[jboss-as7-dev] Metadata for AS

[Jason T. Greene]

On 2/21/11 10:04 PM, Andrew Lee Rubinger wrote:
> On 02/21/2011 10:31 PM, Jason T. Greene wrote:
>> On 2/21/11 1:22 PM, Andrew Lee Rubinger wrote:
>>> Alongside our metadata discussions this morning, I'd like to offer up a
>>> prototype for use as the preferred metadata view for AS subsystems.
>>
>> -snip-
>>
>>> The purpose of the project is to define an object model for
>>> specification and JBoss-specific metadata, with the following driving
>>> tenets:
>>
>> One of the problems of any kind of metadata project, is that the scope
>> is somewhat vague. It can be anything from just spec descriptors all the
>> way to the entire runtime state of server deployment. Without a clear
>> scope, it will likely feature creep towards the latter.
>
> Exactly, the idea behind this one is to offer nothing aside from an
> object view of spec metadata, along with import/export utils.  Its scope
> is intentionally limited.  And as you've noted below, I gave an overview
> of the proposed scope further down the email.
>
>>> * No compile dependencies outside the JDK
>>> * Pluggable backends (ie. DOM, XB, Stax, right now we use DOM and
>>> straight hardcoded parsing logic)
>>
>> Why bother with pluggable backends? In particular whats the point of
>> having both a DOM parser AND a STAX parser?
>
> The design is pluggable, but I've no plans slated ATM for any other
> backends besides DOM.  For instance Remy specifically mentioned StAX, so
> if he were so inclined he could provide that backend and user code
> wouldn't know the difference.  Or, alternatively we could actually use
> the existing jboss-metadata stuff as the backend, and use this as a user
> view atop.  Swapping out jboss-metadata later when we wanted.

OK, well the reason for StAX usage in AS is processing and memory 
efficiency (something we definitely don't want to lose). DOM has the 
negative impact of copying everything into a big object model so that 
you can copy it again into a big object model. Granted its far more 
efficient than JAXB based parsers.

>
>>> * Easy import/export to/from File/URL/String/Stream/bytes
>>>
>>> Other concerns are to be addressed at another level, namely:
>>>
>>> * Merging
>>> * Annotation scanning repo output>  object model view
>>> * Validation
>>> * ...basically anything initiated during deployment.
>>
>> This is a good start on specifying a scope. It sounds like an attempt to
>> do a saner JSR-88 framework.
>
> JSR-88 is a weird beast.  It's a deployment spec which also exposes out
> deployment configuration for tooling?

Yeah its very odd. It allows you to manipulate any descriptor as a chunk 
of xml that is in your deployment based on an xpath mechanism.

>
> The itch which lead to this project's inception was from the testing
> community; users wanted to be able to mock out servlet configs in java
> code, and send 'em off for deployment in ShrinkWrap.

Yeah I can certainly see a lot of value in having a EE standard 
descriptor manipulation API.

>
> But I think we could be well served to adopt it here as the agreed view
> for AS subsystems.  Because the API does little more than expose
> spec/JBoss metadata grammars (and brings in no dependencies), I think
> we've got a good chance of agreeing upon this model.

Yes, I can see a spec defined xml descriptor API having a good potential 
for reuse. JBoss xml descriptors potentially as well, although for 
reasons I get into below, I think it needs to be limited to areas that 
have a very strong and clear contract.

Also, due to the rapid AS schedule, I don't want to create new potential 
blockers, at least in the 7.0 timeframe. When the project is further 
along and the API is solid/battle tested, I think then would be a better 
time to look at including it.

>
>> One of the challenges of using something like this for annotation
>> processing is the assumption that there is a one-to-one relationship
>> with annotations and configurable xml. This is not always the case. CDI
>> for example has a large number of annotations to process that are only
>> relevant to the actual CDI engine itself. This is not to say that there
>> could not be value in some kind of xml override mechanism (originally in
>> the draft specification), just that the consumers of interest are rather
>> limited (just weld cares about this information). Also you probably
>> don't want to have something that produces annotation java output.
>
> This is a problem, but not one for the Descriptors project.  What we'd
> need here is a bridge between the annotation provider/index (I'm aware
> of a few we've mocked up already) and the Descriptor model.
>
> Annotation processor>  index>  bridge>  Descriptor model

The point though is that annotation data does not always have a clear 
correspondence in the XML. So your object model becomes a super set, and 
represents the "merged view" (aka private deployer implementation details).

IMO any kind of descriptor API will be much easier to get right if its 
modeled as a direct one-to-one view of each input.

>
>>> Deployment logic should be, IMO:
>>>
>>> Deployer takes in deployment>  Gen object model>  calls outside layers
>>> for merging/validation/etc>  deployer installs final view
>> That seems inline with what we have at the moment. The key thing though
>> is that a deployer's state be allowed to evolve at the same pace as the
>> deployer.
>
> Can you expand on the "deployer state evolving" bit?  I'm near certain
> Descriptors wouldn't be imposing on anything here, but let's clarify?


So lets say you do the annotation -> bridge - > descriptor. Then we add 
support for a new annotation. Now, how many libraries, APIs, classes, 
releases are involved to make this change?

Another example, is lets say that the object model (and not the 
descriptor) is wrong and uses a single value object instead of a list. 
Since external users are using this, we now have to break an API to 
update something that should just be an implementation change.

The two above reasons are why IMO this should closely reflect the 
existing contracts, and not be a unified state abstraction.

-- 
Jason T. Greene
JBoss, a division of Red Hat