This JIRA is an alternative take on
We propose providing an option of using a RollingFileAppender(RFA)-based implementation of container log appender as means of log size control via mapreduce.task.userlog.limit.kb.
The idea is to use mapreduce.task.userlog.limit.kb as maximumFileSize of RFA. In addition yarn.app.mapreduce.container.log.backups (task attempt containers) and yarn.app.mapreduce.am.log.backups (MR-AM) are passed as maxBackupIndex.
Both current ContainerLogAppender (CLA) and new ContainerRollingLogAppender (CRLA) co-exist. CLA is the default. CRLA is chosen when mapreduce.task.userlog.limit.kb > 0 && *.backups > 0.
1) CRLA output is visible in UI right away. CLA output with mapreduce.task.userlog.limit.kb > 0 is not visible until the task attempt finishes that prevents timely diagnostics.
2) Even with excessive logging and a large mapreduce.task.userlog.limit.kb, no space is taken from the JVM heap.
3) No UI impact, since YARN is already designed to deal with any log name beyond stderr/out, syslog, debug.out, profile.out
1) if the logging is excessive there will be more local filesystem metadata I/O due to roll. That should be negligible in the grand scheme.
Furthermore, to improve log consistency and completeness in the case of JVM crashes and SIGTERMing by NM, we propose to restore the MRv1 behavior of periodic log syncing (every 5s) and having log sync as part of a shutdown hook.