Eli Bendersky's blogpost on the Expression Problem: https://news.ycombinator.com/item?id=45155877

I thought this was a good read, and it’s clearly a problem in an OO language, but I struggled to see how it’s an actual problem in a functional language. Maybe I need to go read that paper he referenced by the Racket guys.

▲

kragen a day ago | parent | next [-]

In my example above, to add a new variant type to bicexpr, you need to edit roughly every function that pattern-matches on bicexprs, which is a good fraction of the Bicicleta interpreter codebase.

▲

sevensor a day ago | parent [-]

Ah, I see. I’ve written that code where you have to add one case to a whole bunch of pattern matches. It’s not terrible; I rather like that everything that does the same thing is in the same place. But I have wondered if there was a better way. Clearly I need to go read the article again.

▲

kragen 19 hours ago | parent [-]

Yeah, I share your assessment that it's not unmanageable.

But it does add friction. I suspect that there are only two real solutions:

- program source code that isn't a linear document but instead can be viewed in multiple different ways: by operation or by operand type, for example. Then you would use the view that best suits the modification you were doing at the time. Smalltalk, for example, does support browsing either all the implementors of a method selector or all the methods of a class. (But in both cases you normally only see one method at a time.)

- Alternating levels of hierarchy. Maybe if you have 63 operations on 63 types, you could arrange it so that you can add a 64th operation or a 64th type by editing only 8 places instead of 63.

▲

sevensor 19 hours ago | parent [-]

> Smalltalk, for example, does support browsing either all the implementors of a method selector or all the methods of a class

That’s pretty neat. I tried using the object browser a million years ago and it confused me, but I didn’t know anything about anything at the time. Maybe Smalltalk deserves another look.

	▲	igouy 2 hours ago \| parent [-]
		Smalltalk IDEs are usually multi-window, so it isn't either/or, it's by method selector and by class instance methods and by …

▲

naasking a day ago | parent | prev [-]

The example in the article was very clear I thought: try to extend an interpreter with a new primitive (function call in the article) without changing the original interpreter's source code. I'm an FP language, you can't extend the core AST without changing the source code for eval.