[
https://issues.apache.org/jira/browse/FLINK-5985?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15904869#comment-15904869
]
Ufuk Celebi commented on FLINK-5985:
------------------------------------
Thanks again for reporting this. You are right, we accidentally changed the
behaviour between 1.1 and 1.2. This is a critical issue.
In 1.1 we did not serialize stateless tasks as part of the savepoint and
therefore never took them into account on savepoint loading. With the recent
refactorings in 1.2 we now serialize stateless tasks this as "zero length
state". This makes the stateless operators part of the savepoint we try to load
the state back to the new job.
I think there a couple of easy ways to fix this, but
[~stefanrichte...@gmail.com] probably has the definitive answer here.
I think this issue warrants a 1.2.1 release asap as users cannot change their
topologies if they didn't specify a UID for each operator.
> Flink treats every task as stateful (making topology changes impossible)
> ------------------------------------------------------------------------
>
> Key: FLINK-5985
> URL: https://issues.apache.org/jira/browse/FLINK-5985
> Project: Flink
> Issue Type: Bug
> Components: State Backends, Checkpointing
> Affects Versions: 1.2.0
> Reporter: Gyula Fora
> Priority: Critical
>
> It seems that Flink treats every Task as stateful so changing the topology
> is not possible without setting uid on every single operator.
> If the topology has an iteration this is virtually impossible (or at least
> gets super hacky)
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
