Details
-
New Feature
-
Status: Resolved
-
Major
-
Resolution: Later
-
None
-
None
-
None
Description
The purpose of this enhancement is to make it easier to collect garbage collection logs and insure that they persist across restarts in the same way that the standard output files of Hadoop daemon JVM's currently does.
Garbage collection logs are a vital debugging tool for administrators and developers. In our production environments, at some point or another, every single type of Hadoop daemon has OOM'ed or experienced other significant issues related to GC and/or lack of heap memory. For the longest time, we have put in garbage collection logs in our HADOOP_NAMENODE_OPTS, HADOOP_JOBTRACKER_OPTS, etc. by using options like "-XX:+PrintGCDateStamps -XX:+PrintGCDetails -Xloggc:$HADOOP_LOG_DIR/jobtracker.gc.log".
Unfortunately, these logs don't survive a restart of the node, so if a node OOM's and then is restarted automatically, or manually by someone who is unaware, we lose the GC logs forever. We also have to manually add GC log options to each daemon. This patch:
1) Creates a single, optional, off by default, parameter for specifying GC logging.
2) If that parameter is set, automatically enables GC logging for all daemons in the cluster. The parameter is flexible enough to allow for the different ways various vendor's JVM's require garbage collection logging to be specified.
3) If GC logging is on, insures that the GC log files for each daemon are rotated with up to 5 copies kept, same as the .out files currently.
We are currently running a variation of this patch in our 0.20 install. This patch actually includes changes to common, mapred, and hdfs, so it obviously cannot be applied as-is, but is included here for review and comments.