Hi Andrew, We are also going to implement an RNS based on the newest spec. The discussion in my group results in the "obvious" implementation not the "entry" implementation. The idea of "entry" implementation and entry_parent is provided by Manuel. It might be better to ask him the rationale, but probably it is not possible. We would like to present our implementation of RNS and GFS naming profile at the GFS-WG GFS Naming Profile session from 11am to 12:30pm on Oct 19. Is it possible to discuss this at the session? Thanks, Osamu On Mon, 10 Sep 2007 10:21:38 -0400 "Andrew Grimshaw" <grimshaw@virginia.edu> wrote:
Osamu, I'm going to include my comments responses below after [Andrew Grimshaw]
-----Original Message----- From: Osamu Tatebe [mailto:tatebe@cs.tsukuba.ac.jp] Sent: Thursday, August 23, 2007 5:07 PM To: Andrew Grimshaw Cc: gfs-wg@ogf.org; ogsa-wg@ogf.org; 'Mark Morgan' Subject: Re: RNS experiences document
Hi Andrew,
I appreciate your work on the experiemce document.
RNS spec basically intends "entry" implementation. This comes from a discussion such that every entry would be pointed by an EPR.
Note that if the "entry" implementation is used, that after the "add" is performed there is no way to subsequently look up or find the EPR of the entry! We therefore use the "obvious" implementation.
How do you solve this problem with the "obvious" implememtation? I guess the difference of the "obvious" implementation and the "entry" implementation is a returned EPR when a virtual directory is created. I think you need some additional arguments for some operations. [Andrew Grimshaw] I'm not sure why we need the fiction of an EPR of the entry itself. Unless you are going to do operations on it - why have an EPR. If you wanted to keep it around, I suppose you could keep it in the {any}* part of the rns:entry, i.e., along with the entry reference and the entry_name.
It seems to me that the only operation you need to perform on the entries themselves in the current spec is the "query" operation. Our approach has been to eliminate "query", and instead use "list" to retrieve lists of <rns:entry>. In the limit, just one in the list. We then parse the XML ourselves to get what we want out.
We still need discussion about this, but, is it enough for list operation to additionally return an EPR of the directory entry itself in case of a virtual directory?
[Andrew Grimshaw] I don't understand a "virtual directory" Do you mean another RNS service instance? That is what the RNS services are!
Thanks, Osamu
On Tue, 21 Aug 2007 12:44:02 -0400 "Andrew Grimshaw" <grimshaw@virginia.edu> wrote:
All,
We have done an implementation of RNS and have several observations before we write our experiences document.
First, the version of the document I am referring to is the September 22, 2006 version.
p.4 Add function
We're fine with this BUT we had to make some assumptions. You'll note that the function returns an EPR.
Our interpretation was:
In the case that no EPR was passed in, i.e., I'm creating and linking a new directory, the function is to return an EPR of a new directory resource.
If the case that an EPR was passed in, i.e., I'm adding a link to an existing EPR, the EPR that was passed in is returned.
We'll call this the "obvious" implementation
An alternative interpretation is that a new resource is created for the DIRECTORY ENTRY ITSELF. In this case the EPR of this entry is returned.
We'll call this the "entry" implementation.
Note that if the "entry" implementation is used, that after the "add" is performed there is no way to subsequently look up or find the EPR of the entry! We therefore use the "obvious" implementation.
Both the "entry" and "obvious" implementations have ramifications and/or inconsistencies with later parts of the spec.
Note on page 7 the pseudo schema for an entry (returned by "list")
<rns:entry>
<rns:entry_name> xsd:string </rns:entry_name>
{any}*
<rns:entry_reference>
wsa:EndpointReferenceType
</rns:entry_reference>
</rns:entry>*
This is consistent with our "obvious" interpretation.
One of the things we have observed is that for very large directories this may create messages that are larger than the SOAP max - we may need to extend with an iterator.
Also note on page 10 the query operation (which takes NO parameters) and returns
<rns:EntryProperties>
<rns:entry_parent>
wsa:EndpointReferenceType
</rns:entry_parent>
<rns:entry_name> xsd:string </rns:entry_name>
{any}*
<rns:entry_reference>
wsa:EndpointReferenceType
</rns:entry_reference>
</rns:EntryProperties>
The implication of query is that the "entry" interpretation is being using because there is no way to tell the directory which entry to return data about therefore the SOAP headers contain the target in the form of a ws-addressing epr. In other words - this query seems to be targeted to an entry rather than a directory.
The objective of the "entry" implementation seems to be to work with query. The only stuff that seems to be added is the concept on a "entry parent" EPR - which in our opinion could be placed in the {any}* part if needed (the whole notion of parents is a bit problematic; it is likely to be a list not a singleton, and using it for garbage collection is difficult in a distributed system.) Thus, query seems mostly to be a highly specific form of list.
Thus, we have chosen the "obvious" implementation. That means that the query operation makes no sense by itself (beyond "list") and we have not implemented it.
Page 8, move
Move takes two parameters, a target directory and a "new" name for the entry, i.e., file1. This seems to imply an "entry" implementation. However, that means the client must be keeping track of the EPR's of the entries in order to do a move. That's the whole point of a directory services. Therefore, we believe that move should be an operation on a directory (i.e., the EPR of the directory is in the SOAP header) and take two parameters, the old name and the new name - in other words this is a rename.
Finally, from a interoperability point of view the fact that the RNS spec does not include a normative specification of the schema/WSDL, we don't believe that it is likely for interoperable implementations to exist solely based on the spec.
Modulo the above changes we have an implementation of RNS (we don't think that given the existing spec it is possible to completely implement RNS as specified) and a set of tools that interact with RNS name spaces, command line tools for ls, mkdir, etc, as well as a Linux file system software shim that intercepts file system calls and redirects (when appropriate) to RNS/ByteIO space.
A