Github user StephanEwen commented on the issue:
https://github.com/apache/flink/pull/5532
In FLIP-6 mode, the right thing is to not have the number of TaskManagers
preconfigured, but to start them when needed. Does that work already?
The Yarn session would then
- Start Dispatcher + ResourceManager (= Yarn App Master)
- Start JM when job i ssubmitted
- Allocate TMs when JobManager requests slots
- TMs are held by the ResourceManager after the JM releases them, and
kept in the session for a certain time before they are returned to YARN.
I am +1 to remove this "auto change slots" code, it really breaks
abstraction, and go for the above model
---
