[ 
https://issues.apache.org/jira/browse/ACCUMULO-4808?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16359034#comment-16359034
 ] 

Keith Turner commented on ACCUMULO-4808:
----------------------------------------

Create table is a FATE operation. FATE operation are persisted in zookeeper and 
therefore should be small.  So it would not be good to include table splits in 
a FATE repo as this could be a large amount of data.  One possible way to avoid 
this is to store the split points in a file in HDFS before the FATE op is 
started.  The master could do this and store it in a accumulo dir in dfs (just 
randomly pick a volume).  The FATE repo would then only need to store the file 
path in ZK.

> Add splits to table at table creation.
> --------------------------------------
>
>                 Key: ACCUMULO-4808
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-4808
>             Project: Accumulo
>          Issue Type: Sub-task
>          Components: master, tserver
>            Reporter: Mark Owens
>            Assignee: Mark Owens
>            Priority: Major
>             Fix For: 2.0.0
>
>
> Add capability to add table splits at table creation. Recent changes now 
> allow iterator and locality groups to be created at table creation. Do the 
> same with splits. Comment below from 
> [ACCUMULO-4806|https://issues.apache.org/jira/browse/ACCUMULO-4806] explains 
> the motivation for the request:
> {quote}[~etcoleman] added a comment - 2 hours ago
> It would go al long way if the splits could be added at table creation or 
> when table is offline.  When the other API changes were made by Mark, I 
> wondered if this task could also could be done at that time - but I believe 
> that it was more complicated.
> The delay is that when a table is created and then the splits added and then 
> taken offline there is a period proportional to the number of splits as they 
> are off-loaded from the tserver where they originally got assigned.  (The 
> re-online with splits distributed across the cluster is quite fast)
> If the splits could be added at table creation, or while the table is offline 
> so that the delay for shedding the tablets could be avoided, then the need to 
> perform the actual import offline would not be as necessary.
>  
> {quote}
>  



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to