[ https://issues.apache.org/jira/browse/FLINK-11771?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Tzu-Li (Gordon) Tai updated FLINK-11771: ---------------------------------------- Issue Type: Bug (was: Improvement) > Serializer snapshot cannot be read if directly upgraded in-place to a > TypeSerializerSnapshot from a TypeSerializerConfigSnapshot written in 1.7+ > ------------------------------------------------------------------------------------------------------------------------------------------------ > > Key: FLINK-11771 > URL: https://issues.apache.org/jira/browse/FLINK-11771 > Project: Flink > Issue Type: Bug > Components: API / Type Serialization System > Reporter: Tzu-Li (Gordon) Tai > Assignee: Tzu-Li (Gordon) Tai > Priority: Blocker > Fix For: 1.8.0 > > > This is a upgrade path that was overlooked in > {{TypeSerializerSnapshot#readVersionedSnapshot}}. > If the serializer snapshot was a {{TypeSerializerConfigSnapshot}} when it was > written in 1.7+ Flink versions, the actual snapshot content will be prefixed > by a magic {{TypeSerializerConfigSnapshot#ADAPTER_VERSION}} int as the > snapshot version, as well as Java-serialized prior serializer. > If when restoring, the serializer snapshot was upgraded in-place to a > {{TypeSerializerSnapshot}} (in-place meaning, same classname, not introducing > a new snapshot class), {{TypeSerializerSnapshot#readVersionedSnapshot}} > doesn't work as expected. > Firstly, the provided {{readVersion}} to the user-implemented > {{TypeSerializerSnapshot#readSnapshot}} method would be the magic > {{ADAPTER_VERSION}}. > Secondly, the remaining stream would contain the Java-serialized prior > serializer, which is no longer relevant for the user (because they already > upgraded to the new abstraction and would have a fully functional > {{restoreSerializer}} implementation). To workaround that, they would have to > read and drop that Java-serialized serializer in user code. -- This message was sent by Atlassian JIRA (v7.6.3#76005)