Option Monads in Rust
hoverbear.org
hoverbear.org
Currently my #1 and #2 languages are Haskell and C. I will definitely jump on the Rust bandwagon with the 1.0 release -- maybe sooner, we'll see.
One painful point i've seen so far is pointer semantics, but it seems pretty much like the only one, and after having read a slidedeck explaining the reason behind them i'm feeling even more excited.
The idea of Rust memory-safety is to totally forbid two simultaneous mutable lifetime whatever the code path could be.
Whatever the code path could be. Do you begin to realize how strong this assumption is ?
Yes, the rust compiler is smart. But sometimes, it's not enough and you have to do the lifetime tracking job by yourself. It's awfully hard (even without recursion). But you know what's even harder ? When all the common code patterns you are aware of for a given task doesn't provide enough warranty for the rust compiler. You know what you want to do. It's indeed very simple (can be an inner swap in some sort of algorithm) but you can't find a way to warranty the compiler that all should be good. Because in its own sort of bizarre logic, it makes sense. Then you're trapped. And you are waiting on #rust@irc.mozilla.org an hint from the bunch of truly marvelous guys who pass their time here.
TLDR: Rust can't express 100% what others languages do (ML or C based). And in the portion where it can, you have to do a lot of gymnastic to make it compile.
And our Memcached, you'll ask ? Well. After 6 days of work, we finally make it compile, and it worked quite fine. But then we got some mysterious tasking errors when reaching 20 clients and gave up at this point.
But even if you are familiar with the patterns we used and can guess the architecture, there will be a lot of unexplained code. I don't think you'll be able to extract a lot of knowledge of it.
Anyway, you should look for io_thread.rs (I think it's the most interesting file). Look how the scope are defined (especially the use and abuse of "{}"), the alternation of &self and &mut self and random methods put out of the implementation cause rustc would accept anything when they where inside.
You can also look at the mixed use of String and str which introduce a lot of chaos. And, last but not least, bonus points if you succeed to understand the why and the how of the explicits lifetimes we had to put.
A memcached client lib implementing both the binary protocol and text protocol, plus sharding to multiple servers. This is simpler in scope. https://github.com/williamw520/rustymem
A gzip implementation and deflate implementation. This is more complicate in scope. https://github.com/williamw520/rustyzip
Oh, nice, I hadn't seen this before. That looks really elegant and idiomatic :)
I definitely asked for a Rust review from the C POV, but I feel like they do a good job marketing the idea that it's not supposed to express 100% of what other languages do.
Or perhaps it's just further from reasonably expressive than I had believed.
You can express everything that other languages can do—we've written a compiler and a Web browser in it, remember. Of course there is a learning curve, but once you've learned the rules I've found that, by and large, people don't have too many problems, and the time saved by not having to track down segfaults in a debugger is a huge productivity booster.
There is unsafe in Rust. There you can go low level as your heart desires.
"Web browser" is an exaggeration for an experimental render engine in an early stage.
You can only take mutability lifetimes if you are alone. That mean for example that you can't insert in a tree without going mad. Yeah, you have to search (non-mut operation), insert (mut operation), and balance your tree (mut operation). Not possible.
How do you do it in the official B-Tree implementation ? You clone all the branch at each insertion[1] so that the insertion don't compete with the search. Oh well. Hopefully it's a B-Tree so it won't be too deep. But for a performance oriented language, it's make me cry a little.
[1]: a) https://github.com/rust-lang/rust/blob/master/src/libcollect... b) https://github.com/rust-lang/rust/blob/master/src/libcollect...
Please understand that by no mean I want to discredit your work. Rust is a great thing, and will be amazing. However, a random programmer alone in the jungle with only the rustc and the doc will had a hard time to express whatever he likes. It will get better, I trust you. Just don't let the hype grow too fast as Rust is not ready yet, not only in features but also in accessibility.
Of course, this is presumably a pedagogical problem (e.g. the language grows and changes so tutorials go out of date, including the official one, and the language is new, so there's not a huge number of examples anyway), but I think it's a little unfair to say that Rust can't express a lot of things, when it's mostly just a question of practice. That is, Rust's semantics are fairly different to a managed language like an ML or a "do whatever you like" one like C, and so it takes effort to really 'get' it, just like, for many people, it takes effort to 'get' Haskell.
> But then we got some mysterious tasking errors when reaching 20 clients and gave up at this point.
Your code has a lot of `unwrap` and `fail!`, it's generally recommended to use a more composable error handling technique (like returning Option and Result) or otherwise any error (either in the input, or a bug in your application) will lead to these "mysterious" errors, and especially avoiding `unwrap`.
edit: Not Monad<M<T>>, but just Monad<M> with the facility to use M<T> in the body of the trait. I don't have a clear idea of how references, lifetimes, etc. would pan out in this case.
That's unfortunate. :( Which version will have a full backwards compatibility guarantee?
Freezing the entire standard library in one fell swoop would freeze design mistakes, and delaying 1.0 until we're sure about every last symbol in every last library would delay 1.0 for no reason. I'm confident in our approach.
For example, unless IO is stable no serious program can be written. But IO would benefit from monads and HKT.
The code computes sqrt(-1 * (log(-1 * (20 * 2)))^2).
The potentially failing functions are log and sqrt. Failure is determined by checking if the result is a "normal" floating point value. But 'double' and 'square' may also produce non-normal values, e.g. squaring a large float may produce infinity. So why doesn't double() return an Option as well?
Furthermore, checking if a value is normal is not correct. If x slightly exceeds 1, then log x is a small positive value, i.e. a denormal, and therefore log() will incorrectly report failure.
Floating point arithmetic already has a serviceable "Option" type in its non-finite values (NaN and infinities), so the best way to write this code is the naive way:
let result: f64 = sqrt(pow(-1 * (log(-1 * (20 * 2))), 2));
let success: bool = is_finite(result);
No it won't, for the neighboring values of 1, which are 1-2^-53 and 1+2^-52, log will return approximately -2^-53 and 2^-52, which are not denormal.
with those in place, some really beautiful api technology becomes possible to port to Rust. I really appreciate that the core dev team has those partly underway, and partly planned for the post 1.0 stuff.
Given Java 8 and Optional, I recently used Either to replace 80 something lines of crazy interconnected try-catch-bullstuff in a server with 4 straight either-based lines, "do this, in good case try this, in good case try that, ... and finally send a good result to the client, or log a bad result and send the error to the client".
Had to rework and introduce some methods to make this work, of course, but those methods turned into try { foo = doWork(); return Either.goodResult( foo ); } catch ( Exception e ) { return Either.badResult(e); }. It's structured, and neat, and readable. :)
Though I don't know what the reasoning is on the naming of that type; for all I know it might make a lot of sense within the context of rust and be very much in line with the conventions of the rest of the standard library.
Rust used also to have an isomorphic type called Either, but error handling was found to be by far the most common use of these two types, so it was decided to standardise on a single one with obvious variant names (i.e. it's clear with Ok/Err which one is the error and which one is the real value, while Either requires getting the Left/Right convention right).
Edit: looks like rust's and_then() is similar to bind, so it's reasonable to call this monadic.
I initially thought the same thing, due to the use of `map` as something of an equivalent of the Haskell `bind` (>>=) operator. It reminded me of just using `fmap`.
But no, they are using it as a full monad. Consider the type of fmap:
Functor f => (a -> b) -> f a -> f b
Monad m => m a -> (a -> m b) -> m b
So don't let Rust's use of `map` deceive you -- it seems to be the same as (>>=).
That being said, I'd be interested to see Rust introduce something akin to do-notation syntactic sugar.
=======
EDIT: I am wrong, the `map` function is basically equivalent to `fmap` and it seems the `and_then` function is similar to (>>=). See user tel and user wyager's responses.
fn some_on_even(val: int) -> Option<int> {
match val {
// Matches an even number.
even if even % 2 == 0 => Some(even),
// Matches anything else.
_ => None
}
}
That seems absurdly confusing to me.
match val {
even => {
if even % 2 == 0 {
Some(even)
} else {
None
}
}
_ => None
}
if val % 2 == 0 {
Some(val)
} else {
None
}
match () {
_ if val % 2 == 0 => Some(val),
_ => None
}
let val: Option<int> = ...;
match val {
// even
Some(x) if x % 2 == 0 => Some(x),
// odd
Some(x) => Some(x + 1),
None => None
}
match value {
pattern(variable-binding) [if guard-condition] => matched result
...
}
The wildcard _ matches whatever left, which is the non-even values of val.
15 points is a fair bit of attention, but 0 comments not so much, so I think this repost is ok.
I could swear I have seen that even when the prior submission was a few years old.
Does the algorithm to decide that take comment activity into account?