Hacker News Reader

Comment by Veraticus | Hacker News Reader

Yeah this was my primary complaint with the inclusion of generics. People will try to be all clever and their code will wind up as an unreadable, unmaintainable disaster. It has definitely led to some cool stuff, but I prefer boring, verbose, and clear any day.

View on HN

Problems are generic, such as having a tree collection. For solving generic problems, you can either use language generics, or reflection, or type erasure, or codegen. The latter three are about as far from 'boring and clear' as you can get. Still verbose, though, but I don't expect that's a benefit.

I would say 95% of people who are trying to solve a complicated problem with reflection, type erasure, or codegen are solving the problem the wrong way. Obviously I'm glad these tools exist, and some problems can indeed only be solved with them. But I think people reach for them as a first solution and make bad code as a result. Go strongly encourages you not to reach for bad solutions, which is one of its bigger advantages, in my opinion.

Collections obviously need to be generic though. Slices and maps were generic from day 1 in go so it's not like this is controversial.

For sure — a lot of language features need to be generic. But most people aren’t coding language features, and giving them the ability to do so can lead to, well, CodeFactoryFactoryMakerGenericMethodHelpersFactory, rather than good, clear, usable code.

I've spent a lot of time writing library-level C++, a reasonable amount of time writing application-level C++, and a reasonable amount of time writing both library and application-level golang.

> But most people aren’t coding language features

It's definitely a tough lever, but supporting generics for the people who are means that the people who aren't can write cleaner code. I think go got it right, eventually.

There is no language feature for a tree map. Many problems require a tree map. Either inclusion of generics was a good thing, or it is worth sacrificing static typing or requiring codegen for these use cases. Repeat for numerous collections and APIs.

Disallowing someone from using easy statically typed tree maps is not accomplishing any of the simplicity virtues people trumpet Go for having. While the much-warned-of castles of inappropriately applied generics have yet to be found in any codebase I've worked with in any language, including Rust.

This is a pretty arbitrary benchmark. I would guess, again, that 95% of programmers have never and will never need to create a tree map. Repeat for numerous collections and APIs that are totally irrelevant to most programmers' actual experience of programming.

Go is optimized for use, not computer science edge cases. And as a result it is widely used, and some of the most complicated and widely-used open-source projects out there are built in it, even before it had generics. For example, Kubernetes.

This is because of Go's simplicity, not in spite of it.

It's not about creating a tree map, it's just about using one. Maybe even one provided by a library, doesn't matter. I'm sure that 95% of the developers will need a tree map at least once in their career.

I've been using Golang professionally for about a decade. In that time the only place I've ever needed to use a tree map was an interview that specifically called for it.

Whose usage are you trying to optimize the language for? Golang is not a good academic language. But it is extremely good for actually solving problems with code.

A tree map is just an example.

More compelling ones are sets, ordered maps, a generic sync.Map, optionals (nulls suck!), abstractions over channels - badly needed!

I wouldn't consider any of those to be academic.

See the example given from Kubernetes. Is that not a real codebase, or what?

    > rg BTreeMap monorepo/ --type rust --no-filename | awk NF | wc -l
    1656

You may not have ever encountered a use case for a tree map, but there is quite a wide gray area between 'computer science edge cases' and cookie-cutter CRUD apps. For example, deterministic map ordering, or sets for map keys, or fast map equality.

Funny that you mention Kubernetes. This is the tree map implementation Kubernetes depends on https://github.com/google/btree

You are writing a Turing complete language. If someone wants to be stupid architecture astronauts, you can’t help them, beside better filtering people who work on your project.

Some problems benefit more from a solution using generics than others. Collection types, as you note, are a good example because you will have heavy re-use and the way you use generics is simple. It is worth a little bit of extra effort and it isn't going to demand much from the programmer using it.

But in the hands of people who love to go bananas, generics can result in hellish code.

My favorite example of a problem where generics is not useful is the OpenSAML implementation (Java). For all I know it may have been sanitized since I tried to make use of it about a decade ago, but that thing was, for lack of a better word, a pointless exercise in type system wankery. It was so bad it took us less time to implement what we needed from scratch than to figure out how to use the library correctly. (The code I wrote a decade ago has been in production, for a decade on a system that is used by ~150-200M users. Not just because it works, but because it can be maintained by people who aren't really that interested in spending weeks understanding it)

Is writing the same function n times more readable and more maintainable? Is it more maintainable to entirely circumvent the type system with interface{}?

Reply on news.ycombinator.com