[ 
https://issues.apache.org/jira/browse/FLINK-11912?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16808541#comment-16808541
 ] 

aitozi commented on FLINK-11912:
--------------------------------

Hi [~suez1224] , [~tzulitai]

I have tried to solve this bug in a very old flink version by tryRegister 
metric for a time after the poll message from kafka as you did in the 
experimental change you posted . But in production, I find there is also a 
possibility that we can lose the per partition lag, because may be not all 
partition lag metric are registered after a poll call, as i mentioned in 
[FLINK-7945|https://github.com/apache/flink/pull/4935#issuecomment-348252535] . 
 

> Expose per partition Kafka lag metric in Flink Kafka connector
> --------------------------------------------------------------
>
>                 Key: FLINK-11912
>                 URL: https://issues.apache.org/jira/browse/FLINK-11912
>             Project: Flink
>          Issue Type: New Feature
>          Components: Connectors / Kafka
>    Affects Versions: 1.6.4, 1.7.2
>            Reporter: Shuyi Chen
>            Assignee: Shuyi Chen
>            Priority: Major
>
> In production, it's important that we expose the Kafka lag by partition 
> metric in order for users to diagnose which Kafka partition is lagging. 
> However, although the Kafka lag by partition metrics are available in 
> KafkaConsumer after 0.10.2,  Flink was not able to properly register it 
> because the metrics are only available after the consumer start polling data 
> from partitions. I would suggest the following fix:
> 1) In KafkaConsumerThread.run(), allocate a manualRegisteredMetricSet.
> 2) in the fetch loop, as KafkaConsumer discovers new partitions, manually add 
> MetricName for those partitions that we want to register into 
> manualRegisteredMetricSet. 
> 3) in the fetch loop, check if manualRegisteredMetricSet is empty. If not, 
> try to search for the metrics available in KafkaConsumer, and if found, 
> register it and remove the entry from manualRegisteredMetricSet. 
> The overhead of the above approach is bounded and only incur when discovering 
> new partitions, and registration is done once the KafkaConsumer have the 
> metrics exposed.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to