[
https://issues.apache.org/jira/browse/FLINK-2268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16155078#comment-16155078
]
ASF GitHub Bot commented on FLINK-2268:
---------------------------------------
Github user aljoscha commented on the issue:
https://github.com/apache/flink/pull/4636
Thanks for reviewing @EronWright and @zentol. I pushed some more commits
that address your comments.
I did check the approach of first using `Class.forName()` and then using
the class normally by building a Hadoop-free Flink and running a cluster and
some examples. I think the class loader only loads classes if they appear in
method signatures or fields, not when classes only appear in code.
@EronWright Yes, your hunch is correct and I did check this on GCE
(dataproc)and AWS(EMR). This is actually quite nice because you can now build a
Hadoop-free Flink and only use the Hadoop dependencies provided by your distro.
> Provide Flink binary release without Hadoop
> -------------------------------------------
>
> Key: FLINK-2268
> URL: https://issues.apache.org/jira/browse/FLINK-2268
> Project: Flink
> Issue Type: Improvement
> Components: Build System
> Reporter: Robert Metzger
> Assignee: Aljoscha Krettek
>
> Currently, all Flink releases ship with Hadoop 2.3.0 binaries.
> The big Hadoop distributions are usually not relying on vanilla Hadoop
> releases, but on custom patched versions.
> To provide the best user experience, we should offer a Flink binary that uses
> the Hadoop jars provided by the user (=hadoop distribution)
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
