[ https://issues.apache.org/jira/browse/PARQUET-118?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16928131#comment-16928131 ]
Mitesh edited comment on PARQUET-118 at 9/12/19 1:43 AM: --------------------------------------------------------- Any update on this? I am hitting this with Spark when a column is a struct that is very deep. Seems like the entire thing gets buffered at one time, so I have to set {{-XX:MaxDirectMemorySize}} to a very large number (biggest size of column * num rows in partition * num partitions processed by a single JVM). It would be great to have a config to force on-heap buffer usage, even if there is a latency hit. Netty provides this functionality via {{-Dio.netty.noUnsafe}} flag and I think it was a wise decision by them. cc [~nongli] was (Author: masterddt): Any update on this? I am hitting this with Spark when a column is a struct that is very deep. Seems like the entire thing gets buffered at one time, so I have to set {{-XX:MaxDirectMemorySize}} to a very large number (biggest size of column * num rows in partition * num partitions processed by a single JVM). It would be great to have a config to force on-heap buffer usage, even if there is a latency hit. Netty provides this functionality via {{-Dio.netty.noUnsafe}} flag and I think it was a wise decision by them. > Provide option to use on-heap buffers for Snappy compression/decompression > -------------------------------------------------------------------------- > > Key: PARQUET-118 > URL: https://issues.apache.org/jira/browse/PARQUET-118 > Project: Parquet > Issue Type: Improvement > Components: parquet-mr > Affects Versions: 1.6.0 > Reporter: Patrick Wendell > Priority: Major > > The current code uses direct off-heap buffers for decompression. If many > decompressors are instantiated across multiple threads, and/or the objects > being decompressed are large, this can lead to a huge amount of off-heap > allocation by the JVM. This can be exacerbated if overall, there is not heap > contention, since no GC will be performed to reclaim the space used by these > buffers. > It would be nice if there was a flag we cold use to simply allocate on-heap > buffers here: > https://github.com/apache/incubator-parquet-mr/blob/master/parquet-hadoop/src/main/java/parquet/hadoop/codec/SnappyDecompressor.java#L28 > We ran into an issue today where these buffers totaled a very large amount of > storage and caused our Java processes (running within containers) to be > terminated by the kernel OOM-killer. -- This message was sent by Atlassian Jira (v8.3.2#803003)