I do wonder if Rust 1.0 is being branched just a little bit too early. I understand that you have to draw a line in the sand somewhere, and the rolling release model means new features can be rolled out quickly. On the other hand, I'm rather disappointed that the ergonomics of error handling may change substantially after 1.0. This is something that is going to affect every single Rust program written, and people may be surprised if 'the best way' to do this changes after 1.0.
There's always going to be some feature that would make something nicer and that seems worth waiting for. At the limit we'd be holding off 1.0 until Rust has higher-kinded types (which has never had a concrete proposal nor is anyone really sure it can be incorporated nicely with the rest of Rust's type system), and that's a feature that would absolutely change the paradigm of error handling overnight.
The time is soon. I as much as anyone want error handling to be flawless, but I don't think that's enough of a reason to hold back the rest of the language. If we have to evolve our approach to error handling a ways down the road, so be it. Other languages have endured similar paradigm shifts.
However, this issue will be inherently subjective... all I can say for sure is, if nobody at all were saying "no, it's too early!", it would certainly be too late. :)
I do get that, but the distinction I'm trying to make between a feature like this and HKT is that you would expect changes in this area to impact practically every Rust program in existence, while the same is not necessarily true for HKT. Sure the changes can be made backwards compatibly, but it still means that the definition of 'normal Rust code' will be quite different, which people may find odd for a language that just hit 1.0.
Ultimately I do appreciate reaching 1.0 is a difficult set of tradeoffs and everyone has their pet feature they're sad isn't going to make it in. Bring on 2.0!
By far the biggest demand for HKT is from people who are requesting Haskell-like do-notation for error handling. HKT would affect everything! Given this, I absolutely agree with the notion that HKT is a 2.0 feature rather than a 1.0 feature.
> By far the biggest demand for HKT is from people who are requesting Haskell-like do-notation for error handling.
That hasn't been my experience -- we are constantly running into cases in the libraries where we need HKT to express signatures in traits. The simplest example is talking about iterator return types, which often need to depend on a lifetime. See the collections reform RFC (https://github.com/aturon/rfcs/blob/collections-conventions/...) for details on that example. There are many, many others.
I suspect that HKT's role in Rust will be much more about extending the trait system for generic programming, rather than full-blown monadic programming.
Thanks for posting this link. I struggled with this problem all yesterday, and hadn't realized that rust can't handle lifetimes over traits until you posted this.
Unfortunately without HKT you cannot model monads, abstractions that are useful for a wide range of use-cases, one of which is error handling.
In my opinion it is a pity that HKT wasn't there from day one. I had high hopes for Rust, because in comparison with languages like Go, it didn't take a stance like generics are hard, lets go shopping. And personally I can see problems with such pragmatic approaches, because such features have to be baked in, otherwise they'll either never make it or they'll suck.
And then we are back to square one - C++ and Java forever.
> In my opinion it is a pity that HKT wasn't there from
> day one.
I think this shows a misunderstanding of the genesis of Rust. :P Literally the only feature of early Rust that has survived to the present day has been the use of `unsafe` blocks to cordon off bits of code that cannot be statically checked for safety (I don't count things like generics as having survived from early Rust, because early Rust used Java-style virtual dispatch for generics rather than C++-style monomorphization (nor did it have typeclasses, because traits weren't conceived until later)). From day one, the goal has simply been to make a systems language that was safer than C++, and it's taken years to figure out how to do that effectively. And as a goal-oriented language, features are cheap: Rust has thrown out more features than most languages will ever have. The fact that HKT aren't there yet is because they're merely a nice-to-have rather than strictly necessary (in the way that lifetimes are strictly necessary), and because Rust has never particularly valued maximal expressivity in the way that e.g. Scala has. In the meantime, just because generic monads aren't possible doesn't mean that monads can't be approximated for specific types like Result and Option, and the lack of them isn't going to prevent Rust from its goal of being a replacement for C++.
We do plan to add some version of this RFC, but made a careful decision about not including it for 1.0. It will not completely change the error-handling design, but rather empower it with additional, finer-grained control.
It's possible that we will add at least the `?` notation to replace `try!` before 1.0 final; it remains to be seen.
Are they really ready for a 1.0? I'm looking at the standard library, and things (or core functionality of things) like file-system I/O, vectors, strings, _iterators_ … are marked as "experimental"; I was particularly curious as to how they were going to deal with iterators, since I think this is a particularly hard problem. (In many languages these are either pointers (C++) or close to it.¹)
I wouldn't really consider a language without a standard library "complete", and a good standard library is important to me in a language. (As it's going to help you get stuff done.)
¹Python, for example, doesn't specify iterator invalidation rules AFAICT; you just have to look at what CPython does. And in CPython, there are ways to make an iterator return a sequence of items that does not make any sense. (e.g., the same item twice, or skipping items…) In C++, invalidation is well-specified, but the compiler doesn't really help catch errors.
I completely agree that a strong standard library is essential, and we've been working very hard over the last months to set explicit conventions, revamp APIs, and generally get Rust's standard library into a 1.0 state.
You shouldn't read much into the "experimental" tags right now; they're part of an API stability tracking system that is heavily in flux at the moment. In particular, for most of the modules you mentioned, only the module name itself has not been marked "stable", while the contents have been. For example, if you look at the Vec type itself:
you'll see that most methods are unstable or stable, rather than experimental. (The "experimental" status is currently the default, to help us keep track of API "stabilization" progress.)
This stability tracking is part of our general story about API compatibility; you can read more at
These markers won't take on their full meaning until we ship 1.0.
For 1.0, all of the APIs you mentioned will be "stable", and indeed most of their internals already are. Over the next few weeks, we'll be updating stability markers to reflect the API scrutiny and design work that we've been doing.
More generally, for all but a couple of the modules listed at
we have already performed a detailed scrutiny of the API, and have either landed changes needed for 1.0 or will do so very soon. Most truly experimental APIs have already been deprecated or removed.
For IO in particular, you can read a very recent RFC that proposes the final set of API changes for 1.0:
1) in general, Rust 1.0 is about language stability. This means that any library not of quality will be pulled out.
2) this means Rust will have a small standard library. But we already have Cargo. So you end up getting lots of packages, they just aren't built-in: they require adding one line to a config file instead.
3) iterator invalidation is a compile-time error in Rust. Rust relies on Iterators quite heavily.
I worry about that too. In addition, I haven't seen any confidence that HKT can be implemented in a backwards compatible way, and that is one of the most-requested features (along with Async IO). I would hate to have to wait around for a 2.0 for this stuff.
The team has actually been pretty careful to design in a HKT-compatible way. While there isn't a fully-fleshed out design yet for Rust, there are designs in other languages, and we've looked at their pitfalls to make sure we can avoid them.
Can you elaborate? A panic is a panic, and it will either unwind or abort if unwinding is interrupted with a second panic. There's nothing silent about that.
If you do this:
spawn(proc() { panic!("hmmph"); });
Then "nothing" happens. The default, correct, behaviour for "unhandled" errors is to abort the program. Rust apparently thinks it doesn't need this, because "poisoning". Which is a great thing, except it still doesn't excuse silently swallowing errors. It's one way to end up with a half-dead process. And considering Rust's overall strict handling of errors, it seems bizarrely counter-intuitive to go with the silent route here.
The .NET team went through the exact same process and ended up taking a breaking change in v2 to fix their handling of background exceptions (so now they crash the app unless you opt-in to handling them).
Ah, I see, I wasn't aware that a panic in a child task no longer panicked the parent task. That was surely the case in older versions of Rust, but I have no idea when it changed. Thanks for the RFC!
Well it can't just panic the "parent" - the parent is an arbitrary thread running so you'd have to find a way to inject faults into the parent. I'm guessing with Rust's older task system, you could handle it there, since you owned the scheduler?
With the new, threads-only system, I think the only way out is abort.
It doesn't follow that it's not branching a bit early. Especially since the language five years ago bore little resemblance to what it is currently. I understand the justification that if Rust waits too long it will become irrelevant, but there are certainly plenty of reasons to think it should hold off. For example, the macro system is going to need to be completely rewritten after 1.0.
> I understand the justification that if Rust waits too long it will become irrelevant, but there are certainly plenty of reasons to think it should hold off. For example, the macro system is going to need to be completely rewritten after 1.0.
There are many, many people who have been waiting literally years for Rust to stabilize to start using it. (Just look at the comments in any other HN thread about Rust.)
Rough edges in the macro system (which can be fixed post-1.0 and doesn't need to be completely rewritten) and error handling changes aren't enough reason to force them to keep waiting. It would take years to get every single feature in that people want and it'd be doing a disservice to our community to force that huge group of people to wait even longer. And I feel it too—as someone who uses Rust every day, the issues that keep coming up are minor annoyances compared to the major pain caused by lack of stability.
https://github.com/rust-lang/rfcs/pull/243