Dan Diephouse wrote:
Looking good. But even more new comments! Hopefully you aren't sick of me yet :-)


Never!  Of course, large payments of small unmarked bills would help ;-)

Can we collapse ServiceContext and Provider? I would propose then that Provider gets getters/setters for Resolver<Subject>/<Target>. If we did this we could get rid of ServiceContext too I think -- which would remove one more step for the user. I did a similar thing with when I made it implement Resolver<Target>. While I think this should be pluggable, I really think users should not have to configure a resolver by default.


Quite possibly. Will take a look today.

Along these lines I think we should remove the instantiation code which is in DefaultServiceContext. IMO this is the job of a container of some sort. I mean container in the broadest sense here. For instance, I think it could be wrapped up in something like a PropertiesProviderBuilder. It could then configure a whole provider instance - including resolvers, workspace manager, and collectionadapters.


Agreed.

Transactional.begin/end needs to take RequestContext so I can store a JCR session in it.


Ok, will add that today.

The JCR tests provide a good means to test if the basic ServiceProvider stuff I wrote still roughly works.


Well, it would need to be ported to the new model as the interfaces have changed. Would you like to take a stab at that? Doing so would help you get a better feeling for the new design.

My feeling is that CollectionInfo should not extend CollectionAdapter. While I agree that people will probably implement both CollectionInfo and CollectionAdapter by default, sometimes you want to just return the metadata.


Agreed. CollectionInfo can have a getCollectionAdapter method to connect the two.

Re: BasicAdapter - I'm still unconvinced about this API and its ability to support the necessary semantics for generic APP stores.


Ditto.

CollectionAdapterManager should be in the basic package if we keep the basic stuff around.


Take another look at CAM; I modified it so that it works generically with any CollectionAdapter, not just BasicAdapter. This would allow any Provider implementation to use the properties file deployment mechanism independently of whether the BasicAdapter interface is used.

I get the feeling that WorkspaceManager.checkMethod/checkTarget/getMethods can all probably go. Could we not work that logic into the CollectionAdapter.extensionRequest?


Not sure about this.

I still don't think its a good idea for a Provider to act as its own WorkspaceManager per my previous email.


Noted, but it does make things easier in some cases (see the Simple example). However, the design assumes they are separate things.

Shouldn't the WorkspaceInfo logic in DefaultProvider be in WorkspaceManager?


Possibly; will take another look.

- James

- Dan
James M Snell wrote:
Ok, the server refactoring is coming along very well I think. Thus far, I've been able to greatly simplify the overall design, as well as integrate the approaches implemented by Dan and the google feed-server team while still maintaining the flexibility of the original framework.

The main components of the refactored design are:

 * AbderaServlet
 * ServiceContext
 * Provider
 * WorkspaceManager
 * CollectionAdapter
 * RequestContext
 * ResponseContext
 * Target
 * Resolver<Target>
 * Resolver<Subject>

For the overwhelming majority of implementations, all an end developer will need to do is implement a CollectionAdapter, provide a TargetResolver, and glue the two together using a ServiceContext.

The examples in the server_refactor test cases illustrates the three fundamental approaches.

The application flow is simple:

 * The AbderaServlet receives a request and creates a ServiceContext
 * The ServiceContext creates a Provider instance
 * The Provider instance uses it's associated WorkspaceManager to
   select a CollectionAdapter.
 * Once the CollectionAdapter is selected, the Provider forwards the
   request on to the appropriate method

In some cases, the Provider and the WorkspaceManager will be the same object. In other cases, the Provider and the WorkspaceManager will be separate.

Simple
------

In the Simple example, the developer implements a CollectionAdapter, a Provider and a ServiceContext.

The CollectionAdapter provides the guts of the implementation, providing implementations for each of the main prototocol operations such as getFeed, getEntry, etc. These are the methods that used to exist on the old style provider interface.

The Provider extends AbstractWorkspaceProvider which means the Provider is acting as it's own WorkspaceManager.

The Provider/WorkspaceManager dispatches requests to the SimpleAdapter implementation.

The SimpleServiceContext glues the Provider to the appropriate target resolver.

Default
-------

In the Default example, the developer uses the same CollectionAdapter implementation used in the simple example but uses the DefaultProvider class.

The Default example is similar in nature to the CollectionProvider stuff implemented by Dan in that it is the DefaultProvider's job to handle things like the creation of the Service document based on metadata provided by the developer.

The simple and default examples are nearly identical with the exception that the default example supports multiple workspaces and collections.

Basic
-----

The basic example (which is poorly named, I know) is essentially the google feed-server stuff, slightly modified so that it sits on top of the default provider implementation.

The properties file based deployment is kept intact, but the Adapter interface is replaced by the BasicAdapter abstract class, which is an abstract CollectionAdapter implementation that defines the same abstract methods defined by the google feed-server Adapter interface. Existing google feed-server Adapter implementations can be ported to this design simply by replacing "implements Adapter" with "extends BasicAdapter".

The implementation has been further modified to support the creation of a service document. The code will read all of the *.properties files for the adapters and will generate a service document with one workspace and one collection per configured adapter.

Service Documents
-----------------

In this design, the Provider is responsible for serving the service documents. This means that the service document support will vary depending on the capabilities of the Provider implementation.

CollectionAdapters
------------------

In each of the three approaches, the same CollectionAdapter interface is used. Also, with the possible exception of URI patterns used for links, CollectionAdapters should be independent of the Provider implementation used. We need to find a more elegant way of tying Target Resolvers and Collection Adapters together while at the same time making it easier to manage links, but that can come later. The more important thing is that once a CollectionAdapter is implementation, we should be able to use it regardless of which of the three models are selected.

Media Collections
-----------------

The basic CollectionAdapter interface does not support media operations. If you want to implement support for media collections, the CollectionAdapter has to implement the MediaCollectionAdapter interface. The reason for this separation is to reduce the complexity of the simplest Atompub implementations that will only ever support Atom entries.

Transactional
-------------

CollectionAdapter implementations can implement the Transactional interface if they wish to support transactional start/end/compensate semantics. Provider implementations SHOULD call Transactional.start() before delegating to the CollectionAdapter method and Transactional.end() after delegating. If an error occurs, the Provider SHOULD call Transactional.compensate().

- James


Reply via email to