[FLINK-9080] Flink Scheduler goes OOM, suspecting a memory leak

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Won't Fix
• Affects Version/s: 1.4.0
• Component/s: Runtime / Coordination
• Assignee: Stefan Richter
• Reporter: Rohit Singh
• Created: 26/Mar/18 13:16
• Updated: 29/Jan/21 11:29
• Resolved: 29/Jan/21 11:29

Description

Running FLink version 1.4.0. on mesos,scheduler running along  with job manager in single container, whereas task managers running in seperate containers.

Couple of jobs were running continously, Flink scheduler was working properlyalong with task managers. Due to some change in data, one of the jobs started failing continuously. In the meantime,there was a surge in  flink scheduler memory usually eventually died out off OOM

 

Memory dump analysis was done, 

Following were findings

•  Majority of top loaded packages retaining heap indicated towards Flinkuserclassloader, glassfish(jersey library), Finalizer classes. (Top level package image)
• Top level classes were of Flinkuserclassloader, (Top Level class image)
• The number of classes loaded vs unloaded was quite less  PFA,inspite of adding jvm options of -XX:+UseConcMarkSweepGC -XX:+CMSClassUnloadingEnabled , PFAclassloaded vs unloaded graph, scheduler was restarted 3 times
• There were custom classes as well which were duplicated during subsequent class uploads

PFA all the images of heap dump.  Can you suggest some pointers on as to how to overcome this issue.

 

 

Issue Links

• relates to: FLINK-11205 Task Manager Metaspace Memory Leak
• relates to: FLINK-10317 Configure Metaspace size by default