Zhu Zhu created FLINK-14236:
-------------------------------
Summary: Make LazyFromSourcesSchedulingStrategy do lazy scheduling
only based on partition state
Key: FLINK-14236
URL: https://issues.apache.org/jira/browse/FLINK-14236
Project: Flink
Issue Type: Sub-task
Components: Runtime / Coordination
Affects Versions: 1.10.0
Reporter: Zhu Zhu
Fix For: 1.10.0
The philosophy of lazy_from_sources scheduling is to schedule a task when its
inputs are ready, i.e. when needed result partitions are consumable. It
actually does not directly care whether a task is finished, though a finished
task will lead its partitions to be consumable.
The {{LazyFromSourcesSchedulingStrategy}} currently does lazy scheduling in two
ways:
1. trigger scheduling on FINISHED execution state for blocking partition
consumers
2. trigger scheduling on partition consumable events for pipelined partition
consumers
This makes the scheduling decision a bit messy. And it also requires the
{{LazyFromSourcesSchedulingStrategy}} to manage the partition state by itself,
to decide when a partition is consumable, which is not good.
I'd propose to make {{LazyFromSourcesSchedulingStrategy}} do lazy scheduling
only based on partition state. This can make the scheduling decision making
clearer and relieve {{LazyFromSourcesSchedulingStrategy}} from partition state
management.
This work requires FLINK-14234 so that all partition consumable events can be
notified.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)