deploying flink in AWS - some teething issues

[Chakravarthy varaga]

Hi Team,

   If the flink cluster is conainerized and managed through by a container orchestrator,  

    1.  the orchestrator allocates resources for each JM. TM etc., say if the container (TM) needs to run with 2G RAM, how should this allocation be honoured by the TM when its JVM starts. I'm thinking of wrapping up a script that determines the resource allocation for the container and writes the flink-conf.yaml before the TM starts the process. Is this the way to go?

    2. The container orchestrator looks at health of the containers and is however unaware of the job health status runnning inside the container/cluster. How should this be determined?

Best Regards

CVP

[Chakravarthy varaga]

Hi,

    Any updates here? I'm sure many would have faced similar issues like these, any help here is highly appreciated.

Best Regards

CVP

On Tue, Mar 28, 2017 at 5:47 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Team,

   If the flink cluster is conainerized and managed through by a container orchestrator,  

    1.  the orchestrator allocates resources for each JM. TM etc., say if the container (TM) needs to run with 2G RAM, how should this allocation be honoured by the TM when its JVM starts. I'm thinking of wrapping up a script that determines the resource allocation for the container and writes the flink-conf.yaml before the TM starts the process. Is this the way to go?

    2. The container orchestrator looks at health of the containers and is however unaware of the job health status runnning inside the container/cluster. How should this be determined?

Best Regards

CVP

[Patrick Lucas]

For 1., I think the standard approach would be to specify from without what the heap size should be. If you want an x MB heap, you could set your container memory limit to 1.3 * x or so (to account for overhead) and set taskmanager.heap.mb: x in your config.

The other way around—e.g. from inside the container determine its memory limit and divide it by 1.3—sounds interesting though, so please share if you have success with that.

For 2. I don't think there's really a good way yet to monitor the health of containerized jobs directly, so probably your best bet is to watch the job's metrics from outside the Flink cluster.

On Wed, Mar 29, 2017 at 10:58 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi,

    Any updates here? I'm sure many would have faced similar issues like these, any help here is highly appreciated.

Best Regards

CVP

On Tue, Mar 28, 2017 at 5:47 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Team,

   If the flink cluster is conainerized and managed through by a container orchestrator,  

    1.  the orchestrator allocates resources for each JM. TM etc., say if the container (TM) needs to run with 2G RAM, how should this allocation be honoured by the TM when its JVM starts. I'm thinking of wrapping up a script that determines the resource allocation for the container and writes the flink-conf.yaml before the TM starts the process. Is this the way to go?

    2. The container orchestrator looks at health of the containers and is however unaware of the job health status runnning inside the container/cluster. How should this be determined?

Best Regards

CVP

[Chakravarthy varaga]

Hi,

    With regards to logging (both Flink & application specific logs) within the container, are there best practices that you know of to get the logs to a centralized locations..

    For e.g. the flink TM's log are local inside the container and I don't wish to write to shared/mounted volumes, this means that I have to run separate daemon running inside the container along with the TM to transport these logs to another server or so...

    + I don't see that flink provides support for syslog to be able to connect to rsyslog etc.,

    Can you please advice a way to go here?

Best Regards

CVP

On Wed, Mar 29, 2017 at 11:33 AM, Patrick Lucas <[hidden email]> wrote:

For 1., I think the standard approach would be to specify from without what the heap size should be. If you want an x MB heap, you could set your container memory limit to 1.3 * x or so (to account for overhead) and set taskmanager.heap.mb: x in your config.

The other way around—e.g. from inside the container determine its memory limit and divide it by 1.3—sounds interesting though, so please share if you have success with that.

For 2. I don't think there's really a good way yet to monitor the health of containerized jobs directly, so probably your best bet is to watch the job's metrics from outside the Flink cluster.

On Wed, Mar 29, 2017 at 10:58 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi,

    Any updates here? I'm sure many would have faced similar issues like these, any help here is highly appreciated.

Best Regards

CVP

On Tue, Mar 28, 2017 at 5:47 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Team,

   If the flink cluster is conainerized and managed through by a container orchestrator,  

    1.  the orchestrator allocates resources for each JM. TM etc., say if the container (TM) needs to run with 2G RAM, how should this allocation be honoured by the TM when its JVM starts. I'm thinking of wrapping up a script that determines the resource allocation for the container and writes the flink-conf.yaml before the TM starts the process. Is this the way to go?

    2. The container orchestrator looks at health of the containers and is however unaware of the job health status runnning inside the container/cluster. How should this be determined?

Best Regards

CVP

[Patrick Lucas]

I think Log4j includes a Syslog appender—the log4j config included with Flink just logs to the logs/ dir, but you should just be able to modify it (log4j.properties) to suit your needs.

On Thu, Mar 30, 2017 at 2:39 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi,

    With regards to logging (both Flink & application specific logs) within the container, are there best practices that you know of to get the logs to a centralized locations..

    For e.g. the flink TM's log are local inside the container and I don't wish to write to shared/mounted volumes, this means that I have to run separate daemon running inside the container along with the TM to transport these logs to another server or so...

    + I don't see that flink provides support for syslog to be able to connect to rsyslog etc.,

    Can you please advice a way to go here?

Best Regards

CVP

On Wed, Mar 29, 2017 at 11:33 AM, Patrick Lucas <[hidden email]> wrote:

For 1., I think the standard approach would be to specify from without what the heap size should be. If you want an x MB heap, you could set your container memory limit to 1.3 * x or so (to account for overhead) and set taskmanager.heap.mb: x in your config.

The other way around—e.g. from inside the container determine its memory limit and divide it by 1.3—sounds interesting though, so please share if you have success with that.

For 2. I don't think there's really a good way yet to monitor the health of containerized jobs directly, so probably your best bet is to watch the job's metrics from outside the Flink cluster.

On Wed, Mar 29, 2017 at 10:58 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi,

    Any updates here? I'm sure many would have faced similar issues like these, any help here is highly appreciated.

Best Regards

CVP

On Tue, Mar 28, 2017 at 5:47 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Team,

   If the flink cluster is conainerized and managed through by a container orchestrator,  

    1.  the orchestrator allocates resources for each JM. TM etc., say if the container (TM) needs to run with 2G RAM, how should this allocation be honoured by the TM when its JVM starts. I'm thinking of wrapping up a script that determines the resource allocation for the container and writes the flink-conf.yaml before the TM starts the process. Is this the way to go?

    2. The container orchestrator looks at health of the containers and is however unaware of the job health status runnning inside the container/cluster. How should this be determined?

Best Regards

CVP