Custom(application) Metrics - Piggyback on Flink's metrics infra or not?

[Eswar Reddy]

Hi,

I see Flink support's built-in metrics to monitor various components of Flink. In addition, one can register application specific(custom) metrics to Flink's built-in metrics infra. The problem with this is user has to develop his custom metrics using Flink's metrics framework/API rather than a generic framework such as dropwizard. Alternatively, user can follow this approach where his   dropwizard metrics push code is co-located with actual app code within each Task and metrics are directly pushed to a backend writer(say, Graphite) from each Task.

In this alternative, I am aware of having to handle mapping spatial granularity of Flink's run-time with metrics namespace, but doing it myself should not a big effort. Fault-tolerance comes automatically since app code and metrics push code are co-located in the Task. Is there anything else Flink's metrics infra handles automatically? Based on this I'd weigh using good old dropwizard vs Flink specific metrics framework.

Finally, I guess feasibility an automatic dropwizard-to-flinkmetrics translation utility can be checked out, but I would like to first understand additional benefits of using flink's infra for custom metrics.

Thanks,

Eswar.

[Chesnay Schepler]

Hello Eswar,

as far as I'm aware the general structure of the Flink's metric system is rather similar to DropWizard. You can use DropWizard metrics by creating a simple wrapper, we even ship one for Histograms. Furthermore, you can also use DropWizard reporters, you only have to extend the DropWizardReporter class, essentially providing a factory method for your reporter.

Using Flinks infrastructure provides the following benefits:
* better resource usage, as only a single reporter instance per taskmanager exists
* access to system metrics
* namespace stuff; you cannot access all variables yourselves from a UDF without modifying the source of Flink; whether this is an advantage is of course dependent on what you are interested in

Regards,
Chesnay

On 20.09.2016 08:29, Eswar Reddy wrote:

Hi,

I see Flink support's built-in metrics to monitor various components of Flink. In addition, one can register application specific(custom) metrics to Flink's built-in metrics infra. The problem with this is user has to develop his custom metrics using Flink's metrics framework/API rather than a generic framework such as dropwizard. Alternatively, user can follow this approach where his   dropwizard metrics push code is co-located with actual app code within each Task and metrics are directly pushed to a backend writer(say, Graphite) from each Task.

In this alternative, I am aware of having to handle mapping spatial granularity of Flink's run-time with metrics namespace, but doing it myself should not a big effort. Fault-tolerance comes automatically since app code and metrics push code are co-located in the Task. Is there anything else Flink's metrics infra handles automatically? Based on this I'd weigh using good old dropwizard vs Flink specific metrics framework.

Finally, I guess feasibility an automatic dropwizard-to-flinkmetrics translation utility can be checked out, but I would like to first understand additional benefits of using flink's infra for custom metrics.

Thanks,

Eswar.

[Chawla,Sumit]

In addition, It supports enabling multiple Reporters.  You can have same data pushed to multiple systems.  Plus its very easy to write new reporter for doing any customization.

On Tue, Sep 20, 2016 at 2:10 AM, Chesnay Schepler <[hidden email]> wrote:

Hello Eswar,

as far as I'm aware the general structure of the Flink's metric system is rather similar to DropWizard. You can use DropWizard metrics by creating a simple wrapper, we even ship one for Histograms. Furthermore, you can also use DropWizard reporters, you only have to extend the DropWizardReporter class, essentially providing a factory method for your reporter.

Using Flinks infrastructure provides the following benefits:
* better resource usage, as only a single reporter instance per taskmanager exists
* access to system metrics
* namespace stuff; you cannot access all variables yourselves from a UDF without modifying the source of Flink; whether this is an advantage is of course dependent on what you are interested in

Regards,
Chesnay

On 20.09.2016 08:29, Eswar Reddy wrote:

Hi,

I see Flink support's built-in metrics to monitor various components of Flink. In addition, one can register application specific(custom) metrics to Flink's built-in metrics infra. The problem with this is user has to develop his custom metrics using Flink's metrics framework/API rather than a generic framework such as dropwizard. Alternatively, user can follow this approach where his   dropwizard metrics push code is co-located with actual app code within each Task and metrics are directly pushed to a backend writer(say, Graphite) from each Task.

In this alternative, I am aware of having to handle mapping spatial granularity of Flink's run-time with metrics namespace, but doing it myself should not a big effort. Fault-tolerance comes automatically since app code and metrics push code are co-located in the Task. Is there anything else Flink's metrics infra handles automatically? Based on this I'd weigh using good old dropwizard vs Flink specific metrics framework.

Finally, I guess feasibility an automatic dropwizard-to-flinkmetrics translation utility can be checked out, but I would like to first understand additional benefits of using flink's infra for custom metrics.

Thanks,

Eswar.

[Eswar Reddy]

Thank you Chesnay. Good to know there are few wrappers available to get best of both worlds. I may mostly go without piggybacking though to have more control and learning for now, but I will keep an eye for new benefits I will get in future via piggybacking.
The UDF point looks like a deal breaker, I will spend some more time understanding it( we can get Flink's runtime using this inside the UDF so by 'variables' you must have meant the metrics object that gets passed around)

@Sumit, Adding backend writers(reporter) is as much simple in DropWizard as well. Thanks for bringing it up though.

Thanks,

Eswar.

On Tue, Sep 20, 2016 at 11:33 PM, Chawla,Sumit <[hidden email]> wrote:

In addition, It supports enabling multiple Reporters.  You can have same data pushed to multiple systems.  Plus its very easy to write new reporter for doing any customization.

On Tue, Sep 20, 2016 at 2:10 AM, Chesnay Schepler <[hidden email]> wrote:

Hello Eswar,

as far as I'm aware the general structure of the Flink's metric system is rather similar to DropWizard. You can use DropWizard metrics by creating a simple wrapper, we even ship one for Histograms. Furthermore, you can also use DropWizard reporters, you only have to extend the DropWizardReporter class, essentially providing a factory method for your reporter.

Using Flinks infrastructure provides the following benefits:
* better resource usage, as only a single reporter instance per taskmanager exists
* access to system metrics
* namespace stuff; you cannot access all variables yourselves from a UDF without modifying the source of Flink; whether this is an advantage is of course dependent on what you are interested in

Regards,
Chesnay

On 20.09.2016 08:29, Eswar Reddy wrote:

Hi,

I see Flink support's built-in metrics to monitor various components of Flink. In addition, one can register application specific(custom) metrics to Flink's built-in metrics infra. The problem with this is user has to develop his custom metrics using Flink's metrics framework/API rather than a generic framework such as dropwizard. Alternatively, user can follow this approach where his   dropwizard metrics push code is co-located with actual app code within each Task and metrics are directly pushed to a backend writer(say, Graphite) from each Task.

In this alternative, I am aware of having to handle mapping spatial granularity of Flink's run-time with metrics namespace, but doing it myself should not a big effort. Fault-tolerance comes automatically since app code and metrics push code are co-located in the Task. Is there anything else Flink's metrics infra handles automatically? Based on this I'd weigh using good old dropwizard vs Flink specific metrics framework.

Finally, I guess feasibility an automatic dropwizard-to-flinkmetrics translation utility can be checked out, but I would like to first understand additional benefits of using flink's infra for custom metrics.

Thanks,

Eswar.

[Chesnay Schepler]

Actually i was wrong on the UDF point. By variables i meant the information that is encoded in the scope, like the subtask index, task name, taskmanager ID etc., however all these can be accessed from the MetricGroup that is returned by RuntimeContext#getMetricGroup(), which you can of course use in your UDF.

On 22.09.2016 05:47, Eswar Reddy wrote:

Thank you Chesnay. Good to know there are few wrappers available to get best of both worlds. I may mostly go without piggybacking though to have more control and learning for now, but I will keep an eye for new benefits I will get in future via piggybacking.
The UDF point looks like a deal breaker, I will spend some more time understanding it( we can get Flink's runtime using this inside the UDF so by 'variables' you must have meant the metrics object that gets passed around)

@Sumit, Adding backend writers(reporter) is as much simple in DropWizard as well. Thanks for bringing it up though.

Thanks,

Eswar.

On Tue, Sep 20, 2016 at 11:33 PM, Chawla,Sumit <[hidden email]> wrote:

In addition, It supports enabling multiple Reporters.  You can have same data pushed to multiple systems.  Plus its very easy to write new reporter for doing any customization.

On Tue, Sep 20, 2016 at 2:10 AM, Chesnay Schepler <[hidden email]> wrote:

Hello Eswar,

as far as I'm aware the general structure of the Flink's metric system is rather similar to DropWizard. You can use DropWizard metrics by creating a simple wrapper, we even ship one for Histograms. Furthermore, you can also use DropWizard reporters, you only have to extend the DropWizardReporter class, essentially providing a factory method for your reporter.

Using Flinks infrastructure provides the following benefits:
* better resource usage, as only a single reporter instance per taskmanager exists
* access to system metrics
* namespace stuff; you cannot access all variables yourselves from a UDF without modifying the source of Flink; whether this is an advantage is of course dependent on what you are interested in

Regards,
Chesnay

On 20.09.2016 08:29, Eswar Reddy wrote:

Hi,

I see Flink support's built-in metrics to monitor various components of Flink. In addition, one can register application specific(custom) metrics to Flink's built-in metrics infra. The problem with this is user has to develop his custom metrics using Flink's metrics framework/API rather than a generic framework such as dropwizard. Alternatively, user can follow this approach where his   dropwizard metrics push code is co-located with actual app code within each Task and metrics are directly pushed to a backend writer(say, Graphite) from each Task.

In this alternative, I am aware of having to handle mapping spatial granularity of Flink's run-time with metrics namespace, but doing it myself should not a big effort. Fault-tolerance comes automatically since app code and metrics push code are co-located in the Task. Is there anything else Flink's metrics infra handles automatically? Based on this I'd weigh using good old dropwizard vs Flink specific metrics framework.

Finally, I guess feasibility an automatic dropwizard-to-flinkmetrics translation utility can be checked out, but I would like to first understand additional benefits of using flink's infra for custom metrics.

Thanks,

Eswar.