Hello, Generally, 1gb is a good starting point for JM memory. If a JM is managing multiple pipelines, with a large number of operators and tasks, JM needs more memory.
On Tue, Dec 9, 2025 at 5:57 AM Jean-Marc Paulin <[email protected]> wrote: > Hi, > > We have a flink 1.20.3 streaming job, using hashmap for the state. We > normally run with a job parallelism of 6, spread over 6 task managers, and > that normally behaves correctly. The jobmanagers have a heap size of 300mb > and we never had an issue. We also have HA with zookeeper > > For one particular scenario, we increased the job parallelism to 10. We > also adjusted the taskmanager parallelism and memory with our expectation, > but the jobmanagers are now OOM. We increased their heap size to 400 mb, > which seems to help, but we still see the odd restarts, > > Q: Is there a correlation between the parallelism, number of task > managers, and the memory usage for the jobmanagers...is there any tips on > how to configure it (other than try/watch/repeat)? what should we take into > account to size the jobmanager memory? > > Thanks > > JM > >
