[
https://issues.apache.org/jira/browse/HADOOP-6356?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12918313#action_12918313
]
Sanjay Radia commented on HADOOP-6356:
--------------------------------------
Dhruba says > I think it is appropriate to do the compatible-version-check for
each RPC instead of doing it per creation-of-FileSystem object.
Having a FileSystem cache => RPCProxy is also cached => version check is done
once per file system object. Dhruba is right in that this is incorrect since
the NN may reboot with a different version.
However we will see a more serious performance problem with FileContext under
certain usage patterns. FileContext keeps a pointer to the default file system;
hence all file operations to the default filesystem will not require a version
check (but it is susceptible to the NN reboot with different verison problem).
However operations to other file systems such as fc.open("hdfs://nn/foo/bar)
will create a new AbstractFileSystem object that will then create a new
dfsClient which will create a new rpcProxy which will do a getVersion() - this
is an extra round trip. Good news is that the connection at the lowest layer is
cached.
We can fix this by having each rpc call also piggy back the version number so
that the version check is not needed when a proxy is created. This however will
break wire compatibility.
> Add a Cache for AbstractFileSystem in the new FileContext/AbstractFileSystem
> framework.
> ---------------------------------------------------------------------------------------
>
> Key: HADOOP-6356
> URL: https://issues.apache.org/jira/browse/HADOOP-6356
> Project: Hadoop Common
> Issue Type: Improvement
> Components: fs
> Affects Versions: 0.22.0
> Reporter: Sanjay Radia
> Assignee: Sanjay Radia
> Fix For: 0.22.0
>
>
> The new filesystem framework, FileContext and AbstractFileSystem does not
> implement a cache for AbstractFileSystem.
> This Jira proposes to add a cache to the new framework just like with the old
> FileSystem.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
