accumulo-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Eric Newton (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (ACCUMULO-2842) Allow for use of hflush instead of hsync
Date Tue, 17 Jun 2014 15:25:34 GMT

    [ https://issues.apache.org/jira/browse/ACCUMULO-2842?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14033916#comment-14033916
] 

Eric Newton commented on ACCUMULO-2842:
---------------------------------------

Orthogonal.

hflush: get the data in the various client buffers out to the datanodes
hsync: hflush + datanode syncs the data to disk

When doing compactions, we don't flush anything... let the buffers do their magic.  But when
we close the file, we expect to be able to read it again even if the power goes out.  That's
what {{dfs.datanode.synconclose}} does.  When we're writing to the WAL, we're flushing (and
syncing) constantly. 

> Allow for use of hflush instead of hsync
> ----------------------------------------
>
>                 Key: ACCUMULO-2842
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-2842
>             Project: Accumulo
>          Issue Type: Improvement
>          Components: logger
>            Reporter: Josh Elser
>            Assignee: Josh Elser
>             Fix For: 1.5.2, 1.6.1, 1.7.0
>
>
> We keep getting onto the discussion on hsync versus hflush performance; however, we lack
a way to actually test this.
> Make the invocation of hflush or hsync configurable. It should not change the default,
so users on existing major releases should not be affected unless they choose to change their
configuration.



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

Mime
View raw message