[jboss-jira] [JBoss JIRA] (WFCORE-1556) Poor handling of 'required', 'nillable' and 'alternatives' in AttributeDefinition

[Brian Stansberry (JIRA)]

    [ https://issues.jboss.org/browse/WFCORE-1556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13239726#comment-13239726 ] 

Brian Stansberry commented on WFCORE-1556:
------------------------------------------

I don't think the AD can ever be responsible for validating "required but not if alternatives are present". The AD is asked to validate (i.e. a call to validateAndSet) during execution of a particular step, but that step may be part of a composite, and it's only when the composite is complete that it's ok to validate the resulting model. For example the composite may have two write-attribute steps, one of which sets alternative a and the other undefines alternative b. The overall model state will never be valid after just one step.

There are other situations where things may be in flux, e.g. validating during parsing when some alternatives haven't been parsed yet.

In theory we could be lenient in some situations like parsing and write-attribute and stricter in others, like an add op, but that requires changes to AD that some OSHs may not use, it's complex and probably buggy, and it still doesn't allow for weird corner cases like a composite with a broken add step followed by a correcting write-attribute step.

It seems like we should be able to get rid of the need for setValidating(false) though and just detect the combination of required + alternatives and be lenient.

> Poor handling of 'required', 'nillable' and 'alternatives' in AttributeDefinition
> ---------------------------------------------------------------------------------
>
>                 Key: WFCORE-1556
>                 URL: https://issues.jboss.org/browse/WFCORE-1556
>             Project: WildFly Core
>          Issue Type: Bug
>          Components: Domain Management
>            Reporter: Brian Stansberry
>
> The handling of the notions of 'require'd and 'nillable' don't comply with the specs in https://docs.jboss.org/author/display/WFLY10/Description+of+the+Management+Model, particularly the "Description of an Attribute" part, where the 'required' and 'alternatives' fields are well described, with their relationship spelled out, while 'nillable' doesn't appear at all. Then in "Description of an Operation Parameter or Return Value" nillable is specified, although I think those descriptions could be tightened up in general.
> The read-resource-description output for an attribute doesn't show "required" at all.
> After a fair bit of thinking, I've finally recalled why the two separate notions of "required" and "nillable" exist in the first place.
> The "required" and "alternatives" pieces go together. Is something required? And then if it is required, an alternative satisfies. So you can have two attributes/params, both required, but they are alternatives so one is defined and the other is not. And this is an ok thing.
> And then 'nillable' comes in to help clients understand in a simple way whether they need to account for the possibility an undefined value. Basically:
> nillable = !required || alternatives != null
> Right now, nillable is implemented as
> nillable = !required
> There are a number of problems I see with AttributeDefinition:
> 1) We don't output the 'required' metadata unless it's an operation param being described.  This is contrary to spec. However we we shouldn't fix this unless we can have meaningful values for 'required', ones where the value can be true but the attribute/param can still have an undefined value, due to an alternative being present. If we can't fix that, there's no point outputting required as it's just redundant with what we output for 'nillable'.
> 2) We do output the 'nillable' metadata, even for attribute descriptions, where it isn't in the spec. But in this case I think we change the spec, as dropping nillable would be an incompatible change.
> 3) We don't properly validate the "required but has alternatives case." This can't be validated solely with ParameterValidator impls as those only see a single attribute value and don't have contextual information about other attributes/params. To get around this limitation, devs are saying that attributes with alternatives "allowNull" which is equivalent to saying they are not required. But they are required! So I think a fix for this will require AttributeDefinition itself to validate the overall resource model or operation object, and skip calling the ParameterValidator if the attribute/param is undefined and an alternative is defined.
> 4) AttributeDefinition and AbstractAttributeDefinitionBuilder unfortunately have a getter/setter/constructor param for property "allowNull" instead of a property named "required". This is unfortunate because "allowNull" intuitively maps to "nillable", but "required" is a much more useful thing to set. The value of "nillable" can be derived from a "required" setting and an "alternatives" setting, but the value of required cannot be so derived.
> I think a fix for this would probably start from 4), deprecating setAllowNull, adding get/setRequired and changing the meaning of the AD(Builder) constructor param to "required". The effect of this would be that all current code setting "allowNull" would now be setting a new "required" field. This should be a non-breaking change as in current code that's the effect anyway.
> Next would be to fix 3), by changing how AD validates.
> Next would be to change the metadata we output, such that "required" is always present and the "nillable" value is !required || alternatives != null. And change the spec accordingly.
> Last, in a separate task, would be to find attributes that were configuring "allowNull = true" solely to allow validation to pass when alternatives are present, and fix them to say "required=false".



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)