[FLINK-32746] Using ZGC in JDK17 to solve long time class unloading STW

• Type: Sub-task
• Status: Closed
• Priority: Major
• Resolution: Done
• Component/s: Table SQL / Runtime
• Assignee: Unassigned
• Reporter: xiangyu feng
• Created: 04/Aug/23 02:16
• Updated: 12/Oct/23 11:11
• Resolved: 12/Oct/23 11:11

Description

In a OLAP session cluster, a TM need to frequently create new classloaders and  generate new classes. These classes will be accumulated in metaspace. When metaspace data usage reaches a threshold, a FullGC with a long time Stop-the-World will be triggered. Currently, both SerialGC, ParallelGC and G1GC are doing Stop-the-World class unloading. Only ZGC supports concurrent class unload, see more in https://bugs.openjdk.org/browse/JDK-8218905.

 

In our scenario, a class unloading for a 2GB metaspace with 5million classes will stop the application more than 40 seconds. After switch to ZGC, the maximum STW of the application has been reduced to less than 10ms.