Currently, eager scheduling immediately triggers the scheduling for all vertices and their subtasks in topological order.
This has two problems:
- This works only, as long as resource acquisition is "synchronous". With dynamic resource acquisition in FLIP-6, the resources are returned as Futures which may complete out of order. This results in out-of-order (not in topological order) scheduling of tasks which does not work for streaming.
- Deploying some tasks that depend on other tasks before it is clear that the other tasks have resources as well leads to situations where many deploy/recovery cycles happen before enough resources are available to get the job running fully.
For eager scheduling, we should allocate all resources in one chunk and then deploy once we know that all are available.
As a follow-up, the same should be done per pipelined component in lazy batch scheduling as well. That way we get lazy scheduling across blocking boundaries, and bulk (gang) scheduling in pipelined subgroups.
This also does not apply for efforts of fine grained recovery, where individual tasks request replacement resources.