[
https://issues.apache.org/jira/browse/HUDI-3026?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Raymond Xu updated HUDI-3026:
-----------------------------
Sprint: Cont' improve - 2022/02/07, Cont' improve - 2022/02/14, Cont'
improve - 2022/03/01 (was: Cont' improve - 2022/02/07, Cont' improve -
2022/02/14)
> HoodieAppendhandle may result in duplicate key for hbase index
> --------------------------------------------------------------
>
> Key: HUDI-3026
> URL: https://issues.apache.org/jira/browse/HUDI-3026
> Project: Apache Hudi
> Issue Type: Bug
> Reporter: ZiyueGuan
> Assignee: ZiyueGuan
> Priority: Critical
> Labels: pull-request-available
> Fix For: 0.11.0
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> Problem: a same key may occur in two file group when Hbase index is used.
> These two file group will have same FileID prefix. As Hbase index is global,
> this is unexpected
> How to repro:
> We should have a table w/o record sorted in spark. Let's say we have five
> records with key 1,2,3,4,5 to write. They may be iterated in different order.
> In the first attempt 1, we write three records 5,4,3 to
> fileID_1_log.1_attempt1. But this attempt failed. Spark will have a try in
> the second task attempt (attempt 2), we write four records 1,2,3,4 to
> fileID_1_log.1_attempt2. And then, we find this filegroup is large enough by
> call canWrite. So hudi write record 5 to fileID_2_log.1_attempt2 and finish
> this commit.
> When we do compaction, fileID_1_log.1_attempt1 and fileID_1_log.1_attempt2
> will be compacted. And we finally got 543 + 1234 = 12345 in fileID_1 while we
> also got 5 in fileID_2. Record 5 will appear in two fileGroup.
> Reason: Markerfile doesn't reconcile log file as code show in
> [https://github.com/apache/hudi/blob/9a2030ab3190acf600ce4820be9a08929595763e/hudi-client/hudi-client-common/src/main/java/org/apache/hudi/table/HoodieTable.java#L553.]
> And log file is actually not fail-safe.
> I'm not sure if [~danny0405] have found this problem too as I find
> FlinkAppendHandle had been made to always return true. But it was just
> changed back recently.
> Solution:
> We may have a quick fix by making canWrite in HoodieAppendHandle always
> return true. However, I think there may be a more elegant solution that we
> use append result to generate compaction plan rather than list log file, in
> which we will have a more granular control on log block instead of log file.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
