I once wrote a unifying parent class for several client-specific reports we had. Think the parent class implementing the top-level control flow/logic and the child implementations having methods that it calls into, sometimes once (e.g. for the header) and sometimes per-row.
Specific reports needed various kinds of customization. Some needed to call the client's web API to get some extra data for each row (and since this was across the internet, it needed to be async). Some needed to accumulate statistics on each row and sum them over the whole report at the end. One needed to query an ancillary database for each row, but all those queries had to be part of the same transaction so that they would be consistent.
Now in theory you can do ad-hoc things for each of those cases. You could make the per-row method always async (i.e. return Future), so that you can override it to do a web API call sometimes. You could stash the statistics in a mutable variable in the report that accumulates them, remembering to do locking. You could use a session on the ancillary database bound to a threadlocal to do the transaction management (most database libraries assume that's how you're doing things anyway), and as long as your returned Futures are never actually async then it would probably work. But realistically it would be very hard to do safely, and I'd never have spotted the underlying symmetry that let me pull out the high-level logic. More likely we'd've stuck with three distinct copy-pasted versions of the report code, with all the maintenance burden that implies.
In principle an abstraction never tells you something you didn't already know - whatever properties you're using, you could've always worked them out "by hand" for that specific case. Like, imagine programming without the concept of a "collection" or any idea of the things you can do on a collection generically (such as traverse it) - instead you just figure out that it's possible to traverse a linked list or a red-black tree or an array, so you write the code that works on all of them when you need it. That's absolutely a way that you can program. But if you don't have this vocabulary of concepts and patterns then realistically you miss so many chances to unify and simplify your code. And if you have the rigorous category-theory concepts, rather than fuzzier design patterns, then you have quick, objective rules for figuring out when your patterns apply - and, even more importantly, when they don't. You can use library code for handling those concepts with confidence, instead of e.g. wondering whether it's ok for your custom iterator to expect the library to always call hasNext() before it calls next(). It's a huge multiplier in practice.
Specific reports needed various kinds of customization. Some needed to call the client's web API to get some extra data for each row (and since this was across the internet, it needed to be async). Some needed to accumulate statistics on each row and sum them over the whole report at the end. One needed to query an ancillary database for each row, but all those queries had to be part of the same transaction so that they would be consistent.
Now in theory you can do ad-hoc things for each of those cases. You could make the per-row method always async (i.e. return Future), so that you can override it to do a web API call sometimes. You could stash the statistics in a mutable variable in the report that accumulates them, remembering to do locking. You could use a session on the ancillary database bound to a threadlocal to do the transaction management (most database libraries assume that's how you're doing things anyway), and as long as your returned Futures are never actually async then it would probably work. But realistically it would be very hard to do safely, and I'd never have spotted the underlying symmetry that let me pull out the high-level logic. More likely we'd've stuck with three distinct copy-pasted versions of the report code, with all the maintenance burden that implies.
In principle an abstraction never tells you something you didn't already know - whatever properties you're using, you could've always worked them out "by hand" for that specific case. Like, imagine programming without the concept of a "collection" or any idea of the things you can do on a collection generically (such as traverse it) - instead you just figure out that it's possible to traverse a linked list or a red-black tree or an array, so you write the code that works on all of them when you need it. That's absolutely a way that you can program. But if you don't have this vocabulary of concepts and patterns then realistically you miss so many chances to unify and simplify your code. And if you have the rigorous category-theory concepts, rather than fuzzier design patterns, then you have quick, objective rules for figuring out when your patterns apply - and, even more importantly, when they don't. You can use library code for handling those concepts with confidence, instead of e.g. wondering whether it's ok for your custom iterator to expect the library to always call hasNext() before it calls next(). It's a huge multiplier in practice.