Hi everyone, In the k8s operator, I see that the session cluster spawns dynamically Task Managers as needed to handle new jobs. But the startup time of these Task Managers is quite important which kind of defeats the interest of using a session cluster when you want low latency (eg for interactive queries). What would be the strategy to reduce this time / have warm TMs ready to handle jobs ? It seems there is a grace period after the end of a job before a TM is destroyed. Is this period configurable ? Would a strategy of regularly submitting empty jobs just to keep the TM warm look like a possibility ? Ideally this would be handled by the operator itself.

1. You can use

standalone

mode if you don't want to spin up new TM every time. 2. On demand TM has configurable property for grace period https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/config/#resourcemanager-taskmanager-timeout

1. On demand TM has configurable property for grace period https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/config/#resourcemanager-taskmanager-timeout

Exactly what I was looking for. Thanks a lot !🙏

Correct, the number of TMs same as taskmanager.replicas in flinkdeployment will be always up and running.

Cool ! It would be a great feature to have the possibility to configure a number of idle TMs in native mode. What is the best way to discuss this with the community ? Jira or mailing list ? I'd be OK to create a FLIP but I don't know enough Flink's internal so I'd need help.

I am also just a flink user, not aware of the best way to discuss this. Someone from flink maintainer's team can answer it probably