hadoop-common-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chris Nauroth <cnaur...@hortonworks.com>
Subject Re: Found weird issue with HttpFS and WebHdfsFileSystem
Date Mon, 20 Apr 2015 17:03:02 GMT
Hi Bram,

Your gut feeling is correct.  These 2 properties are used in private implementation details
of cluster communication.  I believe these 2 properties are currently the only difference
compared to the public REST API.

Chris Nauroth
Hortonworks
http://hortonworks.com/


From: Bram Biesbrouck <b@beligum.com<mailto:b@beligum.com>>
Reply-To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Date: Monday, April 20, 2015 at 4:19 AM
To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Subject: Re: Found weird issue with HttpFS and WebHdfsFileSystem

Hi Chris,

Thanks for your insights. Last question: can you tell me the main differences (from a Hadoop
dev point of view) between the public REST api and the HDFS wire protocol?
My gut feeling tells me hdfs is mainly used in cluster communication and the public one is,
well, for public api's. But maybe I'm missing some more subtle differences?

cheers,

b.

On Fri, Apr 17, 2015 at 7:28 PM, Chris Nauroth <cnauroth@hortonworks.com<mailto:cnauroth@hortonworks.com>>
wrote:
Hello Bram,

I'm glad to hear the information was helpful.

If you'd like to request access to childNum as part of a guaranteed public API, then I encourage
you to create a jira issue in the HDFS project.  We could consider it for the future.

HdfsFileStatus is a representation of the HDFS wire protocol, and it's intended to be decoupled
from the public API FileStatus object so that the two can evolve independently.  From a pure
code reuse perspective, I suppose the two could share a common base class, but then that common
base class would need to creep into the public API too.

Currently, there is no guarantee about the availability of these fields in the public REST
API.  We're going to remove mention of them in the documentation.  We're not necessarily planning
to remove the fields from the JSON immediately, but there is also no guarantee that they'll
stay there.

Chris Nauroth
Hortonworks
http://hortonworks.com/


From: Bram Biesbrouck <b@beligum.com<mailto:b@beligum.com>>
Reply-To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Date: Friday, April 17, 2015 at 5:26 AM
To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Subject: Re: Found weird issue with HttpFS and WebHdfsFileSystem

Hi Chris,

Thanks for this reply. I thought something funny was happening.

The childNum field is actually very useful (eg for (not) rendering a expansion marker next
to a folder in a GUI when it has children), so it's a pity the info is there, but get's "eaten
up" by the general interface, only to be re-calculated later on.
It would be nice to have the info as an optional field in the FileStatus class (initialized
to -1 like it is right now), so we can use it if it's there or just ignore it when not initialized.
While I'm ranting, HdfsFileStatus should override from FileStatus because it's 95% the same
code anyway.

If I read your reply correctly, I assume the fields will be deleted from the webhdfs JSON
responses as well in the future?

Thanks again for the extensive reply, very useful and appreciated.

cheers,

b.



On Fri, Apr 17, 2015 at 12:59 AM, Chris Nauroth <cnauroth@hortonworks.com<mailto:cnauroth@hortonworks.com>>
wrote:
Hello Bram,

There are a few Apache jiras with background discussion of the introduction of these fields
in WebHDFS.

https://issues.apache.org/jira/browse/HDFS-4502

https://issues.apache.org/jira/browse/HDFS-4772

https://issues.apache.org/jira/browse/HDFS-4969

The new fields could not be supported in HTTPFS (only WebHDFS), and they were not intended
to be guaranteed in the public REST API.  Unfortunately, the fields were added to the documentation
mistakenly in Apache Hadoop 2.5.0.

https://issues.apache.org/jira/browse/HDFS-6153

We're going to revert that documentation change in Apache Hadoop 2.8.0.  I suggest that your
application does not rely on these fields, or at least includes fallback logic to keep working
as best as it can if the fields are not present.  Another way to determine the number of children
would be to make a subsequent LISTSTATUS call on the child path.

I apologize if this caused any inconvenience, and I hope the information helps.

Chris Nauroth
Hortonworks
http://hortonworks.com/


From: Bram Biesbrouck <b@beligum.com<mailto:b@beligum.com>>
Reply-To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Date: Thursday, April 16, 2015 at 7:58 AM
To: "user@hadoop.apache.org<mailto:user@hadoop.apache.org>" <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Subject: Found weird issue with HttpFS and WebHdfsFileSystem

Hi all,

I'm experiencing something strange while developing against the HttpFS front-end webapp on
Hadoop 2.6.0.

I'm currently digging into WebHdfsFileSystem and HttpFS to understand it better and understand
how the rest api works. I've setup a local single node Hadoop instance, which I can query
successfully with eg. http://localhost:50070/webhdfs/v1/?op=LISTSTATUS
Returning eg. this FileStatus object:

{
accessTime: 0,
blockSize: 0,
childrenNum: 0,
fileId: 16386,
group: "supergroup",
length: 0,
modificationTime: 1417964248854,
owner: "hadoop",
pathSuffix: "user",
permission: "755",
replication: 0,
storagePolicy: 0,
type: "DIRECTORY"
}

Now, when I start HttpFS and ask for the same data over it's interface (http://localhost:14000/webhdfs/v1/?op=LISTSTATUS),
I get a different reply. Especially, the childrenNum and fileId fields are missing, compared
to the first result (same file or directory):

{
pathSuffix: "user",
type: "DIRECTORY",
length: 0,
owner: "hadoop",
group: "supergroup",
permission: "755",
accessTime: 0,
modificationTime: 1417964248854,
blockSize: 0,
replication: 0
}

Since I need the childrenNum property, I started digging into the code to see where it's "lost"
and found that WebHdfsFileSystem performs a makeQualified() step (around line 1287 in WebHdfsFileSystem.java),
just before the list of filestatuses is returned. Basically, it converts HdfsFileStatus objects
into FileStatus objects, effectively chopping off those two properties.

The sources for HdfsFileStatus clearly state that it's an "Interface that represents the over
the wire information for a file.", so I wonder why this happens, since the HdfsFileStatus
contains all the right properties, according to the docs at http://hadoop.apache.org/docs/r2.6.0/hadoop-project-dist/hadoop-hdfs/WebHDFS.html#List_a_Directory

It feels like the FileStatus class hasn't been updated to match the HdfsFileStatus class,
but since they don't share any interfaces or superclasses I get the feeling it's intentional,
but I just can't find or figure out why.

Can somebody help or shed some light?

thanks,

b.
--

[http://dl.beligum.com/pub/beligum_flag_5dpi.png] Bram Biesbrouck - 0486/118280<tel:0486%2F118280>
- www.beligum.com<http://www.beligum.com> -  the republic of reinvention



--

[http://dl.beligum.com/pub/beligum_flag_5dpi.png] Bram Biesbrouck - 0486/118280<tel:0486%2F118280>
- www.beligum.com<http://www.beligum.com> -  the republic of reinvention



--

[http://dl.beligum.com/pub/beligum_flag_5dpi.png] Bram Biesbrouck - 0486/118280 - www.beligum.com<http://www.beligum.com>
-  the republic of reinvention

Mime
View raw message