flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jiayi Liao (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-10348) Solve data skew when consuming data from kafka
Date Mon, 17 Sep 2018 18:19:00 GMT

    [ https://issues.apache.org/jira/browse/FLINK-10348?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16617952#comment-16617952

Jiayi Liao commented on FLINK-10348:

Ok, what you said makes more sense to me, I can agree with your first opnion.
But for the second one, there always be a partition with the lowest watermark, it's very hard
for you to make a strategy to do "selectively forward messages" unless you want to consume
one partition at a time, which is very bad.
I can't figure out any better ideas:(.

> Solve data skew when consuming data from kafka
> ----------------------------------------------
>                 Key: FLINK-10348
>                 URL: https://issues.apache.org/jira/browse/FLINK-10348
>             Project: Flink
>          Issue Type: New Feature
>          Components: Kafka Connector
>    Affects Versions: 1.6.0
>            Reporter: Jiayi Liao
>            Assignee: Jiayi Liao
>            Priority: Major
> By using KafkaConsumer, our strategy is to send fetch request to brokers with a fixed
fetch size. Assume x topic has n partition and there exists data skew between partitions,
now we need to consume data from x topic with earliest offset, and we can get max fetch size
data in every fetch request. The problem is that when an task consumes data from both "big"
partitions and "small" partitions, the data in "big" partitions may be late elements because
"small" partitions are consumed faster.
> *Solution: *
> I think we can leverage two parameters to control this.
> 1. data.skew.check // whether to check data skew
> 2. data.skew.check.interval // the interval between checks
> Every data.skew.check.interval, we will check the latest offset of every specific partition,
and calculate (latest offset - current offset), then get partitions which need to slow down
and redefine their fetch size.

This message was sent by Atlassian JIRA

View raw message