[jira] [Updated] (FLINK-6285) resolve hadoop-compatibility confusion

["Petr Novotnik (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/FLINK-6285?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Petr Novotnik updated FLINK-6285:
---------------------------------
    Description: 
As of Flink 1.2.0, the binary distribution does not include classes from the {{hadoop-compatibility}} dependency anymore.

{noformat}
flink-1.2.0> for i in lib/*.jar; jar tf $i | grep WritableTypeInfo; end
flink-1.2.0 [1]> # the above finds nothing
{noformat}

Therefore, it is necessary to copy the compatibility jar to flink's installation `lib/` directory (or a sub-directory) if one wishes to use hadoop input formats. Merely packaging the compatibility jar as part of an application's "fat jar" does not suffice, as code in [TypeExtractor#createHadoopWritableTypeInfo|https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/api/java/typeutils/TypeExtractor.java#L1988]'s relies on being able to see the compatibility classes through the classloader {{TypeExtractor}} itself was loaded by. On yarn this seems not to be the case (e.g. when running the application through {{flink run -m yarn-cluster ...}}).

Ideally, we'd fix the class loading issue, such that flink's installation does not need to be altered, due to the need of a particular application. Alternatively, we could include the hadoop-compatibility jar as part of the binary distribution and provide corresponding instructions, [1|https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/hadoop_compatibility.html] and [2|https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/migration.html] seem to be good places.


  was:
As of Flink 1.2.0, the binary distribution does not include classes from the `hadoop-compatibility` dependency anymore.

```
flink-1.2.0> for i in lib/*.jar; jar tf $i | grep WritableTypeInfo; end
flink-1.2.0 [1]> # the above finds nothing
```

Therefore, it is necessary to copy the compatibility jar to flink's installation `lib/` directory (or a sub-directory) if one wishes to use hadoop input formats. Merely packaging the compatibility jar as part of an application's "fat jar" does not suffice, as code in [TypeExtractor#createHadoopWritableTypeInfo](https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/api/java/typeutils/TypeExtractor.java#L1988)'s relies on being able to see the compatibility classes through the classloader `TypeExtractor` itself was loaded by. On yarn this seems not to be the case (e.g. when running the application through `flink run -m yarn-cluster ...`).

* Ideally, we'd fix the class loading issue, such that flink's installation does not need to be altered, due to the need of a particular application.
* Alternatively, we could include the hadoop-compatibility jar as part of the binary distribution and provide corresponding instructions, [1] and [2] seem to be good places.

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/hadoop_compatibility.html
[2] https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/migration.html



> resolve hadoop-compatibility confusion
> --------------------------------------
>
>                 Key: FLINK-6285
>                 URL: https://issues.apache.org/jira/browse/FLINK-6285
>             Project: Flink
>          Issue Type: Bug
>    Affects Versions: 1.2.0
>            Reporter: Petr Novotnik
>
> As of Flink 1.2.0, the binary distribution does not include classes from the {{hadoop-compatibility}} dependency anymore.
> {noformat}
> flink-1.2.0> for i in lib/*.jar; jar tf $i | grep WritableTypeInfo; end
> flink-1.2.0 [1]> # the above finds nothing
> {noformat}
> Therefore, it is necessary to copy the compatibility jar to flink's installation `lib/` directory (or a sub-directory) if one wishes to use hadoop input formats. Merely packaging the compatibility jar as part of an application's "fat jar" does not suffice, as code in [TypeExtractor#createHadoopWritableTypeInfo|https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/api/java/typeutils/TypeExtractor.java#L1988]'s relies on being able to see the compatibility classes through the classloader {{TypeExtractor}} itself was loaded by. On yarn this seems not to be the case (e.g. when running the application through {{flink run -m yarn-cluster ...}}).
> Ideally, we'd fix the class loading issue, such that flink's installation does not need to be altered, due to the need of a particular application. Alternatively, we could include the hadoop-compatibility jar as part of the binary distribution and provide corresponding instructions, [1|https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/hadoop_compatibility.html] and [2|https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/migration.html] seem to be good places.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)