[
https://issues.apache.org/jira/browse/FLINK-9031?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16408636#comment-16408636
]
ASF GitHub Bot commented on FLINK-9031:
---------------------------------------
GitHub user fhueske opened a pull request:
https://github.com/apache/flink/pull/5742
[FLINK-9031] Fix DataSet Union operator translation bug.
## What is the purpose of the change
- Fixes a bug reported in
[FLINK-9031](https://issues.apache.org/jira/browse/FLINK-9031)
- Union nodes had a partitioning strategy on the outgoing channel, that was
(intentionally) not translated by the `JobGraphGenerator` because the
`JobGraphGenerator` assumed that Union nodes would always have outgoing FORWARD
strategies.
- Not translating the partitioning resulted in an incorrect result because
data was not correctly distributed.
## Brief change log
- Add a check in `JobGraphGenerator` to fail if a union node with
non-FORWARD outgoing strategy is found.
- Add a pre-optimization plan traversal that fixes the strategy of union
outputs to FORWARD.
- Add a test based on a simplified version of the reported program.
## Verifying this change
- Run the added test.
## Does this pull request potentially affect one of the following parts:
- Dependencies (does it add or upgrade a dependency): **no**
- The public API, i.e., is any changed class annotated with
`@Public(Evolving)`: **no**
- The serializers: **no**
- The runtime per-record code paths (performance sensitive): **no**
- Anything that affects deployment or recovery: JobManager (and its
components), Checkpointing, Yarn/Mesos, ZooKeeper: **no**
- The S3 file system connector: **no**
## Documentation
- Does this pull request introduce a new feature? **no**
- If yes, how is the feature documented? **n/a**
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/fhueske/flink dataSetUnionBug
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/5742.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 #5742
----
commit 3939f06bacaf1595844de56fe58651732b89592e
Author: Fabian Hueske <fhueske@...>
Date: 2018-03-21T19:54:05Z
[FLINK-9031] Fix DataSet Union operator translation bug.
- Adds a pass over the pre-optimized plan that fixes the output strategy of
union nodes to FORWARD.
----
> DataSet Job result changes when adding rebalance after union
> ------------------------------------------------------------
>
> Key: FLINK-9031
> URL: https://issues.apache.org/jira/browse/FLINK-9031
> Project: Flink
> Issue Type: Bug
> Components: DataSet API, Local Runtime, Optimizer
> Affects Versions: 1.3.1
> Reporter: Fabian Hueske
> Priority: Critical
> Attachments: Person.java, RunAll.java, newplan.txt, oldplan.txt
>
>
> A user [reported this issue on the user mailing
> list|https://lists.apache.org/thread.html/075f1a487b044079b5d61f199439cb77dd4174bd425bcb3327ed7dfc@%3Cuser.flink.apache.org%3E].
> {quote}I am using Flink 1.3.1 and I have found a strange behavior on running
> the following logic:
> # Read data from file and store into DataSet<POJO>
> # Split dataset in two, by checking if "field1" of POJOs is empty or not, so
> that the first dataset contains only elements with non empty "field1", and
> the second dataset will contain the other elements.
> # Each dataset is then grouped by, one by "field1" and other by another
> field, and subsequently reduced.
> # The 2 datasets are merged together by union.
> # The final dataset is written as json.
> What I was expected, from output, was to find only one element with a
> specific value of "field1" because:
> # Reducing the first dataset grouped by "field1" should generate only one
> element with a specific value of "field1".
> # The second dataset should contain only elements with empty "field1".
> # Making an union of them should not duplicate any record.
> This does not happen. When i read the generated jsons i see some duplicate
> (non empty) values of "field1".
> Strangely this does not happen when the union between the two datasets is
> not computed. In this case the first dataset produces elements only with
> distinct values of "field1", while second dataset produces only records with
> empty field "value1".
> {quote}
> The user has not enable object reuse.
> Later he reports that the problem disappears when he injects a rebalance()
> after a union resolves the problem. I had a look at the execution plans for
> both cases (attached to this issue) but could not identify a problem.
> Hence I assume, this might be an issue with the runtime code but we need to
> look deeper into this. The user also provided an example program consisting
> of two classes which are attached to the issue as well.
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
