Chesnay Schepler created FLINK-23893:
----------------------------------------
Summary: Fine-grained resource allocation may fail depending on
slot allocation order
Key: FLINK-23893
URL: https://issues.apache.org/jira/browse/FLINK-23893
Project: Flink
Issue Type: Bug
Components: Runtime / Coordination
Reporter: Chesnay Schepler
Fix For: 1.14.0
The fine-grained slot management is relatively simple in that it iterates over
each registered TM in order for each required slot.
This means that you can pretty easily create scenarios where, depending on the
slot order, the allocation of a slot may fail.
A trivial example, only using memory for conciseness:
2 TMs with 3 MB of memory each
2 slot-sharing groups with parallelism=2 and the following requirements:
1) 1 MB memory
2) 2 MB memory
If both sub-tasks of 1) are scheduled first on a single TM then the job cannot
be scheduled.
It's not terrible for this limitation to exist in the first version, but it
should be _explicitly_ documented.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
