[
https://issues.apache.org/jira/browse/HDFS-6371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13995293#comment-13995293
]
Haohui Mai commented on HDFS-6371:
----------------------------------
In HA design the client always takes responsibility of retries instead of the
server. In the use case mentioned in HDFS-6334, the client and the sever might
have different views on HA, thus it looks to me that retrying on the
client-side might be the only viable solution.
> In HA setup, the standby NN should redirect WebHDFS write requests to the
> active NN
> -----------------------------------------------------------------------------------
>
> Key: HDFS-6371
> URL: https://issues.apache.org/jira/browse/HDFS-6371
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: webhdfs
> Reporter: Tsz Wo Nicholas Sze
> Assignee: Tsz Wo Nicholas Sze
>
> The current WebHDFS implementation in namenode does not check its HA state --
> it does the same thing no matter it is active or standby.
> Suppose a http client talk to the standby NN via WebHDFS. For the read
> operations, there is no problem. For the write operations, if the operation
> requires http redirect (e.g. creating a file), it will work since the standby
> NN will also redirect the client to a DN. When the client connect to the DN,
> the DN will fulfill the request with the active NN. However, for the write
> operations not requiring http redirect (e.g. mkdir), the operation will fail
> with StandbyException since it will be executed with the standby NN.
> There are two solutions:
> # The http client could catch StandbyException and then retries with the
> other NN in this case.
> # The standby NN redirects the request to the active NN.
> The second solution seems better since the client does not need to know both
> active NN and standby NN.
> Note that WebHdfsFileSystem is already able to handle HA failover. The JIRA
> is for other http clients.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
