You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@crunch.apache.org by "Micah Whitacre (JIRA)" <ji...@apache.org> on 2016/07/05 21:06:11 UTC
[jira] [Updated] (CRUNCH-609) KafkaSource could skip data with slow
broker response
[ https://issues.apache.org/jira/browse/CRUNCH-609?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Micah Whitacre updated CRUNCH-609:
----------------------------------
Attachment: CRUNCH-609d.patch
Ok last patch I promise. Fixes a bug in the polltime backoff. Specifically last patch multiplied the backoff too many times. Talking with colleagues the backoff isn't really necessary b/c the Kafka Consumer will be already retrieving in a background thread so should eventually return regardless of how long sleeping. So simplified that part of the code.
> KafkaSource could skip data with slow broker response
> -----------------------------------------------------
>
> Key: CRUNCH-609
> URL: https://issues.apache.org/jira/browse/CRUNCH-609
> Project: Crunch
> Issue Type: Bug
> Components: IO
> Reporter: Micah Whitacre
> Assignee: Micah Whitacre
> Attachments: CRUNCH-609.patch, CRUNCH-609b.patch, CRUNCH-609c.patch, CRUNCH-609d.patch
>
>
> In the case of the consumer.poll timing out, the KafkaRecordReader could return with no values in the records iterator which then causes the KafkaRecordReader to exit prematurely[1]. To fix this we need to probably better track the start/end offset and that progress vs the null value.
> Additionally if we wanted to be smarter we could do some "backoff" for the timeout or just have consumer specify a larger value.
> [1] - https://github.com/apache/crunch/blob/360d72a4f887505e020fdb8f99c3ccb1800693f6/crunch-kafka/src/main/java/org/apache/crunch/kafka/inputformat/KafkaRecordReader.java#L122
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)