When most practitioners think of "typechecking", they typically think about proving properties about programs statically. This project seems equivalent to adding a runtime check at each call site to ensure the arguments and return values are the correct type.
This is certainly useful sometimes, as it gives programs the desirable "fail fast" property. But it isn't "typechecking" as most engineers understand it. Or at least, it should be clarified that this is run-time typechecking. As such, it negatively impacts runtime performance, unlike compile-time typechecking.
This project also seems to miss the primary opportunity of run-time type checking: checking properties that are difficult to prove statically! For example, checking that a number is even, that a string is lowercase, that an index is within the bounds of an array, etc. These exotic types require a dependent type system to be checked statically, but in a dynamic environment they are trivial to verify.
Two suggestions for improvement: 1) add "sum" types (i.e., discriminated unions), and 2) let the user define their own types via lambdas, such as PrimeNumber.
> When most practitioners think of "typechecking", they typically think about proving properties about programs statically. This project seems equivalent to adding a runtime check at each call site to ensure the arguments and return values are the correct type.
It basically seems equivalent (albeit somewhat nicer in a few ways, ie. scalar types are supported) to PHP's "type hinting"[0]. I will say that it is useful, in combination with unit testing, as you can pick up on problems with things slightly easier than hoping for the best in production and tailing error logs, but proper static type analysis is more useful.
editor support for php type hinting and phpdoc method comments goes a long way. intellij can do significant refactors and and searching. you'll get real time feedback for violating types. it's not quite the same as a explicit compile step but pretty close
Komodo IDE helps with that, but I personally switched to HHVM and have a simple bash script that swaps the `<?php` for `<?hh` and back, so that `hh_client` can statically tell me about problems, even when I'm not using any of the Hack features. Works a treat, to be honest!
One thing we've applied on a Python(2, unfortunately, so can't go deep on the static type analysis) is enable this sort of type hinting when running our app in debug or test mode. Obviously can't catch everything, but it catches some of the sillier bugs.
For example, in our application code, we'll use a warning for some type coercion for example
if isinstance(obj,Bar):
warnings.warn('Using a Bar as a Foo!')
In production, this will just print a message to the log. So this is useful for things like phasing out old behavior, where we try to stop using it but don't want hard failure if it's used for now.
But in our tests, we do want hard failure (so we can phase out the behavior as much as possible before removing the functionality). So we do something like the following:
if TEST:
warnings.filterwarning('error','Using a Bar as a Foo!',RuntimeWarning)
Which basically says "if I receive a warning matching the regex given, then raise an error instead of just logging". So if I hit that code path during tests it will fail the test.
It's very useful, even if warnings is a bit too much about regex matching for my tastes
It's not pedantic at all. It's a very useful distinction. Personally I thought this was static checking and got the wrong idea from the submission title. I can see some other comments here have made a similar mistake.
Perhaps it's true that many programmers don't realize the fundamental connection between typechecking and theorem proving (i.e., that they are the same thing). But that wasn't meant to be the point. The point is that when most programmers hear the word "typechecking", they think of compile-time typechecking. This word is usually used in a static context.
I'm not saying that types or typechecking are inherently static concepts. I'm also not saying that this library should do static typechecking—that would be an absurd demand. I only meant that the wording is misleading.
Even more misleading is author's use of the term "gradual type checking", which has a well-understood meaning: the ability to add static checks to an otherwise dynamically-typed program.
> The point is that when most programmers hear the word "typechecking", they think of compile-time typechecking. This word is usually used in a static context.
Most programmers don't think "dynamic type checking" is a misnomer, while it is true that "type checking" itself leans towards a static connotation.
I've seen gradual type checking used both ways in the literature, actually. Wiki has the term defined for dual phase:
> Gradual typing is a type system in which variables may be typed either at compile-time (which is static typing) or at run-time (which is dynamic typing), allowing software developers to choose either type paradigm as appropriate, from within a single language.
I'm sure this is just poor writing (as Siek defines it, it should be from dynamic to static), but there is enough confusion here where you might bother using the term for adding stronger dynamic type checks to an otherwise less dynamically typed language (if you admit dynamic types as types, of course). Also, a dynamically checked type signature is the first step to a statically typed one (as long as your types remain weak enough that static typing is achievable in the future). So in that sense, it is "gradual" typing, just from a meta perspective :)
> This gem brings you advantage of type without changing existing code's behavior.
I'm fairly sure it changes the performance characteristics of code it is applied on. I'd recommend adding benchmarks to the README so prospective users might be aware of this beforehand.
As someone working on a Ruby AOT compiler (whether it'll ever be finished is another matter...), I'd love to get some spec'ed out extensions for things like this with semantics that explicitly allows implementations to recognise the extension and decide to ignore the default implementation and instead provide it's own version that could e.g. do compile time checks instead of runtime checks, or hoist checks, whenever possible. Or use it for optimisations (e.g. by hoisting out and eliminating other checks and caching method lookups)
Ruby is flexible enough to allow quite a bit of experimentation with stuff like that. Especially because classes can be re-opened, so if worried about performance, it'd be easy enough to allow type annotations (or whatever other extensions people come up with) to live in separate files if people want them to, and conditionally include them only as needed/wanted.
Personally I don't get contracts when they aren't used for static verification.
It's much easier to use defensive coding techniques at runtime without having yet another framework in there that anyone who wants to read the code has to learn. Typing a few lines less is not really a benefit when it basically increases code complexity.
The Contracts gem is very easy to use. I introduced it at work recently, and everyone got the hang of it in about 20 minutes.
A lot of languages (Java, C#, Go, etc.) have very inexpressive type systems, and that tends to make many dynamic language developers wary of them. Contracts has a very expressive type system though. It understands duck typing, adhoc union types, and a bunch of other stuff that makes it a great fit for a dynamic language.
Plus, we've disabled it in production, so we get coverage in test/development mode, but no performance penalty in prod. I look at it as an extended form of testing. Type contracts are very quick and easy to add, and they help tests catch additional errors. They're also useful in code reviews. Sometimes I'll have to search for 5 or 10 minutes to figure out what kind of thing is getting passed in, where the contract makes it perfectly clear.
The best part is that the contracts are optional. If some method takes wildly differing arguments and it's going to be a pain to give it a contract, then don't.
This is the best tool I've added to my Ruby arsenal in some time. I strongly recommend taking it for a spin.
What I meant was that contracts in a language like C# give me the ability to statically verify the program at compile time. It is not a "nicer way to add conditional testing logic" but a way to formally prove correctness in my programs.
I find that putting the method signature at the end of the method definition can become unreadable pretty fast. Also, the fact that it doesn't offer any performance improvement (most probably, this will actually degrade performance) makes me see this as a cool trick, but not really recommended in production.
I wonder if it would be possible to configure this gem with "environments" in the same way that Rails applications have environments. Then, in development/test/CI-like environments, very strict checking could be applied based on the specified types. Otherwise, in production-like environments where "performance" may be more important, the type-checking could be looser or simply pass-through.
Really cool idea! I like that you can specify a `#respond_to?` constraint instead of a class. Not sure if OP is the author, but here's some feedback:
* It would be better if this didn't pollute the global namespace by defining `#typesig` in `Module` [0] -- perhaps consider refactoring that method into a module which the user may extend. Doing so would also get you out of needing to define `Module#prepend` for older versions of Ruby.
* Perhaps allow the user to enable/disable type checking at a global/class level. For example, then users could only enable type checking during specs if they wanted.
* Instead of using class-level variables, try using class level instance variables. They have less odd behavior when dealing with subclasses [1].
The OP is eduardordm, which can be found right underneath the article link. If you look on his HN profile, his GitHub name is also eduardordm.
The repo author is gogotanaka.
So, no. He's not. And the repo author is consequently unlikely to ever see your comments.
Not only are these simple checks to make, but it's extremely bad practice to scatter commentary about a project everywhere across the internet; people who only know about this project through GitHub would not see your comments either.
What are you suggesting, that there never be any discussion of code that lives in a github repo on HN, but instead all discussion take place... in the GH issues?
I've been writing Ruby for a few years on a number of production applications.
Recently I've had to pickup Hack for work, and if there's one thing I really like about it is the type hinting. The best part is that it helps you handle nullable types (not sure if it's done here).
When I switch back to Ruby from Hack, I find it harder to reason about my program.
It's definitely me being to naive, however I would type all the expression, in either the source code (into comment) or in a separate file, and all the statement.
Now I do have a table of symbols and their type.
At this point its just a matter of using the AST to being sure that everything is correctly typed...
IMHO the state of static type checking/code analysis in ruby is still deplorable and this (well known, rather trivial) approach won't ameliorate the situation. Even javascript has more to offer in this respect. Who would have suspected that 5 years ago.
Since I still like ruby's syntax, my hopes are that crystal (http://crystal-lang.org/) will one day become more mainstream (and maybe be adapted to some extent in mainstream ruby).
Is there an error in the section named "Typed method can coexist with non-typed method"? There's a line that has `typesig :sum` but sum is never defined.
This is certainly useful sometimes, as it gives programs the desirable "fail fast" property. But it isn't "typechecking" as most engineers understand it. Or at least, it should be clarified that this is run-time typechecking. As such, it negatively impacts runtime performance, unlike compile-time typechecking.
This project also seems to miss the primary opportunity of run-time type checking: checking properties that are difficult to prove statically! For example, checking that a number is even, that a string is lowercase, that an index is within the bounds of an array, etc. These exotic types require a dependent type system to be checked statically, but in a dynamic environment they are trivial to verify.
Two suggestions for improvement: 1) add "sum" types (i.e., discriminated unions), and 2) let the user define their own types via lambdas, such as PrimeNumber.