On Apr 24, 2009, at 11:09 AM, Guillaume Nodet wrote:
The only problem is that I need the object to be fully initialized (including the call to the init method afaik) before being injected. If this is not possible (due to cyclic dependencies), a partially object might be injected.
I had a chat with Rick about that requirement. It's a difficult one to satisfy as is and is one area of the blueprint spec I hope to see improved.
If you expand the reference semantics so that you can differentiate between simply having a reference to something and having a reference to something you need to be initialized by the time your init method is called (let's call it a dependency), so many things become easier.
For one, you could pretty much ignore any circular references, it's only circular dependencies you need to reject. So complex graphs of objects that all use each other at runtime (post initialization) are no problem at all in any fashion as long as there aren't any cyclic dependencies on the init phase.
Second, for dependencies, you wouldn't need to require that init methods be called prior to injection, only that they are called prior to your init method. So then you can construct your graph in one step, then sort it as a list in dependency order, and then simply iterate over it and call init on each object. Very easy. You can even get fancy with what you offer components in their init methods, such as transaction guarantees and other heavier things.
Another benefit, perhaps touchy-feely, is that your entire tree is constructed and issues relating to referring to things that don't exist or other invalid references, or data that is to be injected that can't be converted to the target type, etc., are all flushed out before any initialization is done. This is good because init methods can have some pretty heavy logic and it would be really be terrible to get nearly all the way through a large graph only to fail for what is essentially a syntax error. Whether or not you think that is important really depends on if you're more of an interpreted language person or a compiled language person. IMO, being able to provide a guarantee that the graph has been fully constructed and all is well before any init methods are called is a very good guarantee to offer people.
-David
