[
https://issues.apache.org/jira/browse/HIVE-25777?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17454152#comment-17454152
]
Gopal Vijayaraghavan edited comment on HIVE-25777 at 12/6/21, 5:57 PM:
-----------------------------------------------------------------------
Assumption is that that fix is somewhere here to ensure the old base dir gets
thrown out (if write-id == write-id, but txn-id is valid + present for only one)
https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/io/AcidUtils.java#L1815
was (Author: gopalv):
This fix needs to happen when AcidUtils.ParseDeltaLight is getting built using
the visibilityTxnId field in it.
> ACID: Pick the compactor transaction over insert dir
> ----------------------------------------------------
>
> Key: HIVE-25777
> URL: https://issues.apache.org/jira/browse/HIVE-25777
> Project: Hive
> Issue Type: Bug
> Components: Transactions
> Affects Versions: 3.1.2, 4.0.0
> Reporter: Gopal Vijayaraghavan
> Priority: Major
>
> If there are two competing versions of a particular write-id, one from the
> compactor and another from the original insert, always pick the compactor one
> once it is committed.
> If the directory structure looks like
> {code}
> base_000011/
> base_000011_v192/
> {code}
> Then always pick the v192 transaction if txnid=192 is committed.
> This is required to ensure that the raw base_ dir can be deleted safely on
> non-atomic directory deletions (like s3), without a race condition between
> getSplits and the actual file-reader.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
