[ 
https://issues.apache.org/jira/browse/HADOOP-12091?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14587793#comment-14587793
 ] 

Gil Vernik commented on HADOOP-12091:
-------------------------------------

It has absolutely nothing to do to Jenkins and submissions or dashboards and 
activities.
I thought to separate them in two patches since it will be easier to integrate 
into trunk...
I don't mind to have one patch for this. By the way, it's different things a 
little bit...

Currently isDirectory return true when file have 0 size... This is one problem 
which should be resolved. I can have file with 0 size in Swift. Hadoop commons 
crash because of this when Swift container has objects and one of them is of 
size 0.

Another problem that mkdirs and create methods - create recursively the whole 
parent structure, which is another issue.

But i am fine to have one Jira and one patch to resolve both of them and even 
other issues.

> Issues with directories handling in Swift
> -----------------------------------------
>
>                 Key: HADOOP-12091
>                 URL: https://issues.apache.org/jira/browse/HADOOP-12091
>             Project: Hadoop Common
>          Issue Type: Improvement
>          Components: fs/swift
>            Reporter: Gil Vernik
>            Assignee: Gil Vernik
>
> OpenStack Swift doesn't have notion of directories. In Swift everything is 
> object, stored in some container, that belongs to account.
> Current implementation contains many code that handle directories structure 
> in Swift, in particular functions with code that zero length object is a 
> directory.
> While it might be true for certain cases, there are also many cases where 
> "directory" handling cases problems and highly reduce performance.
> For example, if Swift's container has dozens of objects and one of them has 0 
> length size, than Swift driver thinks it's a directory and report it to
> upper layer as directory. In consequence, this leads to various exceptions 
> and crashes in client side / upper Hadoop layer.
> The propose of this Jira topic is to make directories handling in driver as 
> an optional and configurable. The driver will behave the same, but there
> will be a configurable option that will disable directories handling and so 
> everything will be objects, even those with 0 length size.
> This will cover cases, where clients doesn't care about directories 
> structures.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to