Using narrower property validation definitions in subclasses + using an abstract class for property validation

Hello,

BLUF: It is not clear to me why I can overload methods, but not property validation/class assertions, in subclasses.

I'll lead off by saying that I don't believe what I want to do right now is possible in Matlab 2021a. I do want to know:

1. Is there a best practices workaround
2. Is there a good reason it isn't supported
3. Is this something we should get into the feature request pipeline

Let's start with a refrigerator analogy; I initially want to create an abstract superclass called container, that has a bunch of methods for e.g. adding contents, removing contents, and checking whether the container is full.

classdef (Abstract) container < handle
    properties (Abstract)
        contents 
        capacity
    end
    methods
        tf = addContent(this,content)   % e.g. add a brownie to the top of the stack, return false if the container is full
        content = rmContent(this)       % e.g. take a brownie off the top of the stack
        tf = isFull                     % return true if numel(contents) == capacity
    end
end

Now, let's say I want to make a specific container that can hold eggs. Seems simple enough, and I can even overload the rmContent so that I can grab a specific egg, rather than the top of the stack. I am also going to make an abstract class called "eggShapedItems", since my container could hold chicken eggs, cadbury eggs, easter eggs, etc. I will also make a concrete egg class

classdef eggCarton < container
    properties
        contents eggShapedItem
        capacity = 12
    end
    methods
        content = rmContent(this,index)
    end
end

classdef (Abstract) eggShapedItem < matlab.mixin.heterogeneous
    
end

classdef chickenEgg < eggShapedItem

end

I have gone with heterogeneous so that my container can hold a mixed array of eggShapedItems.

But oh no! I immediately have a problem instantiating an eggCarton: "Using size and validation functions on an inherited property is not supported. Property 'contents' is defined as abstract property in superclass 'container".

Hmm, what if I change contents to concrete in the superclass:

classdef (Abstract) container < handle
    properties 
        contents
    end
end

New error: "Cannot define property 'contents' in class 'eggCarton' because the property has already been defined in the superclass 'container'"

In other words, while I can overload methods in the subclass, I cannot overload properties (or their validation).

Ok...so I can get around this by not defining the property contents at all in the superclass 'container' I am not happy about this for two reasons***: ***(there is another issue, see at bottom)

1. The concrete methods of 'container' still assume that there is in fact a property called 'contents', even if it isn't defined in the superclass. I want these methods to be inherited by all containers, and redefining 'contents' in every subclass is bad practice.
2. I am blocked from further narrowing the definitions of either 'eggShapedItems' or 'eggCarton'. e.g.:

classdef smallEggCarton < eggCarton
    properties
         contents smallEggShapedItem
    end
end

classdef smallEggShapedItem < eggShapedItem

end

While this may be a goofy example, I think it is clear that there would be utility in being able to progressively overload the property validations through a subclass tree. In this case, I have made each class of 'contents' a subclass of the one used in the superclass; however, there would also be utility in using easily converted classes, e.g.:

classdef eggCarton < container
    properties 
        capacity uint16
    end
end

classdef smallEggCarton < eggCarton
    properties 
        capacity uint8
    end
end

If we could overload property definitions, onus is then on me to make sure that inherited methods are compatible, or that I overload them in the subclass to be compatible. I can live with that. AFAIK, that is not possible right now; the property definition is locked in by the superclass - only the default value can change, and that only if the property is abstract in the superclass.

Finally, I noted that there is another issue when trying to intantiate an eggCarton: "Error defining property 'contents' of class 'eggCarton'. Class eggShapedItem is abstract. Specify a default value for property contents."

This is a seperate, but also frustrating restriction on the property validation in Matlab; in this example, the contents of my eggCarton must be instances subclassed to 'eggShapedItem', but I would never make a concrete 'eggShapedItem', it should be an Abstract class.

As before, there are workarounds; make eggShapedItem concrete, or remove the class from the property definition entirlley, and move the validation into an overloaded set.contents(this,content) method. Again, there are drawbacks:

• I lose all of the code completion help that comes with defining the property's class in the prop block.
• This also strikes me as sloppy coding
• I could understand the issue if I tried to create an array on instantiation (e.g. contents (12,:) eggShapedItem). and I do get that Matlab is trying to put in an empty (e.g. this.contents = eggShapedItem.empty), which returns an error for abstract classes.

Thanks,

Dan

[Matt J]

I could understand the issue if I tried to create an array on instantiation (e.g. contents (12,:) eggShapedItem). and I do get that Matlab is trying to put in an empty (e.g. this.contents = eggShapedItem.empty), which returns an error for abstract classes.

That doesn't sound like a drawback. It sounds like you understand why it's not allowed! In any case, a concrete property needs to be able to hold a value at all times, so you have to either specify a default eggShapedItem object or reframe the property validation so as to make an [] value legitimate.

[Daniel Plotnick]

Thanks. I think one way to make sure you cannot actually instantiate an eggShapedItem is to protect the constructor, rather than make it abstract:

classdef eggShapedItem
    methods (Access = 'protected')
        function this = eggShapedItem()

        end
    end
end

This ensures that subclasses that are meant to be concrete can still properly be instantiated, while preventing the instantiation of a pure superclass object.

Answer 1

[Matt J]

There is a lot to unpack in your question, but basically the solution I would recommend is that you make the abstract property a Dependent property in the subclass, like the example below. In the subclass, 'contents' is just an alias for a hidden concrete property 'contents_' which can be given any validation scheme you wish (EDIT: in addtion to those already imposed for 'contents' in the superclass).

As for the why's - basically the design philosophy followed by MATLAB OOP is that what is valid/invalid for a concrete property in a superclass must remain valid/invalid in any subclass. Therefore, you are prevented from overloading validation, and similarly set.property() and get.property() methods, as a way of ensuring that a subclass object remains a valid example of a superclass object. It makes a certain amount of sense because non-overloaded superclass methods often rely on input objects satisfying superclass property restrictions. If you could overload property validation to violate those restrictions, superclass methods may no longer work for child class objects.

In any case, the idea that you should be able to overload these things has been raised before and rejected by MathWorks, so I doubt it would be worthwhile to propose it to them again.

classdef myclass

    properties (Abstract)
        contents uint32 {mustBeNonnegative}
    end

end

 classdef mysubclass<myclass

    properties (Dependent)
        contents 
    end

    properties (Hidden)
       contents_ {mustBeGreaterThan(contents_,10)}  = 12 
    end
    
    methods
       
        function val=get.contents(obj)
            val=obj.contents_;
        end
        
        function obj=set.contents(obj,val)
            obj.contents_=val;
        end
        
    end
    
    
 end

[Daniel Plotnick]

Excellent! I played around a bit with your suggestion and managed to use a variation to handle the validation:

classdef myClass < handle
    properties (Hidden,SetAccess = 'protected')
        contents_
    end
end

classdef mySubclass < myClass
    properties 
        contents validContents
    end
    methods
        function set.contents(this,v)
            this.contents_ = v;
        end
        
        function v = get.contents(this)
            v = this.contents_;
        end  
    end
end

classdef validContents
    
end

c = validContents;
b = 1;
s = mySubclass
s.contents = c
s.contents = b

It turns out that if I do not define contents in mySubClass as (Dependent), it will perform the property validation during the call to 'set.contents', throwing an error if contents is not a 'validContents'.

I had expected this to throw me a linting warning about not accessing other properties during a set/get call, but it doesn't.

This will let me do exactly what I wanted. All of the higher level methods can be left in the superclass and reference contents_.

The only downside is that you will need to add the properties/methods blocks from mySubclass to every other subclass of myClass (although I can also modify the property validation then). Not a huge burden.

One final note; I can also add a property validation to the contents_ in the superclass, if I wanted to maintain a broad restriction on what the contents can be:

classdef myClass < handle
    properties (Hidden,SetAccess = 'protected')
        contents_ validContentsSuper
    end
end
classdef validContents < validContentsSuper
    
end

Thanks for your help. I don't know if you want to add any of this comment to your above Answer, since I think the combination does provide the workaround I was looking for. I'll go ahead and Accept it later today.

[Matt J]

I had expected this to throw me a linting warning about not accessing other properties during a set/get call, but it doesn't.

I had expected the same, and worry that it may not be safe to do this. Assuming it's not an oversight on the part of MathWorks, I suppose the reason it might be safe is that contents_ (as defined in your example) can always be counted upon, as a superclass property, to receive its initializing value before contents ??

I still recommend using a Dependent property, but I have embellished my original example to capture the point you raised, that both broad and specific property validation can be imposed in the superclass and subclasses respectively.

[Austin Spencer]

As for the why's - basically the design philosophy followed by MATLAB OOP is that what is valid/invalid for a concrete property in a superclass must remain valid/invalid in any subclass.

I agree that this is an important principle of OOP. However, I don't believe the OP's original example violates this rule. The class of contents is unspecified in the superclass (i.e., any class is allowed) and defined as eggShapedItem in the subclass. Therefore, eggCarton doesn't violate the class validation of container.

It would add so much to MATLAB's OOP capability (especially when using composition) if subclasses could further restrict the type of superclass properties. As long as the property's type in the subclass is itself a subclass of the property's type in the superclass, it should be valid. To me, this follows the principle that a subclass object can always be used in place of it's superclass.

[Matt J]

Yeah. Well, there are backdoor methods to accomplish this, like in the following:

classdef (Abstract) container < handle
    properties (Abstract)
        contents 
    end
    methods (Static, Abstract)
        tf=contentsValid(val);
    end
    methods
        function obj=set.contents(obj,val)
            assert( obj.contentsValid(val), 'Invalid contents');
            obj.contents=val;
        end
    end
end

classdef eggCarton < container
    properties
        contents
    end
    methods (Static)
        function tf=contentsValid(val)
           tf=isa(val,'eggShapedItem');
        end
    end
end

classdef smallEggCarton < container
    methods (Static)
        function tf=contentsValid(val)
           tf=isa(val,'chickenEgg');
        end
    end
end

[Austin Spencer]

Yeah. Well, there are backdoor methods to accomplish this, like in the following:

I totally agree, there are ways to enforce data types manually, like you show. I think these are valid approaches for achieving certain goals and they have existed since long before property validation what a thing.

What I would love to see though (and this comment is more aimed at Mathworks) is type validation via definition (or what you might call "at compile time") rather than "check at run time". In my view, one of the main benefits of property validation is to provide a clear definition of acceptable data type for those using, modifying, or subclassing the class. It's basically a form of "structural documentation" that the person reading the code (or MATLAB's linter) can quickly see and interpret. Otherwise, there's not much marginal benefit of property validation other than making type checking slightly more succinct. Maybe people will think that this borders too closely on expecting MATLAB to have strong typing, but I think this is more in the realm of having a complete and effective OOP system.

[Jan Kappen]

I fully agree and would love to see any updates here.

[Jan Kappen]

Simple example why this would be highly valuable:

classdef dataLoaderBase < handle
    properties
        DataFormatVersion (1,1) string % must be string scalar
    end
end

classdef concreteDataLoader < dataLoaderBase
    properties
        DataFormatVersion (1,1) string {mustBeMember(DataFormatVersion, ["v1", "v2"])}
    end
end

This does not work; but I can't use mustBeMember in the base class for obvious reasons. The only way is now to use validation functions in the set/get functions as shown above, but the key issue with that is that all the nice auto completion features are gone. One of the best features in latest MATLAB versions. Also, the concreteDataLoader clearly indicates that there are only two data format versions available, it's a discrete set I can choose from. That's very nice when parsing the class and visualizing it in a user interface - it directly maps to a dropdown, not to a generic string.

Any workarounds to make that happen?