Github user StephanEwen commented on the issue:
https://github.com/apache/flink/pull/6353
Maybe we can solve this simpler? Avoiding concurrency in the execution
graph creation makes the code simpler and more robust - very desirable for an
already fairly complex construct.
The issue here is the time it takes to create the splits, so we could see
if we parallelize that, rather than parallelizing the job vertex creation.
I would think in the direction of having a Future that supplies the input
splits and computing the future in the IOExecutor. That would parallelize the
core problem and leave the executiongraph as it is.
---
