hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sergey Shelukhin (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HIVE-5705) TopN might use better heuristic for disable
Date Thu, 31 Oct 2013 02:44:25 GMT

     [ https://issues.apache.org/jira/browse/HIVE-5705?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Sergey Shelukhin updated HIVE-5705:
-----------------------------------

    Description: 
Right now, if TopN overruns memory threshold it disables itself if it couldn't directly exclude
rows as they are sent; it doesn't count evictions that were initially put in the heap and
then superceded for this purpose. 
It's reasonable in most cases, but if N is relatively small, and map output is large, the
cost could still be worth it even if rows don't get excluded immediately and are only evicted
after being stored for some time. So we'd pay some memory copies but emit much less rows.

  was:
Right now, if TopN overruns memory threshold it disables itself if it couldn't directly exclude
rows as they are sent; it doesn't count evictions that were initially put in the heap and
then superceded for this purpose. 
It's reasonable in most cases, but if N is relatively small, and map output is large, the
cost could still be worth it even if rows don't get excluded. So we'd pay some memory copies
but emit much less rows.


> TopN might use better heuristic for disable
> -------------------------------------------
>
>                 Key: HIVE-5705
>                 URL: https://issues.apache.org/jira/browse/HIVE-5705
>             Project: Hive
>          Issue Type: Improvement
>            Reporter: Sergey Shelukhin
>            Priority: Minor
>
> Right now, if TopN overruns memory threshold it disables itself if it couldn't directly
exclude rows as they are sent; it doesn't count evictions that were initially put in the heap
and then superceded for this purpose. 
> It's reasonable in most cases, but if N is relatively small, and map output is large,
the cost could still be worth it even if rows don't get excluded immediately and are only
evicted after being stored for some time. So we'd pay some memory copies but emit much less
rows.



--
This message was sent by Atlassian JIRA
(v6.1#6144)

Mime
View raw message