tuscany-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jean-Sebastien Delfino <jsdelf...@apache.org>
Subject Re: [BRAINSTORM] Flexibility in distributed operation and extension implementations - was: Re: Request to propogate the value of a references target= attribute on its associated bindings model object
Date Thu, 01 May 2008 19:27:57 GMT
Simon Laws wrote:
> On Tue, Apr 22, 2008 at 10:42 PM, Jean-Sebastien Delfino <
> jsdelfino@apache.org> wrote:
> 
>> Simon Laws wrote:
>>
>>> On Mon, Apr 21, 2008 at 4:49 PM, Simon Laws <simonslaws@googlemail.com>
>>> wrote:
>>>
>>>
>>>> On Tue, Apr 15, 2008 at 6:10 PM, Yang Lei <genially@gmail.com> wrote:
>>>>
>>>>  I agree with Simon's emphases on the point of view. I understand
>>>>> Tuscany may prefer one solution over the other. However from
>>>>> extensibility perspective, there need some extension points to
>>>>> enable
>>>>> Tuscany adapters to overwrite the default behavior. I think the
>>>>> thread
>>>>> discussion on reference target and the comparing of 1 and 2 showcase
>>>>> one of the extensibility area :  how to resolve reference target for
>>>>> different bindings.
>>>>>
>>>>> I am actually looking beyond just reference target, I see the
>>>>> extensibility in the following areas:
>>>>>
>>>>> 1. When/How to enable a binding to resolve the target endpoint .
>>>>> This
>>>>> include the case to support reference target, and beyond, such as
>>>>> supporting wireByImpl or autoWire. This also include distributed
>>>>> support in case adapters have different ways to support distributed
>>>>> contributions for a given virtual domain.
>>>>>
>>>>> I understand Tuscany has workspace discussions. It may potentially
>>>>> be
>>>>> a solution.I am still waiting to see how workspace is intending to
>>>>> support distributed scenarios or how it can enable late binding on
>>>>> resolving target endpoint. Regardless workspace is the solution or
>>>>> not, we need the flexibility and extensibility to overwrite
>>>>> Tuscany's
>>>>> default behavior on binding end point resolving.
>>>>>
>>>>> 2. When/How the binding resolvable is in used,
>>>>>
>>>>> Some part of the Tuscany code is using binding resolved or not to
>>>>> have
>>>>> different process  (see point 3). I think if certain logic outside
>>>>> binding needs to understand if a binding is resolvable, we should
>>>>> make
>>>>> it clear which method achieve it so binding implementations know
>>>>> what
>>>>> to expect.
>>>>>
>>>>> I can see Tuscany code uses binding's URI and targetComponentService
>>>>> today, I think it should be limited to one method only, I am not
>>>>> sure
>>>>> overloading URI is  good .
>>>>>
>>>>> 3. When/How to make binding selections on the reference side.
>>>>>
>>>>> I can see Tuscany is trying to remove the unresolvable bindings
>>>>> first
>>>>> from the reference side , then use some algorithm to either pick the
>>>>> default binding if it exists or pick the first on the list.
>>>>>
>>>>> I think we need some plug in point in Tuscany to enable different
>>>>> algorithm from the above default behavior. And the plugin point need
>>>>> to enable late binding so during reference's execution time we can
>>>>> determine a binding is resolvable or not and then use some own
>>>>> prioritizing rules to select the right bindings.
>>>>>
>>>>>
>>>>> I would like to see these discussions concluded with a set of API
>>>>> and
>>>>> some form of API interaction diagrams in the end.
>>>>>
>>>>> Thanks.
>>>>>
>>>>> Yang
>>>>>
>>>>>
>>>>>
>>>>> I can see a couple of scenarios:
>>>>>
>>>>>
>>>>>
>>>>> I thinkand binding selection that we need to enable some extension
>>>>> points for others using other algorism or other
>>>>>
>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe, e-mail: tuscany-dev-unsubscribe@ws.apache.org
>>>>> For additional commands, e-mail: tuscany-dev-help@ws.apache.org
>>>>>
>>>>>
>>>>>  I've been thinking about this issue for a few days on and off and
>>>> it seems
>>>> to me that the key to this is in the way that we store bindings that
>>>> have
>>>> been read in from a composite file.
>>>>
>>>> The assembly model starts out with each reference holding all the
>>>> bindings
>>>> it is configured with in the composite file.
>>>>
>>>> During model build the set of bindings is matched with targets and the
>>>> resulting list represents the set of resolved bindings complete with
>>>> URIs
>>>> identifying target services. These bindings represent the runtime
>>>> configuration and are used to generate wires.
>>>>
>>>> To do late binding we have to maintain the original set of bindings as
>>>> well as any bindings that have been fully resolved. In this way the
>>>> reference can resolve targets at runtime with all the information that
>>>> is
>>>> used to resolve them at build time.
>>>>
>>>> During the first domain implementation I ran across this problem and
>>>> stored the original list of bindings on the dummy target service that
>>>> is
>>>> created for each target. However this is less than satisfactory as
>>>> this list
>>>> is not persisted by the processors should the composite be written out
>>>> again.
>>>>
>>>> If we reorganize the bindings such that we have a notion of candidate
>>>> bindings and resolved bindings then candidate bindings can be used at
>>>> a
>>>> later point to create resolved bindings.
>>>>
>>>> Much of the builder processing can be done early to associate policy
>>>> with
>>>> bindings etc. But the wiring processing needs a bit of thinking about.
>>>> Anyhow this is not a fully formed thought but I'm throwing this out
>>>> there as
>>>> I want to spend some time on this over the next few days and welcome
>>>> any
>>>> input.
>>>>
>>>> Regards
>>>>
>>>> Simon
>>>>
>>>>
>>> I've given this some more thought. It's a long post, sorry about that,
>>> but
>>> just diving in and changing code is going to complicate matters on this
>>> one.
>>>
>>>
>>> Requirements
>>> ------------
>>>
>>> A - late target service location
>>> B - late binding selection
>>> C - late autowire
>>>
>>> A and B could be achieved at the time when a proxy is first called on
>>> the
>>> understanding that an empty proxy can be first generated to represent
>>> these
>>> unresolved wires.This is different from the case where no proxy is
>>> provided
>>> for an unwired reference.
>>>
>>> C is more difficult as it involves the injection of one or more new
>>> proxies
>>> depending on what the autowire process discovers. I suggest
>>> we defer this requirement.
>>>
>>> Inputs and Outputs
>>> ------------------
>>>
>>> Currently the list of bindings that remain on a reference after the
>>> assembly
>>> model is "built" is what is used to drive the creation of the runtime.
>>> The
>>> following is a little strange but what I'm trying to show is the general
>>> configuration of a reference, e.g. "Target = set" means that the
>>> reference
>>> has one or more targets and "binding uri = set" means that the reference
>>> has
>>> bindings where the binding URI is set.
>>>
>>> Input from .composite file:
>>>
>>> Target = null, binding uri = null
>>> Target = null, binding uri = set
>>> Target = set,  binding uri = null
>>> Target = set,  binding uri = set  - this is an error case
>>>
>>> Output from build phase
>>>
>>> Target = null, binding uri = null - unwired reference, i.e. autowire
>>> hasn't
>>> matched up the reference
>>> Target = null, binding uri = set  - the binding is used to generate a
>>> provider and a wire
>>> Target = set,  binding uri = null - A candidate for late binding (this
>>> is
>>> currently represented by storing bindings on the dummy target service)
>>> Target = set,  binding uri = set  - this is an error case but
>>> interestingly
>>> our workspace will generate these now
>>>
>>> Current Binding Lifecycle
>>> -------------------------
>>>
>>> Binding lifecycles depending on input. Note. These stages are not
>>> represented explicitly in the binding model.
>>>
>>> Wired Reference (Target = set,  binding uri = null)
>>>  Unresolved - not associated with service and a candidate for resolution
>>> against a service binding
>>>     (All candiate/unresolved bindings are removed from the
>>> reference.bindings list after the build phase)
>>>  Resolved - associated with service  - make better use of the resolved
>>> flag
>>> on binding
>>>  Activated - binding provider created
>>>  Started - wire created and provider started
>>>
>>> Unwired reference (Target = null, binding uri = set)
>>>  Resolved - assumed resolved due to user input
>>>  Activated - binding provider created
>>>  Started - wire created and provider started
>>>
>>> Unwired reference (Target = null, binding uri = null)
>>>  Unresolved - no proxy should be injected unless autowire matches them
>>>  Resolved - autowire has matched the reference
>>>  Activated - binding provider created
>>>  Started - wire created and provider started
>>>
>>> Once past the resolved point point the final set of bindings are known
>>> (the
>>> reference.bindings list) and the runtime is constructed from this list.
>>>
>>> Suggested Changes
>>> -----------------
>>>
>>> Tidy the binding model
>>>
>>> A) remove the concept of Optimizable binding
>>>     Bindings will not have access directly to the target service that
>>> they
>>> refer to through this route
>>>
>>> B) add a direct reference between a binding and the target service model
>>> that the binding is configured to represent
>>>     We can tell which binding is related to which target
>>>     we can tell which bindings are candidate bindings as they are not
>>> already associated with a target model
>>>
>>> C) make clear use of the binding.unresolved flag to indicate true = not
>>> pointing at a service in the domain
>>>
>>> Maintain candidate bindings
>>>
>>> D) retain the allocation of all bindings to unresolved services
>>>     These bindings should all be fully configured in terms of available
>>> policy
>>>     There is potentially a gap here if we don't feel the full policy set
>>> is
>>> available in the domain when these bindings are configured
>>>
>>> OR (not sure which one is best without doing the code but the first is
>>> there
>>> already)
>>>
>>> E) retain all candidate bindings on the bindings list but with URI=null
>>> and
>>> marked as unresolved
>>>
>>> OR
>>>
>>> F) create two lists, candidate and resolved
>>>
>>> Propogate unresolved targets to later processing stages
>>>
>>> G) Allow wires to be created for unresolved services (this is different
>>> from
>>> unwired references)
>>>     This will allow proxies to be generated in this case
>>>     Proxies must have access to the unresolved binding list through the
>>> dummy target service (or whatever storage is chosen for this)
>>>     Not clear how this will work but will have to allow it to happen
>>> without selecting a binding
>>>
>>> Late resolution
>>>
>>> H)  When the proxy is called.
>>>    If the wire target is unresolved then call out to a binding resolver,
>>> which could be pluggable, in order to:
>>>       Find target service by look up in some registry, for example
>>>       Find a binding in common with the target service - as in builder
>>>       Configure the binding with the target URI and copy it into the
>>> reference
>>>       Create a provide for the binding - as in activate
>>>       Create a wire for the binding - as in start
>>>       Start the binding provider
>>>
>>>    If successful send the message down the wire
>>>
>>>    If not successful throw a service not found exception
>>>
>>>
>>> Any comments gratefully accepted
>>>
>>> Simon
>>>
>>>
>> So if I understand correctly:
>>
>> - You want your binding invoker code to resolve a target and pick a
>> suitable binding at invoke time, as for example that target was not
>> resolvable at build time.
>>
>> - The composite builder currently keeps unresolved targets as Service
>> proxies in the reference.target list, with each proxy holding a list of the
>> candidate bindings initially specified on the reference
>>
>> - But unresolved targets are not properly kept in the reference.bindings
>> list used by the composite activator and the binding providers at runtime,
>> making difficult to work with these unresolved targets then.
>>
>> I'd like suggest the following, a simple non-breaking change:
>>
>> - For each unresolved target, store an instance of SCABinding in the
>> reference.bindings list with targetComponentService pointing to the target.
>>
>> - In your SCABinding invoker implementation, access the
>> targetComponentService, resolve its location using whatever smart resolution
>> mechanism you want to use, pick the candidate binding that suits it, create
>> that binding's invoker and dispatch to it.
>>
>> Or with a twist, if you don't want to do that in SCABinding, create a new
>> SuperSmartBindingThatResolvesThingsAtInvocationTime to handle that smart
>> last-second resolution logic
>>
>> Makes sense? Simple enough?
>> --
>> Jean-Sebastien
>>
> 
> Hi Sebastien
> 
> Thanks for the thought. Sounds good to me. Simpler than my suggestion. The
> basis of this is actually what the code used to do in the old domain
> implementation :-) So not difficult to move back to that. I like your twist
> that binding.sca has make use of the candidate binding list for binding
> matching. Nice one. I'll give it a go and report back.
> 
> Thanks for the input.
> 
> Simon
> 

Simon, your changes under SVN r651490 to represent unresolved targets 
with bindings worked - almost - perfectly :) I only ran into a minor 
issue with tutorial/store-merger where the reference to the 
VegetablesCatalog was overriden by the default SCA binding that was 
created for that reference (as its target was remote).

Unrelated to your changes but contributing to cause that bug - we were 
missing a little bit of code to deal with URIs on SCA bindings. I 
checked in a small fix to BaseWireBuilderImpl in SVN r652631.

All the wiring works now, I'll add some tests to tutorial/domain to make 
sure it continues to work :)

Thanks!
-- 
Jean-Sebastien

Mime
View raw message