Applications Messaging Standard

[Doug Tody]

Hi Mark -

Thanks, I see better now what your motivations are for this mechanism
(John did say "user-friendly" which is why I thought user interface had
something to do with this).

I can see where an automatic auto-connect mechanism such as you
describe could be convenient, but it is unusual.  When software tries
to get smarter than whatever drives it this often leads to problems,
making the software harder to use in some cases; it is better to
tell it what to do.  Hence we use explicit command line arguments,
environment variable triggers/defaults, and so forth.

Off-hand I am not sure how best to solve a problem like this.  A simple
approach might be to require at least an environment variable,
possibly with an optional command line override, to trigger the
connection attempt and provide connection information.   Defining a
standard API would also provide a way to hide whatever is required,
but would require more than just a wire protocol.  We might also
consider whether a simple, general naming service would be worth
defining for this sort of thing (XPA, being based on the X server,
which is unique for a user session, has this by the way; it supports
user sessions, distributed execution, and authentication - of course
it is also obsolete technology, and a bit platform specific).

> could you give an explicit scenario for how you envisage a client
> application could bootstrap its connection to the messaging system?
> Your story about a UI transparently autogenerating arguments seems to 
> describe a skeleton for this but I don't know how you
> see the details being filled in.

It should be easy if the connection information can be optionally
provided on the command line when an application is started.  One just
passes an opaque string containing the necessary connection info.
Something which knows what to do with this information is also
required: either a library with a standard API, or we have to
say something public about how to parse the connection info.
Multiple sessions, distributed execution, etc., are easily supported.
An environment-based approach is similar and in fact they are often
combined.  This solves the problem for the application, but puts it
off on the "session manager" which has to spawn the hub/bus and pass
the info.  This is no problem so long as we have something which
serves this purpose, and it is the more general approach, but it
does require more than the simple auto-connect mechanism.

 	- Doug



On Wed, 14 Feb 2007, Mark Taylor wrote:

> On Tue, 13 Feb 2007, Doug Tody wrote:
>
>> Hi John -
>> 
>> On Tue, 13 Feb 2007, John Taylor wrote:
>>>> BTW, scanning the recent messages I still see a lot about dot files
>>>> etc.  This is an implementation issue; platform specific issues such
>>>> as dot files in U*nix can be dealt with by the implementation, and
>>>> do not have to be addressed by the standard protocol.
>>> 
>>> Yes and no.   How one goes about getting the information required to 
>>> connect could be part of the protocol.  You've already mentioned passing 
>>> this in on the commandline when you start an application.  Another (and 
>>> arguably more user-friendly) way would be to put this information in a 
>>> well-known location on the file system where client applications can find 
>>> it.
>> 
>> The information content and how it is used in operations could be
>> part of the protocol; how it is stored externally is part of the
>> implementation.  User interface issues such as providing a file-based
>> configuration mechanism which the user could edit is an implementation
>> issue; a different implementation could store the same information
>> in a different fashion.  That is not to say that it is not important,
>> just that it is important to separate the two concerns.
>
> I don't think John is talking here about a configuration file which the user 
> would edit (see below for what I think he is talking about).
>
>> Specifying connection information on the command line when an app is
>> started is not normally something one would do in a user interface,
>> rather such arguments would probably be autogenerated transparent
>> to the user, when the actual user interface (e.g., a CLI, some GUI,
>> etc.)  starts up the applications participating in the session.
>
> This model can work if the the "actual user interface" in question 
> understands which applications it can start up will/might participate in the 
> messaging conversation and can transmit information to them about how they 
> should connect to it.  But I don't see how it
> can operate if applications are started up from some third-party environment 
> such as /bin/sh or javaws.
>
>> This is runtime information used to tell apps how to dynamically
>> connect to a session; it is hard to even see how to reliably pass
>> such information via files, even if we ignore issues such as could
>> occur in a distributed or heterogeneous environment.
>
> What is hard to see here?  The idea is that a service (the "hub")
> writes information (such as a port number or URL) in a file at a well-known 
> location (~/.ivoa-messaging) which describes how clients can connect to it. 
> The file effectively acts as a very simple naming service; messaging-aware 
> applications can read the file to find out how they should connect to the 
> current messaging session.  When the
> "hub" shuts down it removes the file.  In this scenario the location and 
> format of the file are part of the protocol not part of the implementation 
> because participants in the protocol need to locate
> and read it to bootstrap their connections.  Admittedly this
> can't be easily extended to, say, multi-user scenarios and so it
> may be that it's insufficient for the requirements of whatever messaging 
> system we want to come up with, but "ignor[ing] issues such as could occur in 
> a distributed or heterogeneous environment" I don't see any problem with it; 
> as you may not be surprised to hear, it is exactly what is done in PLASTIC as 
> well as a number of other systems providing informal interprocess 
> communication I could name.
>
> I feel we may have some implicit misunderstanding or disagreement
> about how connection to a messaging service can work here.  Doug,
> could you give an explicit scenario for how you envisage a client
> application could bootstrap its connection to the messaging system?
> Your story about a UI transparently autogenerating arguments seems to 
> describe a skeleton for this but I don't know how you
> see the details being filled in.
>
> Mark
>
>