NoClassDefFoundError in failing-restarting job that uses url classloader

[Subramanyam Ramanathan]

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

 

[Zhu Zhu]

Hi Subramanyam,

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

This information would be helpful to identify the failure cause.

Thanks,

Zhu Zhu

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

 

[Subramanyam Ramanathan]

Hi.

 

1)      The url pattern example : <a href="file:///\\root\flink-test\lib\dependency.jar">file:///root/flink-test/lib/dependency.jar

2)      I’m trying to simulate the same issue on a separate flink installation with a sample job so that I can share the logs. (However so far I’ve been unable to simulate it. Though on our product setup it can be simulated quite frequently. )

3)      The job is running in standalone mode. We have separate k8s pods with our own images which incorporate the taskmanager and jobmanager for our product. A 3^rd pod connects using k8s and submits the job

4)      Per job mode

 

I’m trying to simulate the issue on a separate flink installation outside of our produce env. I’ll update as soon as I have results.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

 

[Zhu Zhu]

Hi Subramanyam,

I think the standalone per job mode does not invoke PackagedProgram(File jarFile, List<URL> classpaths, String... args) to generate PackagedProgram, and thus does not add extra classpaths to the job.

Regarding the NoClassDefFoundError, there is another possibility that the class file exists but it has some static initialization process which may fail. This can also lead to the class to not be loaded and cause NoClassDefFoundError.

Thanks,

Zhu Zhu

Subramanyam Ramanathan <[hidden email]> 于2019年8月10日周六 下午2:38写道：

Hi.

 

1)      The url pattern example : file:///root/flink-test/lib/dependency.jar

2)      I’m trying to simulate the same issue on a separate flink installation with a sample job so that I can share the logs. (However so far I’ve been unable to simulate it. Though on our product setup it can be simulated quite frequently. )

3)      The job is running in standalone mode. We have separate k8s pods with our own images which incorporate the taskmanager and jobmanager for our product. A 3^rd pod connects using k8s and submits the job

4)      Per job mode

 

I’m trying to simulate the issue on a separate flink installation outside of our produce env. I’ll update as soon as I have results.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

 

[Subramanyam Ramanathan]

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ?

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

 

[Dian Fu]

Hi Subbu,

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

Regards,

Dian

[1] https://issues.apache.org/jira/browse/FLINK-10455

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

[Subramanyam Ramanathan]

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Zhu Zhu]

Hi Subramanyam,

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

Thanks,

Zhu Zhu

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Subramanyam Ramanathan]

Hi Zhu,

 

We also use FlinkKafkaProducer(011), hence I felt this fix would also be needed for us.

 

I agree that the fix for the issue I had originally mentioned would not be fixed by this, but I felt that I should be consuming this fix also.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 6:13 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

 

Thanks,

Zhu Zhu

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Zhu Zhu]

Hi Subramanyam, 

I checked the commits.

There are 2 fixes in FLINK-10455, only release 1.8.1 and release 1.9.0 contain both of them.

Thanks,

Zhu Zhu

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午11:02写道：

Hi Zhu,

 

We also use FlinkKafkaProducer(011), hence I felt this fix would also be needed for us.

 

I agree that the fix for the issue I had originally mentioned would not be fixed by this, but I felt that I should be consuming this fix also.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 6:13 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

 

Thanks,

Zhu Zhu

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Dian Fu]

Hi Subramanyam,

1.7.3 is not released yet. You need cherrypick these fixes if they really need them.

Regards,

Dian

在 2019年9月25日，上午12:08，Zhu Zhu <[hidden email]> 写道：

Hi Subramanyam, 

I checked the commits.

There are 2 fixes in FLINK-10455, only release 1.8.1 and release 1.9.0 contain both of them.

Thanks,

Zhu Zhu

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午11:02写道：

Hi Zhu,

 

We also use FlinkKafkaProducer(011), hence I felt this fix would also be needed for us.

 

I agree that the fix for the issue I had originally mentioned would not be fixed by this, but I felt that I should be consuming this fix also.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 6:13 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

 

Thanks,

Zhu Zhu

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Subramanyam Ramanathan]

Hi Zhu,

 

Thanks a lot !

Since 1.8.2 is also available, would it be right to assume 1.8.2 would also contain the fix ?

 

Thanks,

Subbu

 

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 9:39 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam, 

 

I checked the commits.

There are 2 fixes in FLINK-10455, only release 1.8.1 and release 1.9.0 contain both of them.

 

Thanks,

Zhu Zhu

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午11:02写道：

Hi Zhu,

 

We also use FlinkKafkaProducer(011), hence I felt this fix would also be needed for us.

 

I agree that the fix for the issue I had originally mentioned would not be fixed by this, but I felt that I should be consuming this fix also.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 6:13 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

 

Thanks,

Zhu Zhu

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu

 

[Zhu Zhu]

Yes. 1.8.2 contains all commits in 1.8.1.

Subramanyam Ramanathan <[hidden email]> 于2019年9月25日周三 下午5:03写道：

Hi Zhu,

 

Thanks a lot !

Since 1.8.2 is also available, would it be right to assume 1.8.2 would also contain the fix ?

 

Thanks,

Subbu

 

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 9:39 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam, 

 

I checked the commits.

There are 2 fixes in FLINK-10455, only release 1.8.1 and release 1.9.0 contain both of them.

 

Thanks,

Zhu Zhu

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午11:02写道：

Hi Zhu,

 

We also use FlinkKafkaProducer(011), hence I felt this fix would also be needed for us.

 

I agree that the fix for the issue I had originally mentioned would not be fixed by this, but I felt that I should be consuming this fix also.

 

Thanks,

Subbu

 

From: Zhu Zhu [mailto:[hidden email]]
Sent: Tuesday, September 24, 2019 6:13 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Dian Fu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

I think you do not need the fix in FLINK-10455 which is for Kafka only. It's just a similar issue as you met.

As you said, we need to make sure that the operator/UDF spawned threads are stopped in the close() method. In this way, we can avoid the thread to throw NoClassDefFoundError due to the class loader gets closed.

 

Thanks,

Zhu Zhu

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年9月24日周二 下午8:07写道：

Hi,

 

Thank you.

I think the takeaway for us is that we need to make sure that the threads are stopped in the close() method.

 

With regard to FLINK-10455, I see that the fix versions say : 1.5.6, 1.7.0, 1.7.3, 1.8.1, 1.9.0

 

However, I’m unable to find 1.7.3 in the downloads page(https://flink.apache.org/downloads.html). Is it yet to be released, or perhaps I am not looking in the right place ?

We’re currently using 1.7.2. Could you please let me know what is the minimal upgrade for me to consume the fix for FLINK-10455 ?

 

Thanks,

Subbu

 

From: Dian Fu [mailto:[hidden email]]
Sent: Monday, September 23, 2019 1:54 PM
To: Subramanyam Ramanathan <[hidden email]>
Cc: Zhu Zhu <[hidden email]>; [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subbu,

 

The issue you encountered is very similar to the issue which has been fixed in FLINK-10455 [1]. Could you check if that fix could solve your problem? The root cause for that issue is that the method close() has not closed all things. After the method "close()" is called, the classloader (URLClassloader) will be closed. If there is thread still running after "close()" method is called, it may access the classes in user provided jars. However, as the URLClassloader has already been closed, NoClassDefFoundError will be thrown.

 

Regards,

Dian

 

[1] https://issues.apache.org/jira/browse/FLINK-10455

 

在 2019年9月23日，下午2:50，Subramanyam Ramanathan <[hidden email]> 写道：

 

Hi,

 

I was able to simulate the issue again and understand the cause a little better.

 

The issue occurs when :

-        One of the RichMapFunction transformations uses a third party library in the open() method that spawns a thread.

-        The thread doesn’t get properly closed in the close() method.

-        Once the job starts failing, we start seeing a NoClassDefFound error from that thread.

 

I understand that cleanup should be done in the close() method. However, just wanted to know, do we have some kind of a configuration setting  which would help us clean up such threads ? 

I can attach the code if required.

 

Thanks,

Subbu

 

From: Zhu Zhu [[hidden email]] 
Sent: Friday, August 9, 2019 7:43 AM
To: Subramanyam Ramanathan <[hidden email]>
Cc: [hidden email]
Subject: Re: NoClassDefFoundError in failing-restarting job that uses url classloader

 

Hi Subramanyam,

 

Could you share more information? including:

1. the URL pattern

2. the detailed exception and the log around it

3. the cluster the job is running on, e.g. standalone, yarn, k8s

4. it's session mode or per job mode

 

This information would be helpful to identify the failure cause.

 

Thanks,

Zhu Zhu

 

 

 

 

 

 

 

 

 

 

 

Subramanyam Ramanathan <[hidden email]> 于2019年8月9日周五 上午1:45写道：

 

Hello,

 

I'm currently using flink 1.7.2.

 

I'm trying to run a job that's submitted programmatically using the ClusterClient API.

               public JobSubmissionResult run(PackagedProgram prog, int parallelism)

 

 

The job makes use of some jars which I add to the packaged program through the Packaged constructor, along with the Jar file.

   public PackagedProgram(File jarFile, List<URL> classpaths, String... args)

Normally, This works perfectly and the job runs fine.

 

However, if there's an error in the job, and the job goes into failing state and when it's continously  trying to restart the job for an hour or so, I notice a NoClassDefFoundError for some classes in the jars that I load using the URL class loader and the job never recovers after that, even if the root cause of the issue was fixed (I had a kafka source/sink in my job, and kafka was down temporarily, and was brought up after that).

The jar is still available at the path referenced by the url classloader and is not tampered with.

 

Could anyone please give me some pointers with regard to the reason why this could happen/what I could be missing here/how can I debug further ?

 

thanks

Subbu