Re: [Consulting] Should a Flink class loader be a singleton?

[Ken Krugler <kk...@transpac.com>]

Hi all,

I’m running into the exact same issue, though with Flink 1.15.1. I get a ClassCastException with HoodieWrapperFileSystem being loaded from two different ChildFirstClassLoader instances:

Caused by: java.lang.ClassCastException: class shaded.org.apache.hudi.common.fs.HoodieWrapperFileSystem cannot be cast to class shaded.org.apache.hudi.common.fs.HoodieWrapperFileSystem (shaded.org.apache.hudi.common.fs.HoodieWrapperFileSystem is in unnamed module of loader org.apache.flink.util.ChildFirstClassLoader @52deeb14; shaded.org.apache.hudi.common.fs.HoodieWrapperFileSystem is in unnamed module of loader org.apache.flink.util.ChildFirstClassLoader @5dab96b9)

Switching to parent-first doesn’t solve the problem.

Did this ever get resolved?

When running w/the MiniCluster this doesn’t happen.

Regards,

— Ken


> On Jan 31, 2022, at 5:34 AM, Till Rohrmann <tr...@apache.org> wrote:
> 
> Hi Zhiping,
> 
> I am pulling in Timo who knows the sql client probably best. He might be
> able to help you. In general, this sounds like a bug to me. Maybe you can
> provide us with the full stack trace and maybe even the debug logs of the
> failing run.
> 
> Cheers,
> Till
> 
> On Sun, Jan 30, 2022 at 4:21 AM Zhiping Wu <st...@gmail.com> wrote:
> 
>> Hi there,
>> 
>> I'm trying to consult whether a flink class loader(ParentFirstClassLoader
>> or ChildFirstClassLoader) must be a singleton?
>> 
>> I met a problem with the below error message by using sql-client to submit
>> sql jobs to a standalone session model cluster with Flink 1.13.3.
>> I also tried ParentFirstClassLoader & ChildFirstClassLoader, both of them
>> will meet this problem.
>> 
>> Caused by: java.lang.ClassCastException:
>> class org.apache.hudi.common.fs.HoodieWrapperFileSystem cannot be cast to
>> class org.apache.hudi.common.fs.HoodieWrapperFileSystem
>> (org.apache.hudi.common.fs.HoodieWrapperFileSystem
>> is in unnamed module of loader
>> org.apache.flink.runtime.execution.librarycache.
>> FlinkUserCodeClassLoaders$ParentFirstClassLoader @*384d2142*;
>> org.apache.hudi.common.fs.HoodieWrapperFileSystem
>> is in unnamed module of loader
>> org.apache.flink.runtime.execution.librarycache.
>> FlinkUserCodeClassLoaders$ParentFirstClassLoader @*35e21dd1*)
>> 
>> From this documentation, it shows I might have two versions of the class
>> which are different in application code and flink libraries.
>> But in my case, Flink should don't contains HoodieWrapperFileSystem class,
>> and I didn't add the hudi-flink-bundle jar to flink lib folder, instead, i
>> used sql-client.sh embedded -j xxxx/huid-flink-bundle-xxx.jar.
>> 
>> So I'm wondering if the root reason is that they are loaded by two
>> classloader instances? And it seems that FlinkUserCodeClassLoaders is a
>> singleton, but ParentFirstClassLoader and ChildFirstClassLoader aren't from
>> flink source code.
>> 

--------------------------
Ken Krugler
http://www.scaleunlimited.com
Custom big data solutions
Flink, Pinot, Solr, Elasticsearch