[jira] [Commented] (HIVE-20581) Eliminate rename() from full CRUD transactional tables

["Igor Kryvenko (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-20581?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16677325#comment-16677325 ] 

Igor Kryvenko commented on HIVE-20581:
--------------------------------------

I've been doing an investigation of this JIRA. And some behavior looks stranger to me. 

Suppose we have t1 -  full acid table, and t2 - insert-only table.

And if I run {{Insert overwrite into t1 select * from another_table_1 union all another_table_2}} files are moving inside {{MoveTask}}.
But, in the case {{Insert overwrite into t2 select * from another_table_1 union all another_table_2}} all files are moving inside MR job, and all operations in MoveTask are related only to metadata in the metastore. 

Is it correct behavior?

cc [~ekoifman]

> Eliminate rename() from full CRUD transactional tables
> ------------------------------------------------------
>
>                 Key: HIVE-20581
>                 URL: https://issues.apache.org/jira/browse/HIVE-20581
>             Project: Hive
>          Issue Type: Improvement
>          Components: Transactions
>            Reporter: Eugene Koifman
>            Assignee: Emily lozano d1751740 551913586
>            Priority: Major
>
> The {{MoveTask}} in a query writing to full CRUD transactional table still performs a {{FileSystem.rename()}}.  Full CRUD should follow the insert-only transactional table implementation and write directly to delta_x_x in the partition dir.  If the txn fails, this delta will be marked aborted and will not be read.
> There are several places that rely on this rename.  For example, support for {{Insert ... select ... Union All ... Select }} which creates multiple dirs, 1 for each leg of the union.
> Others?



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)