[
https://issues.apache.org/jira/browse/BEAM-11213?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kyle Weaver updated BEAM-11213:
-------------------------------
Description:
All Beam metrics are visible in the Spark UI in a single accumulator value (in
the "Accumulators" tab), which is a large, hard-to-read blob. Originally, this
blob was rendered in a bespoke format
(https://github.com/apache/beam/blob/ead80b469ffeeddcd8e9e5c8dc462eec0b0ffc6b/sdks/java/core/src/main/java/org/apache/beam/sdk/metrics/MetricQueryResults.java#L63-L72).
I changed the format to JSON so it could be easily deserialized (BEAM-9600).
But then an issue was filed (BEAM-10294) reporting that the new JSON format was
harder to read than the original bespoke format. The temporary fix was to
revert to the bespoke format in Spark, while allowing Flink to continue to use
JSON. However, if Beam metrics in Flink are only visible as an accumulator,
then they are also unreadable because the payloads are in binary form
(BEAM-10719).
Having metrics visible in Spark's "Metrics" tab would A) make metrics easier to
read (even compared to the bespoke accumulator string format), and closer to
what users of Beamless Spark expect, and B) free us to use the accumulator
however we wish for Beam internal purposes, without worrying about readability.
was:
All Beam metrics are visible in the Flink UI in a single accumulator value (in
the "Accumulators" tab), which is a large, hard-to-read blob. Originally, this
blob was rendered in a bespoke format
(https://github.com/apache/beam/blob/ead80b469ffeeddcd8e9e5c8dc462eec0b0ffc6b/sdks/java/core/src/main/java/org/apache/beam/sdk/metrics/MetricQueryResults.java#L63-L72).
I changed the format to JSON so it could be easily deserialized (BEAM-9600).
But then an issue was filed (BEAM-10294) reporting that the new JSON format was
harder to read than the original bespoke format. The temporary fix was to
revert to the bespoke format in Spark, while allowing Flink to continue to use
JSON. However, if Beam metrics in Flink are only visible as an accumulator,
then they are also unreadable because the payloads are in binary form
(BEAM-10719).
Having metrics visible in Flink's "Metrics" tab would A) make metrics easier to
read (even compared to the bespoke accumulator string format), and closer to
what users of Beamless Flink expect, and B) free us to use the accumulator
however we wish for Beam internal purposes, without worrying about readability.
One question I'm not sure about is, why can't we see Beam metrics in the Flink
UI already? I thought we were already translating Beam metrics into Flink
native metrics
(https://github.com/apache/beam/blob/ea2a3f6896b66a2852a2ff3d82f4e1b010013d13/runners/flink/src/main/java/org/apache/beam/runners/flink/metrics/FlinkMetricContainer.java#L108-L109).
Is there something else we need to do to display them in the UI?
> Beam metrics should be displayed in Spark UI
> --------------------------------------------
>
> Key: BEAM-11213
> URL: https://issues.apache.org/jira/browse/BEAM-11213
> Project: Beam
> Issue Type: Wish
> Components: runner-flink
> Reporter: Kyle Weaver
> Priority: P2
> Labels: portability-flink
>
> All Beam metrics are visible in the Spark UI in a single accumulator value
> (in the "Accumulators" tab), which is a large, hard-to-read blob. Originally,
> this blob was rendered in a bespoke format
> (https://github.com/apache/beam/blob/ead80b469ffeeddcd8e9e5c8dc462eec0b0ffc6b/sdks/java/core/src/main/java/org/apache/beam/sdk/metrics/MetricQueryResults.java#L63-L72).
> I changed the format to JSON so it could be easily deserialized (BEAM-9600).
> But then an issue was filed (BEAM-10294) reporting that the new JSON format
> was harder to read than the original bespoke format. The temporary fix was to
> revert to the bespoke format in Spark, while allowing Flink to continue to
> use JSON. However, if Beam metrics in Flink are only visible as an
> accumulator, then they are also unreadable because the payloads are in binary
> form (BEAM-10719).
> Having metrics visible in Spark's "Metrics" tab would A) make metrics easier
> to read (even compared to the bespoke accumulator string format), and closer
> to what users of Beamless Spark expect, and B) free us to use the accumulator
> however we wish for Beam internal purposes, without worrying about
> readability.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
