[
https://issues.apache.org/jira/browse/FLINK-6606?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16016573#comment-16016573
]
ASF GitHub Bot commented on FLINK-6606:
---------------------------------------
Github user tillrohrmann commented on the issue:
https://github.com/apache/flink/pull/3933
I agree with @EronWright that it would be nice to offer all user code the a
similar environment. This will become especially important if a
`MasterTriggerRestoreHook` which is contained in the system class loader tries
to load user code classes dynamically.
I'm wondering whether to follow the approach of the `InputFormatVertex` and
`ExecutionJobVertex` to set the TCCL explicitly when calling the hooks. This
could happen in `MasterHooks#triggerMasterHooks` for example. The advantage
would be that we don't have to reset the TCCL for each hook. On the other hand,
it harbours danger that we will forget to set the TCCL when calling the hooks
from somewhere else in the future.
> Create checkpoint hook with user classloader
> --------------------------------------------
>
> Key: FLINK-6606
> URL: https://issues.apache.org/jira/browse/FLINK-6606
> Project: Flink
> Issue Type: Bug
> Components: State Backends, Checkpointing
> Reporter: Eron Wright
> Assignee: Eron Wright
> Priority: Blocker
> Fix For: 1.3.0
>
>
> Flink should set the thread's classloader when calling the checkpoint hook
> factory's `create` method. Without that, the hook is likely to fail during
> initialization (e.g. using ServiceLoader).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
