[SPARK-12511] streaming driver with checkpointing unable to finalize leading to OOM

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.5.2, 1.6.0
• Fix Version/s: 1.6.1, 2.0.0
• Component/s: DStreams, PySpark

Environment

pyspark 1.5.2
yarn 2.6.0
python 2.6
centos 6.5
openjdk 1.8.0

• Assignee: Shixiong Zhu
• Reporter: Antony Mayi
• Created: 23/Dec/15 22:25
• Updated: 11/Apr/16 20:04
• Resolved: 05/Jan/16 21:49

Description

Spark streaming application when configured with checkpointing is filling driver's heap with multiple ZipFileInputStream instances as results of spark-assembly.jar (potentially some others like for example snappy-java.jar) getting repetitively referenced (loaded?). Java Finalizer can't finalize these ZipFileInputStream instances and it eventually takes all heap leading the driver to OOM crash.

Steps to reproduce:

• Submit attached bug.py to spark
• Leave it running and monitor the driver java process heap
  • with heap dump you will primarily see growing instances of byte array data (here cumulated zip payload of the jar refs):

     num     #instances         #bytes  class name
    ----------------------------------------------
       1:         32653       32735296  [B
       2:         48000        5135816  [C
       3:            41        1344144  [Lscala.concurrent.forkjoin.ForkJoinTask;
       4:         11362        1261816  java.lang.Class
       5:         47054        1129296  java.lang.String
       6:         25460        1018400  java.lang.ref.Finalizer
       7:          9802         789400  [Ljava.lang.Object;
    

  • with visualvm you can see:
    • increasing number of objects pending for finalization
      
    • increasing number of ZipFileInputStreams instances related to the spark-assembly.jar referenced by Finalizer
      
• Depending on the heap size and running time this will lead to driver OOM crash

Comments

• The bug.py is lightweight proof of the problem. In production I am experiencing this as quite rapid effect - in few hours it eats gigs of heap and kills the app.
• If the same bug.py is run without checkpointing there is no issue whatsoever.
• Not sure if it is just pyspark related.
• In bug.py I am using the socketTextStream input but seems to be independent of the input type (in production having same problem with Kafka direct stream, have seen it even with textFileStream).
• It is happening even if the input stream doesn't produce any data.

Issue Links

• is blocked by: SPARK-12652 Upgrade py4j to the incoming version 0.9.1
• is duplicated by: SPARK-11711 Finalizer memory leak is pyspark
• relates to: SPARK-11711 Finalizer memory leak is pyspark
• links to: [Github] Pull Request #10514 (zsxwing)