You are viewing a plain text version of this content. The canonical link for it is here.

Posted to commits@hudi.apache.org by "zhangminglei (Jira)" <ji...@apache.org> on 2021/05/20 15:09:00 UTC

[jira] [Updated] (HUDI-1918) Incorrect keyby field would cause serious data skew

     [ https://issues.apache.org/jira/browse/HUDI-1918?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

zhangminglei updated HUDI-1918:
-------------------------------
    Description: 
The code ([https://github.com/apache/hudi/blob/master/hudi-flink/src/main/java/org/apache/hudi/streamer/HoodieFlinkStreamer.java#L92]) that in the actual data warehouse, partition path is mostly based on log_date or log_hour, so keyBy (HoodieRecord: :getPartitionPath) that would cause serious data skew.

we can actually shuffle data by record key here to avoid multiple subtasks write to a bucket at the same time, just like the pipeline in HoodieTableSink.

  was:
The code (https://github.com/apache/hudi/blob/master/hudi-flink/src/main/java/org/apache/hudi/streamer/HoodieFlinkStreamer.java#L92), in the actual data warehouse, partition path is most based on log_date or log_hour, so keyBy (HoodieRecord: :getPartitionPath) that would cause serious data skew.

we can actually shuffle data by record key here, just like the pipeline in HoodieTableSink.


> Incorrect keyby field would cause serious data skew
> ---------------------------------------------------
>
>                 Key: HUDI-1918
>                 URL: https://issues.apache.org/jira/browse/HUDI-1918
>             Project: Apache Hudi
>          Issue Type: Bug
>          Components: Flink Integration
>            Reporter: zhangminglei
>            Assignee: zhangminglei
>            Priority: Critical
>
> The code ([https://github.com/apache/hudi/blob/master/hudi-flink/src/main/java/org/apache/hudi/streamer/HoodieFlinkStreamer.java#L92]) that in the actual data warehouse, partition path is mostly based on log_date or log_hour, so keyBy (HoodieRecord: :getPartitionPath) that would cause serious data skew.
> we can actually shuffle data by record key here to avoid multiple subtasks write to a bucket at the same time, just like the pipeline in HoodieTableSink.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)