[
https://issues.apache.org/jira/browse/FLINK-18808?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17176925#comment-17176925
]
Piotr Nowojski commented on FLINK-18808:
----------------------------------------
But isn't the whole point of the discussion? The version posted in your PR
counts all of the outputs from 3rd level (including {{ChainingOutput}}), and
this is not what we want to achieve.
Given the following operators in a single task:
{noformat}
/-> SideOutput1
A -> B -> C -> NetworkOutput
\-> SideOutput2
{noformat}
For the task level {{numRecordsOut}} (aka number of records produced by the
task) we want to report the number of records that reached {{SideOutput1 +
SideOutput2 + NetworkOutput}}. That's the number of records produced by this
task. Currently we are ignoring side outputs of {{B}}.
For the hypothetical {{numRecordsSent}} (would be nice to have, but not must
have), we could report the number of records that were sent by {{SideOutput1 +
SideOutput2 + NetworkOutput}} (taking into account that broadcasting multiplies
the number of records).
For the operator level {{numRecordsOut}} we currently have (that should stay
unchanged):
* Operator A: number of records that reached {{ChainingOutput A->B}}
* Operator B: {{SideOutput1 + SideOutput2}} + number of records that reached
{{ChainingOutput B->C}}
* Operator C: number of records that reached {{NetworkInput}}
> Task-level numRecordsOut metric may be underestimated
> -----------------------------------------------------
>
> Key: FLINK-18808
> URL: https://issues.apache.org/jira/browse/FLINK-18808
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Metrics, Runtime / Task
> Affects Versions: 1.11.1
> Reporter: ming li
> Assignee: ming li
> Priority: Major
> Labels: pull-request-available, usability
> Attachments: image-2020-08-04-11-28-13-800.png,
> image-2020-08-04-11-32-20-678.png, image-2020-08-13-18-36-13-282.png
>
>
> At present, we only register task-level numRecordsOut metric by reusing
> operator output record counter at the end of OperatorChain.
> {code:java}
> if (config.isChainEnd()) {
> operatorMetricGroup.getIOMetricGroup().reuseOutputMetricsForTask();
> }
> {code}
> If we only send data out through the last operator of OperatorChain, there is
> no problem with this statistics. But consider the following scenario:
> !image-2020-08-04-11-28-13-800.png|width=507,height=174!
> In this JobGraph, we not only send data in the last operator, but also send
> data in the middle operator of OperatorChain (the map operator just returns
> the original value directly). Below is one of our test topology, we can see
> that the statistics actually only have half of the total data received by the
> downstream.
> !image-2020-08-04-11-32-20-678.png|width=648,height=251!
> I think the data sent out by the intermediate operator should also be counted
> into the numRecordsOut of the Task. But currently we are not reusing
> operators output record counters in the intermediate operators, which leads
> to our task-level numRecordsOut metric is underestimated (although this has
> no effect on the actual operation of the job, it may affect our monitoring).
> A simple idea of mine is to modify the condition of reusing operators
> output record counter:
> {code:java}
> if (!config.getNonChainedOutputs(getUserCodeClassloader()).isEmpty()) {
> operatorMetricGroup.getIOMetricGroup().reuseOutputMetricsForTask();
> }{code}
> In addition, I have another question: If a record is broadcast to all
> downstream, should the numRecordsOut counter increase by one or the
> downstream number? It seems that currently we are adding one to calculate the
> numRecordsOut metric.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
