[jsr-314-open] [Fwd: [NEW] <f:ajax> and <f:validateBean> - consistent wrapping solution needed?]

[Andy Schwartz]

Gang -

Need to bring this up one more time... Last week we found that our choice of implementations for "wrapping" f:ajax/f:validateBean has implications for delta state saving, so wanted to describe the problem here and see whether we can agree to a solution.

Some quick background on the delta state saving strategy - I'll defer to Ed and Ryan on the API specifics and also on current status with respect to 20 specification. The solution that Ryan has been working on is very similar to the Trinidad approach that we've discussed in the past. Roughly speaking, we're looking at:

1. Facelets tag handlers are applied and we build up the component tree.
2. Once a component has been "populated" (attributes set, attached objects attached), we lock the component state down.
3. From that point on through state saving, we keep track of modifications (attributes modified, atached objects added/removed).
4. At state saving, if nothing has changed (the most common case by far), we have no state to save - our saved state is null.
5. If the component has been modified, we save state. The amount of state that we save for the modified component is implementation specific. The goal is to only save the deltas, though there may be cases where we may save the full state for a specific component instance.
6. At restore time, we first build up the component tree by applying the Facelets tag handlers.
7. Components which did not save any state are fully initialized - don't need to restore any state.
8. Components which did save state restore it over the initialized component instance.

That's it in a nutshell. The end result is that the total size of saved state per component/page is reduced significantly. One really nice part of this solution is that this is all completely transparent to the application developer - ie. the page author does not need to make any changes to take advantage of this optimization.

So, what does this have to do with f:ajax/f:validateBean? The issue has to do with the timing of our current implementations.

In the f:ajax implementation, we apply "inherited" AjaxBehavior instances at tag handler execution time. So, by the time that we reach #2 above, components that are wrapped inside of an f:ajax tag have the appropriate AjaxBehavior instances attached.

With our current f:validateBean implementation, we do not apply "inherited" validators until the encode phase, which, of course, is after we've locked down the component state (ie. after #2). As a result, these validators are treated as component modifications/deltas - and as such impact the amount of state that we save.

In order to avoid this penalty, I would like to propose that we standardize on a tag handler-time solution for both f:ajax and f:validateBean - ie. I would like to see f:validateBean migrated over to a strategy that matches our current f:ajax behavior. My guess is that the implementation changes should not be all that challenging, though I have only taken a superficial look at the f:validateBean wiring. The spec changes should be minimal.

Are people okay with this proposal - ie. okay with standardizing on the f:ajax wrapping behavior? Seems to me that this was the EG's preference back when we discussed f:ajax wrapping behavior last year, so hope that this is acceptable now.

Dan -

I think that you are most familiar with the current f:validateBean implementation.  Assuming that you are okay with the proposal, would you be willing to take a look at what we do for f:ajax and maybe sync up with Ryan to discuss what it would take to convert this over to the f:ajax strategy? The f:ajax approach is fairly simple:

- When we encounter f:ajax, we create an AjaxBehavior and push it onto a stack.
- When we encounter a ClientBehaviorHolder component that can accept the AjaxBehavior instance, we create/configure a new AjaxBehavior instance and attach it to the ClientBehaviorHolder
- When we reach the end of the f:ajax tag, we pop the AjaxBehavior stack.

Note that for creating/configuring the AjaxBehavior instances, we use the technique that you hinted at in a previous email - ie. we use the intial AjaxBehavior instance as a template and save/restore its state to configure new instances.

I am actually out of the office this week - just checking email intermittently. I'll do my best to help/answer questions on this, though there may be some gaps between my access to email.

Andy

-------- Original Message -------- Message-ID: 	<49B67A17.8070609 at oracle.com> 
Date: 	Tue, 10 Mar 2009 10:32:55 -0400 
From: 	Andy Schwartz <andy.schwartz at oracle.com> 
User-Agent: 	Thunderbird 2.0.0.19 (Windows/20081209) 
MIME-Version: 	1.0 
To: 	JSR 314 Open Mailing list <JSR-314-OPEN at JCP.ORG> 
Subject: [jsr-314-open] 	[NEW] <f:ajax> and <f:validateBean> - consistent wrapping solution needed? 
Content-Type: 	text/plain; charset=ISO-8859-1; format=flowed 
Content-Transfer-Encoding: 	7bit 

Gang -

In JSF 2.0 we are introducing two core tags that wrap/enhance component 
subtrees: f:ajax and f:validateBean.

f:ajax is used to enable Ajax behavior for a subtree, eg:

> <h:panelGroup>
>   <!-- Turn on Ajax for the two input texts -->
>   <f:ajax event="valueChange">
>       <h:inputText value="foo"/>
>       <h:inputText value="bar"/>
>   </f:ajax>
> </h:panelGroup>


f:validateBean is used to enable bean validation for a subtree, eg:

> <h:panelGroup>
>   <!-- Turn on bean validation for the two input texts -->
>   <f:validateBean validationGroups="Group1, Group2"/>
>       <h:inputText value="foo"/>
>       <h:inputText value="bar"/>
>   </f:validateBean>
> </h:panelGroup>

Of course, both tags can also be used in their nested form to enable 
Ajax/bean validation on a single component.

We currently use different implementation strategies for the "wrapping" 
behavior.

f:ajax uses a handler execution-time solution.  AjaxHandler maintains a 
stack that keeps track of the current nested state of <f:ajax> tags.  
Before apply child handlers, AjaxHandler pushes the current Ajax-related 
info onto this stack.  When child/descendent handlers are applied, 
ComponentTagHandlerDelegateImpl checks this stack to get at the current 
Ajax-related state and performs the necessary work to wire up an 
AjaxBehavior to the component.  Finally, after child handlers have been 
processed, AjaxHandler pops the stack, and we're done.

The end result of the f:ajax solution is that once handlers have 
finished executing, we have a fully configured component tree.  No 
further processing is necessary to apply "inherited" Ajax behavior.

f:validateBean uses a solution which is split between tag/handler 
execution-time, validation-time and render-time.  
BeanValidatorHandler/ValidatorTagHandlerDelegateImpl set up well known 
properties on the parent component's attribute map (eg. 
BeanValidator.VALIDATION_GROUPS_KEY, 
UIInput.DEFAULT_VALIDATOR_IDS_KEY).  These are used during later 
lifecycle phases (eg. BeanValidator.validate(), UIInput.encodeEnd()) to 
determine which inherited validators to apply and how these validators 
should be configured (eg. what validation groups to use).

The end result of the f:validateBean solution is that the handlers only 
do part of the job - we need to do additional work during later 
lifecycle phases to take advantage of the information that has been 
stored away by the handlers.

While these two different strategies may at first seem like 
implementation details, and thus not especially significant, the choice 
of implementation does have impact on both our API and on spec'ed 
behavior (or, on behavior that should be defined by the spec).

In terms of API, for f:validateBean the choice of implementation leaks 
through a bit into the API in the form of the property keys that are 
used to communicate between the handler implementations and the 
component/validator implementations.

In terms of behavior, the different implementation strategies and up 
with subtly different results which may cause confusion.

For example, in the following f:ajax case:

> <h:panelGroup>
>
>   <!-- Turn Ajax on for this one -->
>   <f:ajax event="valueChange">
>       <h:inputText value="foo"/>
>   </f:ajax>
>
>   <!-- But not for this one -->
>   <h:inputText value="bar"/>
>
> </h:panelGroup>


This should work as expected.  That is, Ajax should be enabled for the 
"foo" inputText, but not for the "bar" input text.

In a similar f:validateBean case:

> <h:panelGroup>
>
>   <!-- Turn bean validation on for this one -->
>   <f:validateBean validationGroups="Group1, Group2"/>
>       <h:inputText value="foor"/>
>   </f:validateBean>
>
>   <!-- But not for this one -->
>   <h:inputText value="bar"/>
>
> </h:panelGroup>


Unless I am missing something in the spec/implementation, I believe that 
the validation-related properties are going to end up being attached to 
the parent panelGroup when the BeanValidatorHandler is applied.  Which, 
if I understand this correctly, means that the validation will be 
applied to both the "foo" and the "bar" inputText.  If this is the case, 
this is definitely not the desired behavior.

I suspect that similar problems would arise when we've got multiple 
<f:validateBean> under a single parent, eg:

> <h:panelGroup>
>
>   <!-- Turn on Group1 validation for this one -->
>   <f:validateBean validationGroups="Group1"/>
>       <h:inputText value="foo"/>
>   </f:validateBean>
>
>   <!-- Turn on Group2 validation for this one -->
>   <f:validateBean validationGroups="Group2"/>
>       <h:inputText value="bar"/>
>   <f:validateBean/>
>
> </h:panelGroup>


Though I haven't had a chance to test this out - just based on code 
inspection seems like this would be a problem.  (Though our <f:ajax> 
implementation strategy handles this case just fine.)

Since there may be other subtle inconsistencies, it would make sense to 
try to unify our implementation strategies across these two core tags.  
Of course, I am biased towards the <f:ajax> approach.  Seems to have 
better behavior in terms of targeting only the wrapped subtrees, has 
less exposure in the public API, and also just seems cleaner to me to 
produce a fully configured component tree as a result of handler execution.

Looking at the code, I didn't notice anything that would make it 
difficult to re-factor the <f:validateBean> code  to use a handler-time 
strategy like <f:ajax>.  I think our users would benefit from the 
consistency.  Any thoughts on whether we should/can go this route?  Did 
I miss anything in the <f:validteBean> behavior that would prevent us 
from using a handler-time solution?

Andy