A Sticky String Quandary
stephendiehl.com
stephendiehl.com
My impression is that the Haskell community is very self-critical (which is great), but someone just peeking in from the outside might think that Haskell is still in the random hobby project stage, and that it still hasn't figured out strings.
That's totally not the case though! Strings are solved, just a bit annoying to use sometimes. We're running Haskell in production and it's amazingly stable and hard to break.
I wish the community was bigger though, that's why I'm posting this to encourage everyone to try it.
Use Text https://hackage.haskell.org/package/text
"The Text type represents Unicode character strings, in a time and space-efficient manner. This package provides text processing capabilities that are optimized for performance critical use, both in terms of large data quantities and high speed."
Clarity about what lives where is one area where the Haskell ecosystem is a bit deficient, in my opinion. Documentation in general doesn't seem to be great.
Anecdotally, everything serious I write in Haskell uses Data.Text by default these days, and just converts to and from other representations like String as necessary. It’s mildly inconvenient for things like writing Show instances or integrating with a library that uses a different convention, but still better in almost any context than using String by default IME.
I would make a conjecture that it's also true for Lisps (quality of std lib is poor often) and other powerful libraries.
On other hand, languages of simpler kind, like Java or Python, have more adequate std libs.
It's because, for a really powerful language std lib has to be opinionated. And people understand that but they can't agree on something. So they live with whatever common denominator. And lost a lot of traction there.
A counterexample is Clojure where std lib is very nice if heavily skewed towards FP, reinforcing my point.
My favorite is the treatment of iteration. You have immutable collections that support an iterator interface that allows modification and throws a runtime exception (so much for a type system...).
There is an interface that lets you iterate over the elements of something that supports it without allowing removal, but it's not recommended, and it doesn't enable the enhanced for loop that came with Java 1.5: https://docs.oracle.com/javase/7/docs/api/java/util/Enumerat....
It's as if someone said "how wrong can we get this?"
On the one hand, parts of the Haskell prelude are quite opinionated. String as a linked list? That's an impressive sort of FP purity: you can write tail recursive functions, and there are people out there who find that thrilling. It's just horribly space and time inefficient.
Other parts are just old. I've read that there the reason head throws an exception when applied to an empty list rather than a Maybe was because working with Maybes was much more painful when Haskell was first written. That's just ordinary backwards compatibility pain.
Other parts just seem painful for no reason whatsoever. Lazy and Strict ByteStrings/Text are both named ByteString/Text. They just live in different namespaces, so when you see code that reads "Text -> a", the only way to distinguish it is to look at what's been imported. If you couldn't decide which was preferred, they should be "StrictText"/"LazyText" (and programmers would be free to import them as Text if they so desired).
(Nevermind that Text, while essential, isn't actually part of the Prelude...).
This must be the curse of writing a language that becomes successful enough to grow a significant user base. You will inevitably discover that some of your original decisions about language features or standard library designs aren’t ideal. However, even if you know a much better way now, you can’t just replace the existing version without breaking a lot of existing code. Worse, over time people will start writing their own libraries, and some of them will adopt the same flawed conventions as your standard library by default. C++ has long struggled with strings and text processing for similar reasons.
"Ugh, this length function isn't parameterized over the integrals?"
Or, alternatively:
"Ugh, this length function is parameterized over the foldables?"
People will never agree what's best, but it doesn't really matter. One nice thing about Haskell is that all the "default" functions, types, and data structures are just imported from the Prelude library. You can just import your own version if you want, and people do that.
It's only bad in the sense that it's inefficient. It's preferable to most other string types in many ways. It's also totally reasonable to have in the prelude: [] is there, and Char is there, so really all they've done is slapped the two together. The current situation (importing ByteString or Text for performance) is pretty good IMO. Haskell doesn't have Map or Set or anything in the prelude, so I think it's reasonable to leave powerful packed text types out as well. In fact, I would even support leaving [] out of the prelude and making the import of Data.List explicit.
> lots of partial functions are legitimate red flags
This is true, but the fact that we're even complaining about this is a first-world problem. Other languages have partial behavior built in to the language. Try going to a python or java developer and telling them "array access should return an optional type instead of throwing an exception on out-of-bounds". This is totally reasonable to deal with in Haskell, but inconceivable in other languages.
The current situation (importing ByteString or Text for
performance) is pretty good IMO. Haskell doesn't have Map
or Set or anything in the prelude, so I think it's
reasonable to leave powerful packed text types out as well.
Try going to a python or java developer and telling them
"array access should return an optional type instead of
throwing an exception on out-of-bounds".
Yes, that's exactly what the problem is. String gets used instead of Text because it's the path of least resistance.
It's unfortunate that people use String just because it's there, or Int when they should use Word (because Word wasn't added to the Prelude until recently). For example, it's awful that length returns an Int. It should return a Word or be generic.
Indeed, the Haskell programmer is the one I'd say lives in "the first world".