[ 
https://issues.apache.org/jira/browse/FLINK-27137?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17519852#comment-17519852
 ] 

dengziming commented on FLINK-27137:
------------------------------------

[~arvid] In fact, Spark also use Adminclient to get offset in 
[https://github.com/apache/spark/pull/29729] .

[~gyfora] 

 

> Since Flink will absolutely need the KafkaConsumer API

Event though we are doomed to use KafkaConsumer, however KafkaConsumer is 
created in TaskManager whereas we seem to get offsets in JobManager? so maybe 
we will not create duplicated component here. And KafkaConsumer contains extra 
component, for example `Fetcher` and `ConsumerCoordinator` which may send extra 
rpc requests.

> regular kafka applications to increase compatibility

AdminClient is also supported by Kafka community and thus is compatible and 
robust as KafkaConsumer, it's designed just like the Flink public API.

> In some environments using the AdminClient can be problematic 

If AdminClient has bugs we should report it to Kafka community to help improve 
it.

> Looking at the OffsetRetriever there are 4 requirements here

In fact there maybe new type of OffsetSpec, for example max-timestamp in 
KIP-734, even though it's trivial but indicate AdminClient is better concerning 
offset management.

> Remove usage of AdminClient from KafkaSource logic
> --------------------------------------------------
>
>                 Key: FLINK-27137
>                 URL: https://issues.apache.org/jira/browse/FLINK-27137
>             Project: Flink
>          Issue Type: Improvement
>          Components: Connectors / Kafka
>    Affects Versions: 1.15.0, 1.14.4
>            Reporter: Gyula Fora
>            Priority: Major
>
> Parts of the KafkaSource logic (specifically the KafkaSourceEnumerator) uses 
> the Kafka AdminClient instead of the KafkaConsumer.
> It seems that the KafkaConsumer already provides all the necessary 
> information that the enumerator needs so there is no reason for introducing 
> the AdminClient.
> In some environments using the AdminClient can be problematic even if we are 
> not using certaing features.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)

Reply via email to