[
https://issues.apache.org/jira/browse/HDFS-9853?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15163740#comment-15163740
]
Colin Patrick McCabe commented on HDFS-9853:
--------------------------------------------
Hi [~anu],
Last time we talked, we hadn't decided on whether OZone would use range
paritioning or hash partitioning. I see here that there is a
ListContainerRequestProto:
{code}
197 message ListContainerRequestProto {
198 required Pipeline pipeline = 1;
199 required uint64 start = 2 [default = 0]; // Start Index
200 required uint64 count = 3; // Max Results to return
201 }
{code}
It seems to me that with hash partioning, it will be impossible to correctly
implement this. If the node membership of the cluster changes between one
{{ListContainerRequestProto}} request and the next, what is referred to by
"start" will change, and we will end up with duplicated or missing results.
With range partitioning, it is unclear what "start" refers to. If it is the
position of the key inside the total ordering of all keys, it seems that adding
another key could cause this to refer to something different.
> Ozone: Add container definitions
> --------------------------------
>
> Key: HDFS-9853
> URL: https://issues.apache.org/jira/browse/HDFS-9853
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Anu Engineer
> Assignee: Anu Engineer
> Attachments: HDFS-9853-HDFS-7240.001.patch
>
>
> This patch introduces protoc definitions that operate against the container
> on a datanode.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
