Github user zsxwing commented on the issue: https://github.com/apache/spark/pull/15387 > During the original implementation I had verified that calling pause kills the internal message buffer, which is one of the complications leading to a cached consumer per partition. I observed the same behavior during my debug. I found that the first `poll(0)` will always send a request to prefetch the data. Pausing partitions just prevents the second `poll(0)` from returning anything at here: https://github.com/apache/kafka/blob/0.10.0.1/clients/src/main/java/org/apache/kafka/clients/consumer/internals/Fetcher.java#L527 > You dont want poll consuming messages, its not just about offset correctness, the driver shouldnt be spending time or bandwidth doing that. I think you have agreed that this is impossible via current KafkaConsumer APIs as well. However, the unknown thing to me is that if the first `poll(0)` will return something. I saw the first `poll(0)` will always send the fetching request, but I'm not sure that if it's possible that the response will be processed in the first `poll(0)`. If this could happen, pausing partitions will not help in such case since it's called after the first `poll(0)`. In addition, since it's unclear in javadoc, it could be changed in the future. That's why I decided to manually seek to the beginning in #15397.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org