Type: New Feature
Resolution: Won't Fix
Affects Version/s: None
Fix Version/s: None
Component/s: API / DataStream
Currently the user can set a timeout which will shut down the iteration source/sink nodes if no new data is received during that time to allow program termination in iterative streaming jobs.
This method is used due to the non-trivial nature of termination in iterative streaming jobs. While termination is not a main concern in long running streaming jobs, this behaviour makes iterative tests non-deterministic and they often fail on travis due to the timeout. Also setting a timeout can cause jobs to terminate prematurely.
I propose to remove iteration timeouts and replace it with the following algorithm for detecting termination:
-We first identify loop edges in the jobgraph (the channels from the iteration sources to the head operators)
-Once the head operators (the ones with loop input) finish with all their non-loop inputs they broadcast a marker to their outputs.
-Each operator will broadcast a marker once it received a marker from all its non-finished inputs
-Iteration sources are terminated when they receive 2 consecutive markers without receiving any record in-between
The idea behind the algorithm is to find out when no more outputs are generated from the operators inside an iteration after their normal inputs are finished.