[
https://issues.apache.org/jira/browse/TEZ-3286?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15352491#comment-15352491
]
Siddharth Seth commented on TEZ-3286:
-------------------------------------
Not really called out in this jira; the properties which need to be set on a
per DAG/vertex level are the TEZ_TASK_SCALE* properties. These have a scope of
AM - and need to be changed to Vertex. Can be done in a follow up jira if
required.
- Adding a compatible ctor in TaskSpec would be useful for downstream llap
integration
- toString in TaskSpec too verbose ? Log only the number of entires instead of
all keys. (This could, instead, be logged once per DAG / Vertex in the AM)
- Don't think there's any test to verify TaskSpec serialization, or usage in
TezTaskRunner2. Some of these could be added later.
> Allow clients to set processor reserved memory per vertex (instead of per
> container)
> ------------------------------------------------------------------------------------
>
> Key: TEZ-3286
> URL: https://issues.apache.org/jira/browse/TEZ-3286
> Project: Apache Tez
> Issue Type: Bug
> Affects Versions: 0.8.3
> Reporter: Wei Zheng
> Assignee: Hitesh Shah
> Attachments: TEZ-3286.1.patch
>
>
> tez.task.scale.memory.reserve-fraction can be set by clients to control how
> much memory is available to the processor. Ths values applies at a container
> level though, instead of at a vertex level.
> In case of a hash-join - the processor typically needs more memory. In case
> of a Shuffle join - the processor may not need as much. In DAGs with a mix
> of map joins and shuffle joins - setting this at a container level is
> sub-optimal.
> To a large extent this comes down to propagating vertex configs to the
> container / task.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
