On 12 Apr 2005, at 15:46, Stefano Mazzocchi wrote:
Pier Fumagalli wrote:On 11 Apr 2005, at 15:50, Reinhard Poetz wrote:Again, to sound stupid, but why in the world a cocoon block would contain classes and libraries?Daniel Fagerstrom wrote:
Ok, I had some remembrance that we had decided to have a particular directory structure on the COBs, but I couldn't find any documentation on that, do you have any link or example?
no. But AFAIK there aren't many rules. Those that I know of are:
Blocks have a block.xml descriptor and a COB-INF directory, that can contain a directory "classes" and "lib".
The block's root sitemap (if there exists any) is defined by the block.xml.
I don't know why we named it "COB-INF" but there was (still is?) a good reason for this because I remember some long discussion. Apart from this I'm not sure where to put "block.xml". In analogy to WAR files it should be "COB-INF/block.xml".
If the reasons for COB-INF vanished in the meantime, I propose following directory structure:
-------------------------------------------------------------- [cocoon block] [DIR] | +-- BLOCK-INF [DIR] +-- block.xml +-- classes [DIR] +-- lib [DIR] --------------------------------------------------------------
WDYT?
A block defines both services (thru its sitemap) and components (thru the classloader).
Presence of either is optional. If a block has only classes and libraries (or files in the class paths), it only works as a component provider.
If it has only a sitemap, it will require components from another block.
If it has both, it can expose services that use its own components and expose these components to the outside.
The 'components' to expose will be declared in the block descriptor, this allows us to avoid 'public' classes of the libraries shipped with the block to be misconsidered as components.
Those should be external "components" _required_ by the block, right???
see above: blocks brings components with them.
Also note that an 'interface' block is very likely to contain the component interfaces (and the remaining classes that these interfaces need to function).
So, for the sitemap, the block interface is just a URI, but for the component part, the block interface *contains* the interfaces.
this will require some more complex build systems as even compile time will require discovery and installation of components (this is the same problem that Merlin had and solved with Magic).
Ok, here is where I don't agree...
By adding to blocks the capability of "bringing components" with them, we enter a quite big minefield, imposed by the restrictions of the Java VM. The complexity escalates at this point as now blocks must be aware of their class-loading mechanism, and the component manager must be aware of blocks and interfaces.
For example, classes and libraries included in an interface block must be loaded by the parent ClassLoader of the blocks using (implementing, extending, requiring) them to avoid ClassCastException(s).
So, alongside blocks and their deployment graph (block A extends B, implements C, ... blabla ...) we need to start keep tracking also the classes included in those blocks, and create a completely different tree based on the class loading structure that those represent.
And then you get into the minefield of versioning... If it's easy to determine that two versions of the same block can co-exist (by using two separate class loaders children of the same parent) without creating too many problems, what happens in the case of two interfaces with different versions? The class loading tree should (at this point) become very fine-grained and start building multiple class-loaders for multiple versions of the same interface, and then build as children of those class loaders for the individual blocks "using" them, ...
I can go on for hours.
This, of course, ties in Cocoon with its component manager, and if today we're using ECM+++ (great work Carsten), if we ever decide to use (or integrate) another container, well, I can see how we'll end up having to rewrite 3/4 of the dependancy calculation and object instantiation code (which, BTW, works _right now_).
If on the other hand we separate entirely components and java code from blocks, the implementation becomes _much_ more easy...
My idea would be that a block (for example, our ForrestSkin implementation) _requires_ a component (not a block) that performs XSLT transformations.
Requiring this, it will expose (for example) a "Virtual Transformer" called "document2html" which will be implemented as XSLT Transformer + document2html.xslt.
Another skin for forrest could (at the same time), implement the same interface and provide the same "document2html" transformer by requiring the STX transformer and "joining it up" with "document2html.stx".
From outside, the two blocks _both_ provide a transformer, which does exactly the same job, but rely on the component manager underneath to solve the class loading problem of providing a component with a given role, add their value to it (the XSLT, STX, ... stylesheet) and expose it to the users of the implemented interface.
And this frees up the blocks implementation from the underlying component management, java, class loaders and so on... I feel this a lot more cleaner in terms of separating the concerns of cocoon and blocks versus the concerns of the Java platform and its classloading cumbersomeness...
Pier
