Details
-
Bug
-
Status: Closed
-
Critical
-
Resolution: Fixed
-
1.11.3, 1.12.2, 1.13.0
Description
For the topology with cross-region blocking edges, there are regions that may never be scheduled. The case is illustrated in the figure below.
Let's denote the vertices with layer_number. It's clear that the edge connects v2_2 and v3_2 crosses region 1 and region 2. Since region 1 has no blocking edges connected to other regions, it will be scheduled first. When vertex2_2 is finished, PipelinedRegionSchedulingStrategy will trigger onExecutionStateChange for it.
As expected, region 2 will be scheduled since all its consumed partitions are consumable. But in fact region 2 won't be scheduled, because the result partition of vertex2_2 is not tagged as consumable. Whether it is consumable or not is determined by its IntermediateDataSet.
However, an IntermediateDataSet is consumable if and only if all the producers of its IntermediateResultPartitions are finished. This IntermediateDataSet will never be consumable since vertex2_3 is not scheduled. All in all, this forms a deadlock that a region will never be scheduled because it's not scheduled.
As a solution we should let BLOCKING result partitions be consumable individually. Note that this will result in the scheduling to become execution-vertex-wise instead of stage-wise, with a nice side effect towards better resource utilization. The PipelinedRegionSchedulingStrategy can be simplified along with change to get rid of the correlatedResultPartitions.
Attachments
Attachments
Issue Links
- is duplicated by
-
FLINK-21734 Allow BLOCKING result partition to be individually consumable
- Closed
- is related to
-
FLINK-23833 Cache of ShuffleDescriptors should be individually cleaned up
- Closed
- links to