Retrieve written records of a sink after job

[Flavio Pompermaier]

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

[Chesnay Schepler]

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

[Flavio Pompermaier]

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

[Chesnay Schepler]

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

[Flavio Pompermaier]

The problem here is that I don't know the vertex id of the sink..would it be possible to access the sink info by id? 

And couldn't be all those info attached to the JobExecutionResult (avoiding to set up all the rest connection etc)?

On Wed, Feb 14, 2018 at 12:44 PM, Chesnay Schepler <[hidden email]> wrote:

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

--

[Chesnay Schepler]

Technically yes, a subset of metrics is stored in the ExecutionGraph when the job finishes. (This is for example where the webUI derives the values from for finished jobs). However these are on the task level, and will not contain the number of incoming records if your sink is chained to another operator. Changing this would be a larger endeavor, and tbh i don't see this happening soon.

I'm afraid for now you're stuck with the REST API for finished jobs. (Correction for my previous mail: The metrics REST API cannot be used for finished jobs)

Alternatively, if you rather want to work on files/json you can enable job archiving by configuring the jobmanager.archive.fs.dir directory. When the job finishes this will contain a big JSON file for each job containing all responses that the UI would return for finished jobs.

On 14.02.2018 12:50, Flavio Pompermaier wrote:

The problem here is that I don't know the vertex id of the sink..would it be possible to access the sink info by id? 

And couldn't be all those info attached to the JobExecutionResult (avoiding to set up all the rest connection etc)?

On Wed, Feb 14, 2018 at 12:44 PM, Chesnay Schepler <[hidden email]> wrote:

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

--

[Flavio Pompermaier]

So, if I'm not wrong, the right way to do this is using accumulators..what do you think about my proposal to add an easy way to add to a sink an accumulator for the written/outputed records?

On Wed, Feb 14, 2018 at 1:08 PM, Chesnay Schepler <[hidden email]> wrote:

Technically yes, a subset of metrics is stored in the ExecutionGraph when the job finishes. (This is for example where the webUI derives the values from for finished jobs). However these are on the task level, and will not contain the number of incoming records if your sink is chained to another operator. Changing this would be a larger endeavor, and tbh i don't see this happening soon.

I'm afraid for now you're stuck with the REST API for finished jobs. (Correction for my previous mail: The metrics REST API cannot be used for finished jobs)

Alternatively, if you rather want to work on files/json you can enable job archiving by configuring the jobmanager.archive.fs.dir directory. When the job finishes this will contain a big JSON file for each job containing all responses that the UI would return for finished jobs.

On 14.02.2018 12:50, Flavio Pompermaier wrote:

The problem here is that I don't know the vertex id of the sink..would it be possible to access the sink info by id? 

And couldn't be all those info attached to the JobExecutionResult (avoiding to set up all the rest connection etc)?

On Wed, Feb 14, 2018 at 12:44 PM, Chesnay Schepler <[hidden email]> wrote:

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

--

--

[Fabian Hueske-2]

Hi Flavio,

Not sure if I would add this functionality to the sinks.

You could also add a MapFunction with a counting Accumulator right before each sink.

Best, Fabian

2018-02-14 14:11 GMT+01:00 Flavio Pompermaier <[hidden email]>:

So, if I'm not wrong, the right way to do this is using accumulators..what do you think about my proposal to add an easy way to add to a sink an accumulator for the written/outputed records?

On Wed, Feb 14, 2018 at 1:08 PM, Chesnay Schepler <[hidden email]> wrote:

Technically yes, a subset of metrics is stored in the ExecutionGraph when the job finishes. (This is for example where the webUI derives the values from for finished jobs). However these are on the task level, and will not contain the number of incoming records if your sink is chained to another operator. Changing this would be a larger endeavor, and tbh i don't see this happening soon.

I'm afraid for now you're stuck with the REST API for finished jobs. (Correction for my previous mail: The metrics REST API cannot be used for finished jobs)

Alternatively, if you rather want to work on files/json you can enable job archiving by configuring the jobmanager.archive.fs.dir directory. When the job finishes this will contain a big JSON file for each job containing all responses that the UI would return for finished jobs.

On 14.02.2018 12:50, Flavio Pompermaier wrote:

The problem here is that I don't know the vertex id of the sink..would it be possible to access the sink info by id? 

And couldn't be all those info attached to the JobExecutionResult (avoiding to set up all the rest connection etc)?

On Wed, Feb 14, 2018 at 12:44 PM, Chesnay Schepler <[hidden email]> wrote:

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

--

--

[Flavio Pompermaier]

Hi Fabian,

thanks for the feedback. Right now I'm doing exactly as you said. 

Since I was seeing this as a useful API extension I just proposed this addition and so I asked for feedbacks.

Of course, it doesn't make much sense if I'm the only one asking for it :)

Best,

Flavio

On Mon, Feb 19, 2018 at 10:31 AM, Fabian Hueske <[hidden email]> wrote:

Hi Flavio,

Not sure if I would add this functionality to the sinks.

You could also add a MapFunction with a counting Accumulator right before each sink.

Best, Fabian

2018-02-14 14:11 GMT+01:00 Flavio Pompermaier <[hidden email]>:

So, if I'm not wrong, the right way to do this is using accumulators..what do you think about my proposal to add an easy way to add to a sink an accumulator for the written/outputed records?

On Wed, Feb 14, 2018 at 1:08 PM, Chesnay Schepler <[hidden email]> wrote:

Technically yes, a subset of metrics is stored in the ExecutionGraph when the job finishes. (This is for example where the webUI derives the values from for finished jobs). However these are on the task level, and will not contain the number of incoming records if your sink is chained to another operator. Changing this would be a larger endeavor, and tbh i don't see this happening soon.

I'm afraid for now you're stuck with the REST API for finished jobs. (Correction for my previous mail: The metrics REST API cannot be used for finished jobs)

Alternatively, if you rather want to work on files/json you can enable job archiving by configuring the jobmanager.archive.fs.dir directory. When the job finishes this will contain a big JSON file for each job containing all responses that the UI would return for finished jobs.

On 14.02.2018 12:50, Flavio Pompermaier wrote:

The problem here is that I don't know the vertex id of the sink..would it be possible to access the sink info by id? 

And couldn't be all those info attached to the JobExecutionResult (avoiding to set up all the rest connection etc)?

On Wed, Feb 14, 2018 at 12:44 PM, Chesnay Schepler <[hidden email]> wrote:

The only way to access this info from the client is the REST API or the Metrics REST API.

On 14.02.2018 12:38, Flavio Pompermaier wrote:

Actually I'd like to get this number from my Java class in order to update some external dataset "catalog",

so I'm asking if there's some programmatic way to access this info (from JobExecutionResult for example).

On Wed, Feb 14, 2018 at 12:25 PM, Chesnay Schepler <[hidden email]> wrote:

Do you want to know how many records the sink received, or how many the sink wrote to the DB?
If it's the first you're in luck because we measure that already, check out the metrics documentation.
If it's the latter, then this issue is essentially covered by FLINK-7286 which aims at allowing functions
to modify the numRecordsIn/numRecordsOut counts.

On 14.02.2018 12:22, Flavio Pompermaier wrote:

Hi to all,

I have a (batch) job that writes to 1 or more sinks. 

Is there a way to retrieve, once the job has terminated, the number of records written to each sink?

Is there any better way than than using an accumulator for each sink?

If that is the only way to do that, the Sink API could be enriched in order to automatically create an accumulator when required. E.g. 

dataset.output(JDBCOutputFormat.buildJDBCOutputFormat()

            .setDrivername(...)

            .setDBUrl(...)

            .setQuery(...)

            .addRecordsCountAccumulator("some-name")

            .finish())

Best,

Flavio

--

--

--

--