Last summer I read chromatic’s Modern Perl, and was recommended to default to using Moo or Moose to define classes, rather than writing code to bless things into objecthood myself. At the time the project I was working on needed to avoid any dependencies outside of the Perl core, so I made a mental note of the advice, but didn’t learn how to use Moo or Moose. I do remember feeling like I was typing out a lot of boilerplate, and wishing I could use Moo or Moose to reduce that.

In recent weeks I’ve been working on a Perl distribution which can freely use non-core dependencies from CPAN, and so right from the start I used Moo to define my classes. It seemed like a no-brainer because it’s more declarative; it didn’t seem like there could be any disadvantages.

At one point, when writing a new class, I got stuck. I needed to call one of the object’s methods immediately after instantiation of the object. BUILDARGS is, roughly, the constructor for Moo/Moose classes, so I started there, but you don’t have access to the new object during BUILDARGS, so you can’t simply call its methods on it. So what I needed to do was change my design around so as to be more conformant to the Moo/Moose view of the world, such that the work of the method call could get done at the right time. I mustn’t have been in a frame of mind for that sort of thinking at the time because what I ended up doing was dropping Moo from the package and writing a constructor which called the method on the new object, after blessing the hash, but before returning a hashref to the caller.

This was my first experience of having the call to bless() not be the last line of my constructor, and I believe that this simple dislocation helped significantly improved my grip on core Perl 5 classes and objects: the point is that they’re not declarative—they’re collections of functionality to operate on encapsulated data, where the instantiation of that data, too, is a piece of functionality. I had been thinking about classes too declaratively, and this is why writing out constructors and accessors felt like boilerplate. Now writing those out feels like carefully setting down precisely what functionality for operating on the encapsulated data I want to expose. I also find core Perl 5 OO quite elegant (in fact I find pretty much everything about Perl 5 highly elegant, except of course for its dereferencing syntax; not sure why this opinion is so unpopular).

I then came across the Cor proposal and followed a link to this recent talk criticising Moo/Moose. The speaker, Tadeusz Sośnierz, argues that Moo/Moose implicitly encourages you to have an accessor for each and every piece of the encapsulated data in your class, which is bad OO. Sośnierz pointed out that if you take care to avoid generating all these accessors, while still having Moo/Moose store the arguments to the constructor provided by the user in the right places, you end up back with a new kind of boilerplate, which is Moo/Moose-specific, and arguably worse than what’s involved in defining core Perl 5 classes. So, he asks, if we are going to take care to avoid generating too many accessors, and thereby end up with boilerplate, what are we getting out of using Moo/Moose over just core Perl 5 OO? There is some functionality for typechecking and method signatures, and we have the ability to use roles instead of multiple-inheritance.

After watching Sośnierz talk, I have been rethinking about whether I should follow Modern Perl’s advice to default to using Moo/Moose to define new classes, because I want to avoid the problem of too many accessors. Considering the advantages of Moo/Moose Sośnierz ends up with at the end of his talk: I find the way that Perl provides parameters to subroutines and methods intuitive and flexible, and don’t see the need to build typechecking into that process—just throw some exceptions with croak() if the types aren’t right, before getting on with the business logic of the subroutine or method. Roles are a different matter. These are certainly an improvement on multiple inheritance. But there is Role::Tiny that you can use instead of Moo/Moose.

So for the time being it seems I should go back to blessing hashes, and that I should also get to grips with Role::Tiny. I don’t have a lot of experience with OO design, so can certainly imagine changing my mind about things like Perlish typechecking and subroutine signatures (I also don’t understand, yet, why some people find the convention of prefixing private methods and attributes with an underscore not to be sufficient—Cor wants to add attribute and method privacy to Perl). However, it seems sensible to avoid using things like Moo/Moose until I can be very clear in my own mind about what advantages using them is getting me. Bad OO with Moo/Moose seems worse than occasionally simplistic, occasionally tedious, but correct OO with the Perl 5 core.

Posted Tue 11 Feb 2020 16:23:38 UTC Tags: