[adamnemecek]

Generally, statically typed languages are less bug prone than dynamically typed languages. Dynamically typed languages are OK for smaller projects but as projects grow in size, you have to write a lot of tests to be sure that everything works.

Also, projects written in statically typed languages are easier to read and navigate for humans and it's also easier to write static analysis tools for them.

[dragonwriter]

> Dynamically typed languages are OK for smaller projects but as projects grow in size, you have to write a lot of tests to be sure that everything works.

Or build your large systems as compositions of smaller systems. If dynamically typed languages are really okay for smaller systmes, than avoiding bad architecture of large tightly-coupled systems in favor of loosely coupled compositions of smaller subsystems means that it is also okay for large systems.

> Also, projects written in statically typed languages are easier to read and navigate for humans

This is not my experience.

> and it's also easier to write static analysis tools for them.

Well, yes, since you have to write a static analysis tool for a statically typed language (since the compiler must include such a tool), its not at all surprising that the structure of statically-typed languages is always designed specifically to serve static analysis tools.

IME, that's why they've historically been harder to read and navigate for humans, though some exceptional modern statically typed languages have largely closed that gap (but not reversed it, IMO.)

[imanaccount]

>If dynamically typed languages are really okay for smaller systmes, than avoiding bad architecture of large tightly-coupled systems in favor of loosely coupled compositions of smaller subsystems means that it is also okay for large systems.

It doesn't mean that at all. You might as well say "if shoes are good enough for going to the corner store, then multiple pairs of shoes are good enough for going to the moon". The notion that large systems are no more complex than small systems if you simply make the large system out of small systems is not supported by anything I can find.

>IME, that's why they've historically been harder to read and navigate for humans

That doesn't make any sense at all. How is having information written down instead of having to constantly remember and/or deduce it an impediment to reading?

[dragonwriter]

> It doesn't mean that at all. You might as well say "if shoes are good enough for going to the corner store, then multiple pairs of shoes are good enough for going to the moon".

A large system can be decomposed into a set of loosely-coupled smaller systems (indeed, that's often a preferred architecture for a variety of reasons independent of whether the language is static or dynamic).

A trip to the moon cannot be decomposed into a series of walks equivalent to a walk to the corner store.

Therefore, no, your analogy is not valid.

> That doesn't make any sense at all. How is having information written down instead of having to constantly remember and/or deduce it an impediment to reading?

Visual (and mental, really) noise, mostly -- there is a reason why clear, effective, easy-to-read writing is often not writing that avoids all potential ambiguities and is sure to define as much as possible to avoid people having to deduce things. Making everything explicit is generally generally in tension with readability, not an aid to it.

[fithisux]

Dynamically typed languages are good for very small programs. Typed Lua is going to be a big innovation.

[srean]

If there are consequential performance gains to be had then even better. An interesting case is Dart. The implementers have a strong dynamic typing bias and wanted to keep Dart as dynamically typed as possible. Their position was we dont need any static typing for speed. They have since relented.

You realize that you just planted a good solid troll magnet of the "I never make any type errors" variety. Some people get annoyed by the notion that you can enlist the compiler (using succinct syntax) to write and validate tests that they ought to be writing. They would rather write the tedious tests themselves.

@spankalee They had a blogpost along those lines, sadly not bookmarked, but you might be able to find it.

[spankalee]

What makes you say that Dart "relented" on not using type annotations in the runtime?

The VM basically throws the type annotations away. In production mode, the type annotations can be completely wrong and your program will still function, and still be just as fast.

[fit2rule]

There isn't really much evidence of this out there, other than groupthink, in my opinion.

What makes a language 'okay for smaller projects' but suddenly unsuitable when it gets larger? Having a minimal set of types doesn't necessarily mean you're going to have 'more bugs as the codebase grows' - the only place it really makes sense to say that dynamically-typed languages 'break' is when they don't have the types that the application needs.

In Lua's case, the only thing I can think of it missing is native floating-point support, and that's really it. Everything else: you can get there with nil, number, string, function, CFunction, userdata, and table. (But CFunction and userdata should be enough for anyone! :)

[ufo]

> What makes a language 'okay for smaller projects' but suddenly unsuitable when it gets larger

As projects get larger the benefits of tooling (automated testing, automated type checking, efficient compilation, etc) become more relevant and the costs of using the tools become less relevant compared to the costs of writing the program itself.

> the only place it really makes sense to say that dynamically-typed languages 'break' is when they don't have the types that the application needs

Actually, I don't think any dynamic languages used in practice actually do this right now. Most dynamic languages only check base types (numbers, null, "object") but fail to check for higher order types (functions and objects). For example if you code

    MyClass x = new OtherClass();
    x.foo();

wouldn't it be better to get an error in the first line than an error on the second line? Specially when you consider that the method call can be delayed and only happen much latter, in some other function and without the true guilty line even showing up in the stack trace.

[chriswarbo]

> In Lua's case, the only thing I can think of it missing is native floating-point support, and that's really it. Everything else: you can get there with nil, number, string, function, CFunction, userdata, and table. (But CFunction and userdata should be enough for anyone! :)

I think you're missing the point; static typing doesn't let you write more programs, it specifically lets you write fewer programs, by ruling out ill-typed programs. If I see a Haskell variable of type "Int -> String" then I can be quite confident that it's a function turning Ints into Strings; in, say, Python I can't be confident of much without analysing the entire codebase.

[adamnemecek]

If nothing else, it kind of follows from Curry-Howard correspondence.

What makes it okay? This Python program will not throw an exception when ran

  def f(x):
    return x

  if False:
    f(1,2)

even though the types are wrong. Are you 100% certain that you will never make a bug like this even when there are 10^13 conditions and the call stack is deeper than the Mariana trench? Note that that this is not the only type of bugs that static typing prevents.

[Jtsummers]

> What makes a language 'okay for smaller projects' but suddenly unsuitable when it gets larger?

One and one makes two. Two and two makes four. Four multiplied by eighteen is seventy-two.

Compare to:

  1+1=2
  2+2=4
  4*18=72

Notation, aka language, has scaleability as a feature, whether the designers intended it or not. Dynamically typed languages are nice, I love them for prototyping, and some of them, I think, are suitable for large scale applications. But not all of them. The notations made available in some languages just make large scale development easier or harder (for various reasons). Consider using maps/json objects to produce data constructs, this is a sort of duck typing unless you add some sort of schema checker to it. At which point it's either a runtime error or you're running your code through a static analysis tool. Why not use an actual type system for this that the compiler checks at compile time?

I feel like a broken record this past week.

Dynamic languages are good, I can't imagine doing the sort of exploratory/prototyping stuff I've done in them with typical statically typed languages. I wouldn't even object to using erlang for a large scale project (actually, I'd love to do this), it's got features in both syntax (bit syntax, I love it) and semantics (concurrency) that make it ideal for certain problem domains. It's got dialyzer which gets back a lot of the static analysis that statically typed languages offer. On the other hand, javascript offers duck typing, weak typing, dynamic typing, it's fine for certain applications (and essential for anything living in a web browser these days), but it's type system holds it back for large scale work. Experienced programmers may be able to get passed it with relatively few errors, but it's still going to suffer huge performance issues because there's only so much that can be known about the values passing through a block of code. Does `a + b` mean we're adding two numbers? A number and a string? two strings? The result changes depending on those circumstances and the interpreter/compiler just doesn't know enough to optimize it significantly. Similarly, it doesn't know that one is an error right away, the actual error may start here, but only appear in some function that's a child/parent/cousin in the call tree, obfuscating the cause and creating a great deal of work for the developer and tester.

[dottrap]

Lua already has native floating support (by default). The number type is double by default.

Lua 5.3 (work 2 already available) introduces integer subtypes so now you can have both floating point and integers.

[ufo]

Similarly to Javascript, all numbers in Lua are floating point numbers and all number operations are floating point operations. Integers can be represented accurately up to 2^53.

[nly]

Changing the Lua number type from a float to an integer type is something like a 3 line patch.

[fit2rule]

Right. :)

[taeric]

Is there empirical evidence of your first claim? Or any claim, honestly.

[srean]

Not a rhetorical question: How many flight-control codebases are written in a dynamically typed language ? Or for that matter, life support systems deployed in ICU's, really curious.

About the dynamic vs Static typing debate, I think it progresses firmly along Upton Sinclair'ish lines,

"It is difficult to get a man to understand something, when his salary depends upon his not understanding it!"

[taeric]

I don't know the answer, sadly. My sardonic response is to ask how many are in what would be called a modern statically typed system?

That is, I expect the tooling around them is ridiculously high. Probably done with a model based development system. Probably not, however, what is typically seen as static typing in the modern landscape. (i.e. Haskell)

Honestly, I wouldn't be surprised to find some lisp in there, just because. :)

I'm open to being proven wrong.

[srean]

> what would be called a modern statically typed system?

Now we are splitting hairs and looking for the one true scotsman. You asked for quantitative evidence, but the problem is that you have already made up your mind (perhaps only subconciously) and you will find a way to lookaway from the evidence or logic (which you term just an appeal). Lookup birthers for instance.

I for one have'nt heard of any flight control or life support software written in a dynamically typed language, whereas have heard of plenty written in what would qualify more as a statically typed than dynamically typed language. ADA in particular comes to mind. I dont discount that there might exist some written in a dynamic languages. Lets agree that they are quite hard to find. I wont be surprised with prototypes written in dynamic languages though.

> I wouldn't be surprised to find some lisp in there, [...] I'm open to being proven wrong.

Now that would be the Russel's teapot fallacy. I am not ruling it out, but for such assertions, the burden, I think, lies on the one making the assertion. You handed out an epistemic sucker punch there. The proof that you desire requires presenting the entire set of Lisp programs, or the entire set of flight control programs, both are quite ridiculous. BTW I do love lisp a whole freaking lot.

Curious about this: Say you have to be (god forbid) irradiated. There are two pieces of software written by two groups, one that uses statically typed, another that uses dynamically typed and this is all the information you have. Which one would you pick ? There is no right or wrong answer.

I will pick the one written by the group that uses statically typed language because of their cultural obsession with correctness. That is just a Bayesian prior, there are exceptions of course. Sqlite is indeed afflicted by the same obsession.

EDIT: @taeric

> If they are in a static language, then this should be an easy debate for you. Merely give me some evidence and we are done.

Ah! then lookup ADA's resume

> I'd pick the one that has the better track record and has been proven with more tests.

You are dodging the (silly) question sir :) I said thats the only thing you know. Anyways wasnt an important question and here's wishing that you never need to get irradiated with any such thing, seriously.

[taeric]

:( Why are we replying without going down the chain? I almost didn't even see you put more of a response to me.

Regardless, I'm familiar with Ada. I'm also familiar with some fun disasters using software by Ada.[1] Now, do I blame that it was in Ada? No. However, the attitude you are displaying of "if all you know is it was in Ada, then it is safer than the alternatives" seems to be the exact problem that led to that disaster.

If the only thing I know of two irradiation devices is one was statically typed and the other wasn't, I'd likely pass on both. Or I'd like to know how much radiation each is capable of outputting. Consider, x-ray machines far predate what we realize as programmable computers. And much more goes into the safety of the devices than just the language used.

[1] http://www.adapower.com/index.php?Command=Class&ClassID=FAQ&...

[srean]

> :( Why are we replying without going down the chain?

I dislike deeply nested, indentation cramped and tangential threads, so sometimes if I dont have anything that important to say I inline it. More so when everyone has left the building.

> I'm also familiar with some fun disasters using software by Ada.

I dont think we were ever discussing whether using Ada or some other statically type checked language immunizes a project against all other peripheral errors. That would be a supremely ridiculous and stupid claim to make.

Typechecking proves either that the code is type error free or that the typechecker has a bug. In the case of the latter, you try and fix it and the benefit percolates to every software verified by it. Testing proves that it passes only those specific test cases and only if the programmer tested it at all.

BTW the very first line says the fault had nothing to do with Ada. The source of the error was beyond the scope of typechecking.

To backtrack, our discussion was on whether there are notable examples of software written in dynamically typed languages where the cost of error is high.

Runtime error while spaceshuttle descends backs to earth, no thank you, on a fly-by-wire unit, when the pilot is breaking out of a loop, again no thank you. There are times when, no pun intended, you just cannot bail out and wash your hands off the situation, without terrible consequences. In such cases I would rather have the error surfaced in an environment where I can control the consequences, meaning early, and as exhaustively as possible.

Most importantly typechecking does not preclude/replace testing, it complements it.

Everything cannot be statically verified or insured against, but that does not mean that just because we cannot verify everything statically, we should verify nothing statically.

BTW Typical x-ray machines do not have the power to kill you in a single exposure, radiotherapy units absolutely can, example Therac.

  The accidents occurred when the high-power electron beam
  was activated instead of the intended low power beam, and
  without the beam spreader plate rotated into place.
  Previous models had hardware interlocks in place to
  prevent this, but Therac-25 had removed them, depending
  instead on software interlocks for safety. The software
  interlock could fail due to a race condition.

I would rather have it proved (exhaustively, rather than with typical tests which would be anecdotaly) that such errors cannot happen.

[taeric]

I'm not trying to play the scotsman's game. Though, I can confess that that is the direction I am somewhat steering this. Not my main goal. My aim is more against the claims of the parent post. Note that there is a big difference between static tooling and statically typed languages. I have high regard for the both. And I've seen more evidence of the former in older, less statically typed, languages than I have the latter.

So, part of my point is I haven't ever "heard" what language those systems are written in, period. It isn't that I have heard they are in dynamic or static languages. I flat out never hear. If they are in a static language, then this should be an easy debate for you. Merely give me some evidence and we are done.

I do know that the older languaged programs I have seen are typically not in what one would call a statically typed language. There is plenty of static tooling on them, but to pretend that there is statically "typed" tooling around any software that ran something such as the space program feels like it is reaching. Heavily.

So, to your silly question. I'd pick the one that has the better track record and has been proven with more tests. I don't care if they are dynamic tests, statically typed assertions, statically analyzed assertions on a dynamically typed codebase, whatever. There are too many tools to care. Because I can hazard a guess that all of the things that have irradiated me in my years have not been by the strongest statically typed languages out there. :)

[adamnemecek]

What sort of evidence are you looking for? E.g. http://www.infoq.com/presentations/Types-Tests ?

But googling something like static vs dynamic typing will give you enough results to choose from. Or go to your local Haskell user group meetup :-).

[djur]

Empirical evidence. Studies indicating that software developed in statically-typed languages has fewer defects. Research into the speed of development and the ability to accommodate major changes in requirements and functionality would be useful as well.

Last I looked, there isn't a great deal of literature on this, and what there is is fairly old (dealing with '80s-'90s languages) and inconclusive.

[adamnemecek]

Is the presentation not good enough?

> Research into the speed of development and the ability to accommodate major changes in requirements and functionality would be useful as well.

You kind of slipped those in, we were talking about correctness.

[taeric]

No we weren't. We were talking about "being less bug prone" and "easier to read."

Though, I'm not entirely sure that really changes much. :(

Look, I'm not against static typing. I confess to being somewhat in love with lisp at the moment. And, I'm trying to learn MIXAL for some unholy reason. At the same time, I try to type my systems as well as I can. I am far from convinced that typing will be the way to go.

Especially if you consider auxiliary tools. Having used Coverity some, it is down right impressive the stability you can bring to a C codebase with proper discipline and the appropriate tooling. Compared to using something such as a Scala codebase, where right now your only recourse is the compiler.

That is, static analysis does not begin and end with the compiler. Sadly, in moving to "newer" statically typed languages, you through out many of the tools that currently exist in the older languages.

And this is not just for "correctness". Having finally added "-march=native" to my flags for a build of software, I'm literally amazed at how well optimized a compile can be versus just using "-O3". Optimizations in the modern world are quite amazing without necessarily needing "better types."

[adamnemecek]

Remind me, what's the typing discipline of the languages supported by Coverity?

[djur]

The presentation was interesting, but if there were any studies cited I didn't see them. I'm looking for deductive research into the proposal that using a modern statically-typed language results in a better outcome than a modern dynamically-typed one.

The related research I mentioned is useful because if development in a given language improves correctness but also increases development time by an order of magnitude it might still be more effective to develop software in the "worse" language and spend the additional available time mitigating the risk or building additional features.

[taeric]

No, googling those results doesn't really work. :) You get a ton of dogma with a lot of logical arguments. All of which are very appealing. Appealing arguments worry me, though.

What I have not seen, is evidence that any of your claims are true. I have seen easy to read programs in both dynamic and static languages. So far, I've seen very little correlation between the two as to which is desirable. The static languages are typically a bear on some algorithms because you have to provide so much more to the compiler for it to trust you. This does feel like it would lead to less bugs, but it goes against the "readable" and possibly the "more productive" ideas.

[adamnemecek]

How do you determine the return value of a C function? You look at the function prototype. How do you determine the return value of a Python function? You have to read the whole thing. If you are working with large code bases, reading the whole thing is not really possible.

[taeric]

You are just trying to make an appealing argument. This literally adds nothing to the debate. A simple survey of open source software would be more convincing. And even that wouldn't be definitive.

[adamnemecek]

Where's the appeal in my argument? It was an example.