Application log on Yarn FlinkCluster

[Theofilos Kakantousis]

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is
not included in the Yarn log. Instead, it is only printed in the stdout
from where I run my program.  For the jobmanager, I'm using the
log4j.properties file from the flink/conf directory. Yarn log
aggregation is enabled and the YarnJobManager log is printed in the yarn
log. The application is submitted by a Flink Client to the
FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the
application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[rmetzger0]

Hi Theofilos,

how exactly are you writing the application output?

Are you using a logging framework?

Are you writing the log statements from the open(), map(), invoke() methods or from some constructors? (I'm asking since different parts are executed on the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is not included in the Yarn log. Instead, it is only printed in the stdout from where I run my program.  For the jobmanager, I'm using the log4j.properties file from the flink/conf directory. Yarn log aggregation is enabled and the YarnJobManager log is printed in the yarn log. The application is submitted by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[Theofilos Kakantousis]

Hi Robert,

Thanks for the prompt reply. I'm using the IterateExample from Flink examples. In the yarn log I get entries for the YarnJobManager and ExecutionGraph, but I was wondering if there is a way to push all the logging that the client produces into the yarn log. Including the System.out calls. Is there a way to modify the example to use a logging framework to achieve it?

Also when I submit the program using the Client runBlocking method, although I see in the taskmanager and jobmanager log that the application has finished, the runBlocking method does not return. Should I call it in a separate thread?

Cheers,
Theofilos

On 2016-06-10 22:12, Robert Metzger wrote:

Hi Theofilos,

how exactly are you writing the application output?

Are you using a logging framework?

Are you writing the log statements from the open(), map(), invoke() methods or from some constructors? (I'm asking since different parts are executed on the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is not included in the Yarn log. Instead, it is only printed in the stdout from where I run my program.  For the jobmanager, I'm using the log4j.properties file from the flink/conf directory. Yarn log aggregation is enabled and the YarnJobManager log is printed in the yarn log. The application is submitted by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[Maximilian Michels]

Hi Theofilos,

Flink doesn't send the local client output to the Yarn cluster. I
think this will only change once we move the entire execution of the
Job to the cluster framework. All output of the actual Flink job
should be within the JobManager or TaskManager logs.

There is something wrong with the network communication if the Client
doesn't return from `runBlocking(..)`. Would be interesting to take a
look at the logs to find out why that could be.

Cheers,
Max


On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:

> Hi Robert,
>
> Thanks for the prompt reply. I'm using the IterateExample from Flink
> examples. In the yarn log I get entries for the YarnJobManager and
> ExecutionGraph, but I was wondering if there is a way to push all the
> logging that the client produces into the yarn log. Including the System.out
> calls. Is there a way to modify the example to use a logging framework to
> achieve it?
>
> Also when I submit the program using the Client runBlocking method, although
> I see in the taskmanager and jobmanager log that the application has
> finished, the runBlocking method does not return. Should I call it in a
> separate thread?
>
> Cheers,
> Theofilos
>
> On 2016-06-10 22:12, Robert Metzger wrote:
>
> Hi Theofilos,
>
> how exactly are you writing the application output?
> Are you using a logging framework?
> Are you writing the log statements from the open(), map(), invoke() methods
> or from some constructors? (I'm asking since different parts are executed on
> the cluster and locally).
>
> On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:
>>
>> Hi all,
>>
>> Flink 1.0.3
>> Hadoop 2.4.0
>>
>> When running a job on a Flink Cluster on Yarn, the application output is
>> not included in the Yarn log. Instead, it is only printed in the stdout from
>> where I run my program.  For the jobmanager, I'm using the log4j.properties
>> file from the flink/conf directory. Yarn log aggregation is enabled and the
>> YarnJobManager log is printed in the yarn log. The application is submitted
>> by a Flink Client to the FlinkYarnCluster using a PackagedProgram.
>>
>> Is this expected behavior and if so, is there a way to include the
>> application output in the Yarn aggregated log? Thanks!
>>
>> Cheers,
>> Theofilos
>>
>
>

[Theofilos Kakantousis]

Hi Max,

The runBlocking(..) problem was due to a Netty dependency issue on my
project, it works fine now :)

To pinpoint the logging issue, I just ran a single flink job on yarn as
per the documentation "./bin/flink run -m yarn-cluster -yn 2
./examples/streaming/Iteration.jar" and I have the same issue.During the
job I can see in the containers the taskmanager logs, and a sample
output from the taskmanager.out is the following:
"cat
/srv/hadoop/logs/userlogs/application_1465901188070_0037/container_1465901188070_0037_01_000002/taskmanager.out

2> ((49,1),3)
2> ((25,11),4)
2> ((46,44),2
.."

However, the yarn aggregated log contains only the jobmanager output. Is
this expected or could it indicate a problem with my hadoop logging
configuration not picking up taskmanager logs?

Cheers,
Theofilos

On 6/13/2016 12:13 PM, Maximilian Michels wrote:

> Hi Theofilos,
>
> Flink doesn't send the local client output to the Yarn cluster. I
> think this will only change once we move the entire execution of the
> Job to the cluster framework. All output of the actual Flink job
> should be within the JobManager or TaskManager logs.
>
> There is something wrong with the network communication if the Client
> doesn't return from `runBlocking(..)`. Would be interesting to take a
> look at the logs to find out why that could be.
>
> Cheers,
> Max
>
>
> On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:
>> Hi Robert,
>>
>> Thanks for the prompt reply. I'm using the IterateExample from Flink
>> examples. In the yarn log I get entries for the YarnJobManager and
>> ExecutionGraph, but I was wondering if there is a way to push all the
>> logging that the client produces into the yarn log. Including the System.out
>> calls. Is there a way to modify the example to use a logging framework to
>> achieve it?
>>
>> Also when I submit the program using the Client runBlocking method, although
>> I see in the taskmanager and jobmanager log that the application has
>> finished, the runBlocking method does not return. Should I call it in a
>> separate thread?
>>
>> Cheers,
>> Theofilos
>>
>> On 2016-06-10 22:12, Robert Metzger wrote:
>>
>> Hi Theofilos,
>>
>> how exactly are you writing the application output?
>> Are you using a logging framework?
>> Are you writing the log statements from the open(), map(), invoke() methods
>> or from some constructors? (I'm asking since different parts are executed on
>> the cluster and locally).
>>
>> On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:
>>> Hi all,
>>>
>>> Flink 1.0.3
>>> Hadoop 2.4.0
>>>
>>> When running a job on a Flink Cluster on Yarn, the application output is
>>> not included in the Yarn log. Instead, it is only printed in the stdout from
>>> where I run my program.  For the jobmanager, I'm using the log4j.properties
>>> file from the flink/conf directory. Yarn log aggregation is enabled and the
>>> YarnJobManager log is printed in the yarn log. The application is submitted
>>> by a Flink Client to the FlinkYarnCluster using a PackagedProgram.
>>>
>>> Is this expected behavior and if so, is there a way to include the
>>> application output in the Yarn aggregated log? Thanks!
>>>
>>> Cheers,
>>> Theofilos
>>>
>>

[Maximilian Michels]

Please use the `yarn logs -applicationId <id>` to retrieve the logs. If you have enabled log aggregation, this will give you all container logs concatenated.

Cheers,

Max

On Wed, Jun 15, 2016 at 12:24 AM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Max,

The runBlocking(..) problem was due to a Netty dependency issue on my project, it works fine now :)

To pinpoint the logging issue, I just ran a single flink job on yarn as per the documentation "./bin/flink run -m yarn-cluster -yn 2 ./examples/streaming/Iteration.jar" and I have the same issue.During the job I can see in the containers the taskmanager logs, and a sample output from the taskmanager.out is the following:
"cat /srv/hadoop/logs/userlogs/application_1465901188070_0037/container_1465901188070_0037_01_000002/taskmanager.out
2> ((49,1),3)
2> ((25,11),4)
2> ((46,44),2
.."

However, the yarn aggregated log contains only the jobmanager output. Is this expected or could it indicate a problem with my hadoop logging configuration not picking up taskmanager logs?

Cheers,
Theofilos

On 6/13/2016 12:13 PM, Maximilian Michels wrote:

Hi Theofilos,

Flink doesn't send the local client output to the Yarn cluster. I
think this will only change once we move the entire execution of the
Job to the cluster framework. All output of the actual Flink job
should be within the JobManager or TaskManager logs.

There is something wrong with the network communication if the Client
doesn't return from `runBlocking(..)`. Would be interesting to take a
look at the logs to find out why that could be.

Cheers,
Max


On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Robert,

Thanks for the prompt reply. I'm using the IterateExample from Flink
examples. In the yarn log I get entries for the YarnJobManager and
ExecutionGraph, but I was wondering if there is a way to push all the
logging that the client produces into the yarn log. Including the System.out
calls. Is there a way to modify the example to use a logging framework to
achieve it?

Also when I submit the program using the Client runBlocking method, although
I see in the taskmanager and jobmanager log that the application has
finished, the runBlocking method does not return. Should I call it in a
separate thread?

Cheers,
Theofilos

On 2016-06-10 22:12, Robert Metzger wrote:

Hi Theofilos,

how exactly are you writing the application output?
Are you using a logging framework?
Are you writing the log statements from the open(), map(), invoke() methods
or from some constructors? (I'm asking since different parts are executed on
the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is
not included in the Yarn log. Instead, it is only printed in the stdout from
where I run my program.  For the jobmanager, I'm using the log4j.properties
file from the flink/conf directory. Yarn log aggregation is enabled and the
YarnJobManager log is printed in the yarn log. The application is submitted
by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the
application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[Theofilos Kakantousis]

Hi,

By yarn aggregated log I mean Yarn log aggregation is enabled and the log I'm referring to is the one returned by `yarn logs -applicationId <id>`. When running a Spark job for example on the same setup, the yarn aggregated log contains all the information printed out by the application.

Cheers,
Theofilos

On 6/15/2016 10:14 AM, Maximilian Michels wrote:

Please use the `yarn logs -applicationId <id>` to retrieve the logs. If you have enabled log aggregation, this will give you all container logs concatenated.

Cheers,

Max

On Wed, Jun 15, 2016 at 12:24 AM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Max,

The runBlocking(..) problem was due to a Netty dependency issue on my project, it works fine now :)

To pinpoint the logging issue, I just ran a single flink job on yarn as per the documentation "./bin/flink run -m yarn-cluster -yn 2 ./examples/streaming/Iteration.jar" and I have the same issue.During the job I can see in the containers the taskmanager logs, and a sample output from the taskmanager.out is the following:
"cat /srv/hadoop/logs/userlogs/application_1465901188070_0037/container_1465901188070_0037_01_000002/taskmanager.out
2> ((49,1),3)
2> ((25,11),4)
2> ((46,44),2
.."

However, the yarn aggregated log contains only the jobmanager output. Is this expected or could it indicate a problem with my hadoop logging configuration not picking up taskmanager logs?

Cheers,
Theofilos

On 6/13/2016 12:13 PM, Maximilian Michels wrote:

Hi Theofilos,

Flink doesn't send the local client output to the Yarn cluster. I
think this will only change once we move the entire execution of the
Job to the cluster framework. All output of the actual Flink job
should be within the JobManager or TaskManager logs.

There is something wrong with the network communication if the Client
doesn't return from `runBlocking(..)`. Would be interesting to take a
look at the logs to find out why that could be.

Cheers,
Max


On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Robert,

Thanks for the prompt reply. I'm using the IterateExample from Flink
examples. In the yarn log I get entries for the YarnJobManager and
ExecutionGraph, but I was wondering if there is a way to push all the
logging that the client produces into the yarn log. Including the System.out
calls. Is there a way to modify the example to use a logging framework to
achieve it?

Also when I submit the program using the Client runBlocking method, although
I see in the taskmanager and jobmanager log that the application has
finished, the runBlocking method does not return. Should I call it in a
separate thread?

Cheers,
Theofilos

On 2016-06-10 22:12, Robert Metzger wrote:

Hi Theofilos,

how exactly are you writing the application output?
Are you using a logging framework?
Are you writing the log statements from the open(), map(), invoke() methods
or from some constructors? (I'm asking since different parts are executed on
the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is
not included in the Yarn log. Instead, it is only printed in the stdout from
where I run my program.  For the jobmanager, I'm using the log4j.properties
file from the flink/conf directory. Yarn log aggregation is enabled and the
YarnJobManager log is printed in the yarn log. The application is submitted
by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the
application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[Maximilian Michels]

You should also see TaskManager output in the logs. I just verified this using Flink 1.0.3 with Hadoop 2.7.1. I executed the Iterate example and it aggregated correctly including the TaskManager logs.

I'm wondering, is there anything in the Hadoop logs of the Resourcemanager/Nodemanager that could indicate a transfer failure?

On Wed, Jun 15, 2016 at 10:18 AM, Theofilos Kakantousis <[hidden email]> wrote:

Hi,

By yarn aggregated log I mean Yarn log aggregation is enabled and the log I'm referring to is the one returned by `yarn logs -applicationId <id>`. When running a Spark job for example on the same setup, the yarn aggregated log contains all the information printed out by the application.

Cheers,
Theofilos

On 6/15/2016 10:14 AM, Maximilian Michels wrote:

Please use the `yarn logs -applicationId <id>` to retrieve the logs. If you have enabled log aggregation, this will give you all container logs concatenated.

Cheers,

Max

On Wed, Jun 15, 2016 at 12:24 AM, Theofilos Kakantousis <[hidden email][hidden email]> wrote:

Hi Max,

The runBlocking(..) problem was due to a Netty dependency issue on my project, it works fine now :)

To pinpoint the logging issue, I just ran a single flink job on yarn as per the documentation "./bin/flink run -m yarn-cluster -yn 2 ./examples/streaming/Iteration.jar" and I have the same issue.During the job I can see in the containers the taskmanager logs, and a sample output from the taskmanager.out is the following:
"cat /srv/hadoop/logs/userlogs/application_1465901188070_0037/container_1465901188070_0037_01_000002/taskmanager.out
2> ((49,1),3)
2> ((25,11),4)
2> ((46,44),2
.."

However, the yarn aggregated log contains only the jobmanager output. Is this expected or could it indicate a problem with my hadoop logging configuration not picking up taskmanager logs?

Cheers,
Theofilos

On 6/13/2016 12:13 PM, Maximilian Michels wrote:

Hi Theofilos,

Flink doesn't send the local client output to the Yarn cluster. I
think this will only change once we move the entire execution of the
Job to the cluster framework. All output of the actual Flink job
should be within the JobManager or TaskManager logs.

There is something wrong with the network communication if the Client
doesn't return from `runBlocking(..)`. Would be interesting to take a
look at the logs to find out why that could be.

Cheers,
Max


On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Robert,

Thanks for the prompt reply. I'm using the IterateExample from Flink
examples. In the yarn log I get entries for the YarnJobManager and
ExecutionGraph, but I was wondering if there is a way to push all the
logging that the client produces into the yarn log. Including the System.out
calls. Is there a way to modify the example to use a logging framework to
achieve it?

Also when I submit the program using the Client runBlocking method, although
I see in the taskmanager and jobmanager log that the application has
finished, the runBlocking method does not return. Should I call it in a
separate thread?

Cheers,
Theofilos

On 2016-06-10 22:12, Robert Metzger wrote:

Hi Theofilos,

how exactly are you writing the application output?
Are you using a logging framework?
Are you writing the log statements from the open(), map(), invoke() methods
or from some constructors? (I'm asking since different parts are executed on
the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is
not included in the Yarn log. Instead, it is only printed in the stdout from
where I run my program.  For the jobmanager, I'm using the log4j.properties
file from the flink/conf directory. Yarn log aggregation is enabled and the
YarnJobManager log is printed in the yarn log. The application is submitted
by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the
application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos

[Theofilos Kakantousis]

Great then, I will look into my configuration. Thanks for your help!

Cheers,
Theofilos

On 6/15/2016 2:00 PM, Maximilian Michels wrote:

You should also see TaskManager output in the logs. I just verified this using Flink 1.0.3 with Hadoop 2.7.1. I executed the Iterate example and it aggregated correctly including the TaskManager logs.

I'm wondering, is there anything in the Hadoop logs of the Resourcemanager/Nodemanager that could indicate a transfer failure?

On Wed, Jun 15, 2016 at 10:18 AM, Theofilos Kakantousis <[hidden email]> wrote:

Hi,

By yarn aggregated log I mean Yarn log aggregation is enabled and the log I'm referring to is the one returned by `yarn logs -applicationId <id>`. When running a Spark job for example on the same setup, the yarn aggregated log contains all the information printed out by the application.

Cheers,
Theofilos

On 6/15/2016 10:14 AM, Maximilian Michels wrote:

Please use the `yarn logs -applicationId <id>` to retrieve the logs. If you have enabled log aggregation, this will give you all container logs concatenated.

Cheers,

Max

On Wed, Jun 15, 2016 at 12:24 AM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Max,

The runBlocking(..) problem was due to a Netty dependency issue on my project, it works fine now :)

To pinpoint the logging issue, I just ran a single flink job on yarn as per the documentation "./bin/flink run -m yarn-cluster -yn 2 ./examples/streaming/Iteration.jar" and I have the same issue.During the job I can see in the containers the taskmanager logs, and a sample output from the taskmanager.out is the following:
"cat /srv/hadoop/logs/userlogs/application_1465901188070_0037/container_1465901188070_0037_01_000002/taskmanager.out
2> ((49,1),3)
2> ((25,11),4)
2> ((46,44),2
.."

However, the yarn aggregated log contains only the jobmanager output. Is this expected or could it indicate a problem with my hadoop logging configuration not picking up taskmanager logs?

Cheers,
Theofilos

On 6/13/2016 12:13 PM, Maximilian Michels wrote:

Hi Theofilos,

Flink doesn't send the local client output to the Yarn cluster. I
think this will only change once we move the entire execution of the
Job to the cluster framework. All output of the actual Flink job
should be within the JobManager or TaskManager logs.

There is something wrong with the network communication if the Client
doesn't return from `runBlocking(..)`. Would be interesting to take a
look at the logs to find out why that could be.

Cheers,
Max


On Sat, Jun 11, 2016 at 1:53 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi Robert,

Thanks for the prompt reply. I'm using the IterateExample from Flink
examples. In the yarn log I get entries for the YarnJobManager and
ExecutionGraph, but I was wondering if there is a way to push all the
logging that the client produces into the yarn log. Including the System.out
calls. Is there a way to modify the example to use a logging framework to
achieve it?

Also when I submit the program using the Client runBlocking method, although
I see in the taskmanager and jobmanager log that the application has
finished, the runBlocking method does not return. Should I call it in a
separate thread?

Cheers,
Theofilos

On 2016-06-10 22:12, Robert Metzger wrote:

Hi Theofilos,

how exactly are you writing the application output?
Are you using a logging framework?
Are you writing the log statements from the open(), map(), invoke() methods
or from some constructors? (I'm asking since different parts are executed on
the cluster and locally).

On Fri, Jun 10, 2016 at 4:00 PM, Theofilos Kakantousis <[hidden email]> wrote:

Hi all,

Flink 1.0.3
Hadoop 2.4.0

When running a job on a Flink Cluster on Yarn, the application output is
not included in the Yarn log. Instead, it is only printed in the stdout from
where I run my program.  For the jobmanager, I'm using the log4j.properties
file from the flink/conf directory. Yarn log aggregation is enabled and the
YarnJobManager log is printed in the yarn log. The application is submitted
by a Flink Client to the FlinkYarnCluster using a PackagedProgram.

Is this expected behavior and if so, is there a way to include the
application output in the Yarn aggregated log? Thanks!

Cheers,
Theofilos