+-------+---------+----------+----------+-------------+-----------+------------+ > | planner.slice_target | LONG | SYSTEM | CHANGED | 10 | null > | null | null | > | planner.width.max_per_node | LONG | SYSTEM | CHANGED | 5 | null > | null | null | > +-----------------------------+-------+---------+----------+----------+-------------+-----------+------------+ > 2 rows selected (0.16 seconds) > Reporter: jean-claude > > copy the parquet files in the samples directory so that you have a 12 or so > $ ls -lha /apache-drill-1.4.0/sample-data/nationsMF/ > nationsMF1.parquet > nationsMF2.parquet > nationsMF3.parquet > create a file with a few thousand lines like these > select * from dfs.`/Users/jccote/apache-drill-1.4.0/sample-data/nationsMF` > limit 500; > start drill > $ /apache-drill-1.4.0/bin/drill-embeded > reduce the slice target size to force drill to use multiple fragment/threads > jdbc:drill:zk=local> system set planner.slice_target=10; > now run the list of queries from the file your created above > jdbc:drill:zk=local> !run /Users/jccote/test-memory-leak-using-limit.sql > the java heap space keeps going up until the old space is at 100% and > eventually you get an OutOfMemoryException in drill > $ jstat -gccause 86850 5s > S0 S1 E O M CCS YGC YGCT FGC FGCT > GCT LGCC GCC > 0.00 0.00 100.00 100.00 98.56 96.71 2279 26.682 240 458.139 > 484.821 GCLocker Initiated GC Ergonomics > 0.00 0.00 100.00 99.99 98.56 96.71 2279 26.682 242 461.347 > 488.028 Allocation Failure Ergonomics > 0.00 0.00 100.00 99.99 98.56 96.71 2279 26.682 245 466.630 > 493.311 Allocation Failure Ergonomics > 0.00 0.00 100.00 99.99 98.56 96.71 2279 26.682 247 470.020 > 496.702 Allocation Failure Ergonomics > If you do the same test but do not use the LIMIT then the memory usage does > not go up. > If you add a where clause so that no results are returned, then the memory > usage does not go up. > Something with the RPC layer? > Also it seems sensitive to the number of fragments/threads. If you limit it > to one fragment/thread the memory usage goes up much slower. > I have used parquet files and CSV files. In either case the behaviour is the > same. -- This message was sent by Atlassian JIRA (v6.3.4#6332)