Update (27/11/2005, 15:40): After putting the remote logging framework to use on an actual commercial project I've found that the hack I used for mapping a method to its concrete (nonvirtual) implementation wasn't working properly for explicit interface implementations. One horrible hack (which fortunately never saw the light of day) and a little digging later I've come across a useful trick on Andy Smith's weblog and added some special code to handle interfaces. The new version is available from the download section.

As part of the ongoing project I'm working on I was asked to implement a sort of macro recording feature into the product engine. The "keyboard" (as it were) is implemented as an abstract, remotable CAO interface which is registered in the engine via a singleton factory. I figured that the easiest way to record a macro would be to save a list of all remoting calls from when recording begins to when recording ends; the simplest way would be to add code to the beginning of every remote method implementation to save the call and its parameters. Obviously, that is an ugly solution.

About a year ago when I was working on a different project (a web-service back end for what would hopefully become a very popular statistics-gathering service) I faced a similar issue: the QA guys asked for a log of every single Web Service method call for performance and usability analysis. At the time I implemented this as a SOAP extension which infers some details using Reflection and saves everything to a specific log (via log4net). A couple of hours of research, 30 minutes of coding and I was done - the code even worked on the first try! - which only served to boost my confidence in my choice of frameworks for this project. My point is, I figured I would do something along the same lines for this project, and went on to do some research.

If you're just interested in the class library and instuctions skip to instructions or download.

I eventually settled on writing a Remoting server channel sink to do the work for me. I'll save you the nitty-gritty details; suffice to say that the entire process of Remoting sink providers is not trivial and not in any way sufficiently documented. There are some examples around the internet and even a pretty impressive article by Motti Shaked which proved useful, but didn't solve my problem. I eventually got to the point where I could, for a given type, say which methods on that type were invoked via Remoting and with which parameters, but I couldn't tell the object instance for which they were invoked no matter what. A sink may access the IMethodMessage interface which gives you a wealth of information, including arguments, the MethodBase for the method call etc., but the only indication of which object is actually being called is its URI (the interface's Uri property). Unfortunately I couldn't find any way of mapping a URI back to its ObjRef (something like the opposite of RemotingServices.GetObjectUri), only its type.

After a long period of frustrating research and diving into CLR implementation with the use of Reflector I came to the conclusion that the only way to convert the URI back to an object (outside of messing with framework internals via Reflection) would be to implement ITrackingHandler and maintain a URI-object cache. Another annoying hurdle was that the framework internally adds an application ID to its URIs and forwards the calls accordingly; for example, connecting to a server at URI tcp://localhost:1234/test would in fact connect to the object at local URI /3f4fd025_377a_4fda_8d50_2b76f0494d52/test. It took a bit of further digging to find out that this application ID is available in the property RemotingConfiguration.ApplicationID (obvious in retrospect, but there was nothing to point me in the right direction), which allowed me to normalize the incoming URIs and match them to the cache.

Finally, inferring virtual method implementations and URI server types are relatively costly operations, so I've added a caching mechanism for both in order to cut down on the performance loss. I haven't benchmarked this solution, but I believe the performance hit after the first method call should be negligible compared to "regular" remoting calls.

Obviously the first thing to do would be to download the source and add a reference to the class library. I've built and tested it with Visual Studio 2003; I'm pretty confident that it would work well with 2005 RTM, and probably not work with 1.0 or Mono (although I would be delighted to find out otherwise, if anyone bothers to check...)

Next you must configure Remoting to actually make use of the new provider. You would probably want the logging sink provider as the last provider in the server chain (meaning just before actual invocation takes place); if you're configuring Remoting via a configuration file, this is very easy:

Doing it programmatically is slightly less trivial but certainly possible:

Now it is time to decide which types and/or methods get logged. Do this by attaching a [LogRemoteCall] attribute; you can use this attribute with classes (which would log all method calls made to instances of that class) or with specific methods:

Next you should implement IRemoteLoggingConsumer:

Finally you must register with the remote call logging framework via RemoteLoggingServices.Register. You can register as many consumers as you like; moreover, each consumer can be set to receive notification for all remotable types (for generic logging) or a particular type (for example, macro recording as outlined above).

Download