[
https://issues.apache.org/jira/browse/FLINK-3974?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15300642#comment-15300642
]
B Wyatt commented on FLINK-3974:
--------------------------------
After attempting to produce a minimal repro, I realized that there needs to be
an operation between the source and the parallel maps:
{code:java}
DataStream<A> input = ...
input = input
.map(MapFunction<A,A>...);
input
.map(MapFunction<A,B>...)
.addSink(...);
input
.map(MapFunction<A,C>...)
.addSink(...);
{code}
see attached repro case
> enableObjectReuse fails when an operator chains to multiple downstream
> operators
> --------------------------------------------------------------------------------
>
> Key: FLINK-3974
> URL: https://issues.apache.org/jira/browse/FLINK-3974
> Project: Flink
> Issue Type: Bug
> Components: DataStream API
> Affects Versions: 1.0.3
> Reporter: B Wyatt
> Attachments: ReproFLINK3974.java
>
>
> Given a topology that looks like this:
> {code:java}
> DataStream<A> input = ...
> input
> .map(MapFunction<A,B>...)
> .addSink(...);
> input
> .map(MapFunction<A,C>...)
> .addSink(...);
> {code}
> enableObjectReuse() will cause an exception in the form of
> {{"java.lang.ClassCastException: B cannot be cast to A"}} to be thrown.
> It looks like the input operator calls {{Output<StreamRecord<A>>.collect}}
> which attempts to loop over the downstream operators and process them.
> However, the first map operation will call {{StreamRecord<>.replace}} which
> mutates the value stored in the StreamRecord<>.
> As a result, when the {{Output<StreamRecord<A>>.collect}} call passes the
> {{StreamRecord<A>}} to the second map operation it is actually a
> {{StreamRecord<B>}} and behaves as if the two map operations were serial
> instead of parallel.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
