Philosophies of Rust and Haskell

Rust is a systems programming language following fairly standard imperative approaches and a C-style syntax. Haskell is a purely functional programming language, innovating in areas such as type theory and effect management. Viewed that way, these languages are polar opposites.

And yet, these two languages attract many of the same people, including the engineering team at FP Complete. Putting on a different set of lenses, both languages provide powerful abstractions, enforce different kinds of correctness via static analysis in the compiler, and favor powerful features over quick adoption.

In this post, I want to look at some of the philosophical underpinnings that explain some of the similarities and differences in the languages. Some of these are inherent. Rust's status as a systems programming language essentially requires some different approaches to Haskell's purely functional nature. But some of these are not. It wasn't strictly necessary for both languages to converge on similar systems for Algebraic Data Types (ADTs) and ad hoc polymorphism (via traits/type classes).

Keep in mind that in writing this post, I'm viewing it as a consumer of the languages, not a designer. The designers themselves may have different motivations than those I describe. It would certainly be interesting to see if others have different takes on this topic.

Rust: ownership

This is so obvious that I almost forgot to include it. If there's one thing that defines Rust versus any other language, it's ownership and the borrow checker. This speaks to two core pieces of Rust:

The goal of serving as a systems programming language, where garbage collection is not an option
The goal of providing a safe subset of the language, where undefined behavior cannot occur

The concept of ownership achieves both of these. Many additions have been made to the language to make it easier to work with ownership overall. This hints at the concept of ergonomics, which is fundamental to Rust philosophy. But ownership and borrow checking are also known as the harder parts of the language. Putting it together, we see a philosophy of striving to meet our goals safely, while making the usage of the features as easy as possible. However, if there's a conflict between the goals and ease of use, the goals win out.

All of this stands in stark contrast to Haskell, which is explicitly not a systems language, and does not attempt in any way to address those cases. Instead, it leverages garbage collection quite happily, with the trade-offs between performance and ease-of-use inherent in that choice.

Haskell: purely functional

The underlying goal of Haskell is ultimately to create a purely functional programming language. Many of the most notable and unusual features of Haskell directly derive from this goal, such as using monads to explicitly track effects.

Other parts of the language follow from this less directly. For example, Haskell strongly embraces Higher Order Functions, currying, and partial function application. This combination turns many common structures in other languages (like loops) into normal functions. But in order to make this feel natural, Haskell uses slightly odd (compared to other languages) syntax for function application.

And this gets into a more fundamental piece of philosophy. Haskell is willing to be quite dramatically different from other programming languages in its pursuit of its goals. In my opinion, Rust has been less willing to diverge from mainstream approaches, veering away only out of absolute necessity.

This results in a world where Haskell feels quite a bit more foreign to others, but has more freedom to innovate. Rust, on the other hand, has stuck to existing solutions when possible, such as eschewing monadic futures in favor of async/.await syntax.

Expression oriented

I undervalued how important this feature was for a while, but recently I've realized that it's one of the most important features in both languages for me.

I used to think that the reason I loved both Haskell and Rust so much was their shared strong typing, ADTs, and pattern matching combination.

After a recent discussion, I think it may be more about being expression-oriented languages.
— Michael Snoyman (@snoyberg) January 11, 2021

Instead of relying on declare-then-assign patterns, both languages allow conditionals and other constructs to evaluate to values. This reduces the frequency of seeing mutable assignment and avoids cases of uninitialized variables. By restricting mutable assignment to cases where it's actual mutation, we get to free up a lot of head space to focus on the trickier parts of programming.

Type system

Rust and Haskell have very similar type systems. Both make it easy to create new types, provide for features like newtypes, provide type aliases, and offer a combination of product (struct) and sum (enum) types. Both allow labeling fields or accessing values positionally. Both offer pattern matching constructs. Overall, the similarities between the two languages far outweigh the differences.

I place a large part of the shared interest between these languages at the feet of the type system. Since I started using Haskell, I feel strongly hampered using any language without a rich, flexible, and powerful type system. Rust's embrace of Algebraic Data Types (ADTs) feels natural.

There are some differences between the languages in these topics, but they are mostly superficial. For example, Haskell uses the single keyword data for introducing both product and sum types, while Rust uses struct and enum, respectively. Haskell will allow creation of partial field accessors in sum types, while Rust does not. Haskell allows for partial pattern matches (with an optional warning), and Rust does not.

These are meaningful and affect the way you use the languages, but I don't see them as deeply philosophical. Instead, I see both languages embracing the idea that encouraging programmers to define and use strong typing mechanisms leads to better code. And it's a message I wholeheartedly endorse.

Traits and type classes

In the wide world of inheritance and polymorphism, there are a lot of different approaches. Within that, Rust's traits and Haskell's type classes are far more similar than different. Both of them allow you to separate out functionality (methods) from data (struct/data). Both allow you to create new types or traits/classes yourself and add them on to existing types/traits/classes. Both of them support a concept of associated types, and multiple parameters (either via parameterized traits or multi-param type classes).

There are some differences between the two. For one, Rust doesn't allow orphans. An implementation must appear in the same crate as either the type definition or the trait definition. (The fact that Rust treats an entire crate as a compilation unit instead of a single module makes this restriction less of an imposition.) Also, Haskell supports functional dependencies, but that's not terribly interesting, since that can be closely approximated with associated types. And there are other, more subtle differences, around issues like overlapping instances. Rust's lack of orphans allows it to make some closed world assumptions that Haskell cannot.

Ultimately, the distinctions above don't lend themselves to a deep philosophical difference, but rather minor variations on a theme. There is, however, one major distinction in this area between the two languages: Higher Kinded Types (HKTs). In Haskell, HKTs provide the basis for such typeclasses as Functor, Applicative, Monad, Foldable, and Traversable. In Rust, implementing some kind of traits around these concepts is a bit more complicated.

And this is one of the deeper philosophical differences between the two languages. Haskellers readily embrace concepts like HKTs. The Rust community has adamantly avoided embracing them, due to their perceived complexity. Instead, in Rust, alternative and arguably simpler approaches have been used to solve the same problems these typeclasses solve in Haskell. Which leads us to probably the biggest philosophical difference between the languages.

General vs specific

Let's say I want to have early termination in the case of an error. Or asynchronous coding capabilities. Or the ability to pass information to the rest of a computation. How would I achieve this?

In Haskell, the answer is obviously Monads. do-notation is a general purpose "programmable semicolon." It generally solves all of these cases. And many, many more. Writing a parser? Monad. Concurrency? Maybe Monad, or maybe Applicative with ApplicativeDo turned on. But the common factor: we can express large classes of problems as do-notation.

How about Rust? Well, if you want early termination for errors, you'll use a Result return type and the ? try operator. Async? async/.await syntax. Pass in information? Maybe use method syntax, maybe use thread-local state, maybe something else.

The point is that the Haskell community overall reaches for generalizing a solution as far as possible, usually along the lines of some abstract mathematical underpinning. There are huge advantages to this. We build out solutions to problems we didn't even know we had. We are able to rely on mathematical laws to guide our designs and ensure concepts compose nicely.

The Rust community, instead, favors specific, ergonomic solutions. Error handling is really common, so give it a single character operator. Make sure that it handles common cases, like unifying error types via the From trait. Make sure error messages are as clear as possible. Optimize for the 95%, and don't worry about the 5% yet. (And see the next section for the 5%.)

To me, this is the deepest non-inherent divide between the languages. Sure, ownership versus purity is huge, but it's right there on the label of the languages. This distinction ends up impacting how new language features are added, how people generally think about solutions, and how libraries are designed.

One final point. As much as I've implied that the Rust and Haskell communities are in two camps here, that's not quite fair. There are people in the Haskell community looking to make more specific solutions to some problems. (I'm probably one of them with things like RIO.) And while I can't think of a concrete Rust example to the contrary, I have no doubt that there are cases where people design general solutions when a more specific one would suffice.

Code generation/metaprogramming/macros

Haskell has metaprogramming via Template Haskell (TH). It's almost universally viewed as a necessary evil, but evil nonetheless. It screws up compilation in some cases via stage restrictions, it requires a language pragma to enable, and introduces awkward syntax. Features like deriving serialization instances are generally moving towards in-language features via the Generic typeclass.

Rust's "Hello World" sticks a macro call on the second line via println!. The syntax for calling macros looks almost identical to function calls. Common libraries encourage macro usage all over the place. serde serialization deriving, structopt command line parsing, and snafu/thiserror error type creation all leverage macro attributes and deriving.

This is a fascinating distinction to me. I've been on both sides of the TH divide. Yesod famously uses TH for a lot of code generation, which has earned the ire of many Haskellers. I've since generally avoided using TH when possible in the past few years. And when I picked up Rust, I studiously avoided learning how to create macros until relatively recently, lest I be tempted to slip back into my old, evil ways.

Metaprogramming definitely complicates some things. It makes it harder to debug some problems. Rust does a pretty good job at making sure error messages can be comprehensible. But documentation on macro arguments and return types is still not as nice as functions and methods.

I think I'm still mostly in the Haskell camp of avoiding unnecessary metaprogramming in my API design, but I'm beginning to be more free with it. And I have no reservations in Rust about using macros; they're wonderful. I do wonder if the main issue in Haskell isn't the overall concept of metaprogramming, but the specific implementation with Template Haskell.

Backwards compatibility

Rust overall has a more coherent and consistent story around backwards compatibility. It's almost always painless to upgrade to new versions of the Rust compiler. This puts an extra burden on the compiler team, and constrains changes that can be made to the language. And in one case (the module system update), it required a new edition system to allow for full backwards compatibility.

The Haskell community overall cares less about backwards compatibility. New versions of the compiler regularly break code. New versions of libraries will get released to smooth out rough edges in the APIs. (I used to do this regularly, and now regret that. I've tried hard to keep backwards compatibility in my libraries.)

Overall, I think the Rust community's approach here is better for producing production software. Arguably the Haskell approach allows for much more exploration and attainment of some higher level of beauty. Or as they say, "avoid (success at all costs)."

Optimistic optimizations

GHC has a powerful rewrite rules system, which can rewrite less efficient combinations of functions to more optimized ones. This plays in a big way in the vector package, where rewrite rules implement stream fusion, allowing many classes of vector pipelines to completely avoid allocation. This is a massive optimization. At least when it works. As I've personally experienced, and many others have too, rewrite rules can be finicky. The Haskell approach is to be happy that our code sometimes gets much faster, and that we get to keep elegant, easy-to-understand code.

The Rust approach is the polar opposite. Either code will definitely be fast or definitely be slow. I learned this a while ago when looking into recursive functions and tail call optimization (TCO). The Rust compiler will not perform a TCO, because it's so easy to accidentally change a TCO-able implementation into something that eats up stack space. There are plans to make explicit tail calls possible with the become keyword someday.

More generally, Rust embraces the concept of zero cost abstractions. The idea is that you should be able to abstract and simplify code, when we can guarantee that there is no cost. In the Haskell world, we tend to focus on the elegant abstraction, even if a cost will be involved.

Learning curve

A short one here. Both languages have a higher-than-average learning curve compared with other languages. Both languages embrace their learning curves. As much as possible, we try to make learning and using the languages easy. But neither language shies away from powerful features, even if it will make the language a bit harder to learn.

To quote a Perlism: you'll only learn the language once, you'll use it for the rest of your life.

Explicitly mark things

Both languages embrace the idea of explicitly marking things. For example, both languages encourage (in Haskell's case) or enforce (in Rust's case) marking the type signature of all functions. But that's pretty common. Haskell goes further, and requires that you mark all effectful computations with the IO type (or something similar, like MonadIO). Rust requires than anything which may fail be marked with a Result return value.

You may argue that these are actually a difference in the language, and to some extent that's true. But I think the difference is about what the language considers important. Haskell, for reasons of purity, values deeply the idea that an effect may be performed. It then lumps errors and exceptions into the contract of IO and the concept of laziness (for better or worse). Rust, on the other hand, doesn't care if you may perform an effect, but deeply cares about whether an error may occur.

Type enforce everything?

When I initially implemented Haskell's monad-logger, I provided an instance for IO which performed no output. I received many complaints that people would rather get a compile time error if they forgot to initialize the logging system, and I removed the IO instance. (Without getting into details: this was definitely the right decision for the API, regardless of the distinction with Rust.)

That's why I was so amused when I first used the log crate in Rust, and realized that if you don't initialize the logging system, it produces no output. There's no runtime error, just silence.

Similarly, many functions in the Tokio crate will fail at runtime if run from outside of the context of a Tokio runtime. But nothing in the type system enforces this idea.

And finally, I've been bitten a few times by actix-web's state management. If you mismatch the type of the state between your handlers and your service declaration, you'll end up with a runtime error instead of a compile time bug.

In the Haskell world, the overall philosophy is generally to approach "if it compiles, it works." Haskellers love enforcing almost every invariant at the type level.

I haven't discussed this much with Rustaceans, but it seems to me that the overall Rust philosophy here is slightly different. Instead, we like to express tricky invariants at the type level. But if something is so obviously going to fail or behave incorrectly in the most basic smoke testing, such as a Tokio function crashing, there's no need to develop type-level protections against it.

Conclusion

I hope this laundry list comparison was interesting. I've been meaning to write it down for a while, so I kind of feel like I checked off a New Year's Resolution in doing so. I'd be curious to hear any other points of comparison people have, or disagreements about my assessments.