Mailing-List: contact dev-help@hive.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@hive.apache.org
Date: Wed, 12 Mar 2014 22:57:43 +0000 (UTC)
From: "Gunther Hagleitner (JIRA)" <jira@apache.org>
To: hive-dev@hadoop.apache.org
Message-ID: <JIRA.12700625.1394504595165.55871.1394665063425@arcas>
In-Reply-To: <JIRA.12700625.1394504595165@arcas>
References: <JIRA.12700625.1394504595165@arcas>
Subject: [jira] [Commented] (HIVE-6613) Control when spcific Inputs /
 Outputs are started
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/HIVE-6613?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13932593#comment-13932593 ] 

Gunther Hagleitner commented on HIVE-6613:
------------------------------------------

Looks good. Cool to see this - should really help with memory in reducers.

Couple comments:

- Can you avoid creating a conf in TezCacheAccess? Maybe just pass it in get().
- Have you considered adding the input to the cache key instead of using a Set? That way you can also remove the lock (I'm assuming the ObjectRegistry handles that).
- You can drop the getLocalWork check in the tez hashtable loader. Tez doesn't have local work.
- The javadoc of the init function needs to be updated with your changes.

> Control when spcific Inputs / Outputs are started
> -------------------------------------------------
>
>                 Key: HIVE-6613
>                 URL: https://issues.apache.org/jira/browse/HIVE-6613
>             Project: Hive
>          Issue Type: Improvement
>            Reporter: Siddharth Seth
>            Assignee: Siddharth Seth
>         Attachments: TEZ-6613.1.txt
>
>
> When running with Tez - a couple of enhancement are possible
> 1) Avoid re-fetching data in case of MapJoins - since the data is likely to be cached after the first run (container re-use for the same query)
> 2) Start Outputs only after required Inputs are ready - specifically useful in case of Reduce - where shuffle requires a large memory, and the Output (if it's a sorted output) also requires a fair amount of memory.


--
This message was sent by Atlassian JIRA
(v6.2#6252)