Inverted classloading for client

[PaulLam]

Hi devs,

Flink uses inverted class loading by default to allow a different version of dependencies in user codes, but currently this approach is not applied to the client, so I’m wondering if it’s out of some special reason?

If not, I think it would be great to add inverted class loading as the the behavior for the client and make it respect to the classloading configuration to make the components work in a uniform manner.

Moreover, I’ve seen some jobs failed to upgrade to 1.8, because we added hadoop classpath in the client classpath to compensate the removed convenience hadoop binaries that shades lots of common dependencies, which significantly increases the changes of dependencies conflicts.

Looking forward to your feedbacks.

Best,
Paul Lam

[PaulLam]

Hi,

I’ve created a ticket to track this problem [1]. Any comments will be appreciated.

[1] https://issues.apache.org/jira/browse/FLINK-13749 <https://issues.apache.org/jira/browse/FLINK-13749>

Best,
Paul Lam

> 在 2019年8月9日，11:16，Paul Lam <[hidden email]> 写道：
>
> Hi devs,
>
> Flink uses inverted class loading by default to allow a different version of dependencies in user codes, but currently this approach is not applied to the client, so I’m wondering if it’s out of some special reason?
>
> If not, I think it would be great to add inverted class loading as the the behavior for the client and make it respect to the classloading configuration to make the components work in a uniform manner.
>
> Moreover, I’ve seen some jobs failed to upgrade to 1.8, because we added hadoop classpath in the client classpath to compensate the removed convenience hadoop binaries that shades lots of common dependencies, which significantly increases the changes of dependencies conflicts.
>
> Looking forward to your feedbacks.
>
> Best,
> Paul Lam
>