hive-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Adesh Kumar Rao (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HIVE-15131) Change Parquet reader to read metadata on the task side
Date Wed, 18 Jul 2018 20:47:00 GMT

    [ https://issues.apache.org/jira/browse/HIVE-15131?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16548380#comment-16548380
] 

Adesh Kumar Rao commented on HIVE-15131:
----------------------------------------

Uploaded new patch to fix the parquet_analyze test failure. This was happening because in
case of noscan stats collection, parquet reader gets a dummySplit with start/length set to
0/0 in which case, using the filter api was returning 0 blocks for the dummySplit and hence
wrong stats were updated.

> Change Parquet reader to read metadata on the task side
> -------------------------------------------------------
>
>                 Key: HIVE-15131
>                 URL: https://issues.apache.org/jira/browse/HIVE-15131
>             Project: Hive
>          Issue Type: Bug
>          Components: Reader
>            Reporter: Chao Sun
>            Assignee: Adesh Kumar Rao
>            Priority: Major
>         Attachments: HIVE-15131.1.patch, HIVE-15131.2.patch, HIVE-15131.3.patch, HIVE-15131.4.patch
>
>
> Currently the {{ParquetRecordReaderWrapper}} still uses the {{readFooter}} API without
filtering, which means it needs to read metadata about all row groups every time. This could
some issues when input dataset is particularly big and has many columns.
> [Parquet-84|https://issues.apache.org/jira/browse/PARQUET-84] introduced another API
which allows to do row group filtering on the task side. Hive should adopt this API.



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

Mime
View raw message