Stephan Ewen created FLINK-17674:
------------------------------------
Summary: OperatorCoordinator state in checkpoints should always be
a ByteStreamStateHandle
Key: FLINK-17674
URL: https://issues.apache.org/jira/browse/FLINK-17674
Project: Flink
Issue Type: Bug
Components: Runtime / Checkpointing
Reporter: Stephan Ewen
Assignee: Stephan Ewen
Fix For: 1.11.0
State restore to the task vertices and coordinators (even after loading the
Checkpoint Metadata) happens in the JobManager's main thread and must
consequently not do any potentially blocking I/O operations.
The OperatorCoordinator state is a generic {{StreamStateHandle}} whose state
might require I/O to retrieve. This never happens in the current implementation
(we always use {{ByteStreamStateHandle}}) the signatures and contracts don't
guarantee that and leave this open for a potential future bug.
Typing the OperatorCoordinator state to ByteStreamStateHandle makes sure that
we can always retrieve the data directly without I/O and clarifies that no
arbitrary StreamStateHandle is supported at that point.
If state restoring becomes an asynchronous operation we can relax this
restriction.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
