[
https://issues.apache.org/jira/browse/FLINK-5397?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15782743#comment-15782743
]
ASF GitHub Bot commented on FLINK-5397:
---------------------------------------
GitHub user StefanRRichter opened a pull request:
https://github.com/apache/flink/pull/3050
[FLINK-5397] Do not replace ObjectStreamClass on deserialization of m…
This PR fixes [FLINK-5397] by not replacing the ObjectStreamClass in
``readClassDescriptor()`` during deserialization of classes that should be
replaced from the migration package. Instead, we replace the class directly in
``resolveClass(...)``.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/StefanRRichter/flink
migration-keep-objectstreamclass
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/3050.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #3050
----
commit 01834068b50a9cf4186ce83fa9c57d5b8f2bfc8c
Author: Stefan Richter <s.rich...@data-artisans.com>
Date: 2016-12-28T11:50:00Z
[FLINK-5397] Do not replace ObjectStreamClass on deserialization of
migration package classes, override resolveClass(...) instead
----
> Fail to deserialize savepoints in v1.1 when there exist missing fields in
> class serialization descriptors
> ---------------------------------------------------------------------------------------------------------
>
> Key: FLINK-5397
> URL: https://issues.apache.org/jira/browse/FLINK-5397
> Project: Flink
> Issue Type: Bug
> Components: State Backends, Checkpointing
> Reporter: Xiaogang Shi
> Assignee: Stefan Richter
>
> To restore from the savepoints in previous versions, Flink now keeps all
> classes whose serialization is changed and put them in a separated package
> ("migration").
> When deserializing the old savepoints, flink will look up correct descriptors
> ({{ObjectStreamClass}}) for these classes, without using those ones written
> in serialized data. The implementation however is problematic when there
> exist missing field descriptors in the serialized data.
> When serializing an object, Java will only write the descriptors of those
> non-null fields. But when we look up class descriptors with given classes,
> all fields will be put into the descriptors. As a result, we will deserialize
> the savepoints with incorrect descriptors, leading to serialization
> exceptions.
> A simple resolution is to update the name of read descriptors using
> Reflections, without using different descriptors.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
