cocoon-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Daniel Fagerstrom <dani...@nada.kth.se>
Subject [blocks] Changing component strategy
Date Tue, 24 Jan 2006 19:35:33 GMT
In the current implementation of blocks, each block has an own component 
manager (the choice of which type is configurable) for managing the 
components of the block. The block local CM has an 
InterBlockServiceManager as parent manager and through it it can access 
components from component managers in other blocks that it is *wired* 
to, (and not from the other blocks of the system). Let us call this 
behavior wiring based CM for later reference.

Now I'm starting to think that the above described behavior is 
unnecessarily complicated and has some other problems as well. So if no 
one protests I'm going to change it so that the blocks still has own 
component managers, but that they register their (exposed) components in 
a global registry and that the parent manager of the local component 
managers access components from the global registry and not only from 
the connected blocks.

Consequences
------------

A global registry is much more similar to the situation in our "compile 
time" blocks than the wiring based CM, so it should be easier to 
migrate, furthermore so is a global registry used in OSGi so it will be 
more future safe as well.

A global registry requires less configuration in block.xml and 
wiring.xml. It is enough to declare the dependency on the interfaces of 
the components in the POM. If one want to make certain that a certain 
block is used at run time, a run time dependency can declared in the 
POM. Connections in the block.xml is only used for declaring inter block 
servlet communication.

A possible disadvantage is that role names could collide in the global 
registry but by using URIs or package names it should be possible to 
distinguish between components from different manufacturers.

The global registry approach might give run time errors when components 
are missing instead of deploy time errors. But this depends to a large 
extent on what lookup strategy the components and the local CMs use. If 
most of the component lookup is done from within the components using a 
service manager the setup problems might be defered to runtime. But if 
configuration based dependency injection is used the problems can at 
least in principle be detected early.

Local vs Global CMs
-------------------

A question that not will matter much until we use OSGi is whether the 
CMs are used from within the block or from the outside.

The current design assumes that the CM is internal to each block. The 
reason for this is that with OSGi R3 it was, IIUC, the only way to be 
able to have the implementation classes for the components internal to 
the bundle. A global CM would have requried that all implementation 
packages to be exported.

With R4 there are some new possibilities, one can get a class loader 
from a bundle and use that for constructing components in a global CM. 
OSGi R4 uses this for the new declarative services manager. A bundle 
that want to use the declarative service manager signals that by 
pointing out its component declaration with a special manifest header, 
"Service-Component". We could have a similar global ECM++ manager for 
legacy support. It is rather probable that there will be a OSGi adapted 
Spring CM, following the same principles.

                --- o0o ---

AFAIR, we haven't discussed what lookup strategy we wanted for 
components in blocks, I implemented the wiring based strategy because it 
seemed more natural for me before. The OSGi CM bridge that Sylvain 
implemented used OSGis service registry as a global registry. Gianugo 
tried to convince me that global registry was better at ApacheCon, but I 
was obviously not ready for it then ;)

WDYT?

/Daniel

Mime
View raw message