tajo-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Hyunsik Choi (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (TAJO-88) Separate Tajo's cluster resource manager from Yarn (Umbrella)
Date Thu, 25 Jul 2013 04:23:50 GMT

    [ https://issues.apache.org/jira/browse/TAJO-88?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13719213#comment-13719213
] 

Hyunsik Choi commented on TAJO-88:
----------------------------------

The name 'TajoResourceManager' may make you confused. The main difference of TajoResourceManager
and YarnResourceManager is whether to use node containers allocated from Yarn or to directly
manage resources in workers. 

It is necessary because in the standby mode, one worker per node will be on standby for during
whole uptime. So, TajoMaster cannot use Yarn ResourceManager for allocating resources for
each query.
                
> Separate Tajo's cluster resource manager from Yarn (Umbrella)
> -------------------------------------------------------------
>
>                 Key: TAJO-88
>                 URL: https://issues.apache.org/jira/browse/TAJO-88
>             Project: Tajo
>          Issue Type: Improvement
>          Components: master, worker
>            Reporter: Hyunsik Choi
>            Assignee: hyoungjunkim
>             Fix For: 0.2-incubating
>
>         Attachments: Tajo_Resource_Manager_Design.pdf
>
>
> In the current implementation, the resource management of Tajo is tightly coupled to
Hadoop Yarn. Hadoop YARN has emerged and a nice resource manager for large clusters, but Yarn
is not the best choice for low latency query executions.
> The main purpose of this issue is to separate Tajo's resource manager from Yarn and add
a standby mode. In the standby mode, Tajo workers will be on standby, and they will provide
low-latency responses for query tasks.
> This issue is an umbrella issue, and our subtasks are as follows:
> * We will separate QueryMaster from TajoMaster.
> * We will provide a resource manager interface for Yarn and Tajo's one.
> * We will implement our own resource manager and worker containers for the standby mode.
> Finally, we will provide two modes:
>  (1) On-demand mode (using Yarn): in this mode,  Tajo shares the cluster resource with
another Yarn applications. Per query, Tajo will request necessary cluster resources.
>  (2) Standby mode: in this mode, Tajo master and workers will be on standby.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message