As you say, with ADT's and null safety, you can be a lot more judicious about testing to get to the same level of confidence.

When working with Rust or Swift, I'm honestly usually getting away with some integration tests, and whichever unit tests came about organically when a particular section was particularly well suited to TDD. But "code coverage" is a much less pressing concern.

Even when using modern stuff correctly on a mature codebase fuzz testing will find bugs. So if you don't fuzz test and someone evil has your binary then they'll fuzz test for you...

In contrast, a compiles-to-JS language can produce code that behaves unexpectedly in its non-functional requirements, and in the worst cases crashes where it shouldn't. That will be less of an issue as compilers get better, and WASM should help a lot, but in the current ecosystem TypeScript produces more predictably good results.

It's a very complex type system though.

Anybody who last looked at TypeScript a few years ago and dismissed it needs to have another look.

I should also note, I've attempted OOP in TS and it wasn't pretty. It is much more suited to a functional style IME, although I will concede I have seen some open source projects that use it in a Java-esque OOP style presumably to good effect. Personally it wasn't my cup of tea.

If you are curious about this, you might be interested in checking out Haxe [0]. It has static typing, type inference, pattern matching [1] (maybe not on par with ML languages, but still good), and can transpile to multiple targets [2] including JS.

[0]: https://haxe.org [1]: https://haxe.org/manual/lf-pattern-matching.html [2]: https://haxe.org/documentation/introduction/compiler-targets...

[1] https://rescript-lang.org/

[2] https://github.com/melange-re/melange

[3] https://reasonml.github.io/

[4] https://www.purescript.org/

Expand please?

See what I did there?

Testing can't really substitute for types. They're two diametrically opposed approaches to impeding bugs.

In a dynamic language you end up writing a number of tests that would be covered by a static type system and writing a number of tests that would not be covered by that. Going from dynamic to static typing is a replacement for some of the former, and none of the latter.

Not once have I done this. I've written plenty of invariants that would have been covered by a static type system but no tests.

As far as I'm concerned if you had to write an extra test to "cover for a dynamic type system" that test is a bad smell that should be deleted. A test validates that a method raises an exception when being fed a null is pointless if it's just testing a one line assertion that does the same thing. It's a waste of keystrokes. It doesnt catch bugs.

There may be some exceptions to this rule but they would be under very obscure circumstances, I think.

If you have that “pointless” test, you’ll find out about it. If not, you’ll probably only encounter it at runtime with a program that is subtly wrong.

Whether it's a good or a bad decision is another matter and not something the test will help with.

Tests that exist to check for the presence of a single line of code are deadweight.

Failure to sanity-check a null almost never stemmed from someone deleting the non-null assertion intentionally. It stemmed from someone rewiring the control flow in such a way as to bypass the non-null assertion without intending to (such as early-returning because they think they can shortcut the computation, but the shortcut isn't actually valid if the argument in question is null).

The tests have value precisely because they aren't in the regular control flow of the unit under test and will fail when the code flow is changed and the author didn't realize the change modified the expected behavior of the function.

That reason will be one of two things: (1) They replaced it with something they thought would handle the same functionality (along with possibly some other functionality) in a beter way, or (2) The requirement changed.

> That reason will make them remove the test that makes it fail as well.

If it was #2, sure (and in a test-first workflow, they would take out the test first). In case #1, they won’t take out the test, and if their implementation is wrong, the test will let them know.

> Tests that exist to check for the presence of a single line of code are deadweight.

Unless you have a very weird framework, tests don’t check the existence of lines of code, they check the behavior of units. The fact that the first-pass implementation in the same iteration in which the test was added is trivial doesn’t mean the function isn’t part of the contract that needs verified, and doesn’t mean that no one will ever change the intended implementation to a less direct one.

Unit tests that test the contract of a method (if you put in this argument, then you’ll get this return or side effect) are a good idea in both dynamic and static languages.

> There may be some exceptions to this rule but they would be under very obscure circumstances, I think.

Maybe it depends highly on the specific dynamic and static languages under comparison. My main experience with dynamic languages has been with Groovy (all inherited codebases), and it's an annoyingly frequent occurrence that I've had to deal with PROD issues for things that would have been compile errors in Java.

I believe my favorite is when someone used a String for the counter in a for loop, and it worked...until one day the counter had to go past 9. IIRC, Groovy would convert the String to a char, increment it, then convert that into an Integer. That actually works for 0-9, but incrementing "9" in this way gives you ":" which would then get a NumberFormatException in the last conversion.

Do these invariants have some language-level support?

It catches the bug that gets created when someone does a later refactor, replaces the assertion and some other code with a call to another function that they think throws an exception on null as well as doing some computation needed in the current function, but they are wrong.

Tests validating the contract of the unit are useful even when the trivially pass in the first-pass implementation, because implementations evolve, and the thing people will forget to add tests for with changes are the pre-existing, unmodified expected behavior.

(Also, for workflows where tests are written before the first-pass implementation.)

They obviously work best when "combined with test coverage" but then again, I dont consider test coverage to be "optional" in any language. This idea that "if it compiles it works" that some haskellers seem to believe, for instance, is complete horseshit.

If somebody argues that static typing works great because without tests it catches more bugs than dynamic typing does when you also dont write tests, it's notionally true, yes, but theyre telling you more about themselves than they are about types. Namely that they're not in the habit of writing tests.

If you've got enough of the right kind of testing coverage to validate an app written in a statically typed language you will have enough to trigger invariants in a dynamically typed langauge.

I have successfully had invariants usefully triggered in prod and by manual testers in big balls of mud without test coverage (failing fast isnt just for test environments) but I considered that to be a form of emergency programming CPR and damage limitation that isnt a substitute for fixing your testing "diet".

(... I once had a professor who believed that phrase as nearly a point of religion, and his project work was maddening. No documentation on any functions either, because "the types are self-documenting." Then you'd get hit with a 'division' function that takes (number,number) but he put the denominator as the first argument. Ugh.)

Clojure does have spec (https://clojure.org/about/spec) which is pretty cool and declarative. And, of course, your IDE can help you when writing code that obviously violates a spec declaration. But that's not fundamentally different from an IDE parsing the type-hints of Python, JavaScript, or PHP. It is more powerful than the aforementioned type-hints, though.

At the end of the day, it blurs the line between static typing and dynamic typing. IMO it should be considered static typing because it serves that same purpose, works at "write time" to help the programmers, and can probably absolve you of writing your own type-checking tests.

In some sense, it is all just asserts. Even statically typed languages have unsafe type casting, which can trigger an "assert" at runtime.

I don't understand this. This makes a lot of sense when you make complex enterprise products depending on unstable 3rd-party libraries but for personal project this just inflates the time you need to get the thing done. I just code exactly what I need and check if it works and can't imagine dedicating time to writing additional layer of code to automate this checking. Even when I code pure functions which are a lot easier to test.

A solid type system helps a lot in making refactoring a less fearsome endeavour, but a reasonable test coverage on top of static checking makes it even more so.

Also, automatic perf tests that keep track of regressions and improvements are just awesome.

A lot of times designing your interfaces to support mock objects more easily improves them in other ways; if you pass in a file object instead of a filename, for example, not only does it become easier to pass in a mock object that raises an error, but it also becomes easier to pass in a GzipFile object or an HTTPResponse object, and presto, your function now handles compressed data and data stored on a web server.

Also, though, monkeypatching stuff in tests isn't hard in dynamic languages. Here's an example from https://echohack.medium.com/python-unit-testing-injecting-ex...:

    with patch.object(requests, "get") as get_mock:
        with nose.tools.assert_raises(socket.error):
            get_mock.side_effect = socket.error
            inject_exception.call_api("http://example.com")

If your functions have lots of hidden dependencies and side-effects, it’s hard to test.

If you split concerns properly and keep your glue/IO code separated from the decision-making/business-rules/logic, mocking is quite trivial, and there are advantages other than just testability.

Check out this talk to see examples of that in action: https://www.destroyallsoftware.com/talks/boundaries

Tests have advantages, but they also hinder your agility by increasing the costs of the initial version and of making major changes.

Hard to update? Good tests! Well understood problem sphere and product? Good tests! High cost of failure? Good tests! MVP with no actual current users? No TDD. Exploring the product space and update cost is low? No TDD.

I hate tests with a passion, but sometimes I'll write hundreds even for a tiny personal project - when that tiny personal project is going to be at the core of another thing and I want to make absolutely sure that it is working correctly in every way.

Like that time I wrote a process manager for node.js applications because I realized PM2 is a buggy inconsistent mess with race conditions all the way down. Gotta make sure mine works better than that...

    def ok(a, b): assert a == b, (a, b)

    from .util import ok

    ok(Parse(b'.25').do(0, real), (3, ('real', 0.25)))
    ok(Parse(b'1.5').do(0, real), (3, ('real', 1.5)))
    ok(Parse(b'-1.5').do(0, real), (4, ('real', -1.5)))
    ok(Parse(b'+1.5').do(0, real), (4, ('real', +1.5)))

I mean, I do do some manual testing. Sometimes it's easier to bang on a function and decide whether the output looks right than to figure out what the results should be from first principles. But it doesn't take very many repetitions of a manual test of something like the above before it's the manual testing inflating the time needed, not adding another ok() line.

This is especially true for debugging. Adding automated tests is often a faster way to debug than to add print statements or to step through things in a debugger. Sometimes I take the tests out afterwards, but more commonly I leave them in. And with things like Hypothesis and QuickCheck, tests can have a much higher strength-to-weight ratio than simple tests like the above.

It's also a breath of fresh air when I'm trying to understand what some code is supposed to do, or how to call it, and I come across a one-line test like the above.

Speaking of how to call things, sometimes I do test-driven development so that my internal APIs don't suck. And sometimes I don't and wish I had. There are a couple of lines of code in the same project that say this:

    array = Thunk(lambda: drop_ws(b'[') + Any(pdf_obj) + drop_ws(b']'))
    array.xform = lambda d: ('array', d[0][1])

It would have been less work to fix it at design time, the way my friend Aaron taught me: first, write some application code as if the ideal library existed to write it with; then, write the library code that makes it run. And that's just so much easier when that application code is just a unit test.

However, as with most things, I disagree with Robert Martin's point of view on this. I can easily conceive of writing software without TDD, or even without tests. In fact, I do it frequently. Tests are often valuable, and TDD is sometimes valuable, even in weekend personal projects. But they aren't always worth the cost.

This is entirely dependent on the type of application that you're dealing with.

A web application backed by a database, especially CRUD style, acts as little more than a pass through layer. You're getting everything as a string from the browser, storing it in the database with a SQL string, usually transformed by an ORM layer. In that style of application, the types are maintained in the database itself.

Having a language with strong static typing in that middle layer between the browser and the database only serves to add a huge amount of headaches and extra friction, especially when you're not even using each variable that comes through on its way to the database.

The only thing you're doing in that situation is adding an extra set of type checks that you have to keep in sync between the database and the application layer for virtually no benefit.

In your example, the database and the application layer must be kept in sync regardless of whether you use static typing or not. The benefit of static typing is that it can be enforced before you push to production or write a test for congruency.

This is handled automatically by a large number of ORM tools in dynamic languages (Ruby, PHP, etc). It's a non-issue.

I've spent significantly more time in my career debugging issues in statically typed languages used as an app server in front of a database than I have in dynamically typed languages.

You're doing some processing work in the language? Calculations, ML, etc? Static typing is the obvious way to go.

If you're just sticking things in a database you're going to end up wasting a lot of people's time (and subsequently, the money that goes to pay them). In 20 years, I've never seen it be a good choice but I've seen a lot of people married to the delusion that it is while other people are getting real work done.

This may be one of the differences of requirements that causes disagreement here. I can see how this might work if your app:

* Loads data from the client

* Later displays that data, unchanged, to the client

In that case, treating everything as a string is great, because you never need to use it as anything else.

In my case, I've never actually worked on an app that didn't need to do more than just display the data at some point. In those cases, I don't want to get everything as a string from the browser. My data is often not merely a string, and treating it as such means I have to deal with stupid questions like whether "0" == 0 == "null".

In my work, I have spent many, many hours repairing issues that could have been avoided if the client and server were both statically typed (and on the flip side, in services that are statically typed, rarely run into those same problems). I'm sure that this can be remedied by just "doing it better", but I'd rather operate from a foundation where doing it right is easier than doing it wrong.

Instead of strings, I want my apps to have a statically typed contract between the client and the server that specifies what endpoints are available and the data structures they accept and return. Ideally this is in a shared class/type definition file, but two definitions kept in sync can work as well if they don't share a language.

The common argument against statically typing your network requests at each end is that you're not resilient against unexpected values and/or version differences, but:

a) It's trivial to tell any JSON parser to ignore unknown values, so adding new optional fields to the request is never a problem.

b) If your app ever uses the data received from the client, version differences will eventually break your dynamic code, too. You'll just only be told so at runtime, potentially only after loading the bad data from the db.

If you ever use your app's data for anything other than rendering a template, you have to make breaking changes to your data structures very carefully, and using a strong layer of static types on both client and server helps to ensure that you consider all the breaking changes and either avoid them altogether or somehow mitigate them.

I agree that this is a different scenario. If you're able to have the client control the input to ensure only the correct types are passed in there may be some additional value there to keep things more sane.

However, how much of the data that you're dealing with in the client on the server and in the database are you really working with in the server itself? 1% of the variables? 2%?

Generally any heavy work will be handled in either the database itself or a background job just to reduce wait time for the client.

And if that's the case, you're spending the time defining types for the client, server and database, keeping them in sync, having to change them in each place and creating this exercise for the other 98-99% of variables passed in and out of your server just to gain the static benefit for the 1-2% where you're actually working with it in the server on the web request.

It's significantly more work with questionable benefits IMO.

Same argument could be made the other way, if you haven't used Smalltalk, Clojure (and other languages that are strongly dynamic), you shouldn't discount them.

Clojure has some cases where it's insane how elegant you can make the solution, but frankly static languages with good type systems and tooling come close enough but don't have the scaling downsides.

Yeah, agree. But this is hardly surprising. Learn enough about ANY language on this planet and you'll come to the same conclusion.

The beauty of Clojure is that it mostly provides you with tooling and other facilities that leads to elegant and more importantly simple code.

I'm, not sure if Smalltalk and Clojure do this better, but since I've gotten into static languages, I basically haven't looked back, and I haven't missed the dynamism much at all.

Yes, languages that are truly dynamic (ships with a REPL, has a instant change->run this snippet workflow, can redefine anything in the runtime AT runtime) does work a lot better than Python which is basically comparing the type systems of Java and Haskell.

Give it a try. Worst that can happen is that you now have tried and learned something different, I'm sure it'll give you some knowledge that'll be helpful in any programming language :)

Pharo is a nice implementation of Smalltalk and they have a nice page that describes the top features of the language. Take a look and see if it's different from the developer experience you had with Objective-C: https://archive.is/uAWX5 (linking to a archive via archive.is as some images couldn't load)

Everything Is An Object in Smalltalk, and of course this can't be true in Objective-C due to its mixed heritage.

Que?

  Python 3.9.0 (default, Oct 10 2020, 11:40:52)
  [Clang 11.0.3 (clang-1103.0.32.62)] on darwin
  Type "help", "copyright", "credits" or "license" for more information.
  >>> 1+"1"
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
  TypeError: unsupported operand type(s) for +: 
  'int' and 'str'

  Welcome to Node.js v16.0.0.
  Type ".help" for more information.
  > 1 + "1"
  '11'

JavaScript's more forgiving, and often surprising, implicit conversions are still rigidly-defined .

Remember that implicit conversions should never (ostensibly) throw any exception or cause any errors because implicit conversion should always succeed because it should only be used to represent an obvious and safe narrowing-conversion. A narrowing-conversion necessarily involves data-loss, but it's safe because all of the data needed by the rest of the program is being provably retained.

This doesn't feel like a very useful definition. What languages would fall under the category of "weakly typed" by your definition? I can only think of C, and even that is only true for a subset of the language (when you're messing about with pointers).

> Remember that implicit conversions should never (ostensibly) throw any exception or cause any errors because implicit conversion should always succeed because it should only be used to represent an obvious and safe narrowing-conversion.

    > const fun = (x, y) => (x + y) / y
    > fun(1, 2)
    1.5
    > fun("1", 2)
    6

29.5 Types Versus Safety

"Programming and Programming Languages"

https://papl.cs.brown.edu/2020/safety-soundness.html#%28part...

I think it's more germane to refer to that as weak implicit type casting.

  class A(random.choice([B, C]):
    ...

    class Rectangle:
        var w,h

    class Square(Rectangle if w == h):
        ...

The last thing: being able to edit program code, add new types/members/functionality and so-on while it's running... that's not really a language limitation nor typing-system limitation either: that's just good tooling that is able to hot-swap executable code after the process has already started. Visual Studio has had "Edit & Continue" in debugging sessions for almost 3 decades now, and the runtime build system used by ASP.NET means that source code and raw aspx/cshtml views can still be FTP'd up to production and just work, because .NET supports runtime compilation.

------------

For ordinary business software, where the software's focus is on data-modelling and data-transforming-processes, so all these programs basically just pass giant aggregates of business data around, so strong-typing is essential to ensure we don't lose any properties/fields/values as these aggregate objects travel through the software-proper. In this situation there is absolutely no "expressiveness" that's could be gained by simply giving up on type-correctness and instead rely on unit tests (and hope you have 100% test coverage).

It's more complex than that. The language absolutely has to define these operations for them to make any sense. Modifying some code inside a function is the easy case, and supported by many environments. Modifying types is much more difficult, and usually not supported (.NET and JVM do not support any kind of type modification - neither field nor function). I'm not sure if the C++ debugger can handle this either. Common Lisp can do it, and it actually defines how existing objects behave if their type changes.

There is also ways of dealing with this in dynamic languages. For example, Clojure has clojure.spec that achieves the same results, but without relying on strong typing.

In the end, it's all tradeoffs. There is no "one solution" that will be perfect for everything.

https://vvvvalvalval.github.io/posts/what-makes-a-good-repl....

> If you're not familiar with Clojure, you may be surprised that I describe the REPL as Clojure's most differentiating feature: after all, most industrial programming languages come with REPLs or 'shells' these days (including Python, Ruby, Javascript, PHP, Scala, Haskell, ...). However, I've never managed to reproduced the productive REPL workflow I had in Clojure with those languages; the truth is that not all REPLs are created equal.

Basically, the "REPL" you usually call a REPL is in fact a shell, and not a REPL as normally used in lisp languages.

By contrast, most other languages only have eval() - which takes a string and parses and executes it as code. In Lisp, eval() takes a Lisp list object and executes that.

For eg.

  >> code: load "foreach n [1 2 3 4] [print n * 2]"
  == [foreach n [1 2 3 4] [print n * 2]
  ]

  >> type? code
  == block!

  >> do code
  2
  4
  6
  8

Writing correct, maintainable software is hard. You need both tests and typechecking.

Java can put you off with all the boiler plate. That's not a requirement of static typing.

F# skirts-around some of the limitations with just-in-time design-time type generation ( https://docs.microsoft.com/en-us/dotnet/fsharp/tutorials/typ... ) though Microsoft pitches it as a "better T4" for F# specifically, I think it's far more powerful.

The way things are looking, C# is set up get some form of higher-kinded-types in the next 3-5 years (hurrah!) - but from what Mads and the design-team have said it seems like they'll be implemented without significant changes to the CLR's type-system - so be prepared for everything to be implemented as glorified structs (and don't expect struct inheritance either...) with excessive amounts of implicit-conversion.

The feature you state (F# unions) - that simplification has some positives as well. I can exhaustive pattern match for example. I can always use composition/pattern matching/active patterns to mix in other assemblies types cases. Is it really limiting that it doesn't do this? Would changing this mean more complexity and make other language features harder to deliver? All features introduce some additional complexity, and it more than linear typically. F# also has some interesting patterns that were either introduced first there en masse (e.g Async) or are easier to use there.

I'm a fan of static checking, and conciseness with performance close to the platform it is hosted on and is easy to reason about. I'm also a fan of easy to read and clean code that isn't "scary" to look at. F# seems to strike that balance well compared to other JIT functional languages, at least from where I’m sitting. Too much complexity and you start scaring people away and/or too much abstraction in the codebase starts to occur trading off maintainability.

I don't understand. All languages, without exception, must compile down to a "worthless untyped jargon". Only assembly can execute.

Javascript is a garbage layer of broken standards, a worst of breed language that shouldn't have to exist between a strict typed higher level language and lower level bytecode, but everything on the web now needs to be built on top of it, compile to it first, and run through one of a few JS engines. You can never build a full fledged language on top of something that has to go through JS... so TS, coffescript, et al are always going to be pseudo typed but not truly solid or permanent solutions.

Ultimately I think ECMA standards will finally move toward strict types in JS, and in a decade more we'll have something that for all intents and purposes is exactly what Flash was in 2009.

Sure, but if your typed language is sound you shouldn't have type errors at runtime. Not having type at runtime is not a problem, type erasure is nothing new and works really well in OCaml.

> You can never build a full fledged language on top of something that has to go through JS... so TS, coffescript, et al are always going to be pseudo typed but not truly solid or permanent solutions.

Rescript and Elm are really solid, OCaml can compile to JS too, Scala and F# too. Granted, only Elm has been built on JS at first, others were adapted (Rescript is like half and half), but they're all solid.

> Ultimately I think ECMA standards will finally move toward strict types in JS, and in a decade more we'll have something that for all intents and purposes is exactly what Flash was in 2009.

I think the idea here is to offer WASM as a compilation target. Trying to gradually type JS leads to really complex systems (TS types) because people did all sorts of crazy things with JS. I wouldn't like this to be the base on which we build other languages, or the web in general.

TypeScript, by design, does not produce nor introduce any "noisy defensive code": the point of TypeScript was that type-safety in JavaScript can be achieved without any runtime cost.

The only significant code-gen that TypeScript does do by itself is stuff relating to ECMAScript modules (e.g. wrapping and repackaging your TypeScript module for other module systems).

...with the exception of when you take a modern TypeScript compiler and specifically instruct it to target ES3 (lol) or perhaps some older variants of ES5 - at which point the code `tsc` generates is still so minimal you'll still need to bring-your-own polyfill/monkeypatch libraries. This is stuff like reimplementing the spread-operator or `for(of)`, and so on.

the web runs on JavaScript because it's extremely forgiving and therefore stable. there can be lots of errors and things keep chugging along.

...and when types are within TypeScript's sphere-of-influence the compiler is largely able to deduce types without needing assertions or type-checks.

...or so you think! But there's two problems:

1. TypeScript's type system is unsound: so there are times when the compiler will let you do some (for example) contravariant array mutation, even though that's almost essentially guaranteed to crash at runtime if the rest of the system makes incorrect assumptions about the type-variance of the array/collection. So having the type-guard function available to assert to tsc that whatever array operation you performed is safe, you wouldn't need to do that in pure JS.

2. Type-guards are essential for when you're working with values that pass through an interface (because that necessarily means some loss of typing information has occurred), so the type-guard function is needed to recover that static information - and while this approach does also apply to JS, in some situations where if your code is the factory of subclasses that pass through an interface (e.g. so `MyComplicatedSubclassForRockAndRoll` appears as `MusicSubclass`) then in JS a blind "reinterpret-cast" of the object is fine - but TypeScript really doesn't want you doing that and again, having a type-guard function is just to keep the compiler happy - and even if it is possible to mathematically prove that the `object` type-cast is correct (because theoretically static analysis can see on both sides of an interface) but TypeScript isn't perfect and there are cases where we really do know better than the compiler, but we don't want to just make it unsafe and untyped - instead we want to keep the typing and we want to keep the compiler happy, but doing it this way (a type-guard) is not only easier than suppressing warnings and ignoring your sense of guilt but they're invariably genuinely useful a few months down the line when someone made a change that the type-system spotted and prevented from compiling (and so going into production), whereas it would be very easy to (for example) unintentionally introduce vulnerabilities because even with near-100% test coverage, because adding brand new functionality to a project (even if it said functionality doesn't work) it won't cause any tests to fail (at least, until tests for the new functionality are completed).

If it's an external API, do the same thing except it probably needs a more manual approach.

This is not something that's in anyway unique to typescript. The difference is JavaScript is designed to fail gracefully using coercion whereas most other languages and runtimes abort and throw exceptions.

A common misconception.

In tests, you test that values are correct, the types take care of themselves.

Types being correct does not imply values being correct, so you will have to write those tests anyhow, although you maybe be fooled by the safyness of static typing[1] that you don't have to.

Values being correct does imply that the types are correct, so if your tests pass, then your types are also obviously correct. (If a==3 then I don't have to check if a is of type pink elephant, it obviously isn't).

So that's the in-principle argument. And it turns out that this is completely supported in practice as well: neither the defensive code nor the extra type-based tests that you propose must be the result actually occur in dynamic code-bases with any frequency.

That said, static types certainly have their uses, particularly in being machine-checked documentation and also for producing faster code without going through crazy amounts of extra effort.

[1] https://blog.metaobject.com/2014/06/the-safyness-of-static-t...

> In tests, you test that values are correct, the types take care of themselves.

How can you be sure your function e.g. returns an integer for all values and never undefined for some values? Unit tests cannot check this exhaustively unlike type checkers.

> Types being correct does not imply values being correct, so you will have to write those tests anyhow, although you maybe be fooled by the safyness of static typing[1] that you don't have to.

As I said, types compliment tests, they don't replace them but types do reduce some of the burden on writing tests for you.

To nitpick: type checking can prove values are correct when the type system is powerful enough (i.e. dependent types) but this isn't practical enough for mainstream yet.

Right, which is actually an argument against the claim that you have to write more tests with dynamic typing.

In particular, we can use dependent types to place much tighter constraints on critical pieces of code, but 'cut short' a full correctness proof by simply punting to an error case when it's convenient.

For example, say we need to manipulate the Nth element of a collection. We can express indices with a type 'Fin X', whose values are equivalent to 1, 2, ..., X, e.g.

    f1 : (c: List Foo) -> (n: Fin (length c)) -> BAR

Yet this places a burden on anyone who calls this 'f1' function: they need to prove that the index appears within the list (and, in doing so, prove that the list is non-empty). This might be very difficult in the general case, yet a full correctness proof would require it.

However, we don't need a full correctness proof; so we can just check if we're in bounds, and return an error if not:

    f2 : List Foo -> Int -> Either String BAR
    f2 c i = case mkFin (length c) i of
               Just i  -> Right (f1 c i)
               Nothing -> Left ("Index " + show i + " not in list of length " + show (length c))

      where mkFin : (n: Nat) -> (i: Int) -> Maybe (Fin n)
            mkFin Z     _         = Nothing
            mkFin _     i | i < 0 = Nothing
            mkFin (S n) 0         = Just fZ
            mkFin (S n) i         = map fS (mkFin n (i - 1))

Dependent types are really useful for propagating error conditions 'backwards' through the code like this. When practical, I much prefer this to propagating error conditions 'forwards', by returning Maybe, etc. Most applications start with some sort of parsing step, which is a great place to perform such checks, if we can manage to propagate them back that far. Otherwise, we just handle these error cases just like we would any others (dependent types or not).

Urgh, that reference is such a shallow, straw-man facade of static typing. This is especially evident with its link to https://www.sans.org/top25-software-errors/ and claim that those "don't look like type errors to me", a list which includes:

- 16 even says 'type' in the title: "Unrestricted Upload of File with Dangerous Type"! Phrased as a type error, this would be "savePicture: Required argument of type 'Either[PNG, JPEG]', given 'Array[Byte]' instead"

- 2, 3, 6, 11, 18 are all injection attacks, which are classic type errors (AKA "stringly typed programming"). Phrased as a type error, these would be "+: Cannot append 'Foo' with 'FormInput'", where 'Foo' might be 'SQL', 'ShellCommand', etc.

- NULL pointer dereference is Tony Hoare's 'billion dollar mistake', which as type error would be "Expected 'Foo', given 'null'"

- integer over/underflow, buffer over/underflow, out-of-bounds access, etc. can all be mitigated with types like 'Maybe', 'Try', etc. at the cost of a bounds/flag check. If we don't want runtime checks, we can instead track memory usage bounds in the types, and have the compiler check that we're allocating enough.

So many of these problems are avoidable with types, even before we go into fancy stuff like dependent types, linear types, etc. In particular, most of the burden to getting this right is on languages and libraries, not application developers.

For example, if a database library can only be queried by an 'SQL' type, and the only way to construct that type is via literals and a 'withParam[T]: SQL -> T -> SQL' function, then application developers are unable to introduce injection attacks; the library forces them to do the right thing. Instead we get 'dbQuery: String -> ...', 'runShell: String -> ...', 'response.withBody: String -> ...', etc. and draw the conclusion that types can't help :(

The SQL case is much more complex. SQL is such a non-standardized language in practice, and such a complex one, that trying to offer a generic SQL library that presents an AST as the API for even the most popular RDBMSs quickly gets complex. So, you end up with convenient string->SQL converters, and then people stick untrusted input in those strings. Not to mention, your fancy AST library still needs to spit out SQL text to talk to most DBs, so you end up needing to sanitize identifiers in your own SQL->string conversion.

The problem with stringly-typing is the model, which is SQL as strings. Statically type-checking that all your strings are, in fact, strings does not buy you anything whatsoever in this scenario.

On the other hand, if you have an actual model of the SQL, then it also doesn't matter whether that model is checked at runtime (dynamic typing) or at compile-time (static typing). In either case it will not allow injection attacks, if implemented correctly.

And of course the actual place where nastiness happens is the user input, and I haven't yet seen a place that can statically typecheck users. ;-). So, as you point out, you will need to do dynamic checking and sanitising of input.

User input has type `ByteString`. Not only can we check that statically, we absolutely should; and we should enforce that type, to reject any code which assumes otherwise.

Nastiness happens when developers treat user input as something other than ByteString; e.g. they might try appending it to a fragment of HTML (XSS); or wrap it in quotation marks and send it to a database (SQL injection).

We don't need to 'statically typecheck users'; we do need to statically check that `myApp` has type `ByteString -> Foo`, so we can avoid ever executing implementations which actually have type `HTMLFragment -> Foo` (XSS), or `SQLFragment -> Foo` (SQL injection), or whatever.

It's possible to implement anything correctly with dynamic types but the point is static typing makes it easier to do so. Static types will exhaustively check for you that certain errors aren't possible (vs tests that check a finite number of cases) + compile-time (vs at runtime when the error actually occurs on particular inputs).

> And of course the actual place where nastiness happens is the user input, and I haven't yet seen a place that can statically typecheck users. ;-). So, as you point out, you will need to do dynamic checking and sanitising of input.

It sounds like you're talking about dynamic (type) checking here when you really mean regular runtime behaviour? You absolutely can at compile-time make sure your program is going to respond in a sensible way to problematic user inputs at runtime e.g. the compiler verifies that either the user input string value will get turned into a `valid email` value at runtime or the user is asked to try again. I wouldn't call the condition that checks if the string is a valid email a dynamic type check.

I don't see how user input is different or special from any other value either e.g. it's not like you can predict what exact values are going to come from files, random number generators or from complex calculations during runtime, you just care that they're in the range you expect.

No. Quite the opposite. People here confuse "static typing" with "checking stuff", even such obviously dynamic things as "checking outside input".

> You absolutely can at compile-time make sure your program is going to respond in a sensible way to problematic user inputs at runtime

Right. Sort of. But that is just normal code. Code that absolutely, definitely has to be executed at runtime and at no other time. Because that is when the actual users of your system are going to input the actual data. Not while you're compiling stuff. Your users (and the rest of the world you interact with) aren't there while you're compiling. And your compiler isn't around when your users are inputting data.

> the compiler verifies that either the user input string value will get turned into a `valid email` value at runtime

No, the compiler cannot verify user input, because the compiler isn't running when the user is inputting data. You can implement a model that has the concept of "valid e-mail", and you can implement your system such that it converts input data such as strings into your structured model at the edges and then only deals with "valid e-mail" objects internally. But implementing such a model has nothing whatsoever to do with whether you use a statically type-checked language or a dynamically type-checked language.

It also turns out that such systems tend to be really, really bad. What you actually want to do is keep the user input exactly as it was input, for auditing purposes if nothing else, and build your structures as an enrichment on top of that basic data. Because your idea of what constitutes a "valid" e-mail is almost certainly wrong, and it's better if the system can at least represent data even if it doesn't completely understand it, rather than destroy user data.

> No, the compiler cannot verify user input, because the compiler isn't running when the user is inputting data.

I'm not following why you thought I meant that. It's like you're arguing it's impossible to make any compile-time guarantees because you don't know what exact literal inputs your program will be dealing with at runtime when the field of https://en.wikipedia.org/wiki/Formal_verification exists. E.g. at compile time you can prove a sorting function gives the correct output for all possible inputs without having to run the code - you don't know in advance what the input is going to be (which is normal when coding), but you still know the output will be correct with respect to the function specification.

1. Define SQL in your type system, and force programmer to specify for any piece of user input they want to use in a query what its semantics are supposed to be. This is the SQL DSL approach, such as the short lived LINQ-to-SQL or Haskell's Selda. ORMs also do something similar.

2. Enforce that any string sent to the DB passes through some kind of checker that enures certain properties hold for that SQL. The checker will have to understand all of the semantics of SQL, just like in 1.

There are many libraries that go through path 1, but don't support the full capabilities of SQL (usually they support a tiny subset), even for a single DB.

This is where the argument became circular and we can basically stop.

The parent claim (by chriswarbo) was that, for example, SQL injection attacks were, in fact, incontrovertible proof that SQL injection attacks "are classic type errors". That simply isn't true, at least not in the sense of static type checking (and only made true, as I mentioned above, by conflating "static type", "type" and "model" into one incoherent gooey mush).

chriswarbo's incorrect claim was in response to my referenced article, The Safyness of Static Typing [1], where I (as a prelude to my actual point) look at the evidence for the claim that "static types make it easier to implement something correctly".

There isn't any.

Or to be precise, there isn't any that passes any statistical or other standards. What little evidence there is goes both ways and has, at best, tiny effect sizes.

And yes, that article is somewhat old, but the evidentiary situation has not changed, despite further attempts to make the claim that static typing is provably better. Claims that were resoundingly debunked.

So that's the background. With this background we have the idea that SQL injection attacks are somehow evidence for the problems with dynamic typing, which they are not. I can have a statically checked string type and have exactly the same SQL injection attacks, and in principle, checking code needs to run at runtime against the dynamic input it is represented with. Which I think we agree on:

> It's possible to implement anything correctly with dynamic types

but apparently chriswarbo does not. So that brings us to the circularity of the argument:

> but the point is static typing makes it easier to do so

This hasn't been shown. It was claimed. And it was claimed that SQL injection somehow proves this. Which it doesn't, because we agree that what prevents the attack is the code that runs, the model that executes. It may be that such code is easier to write with static types, but that hasn't been shown, it is just claimed and the claim repeated in support of the claim. Circle.

[1] https://blog.metaobject.com/2014/06/the-safyness-of-static-t...

You mean the link from [1] to "An experiment about static and dynamic type systems" based on "an empirical study with 49 [undergraduate] subjects that studies the impact of a static type system for the development of a parser over 27 hours working time.". I think studies like this are more distracting here than useful (low skill level, not large scale enough, too much noise, toy example). The article here sums up a lot of my thoughts on this: https://news.ycombinator.com/item?id=27892615

And do you really need a study that proves e.g. (to pick a simpler example to summarise) a language that makes it impossible for null deferences to happen at compile-time is going to have less null dereference errors than one that lets those errors happen at runtime? It's like insisting for an empirical study that 2 + 2 = 4.

> It may be that such code is easier to write with static types, but that hasn't been shown,

The typed SQL example might look something like:

   function getUserInput(): string

   function createSanitisedString(s: string): SanitisedString
   function createSafeSqlQuery(s: SanitisedString): SqlQuery

1. At best, the code will fail only after the user input is received by createSafeSqlQuery during runtime and you won't know where the input originated from.

2. At worst, the coder forgets to add a check like `typeof s === SanitisedString` or a call to `createSanitisedString` in `createSafeSqlQuery` and creates an injection attack.

The static type checker clearly wins here for me for safety and ease of implementing correctly. I don't need a study to know that compile-time errors are better than production runtime errors, that automated and exhaustive checks of correctness are better than the coder having to remember to add manual checks, and that it's better to know the exact line the unsanitised input came from over only knowing it comes from somewhere.

What languages have you used that have strong static type systems that you've written large programs in? Have you tried OCaml or Haskell for example?

The problem with the whole argument is that these functions are not actually enough to work with SQL, since they don't allow us to create dynamic SQL from safe strings.

Here are some ideas of why we can' use the functions you proposed:

  user_filter = raw_user_input_col_name ++ " LIKE ?"
  createSafeSQLQuery("SELECT * FROM my_table WHERE " ++ user_filter ++ ";") //type error

  sanitizedQuery = createSanitisedString("SELECT * FROM my_table WHERE" ++ user_filter ++ ";") //should return error or quote everything
  createSafeSQLQuery("SELECT * FROM my_table WHERE" ++ createSanitisedString(user_filter) ++ ";") //filter will be wrong, type error

  function asSanitizedString(s: string) : SanitizedString {
    return s
  }
  user_filter = createSanitizedString(raw_user_input_col_name ) ++ asSanitizedString(" LIKE ?")
  query = asSanitizedString("SELECT * FROM my_table WHERE") ++ user_filter ++ asSanitizedString(";")
  createSafeSqlQuery(query) //works, nice
  createSafeSqlQuery(asSanitizedString("SELECT * FROM my_table WHERE " ++ user_col_name ++ "LIKE ?;")) //oops, this also works

If you don't believe me, go looking for a library that allows arbitrary strings as SQL, but statically ensures user input can't be used to construct an SQL query. I don't know of any one.

My bad for being unclear but I meant you would be sanitising something like the user entering "tea cups" into a shop search input form and searching your shop product database with that. I didn't mean the user would be entering SQL queries.

> But either way, you can't fix it without modeling the entirety of SQL into your type system

Using the type system to check correctness sounds good to me.

This is getting way too into the weeds about SQL anyway. Null dereference checks are a less distracting example to focus on for instance. As long as you can encode it into the type system, all my points are still relevant.

No, I understood that. But my point is that the programmer is going to be composing SQL queries, and they will be doing it based on user input. The SQL API will have a very hard time distinguishing which parts it receives are trusted programmer input and which parts are untrusted user input.

> Using the type system to check correctness sounds good to me.

Sure, but SQL is a huge language with significant differences between any 2 major DB implementations (and that depends on SQL server configurations - e.g. in MySQL you can specify via config file whether "a" is a string or an identifier). I have never seen a full SQL DSL or AST used in any SQL client library, in any language: it's just too much work.

No. I mean, yes, that is one study, but there are a lot more. They all come out essentially the same.

Danluu did a an overview article a while ago:

https://danluu.com/empirical-pl/

Note that the A large scale study of programming languages and code quality in github paper, the one that makes some of the strongest arguments for safety of the bunch, was the one that was later completely taken apart at OOPSLA 2019. Its findings, which also had very small effect sizes and statistics significance, are completely invalid.

> do you really need a study that proves ... less null references errors

1. Languages don't have null errors, programs do.

2. I don't need a study to show that a program has less than or equal null errors in a language with null safety, because the program in a language without might still have zero. If you're going to make a logical claim, then let's stick to what's logically provable. If you're going to hand-wave...well you might as well use a dynamically typed language ;-)

3. I do need a study to show that such differences matter at all. All software has bugs, if this a significant source of bugs, then a mechanism to make me not have them might matter.

4. I do need a study to show that the net effect is positive. For example, the famous Caper Jones study showed that dynamic languages such as Smalltalk and Objective-C were significantly more productive than statically typed languages, including Haskell, C++ and Java. Studies also have long shown that bugs scale linearly with code size, a very strong correlation. So let's say null-references are 1% of you total bugs, but you now write twice the amount of code, that means a move to a null-checked language will significantly increase your total bug count, despite eliminating a certain class of bugs.

(In fact other, older studies showed that type errors accounted for 10% of bugs, so if you can save just 10% of code using a dynamically typed language, you're ahead in terms of bugs).

Of course, I can also not require such studies and make an engineering judgement. And this is fine, we do it all the time because very little in software has been demonstrated much at all. But you then need to be aware that this is a subjective judgement call, and others may reasonably come to a different conclusion.

And being aware of this makes for much, much better engineering judgement, IMHO.

We could take some JavaScript programs —

https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

— and try to transliterate them into Dart —

https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

— and see how much or how little boiler plate is required when a language supports null safety, local type inference, implicit-dynamic: false ?

> 2. I don't need a study to show that a program has less than or equal null errors in a language with null safety, because the program in a language without might still have zero. If you're going to make a logical claim, then let's stick to what's logically provable. If you're going to hand-wave...well you might as well use a dynamically typed language ;-)

I think now you're just nitpicking and being uncharitable for the sake of it instead of responding to the overall point of my message I took care to write.

> Studies also have long shown that bugs scale linearly with code size, a very strong correlation. So let's say null-references are 1% of you total bugs, but you now write twice the amount of code, that means a move to a null-checked language will significantly increase your total bug count, despite eliminating a certain class of bugs.

> (In fact other, older studies showed that type errors accounted for 10% of bugs, so if you can save just 10% of code using a dynamically typed language, you're ahead in terms of bugs).

Most software engineering studies have so many confounding factors even in isolation that combing results from multiple studies like this is nonsensical and misleading. This reads like brilliant satire if I'm honest.

> And being aware of this makes for much, much better engineering judgement, IMHO.

You didn't reply to the question about if you've written large projects with strong static type systems like OCaml or Haskell. If the answer is no, they're worth learning for more awareness instead of relying on likely very limited and/or flawed empirical studies in my opinion.

Have you ever considered the validity of that study: which purports to compare programming languages, apparently without considering obvious differences in available programming tools?

Even back in the late '80s, those Smalltalk implementations provided a very integrated development environment — writing Smalltalk in a simple text editor really isn't the same ;-)

> This is where the argument became circular and we can basically stop.

I think any appeal to "easier" is inherently subjective. I've certainly tried to avoid making any such claims. Instead, I've mostly tried to argue that static types can forbid certain programs from ever running, and we can use that to forbid things like vulnerable SQL queries.

> > It's possible to implement anything correctly with dynamic types

> but apparently chriswarbo does not.

I never said any such thing; besides which, it's trivially the case that anything implemented with static types could also be implemented with dynamic types, since static types only exist at compile time (they are "erased" during compilation), and hence every running program is dynamically typed.

In fact, I don't think I've said anything about correct implementations at all. My point is that static types can forbid incorrect implementations. When it comes to security, that is much more important; i.e. I would much rather bang my head against a screen filled with compiler errors, than expose some insecure program to the harsh reality of the Internet.

> The parent claim (by chriswarbo) was that, for example, SQL injection attacks were, in fact, incontrovertible proof that SQL injection attacks "are classic type errors".

I've not claimed that (in fact, I'm stuggling to parse what that sentence might even mean). I claim that SQL injection attacks are "classic type errors" in the sense that:

- It's a widespread problem, easily encountered, commonly warned about, and most developers have probably done it at some point.

- It's a class of error that can be entirely ruled out at the language/API level, through use of static types. Similar to how segfaults can be entirely ruled out at the language level, using managed memory like in Python/JS/etc.

- Due to the above, it's a common example used to illustrate the usefulness of static types in blog posts, articles, etc. One which sticks in my mind is "A type-based solution to the “strings problem”: a fitting end to XSS and SQL-injection holes?" https://blog.moertel.com/posts/2006-10-18-a-type-based-solut...

- Due to the above, it's considered "folk wisdom" (or a "folk theorem") in the relevant fields (e.g. Programming Language Theory, Formal Methods, etc.)

Other examples of "classic type errors" might include:

- Mixing up integers with strings-of-digits, e.g. 'print("You were born around " + (time.now.year - input("How old are you?"))'

- Forgetting to 'unwrap' some list element, e.g. 'open(listdir("config/")).read()' instead of 'listdir("config/").map(f => open(f).read())'

- Mixing up units of measure, e.g. 'distance = speed + time' instead of 'distance = speed * time'

Basically any common mix-up between values/components in a program, where the computer could help up to spot the mix-up (perhaps entirely automatically, if type inference is involved). In the case of SQL injection, the mix-up is between 'string of bytes' and 'SQL query'.

> And yes, that article is somewhat old, but the evidentiary situation has not changed, despite further attempts to make the claim that static typing is provably better. Claims that were resoundingly debunked.

I very much appreciate when researchers attempt to ground things in a bit more empiricism! Unfortunately those particular studies just aren't looking at the sorts of things that I find relevant; in particular, those studies (and that linked blog post, and many of the comments here), seem overly-preoccupied with mundane trivialities like "string", or "int", or "SQLColumnName".

Personally, I'm much more interested in how types can help me:

- Avoid leaking secret information https://link.springer.com/chapter/10.1007/978-3-030-17138-4_...

- Guarantee big-O resource usage https://twanvl.nl/blog/agda/sorting#a-monad-for-keeping-trac...

- Guarantee the correct order of operations https://docs.idris-lang.org/en/latest/st/machines.html

- Prevent AI agents from losing capabilities http://chriswarbo.net/projects/powerplay

> It's a class of error that can be entirely ruled out at the language/API level, through use of static types.

Repeating this claim doesn't make it true.

Once again: this has nothing to do with static vs. dynamic types, and everything to do with modelling.

To make this clear, let's compare a dynamically and a statically type-checked version of this with the model of SQL as just strings.

Both the dynamically and the statically type-checked version of this program will be susceptible to SQL injection attacks. The statically type-checked version will verify at compile time that all the values are indeed strings.

Now let's compare a dynamically and a statically checked program with a proper model for the SQL, not strings.

Both of these will not be susceptible to SQL injection attacks.

It has nothing to do with static vs. dynamic, and everything with how you model the problem.

My gut tells me that's not quite right (for some reasonable definition of 'usable'), since we can always build elimination forms into the language (after all, 'null' must be built in, and few take objection to building in elimination forms like if/then/else).

> it also requires some kind of pattern matching or monad comprehensions - otherwise, code using Optional is neither type safe nor readable.

I really like comprehensions in Python, but hardly ever use them elsewhere. In fact, when I do use them in Haskell and Scala, I usually find myself having to refactor them into calls to `map`/`join`/etc. soon after, to have more fine-grained scoping or somesuch.

> trying to offer a generic SQL library that presents an AST as the API for even the most popular RDBMSs quickly gets complex

That's orthogonal to anything I said. Represent SQL using arrays of bytes in memory if you like; offer a string-like interface for constructing and manipulating them if you like; just make sure to distinguish them from other string-like types, in a way that's statically enforceable, and where the only API to convert a 'String' to an 'SQL' is the escaping function.

Note that I don't particularly care if an SQL value is valid; it would be nice to statically enforce that, but you're right that (vendor-supplied) SQL is pretty complex in its own right. What's much more important to get right is that SQL is not user-generated.

True, you could go the Go array route and have Optional be a special type that is generic, without otherwise supporting generic data structures.

> I really like comprehensions in Python, but hardly ever use them elsewhere.

I think code like possiblyMissing.map(value => code) introduces lots of unpleasant imbrication, and breaking out functions for every possibly-missing value also seems to me to lead to bad readability. I suppose this may be a matter of taste, or it may be related to other language features.

> just make sure to distinguish them from other string-like types, in a way that's statically enforceable, and where the only API to convert a 'String' to an 'SQL' is the escaping function.

I think this is the part that doesn't really work, because, in order to escape a piece of user input, you need to understand what it's meant to be. For example, a common (unsafe) idiom is:

  sqlFormatString = "SELECT %s FROM %s WHERE %s;";
  filterFormatString = "%s LIKE ?";
  filterString = sprintf(filterFormatString, userInputColumnName);
  sqlQuery = sprintf(sqlFormatString, columnName, tableName, filterString);
  preparedQuery = dbImpl.prepare(sqlQuery, userInputValue);

Now, you could replace this with a type safe dbImpl.prepare(), which only takes a NotUserGeneratedSQL type. But now you need to have some way to build a NotUserGeneratedSQL.

One option is to go all in on an SQL AST library, where you could build the query above like

  query = selectQuery();
  query.columnNames = [ escapeColName(columnName) ];
  query.from = escapeTableName(tableName); 
  queryFilter = equalityComparison();
  queryFilter.left = escapeColName(userInputColumn);
  queryFilter.right = preparedArg();
  query.filter = queryFilter;

  preparedQuery = dbImpl.prepare(query, userInputValue)

However, I don't think there is any alternative that can actually enforce validation of user input in constructed queries. You can help users think about it by forcing them to use some kind of string -> NotUserGeneratedSQL function, but this function can't actually be written to reject the example I gave initially. A function that creates an empty minimal NotUserGeneratedSQL and offers other functions to build it up to a useful query will either accept strings (making it possible to introduce SQL injection) or have to accept structured SQL, making it fall into the AST problem.

Of course, the problem of NotUserGeneratedSQL is easy if you don't want to support any kind of dynamic query, other than prepared statements. But if you want to dynamically generate queries based on user input (e.g. user chooses which columns they want to see, which columns they want to filter by etc.), then I don't think it's possible to statically ensure that SQL injection is not possible without an ORM or SQL AST API.

Edit: if you believe otherwise, please show me a type-safe, SQL-injection-safe library in any language that isn't an ORM/EDSL or a straight up SQL AST manipulator. I am quite certain none exists, but I would be happy to be proven wrong.

    allowedColumns: Map[UserInput, SQL] = {
      "name": "name",
      "age": "age",
      ...
    }

    strict: SQL = strictComparison? "=" : "LIKE"
    
    query: Maybe[SQL] = allowedColumns
      .get(userCol)
      .map(col => "SELECT " + col + " FROM tbl WHERE tbl.foo " + strict + addParameter(":?", userFoo))

- We can write literals which look like strings but have type SQL

- We can append fragments of SQL together (potentially making it invalid; oh well)

- We can include user input via parameterised queries, in locations where arbitrary strings/ints/etc. are allowed

- Anything 'structural', like identifiers, choice of comparison operations, building up sub-expressions, etc. must be done programatically, using the above features. In this case we select the column name (written as a static, literal SQL value) by looking up the user's input in a map. We also allow choosing between the type of comparison to use (again, both are SQL literals).

It seems to me that requirements like 'user chooses which columns they want to see, which columns they want to filter by etc.' is fundamentally incompatible with a safe, string-like representation. Instead, the options are:

- Fully representing the structure of the language. This results in an AST approach, which allows safe dynamic queries. I think any alternative, like tracking the offset of each delimiter in a string, etc. will turn out to be equivalent to maintaining an AST.

- A "flat", string-like representation, whose dynamism is limited to choosing between some combination of pre-supplied fragments. This is what I've shown above. This is safe, but the 'dynamism' is inherently limited up-front (i.e. it's overly conservative).

- A "flat", string-like representation, which has unrestricted dynamism, but hence is also inherently unsafe (i.e. it's overly liberal).

What I'm not clear is: what prevents me from accidentally/stupidly doing:

  filter : SQL = "WHERE " + userInputCol + " = ?"

Do you know of any library that implements this?

Ah, maybe I should have made it clearer that I was overloading the double-quote syntax, so we can write:

    "foo": String
    "bar": SQL
    "baz": UserInput
    "quux": Shell
    etc.

    +: String  -> String  -> String
    +: SQL     -> SQL     -> SQL
    +: Int     -> Int     -> Int
    +: Float   -> Float   -> Float
    +: List[T] -> List[T] -> List[T]
    etc.

    // Code as written
    filter : SQL = "WHERE " + userInputCol + " = ?"

    // Right-hand-side must have type SQL, to match left-hand-side
    "WHERE " + userInputCol + " = ?": SQL

    // Resolving order-of-operations of the two '+' operations
    ("WHERE " + userInputCol) + " = ?" : SQL
  
    // Arguments to outer '+' have same type as return value, which is SQL
    "WHERE " + userInputCol : SQL
    " = ?" : SQL

    // Arguments to inner '+' have same type as return value, which is SQL
    "WHERE " : SQL
    userInputCol : SQL

> The main downside is that you need to work entirely with compile-time constructs - e.g. you can't use something like printf to take a compile-time format string and turn it runtime into a query; and you can't take queries from a separate file, they must be in source code.

Yep, although macros could help with that sort of thing, e.g. Haskell's quasiquotation https://wiki.haskell.org/Quasiquotation

A couple of Google hits for 'haskell quasiquote sql':

https://hackage.haskell.org/package/postgresql-simple-0.6.2/...

https://hackage.haskell.org/package/postgresql-query

> Do you know of any library that implements this?

Not completely. De-coupling double-quoted literal syntax from a single String type can be done with Haskell's OverloadedStrings feature, but that relies on a function 'fromString : String -> t', which is what we're trying to avoid https://hackage.haskell.org/package/base-4.6.0.1/docs/Data-S...

Scala's custom interpolators are similar, but they rely on a function from 'StringContext -> t', and StringContext is easily created from a String ( https://www.scala-lang.org/api/current/scala/StringContext.h... )

To ensure safety, we would need some way to encapsulate the underlying fromString/StringContext implementation to prevent it being called by anything other than the literal-expansion at compile-time.

Of course, if we're willing to use macros then it's pretty easy, like those quasiquote examples above.

Haskell's module system is famously mediocre, so it might be possible to do this overloading + encapsulation with Idris https://idris2.readthedocs.io/en/latest/reference/overloaded...

(Of course, Idris also has an incredibly expressive type system, and a very powerful macro system, AKA "elaborator reflection", so it can definitely be done; but I haven't figured out the cleanest way)

So what you do? I find code that uses excessive type checks too verbose, you end up doing things only to make the compiler happy, even if you know that a particular condition will never happen (e.g. a variable that can be null but given a particular condition you know that it cannot be because it was initialized earlier, but the compiler is not smart enough to know).

So to me using static types makes sense where types are used by the compiler to produce faster code (embedded contexts, basically, of course I don't use Python on a microcontroller) and as a linting of the code (Python type annotations or Typescript).

"SafeStrings: Representing Strings as Structured Data"

https://arxiv.org/abs/1904.11254

Bosque

https://github.com/microsoft/BosqueLanguage/blob/master/docs...

You are confusing "type" with "static type". This is user input. You are not going to statically typecheck your users.

(See other comment for stringly-typing).

You are confusing "type" with "not a function type".

Firstly, I absolutely want to check, statically, that my user input is `ByteString`, or `String`, or some more distinguished newtype/AnyVal wrapper around those. If that doesn't pass the type checker, I would be very worried about ever starting such an executable!

Other than that, I want to check, statically, that my functions conform to their type signatures. When I write `parsePNG userInput: Maybe PNG`, not only do I want to make damn sure that `userInput: ByteString` is definitely a `ByteString`; I also want to make damn sure that `parsePNG: ByteString -> Maybe PNG` is definitely a (partial) function from `ByteString` to `Maybe PNG`. If I can't guarantee, without executing the program, that `parsePNG` can only return `Maybe PNG` values, then all bets are off regarding downstream invariants, and that's not a good situation to be in w.r.t. security vulnerabilities.

My condolences, you must have horribly unreliable user widget/user input libraries. My TextFields always deliver a "String" to me (unless I've set them up to deliver numbers). I don't need to check that statically.

> You are confusing "type" with "not a function type".

No I am not.

You don't seem to understand what I meant with "You are not going to statically type check your users".

The libraries seem pretty reliable; but I'm not (exactly the same as in the SQL example, BTW).

For example, I don't think I've ever managed to write a Python3 program which hasn't crashed at some point due to me mixing-up string/bytes; of course, testing doesn't usually find such problems, even with Hypothesis, since I make exactly the same mistakes when I'm writing the tests. I've also written a whole bunch of Python3 programs which spend ages crunching numbers, only to output '<map object at 0x10199c2e0>' instead of printing a list.

I make exactly the same mistakes when writing Haskell and Scala; but the computer tells me that I've made a mistake, and how to fix it, before the program ever gets a chance to run.

> You don't seem to understand what I meant with "You are not going to statically type check your users".

That might be the case. I can't elaborate on what you meant (since I may have misunderstood)

Not sure what that even is, but I'd point out that dynamically-typed ≠ Python. In particular, the "hash languages" (Python, Ruby, JavaScript) seem to make it particularly easy to make a hash of things, but not due to the dynamic typing.

I've never had that problem with NSString and NSData. They are quite distinct.

> Python3 ... crunching numbers ... map instead of list

??

Why would you output a map when you are crunching numbers?

To get a "list", for example, I'd use an NSArray. It seems tricky to get an NSDictionary instead, since you have to explicitly ask for an NSDictionary in order to get one.

If I generate values within a function that are arguments to a constructor call, and the make the constructor call and return the result with no further manipulation, I can—without exhaustive, or even any—testing be quite confident that the function will return the type for which the constructor so called is the constructor any time it returns, and fail otherwise.

But when you start having loops, conditions, arrays, nullable references etc. you can't be so confident any more unlike with proper type checking where you can be fully confident.

What you have, after manual inspection of the code, is some sort of informal and not written down correctness proof. A machine tested proof would be better and types help with that.

In those cases, the difference between a compile-time check and a run-time check is much smaller.

I also find 'phantom types' incredibly useful. For example, we might have a database or key/value store, where the keys are represented as strings and the values can be strings, integers, floats, booleans, etc. We can 'label' each key with a "phantom type" corresponding to the values. This way, the compiler knows exactly what types are expected, we only have to handle the expected type, we don't need any boilerplate to wrap/unwrap the database results, etc.:

    -- Haskell syntax
    data DBKey t = DBKey String

    dbLookup :: DBKey t -> Maybe t
    dbLookup (DBKey key) = ...

    name = DBKey @String "name"
    age  = DBKey @Int    "age"

    // Scala syntax
    final case class DBKey[T](override val toString: String)

    def dbLookup[T](key: DBKey[T]): Try[T] = ...

    val name = DBKey[String]("name")
    val age  = DBKey[Int   ]("age" )

Of course, we can combine such 'labels' with more meaningful types; so that 'name' looks up a 'Username'; 'age' looks up an 'Age', or a 'Nat', or a 'Duration'; or whatever's convenient. It can also be combined with other type system features like implicits/ad-hoc-/subtype-polymorphism, e.g. to decode JSON as its read from the DB, etc.

I've used the example above in real Scala applications when accessing a key/value store. I've also used phantom types (again in Scala) to help me manipulate Array[Byte] values: in that case the bytes could either be 'Plaintext' or 'Encrypted', they could represent either a 'Payload' or a 'Key', and each 'Key' could be intended for either 'Encryption' or 'Signing'. Under the hood these were all just arrays of bytes, and the runtime would treat them as such (using Scala's 'AnyVal' to avoid actually wrapping values up), but these stronger types made things much clearer and safer.

For example, we can decrypt and verify a message which carries its own 'data keys' (encrypted with a 'master key'):

    def decrypt[T](
      key : Bytes[Plaintext, Key[Encryption]],
      data: Bytes[Encrypted, T],
    ): Try[Bytes[Plaintext, T]] = ...

    def validated[T](
      key : Bytes[Plaintext, Key[Signing]],
      data: Bytes[Plaintext, T],
    ): Try[Bytes[Plaintext, T]] = ...

    def decryptMessage[T](
      master : Bytes[Plaintext, Key[Encryption]],
      dataEnc: Bytes[Encrypted, Key[Encryption]],
      dataSig: Bytes[Encrypted, Key[Signing   ]],
      message: Bytes[Encrypted, T              ],
    ): Try[Bytes[Plaintext, T]] = for {
      encKey <- decrypt(master, dataEnc)
      sigKey <- decrypt(master, dataSig)
      raw    <- decrypt(encKey, message)      
      valid  <- validated(sigKey, raw)
    } yield valid

It is confusing that people talk about C in terms of being a typed language. The types are an annotation so the compiler knows how much memory to allocate / read. All the side benefits of having to spell that out are incidental.

That sort of typing is a distraction. A programmer does not care that they are passing 4 bytes as the first argument. That isn't really trying to squeeze anything useful out of a type system. They care that they are passing something recognised as an ID, not that it has 4 bytes.

Using C/C++ as an example of a typed language doesn't say much. The guarantees are more a culture of how to use the language than in the language itself.

  struct ID { uint32_t val; };

  void check_id(struct ID id);

I never understood this, how is it possible that a developer calling a function doesn't test if they are calling it right. Once they test it, how does the additional compiler test matter?

Unless of course your calling function is mutating the value of a variable across so many types that it's impossible to realistically test it. It would then make sense to use a type to prevent the variable from being something it shouldn't.

But most people, don't have that situation most of the time. Even in a big application.

May be that explains why something like Python has such wide adoption.

Of the all the constraints I've faced while developing software, reusing variables for many purposes in one code block isn't one of them.

Maybe I am just more absent minded than most dynamic programmers, but live IDE diagnostics and autocomplete are very important for me.

And if you have, you probably know that its awesome but for most purposes doesn’t compensate for the weak ecosystems of the languages that have it, compared either to poorer (higher cost to satisfy for lower benefits) statically typed languages or dynamically typed languages with or without available optional static type checking.

Less opinionated/experienced with strong type systems though, I do appreciate the flexibility of being able to hack things together to see if they work - then improving upon the solution. strong types seem to make that harder.

I often build systems, then break them up to take the working parts out, then put a subset back together.

The type system is like magnetic Lego that lets you snap pieces together.

With a dynamic system it is really difficult to break things apart and then put them back together (without ever running the code, and with 100% accuracy and immediate feedback).

Dynamic languages can be very productive for small teams.

I definitely liked the productivity of F#, but the ecosystem was lacking. Often in dynamic languages the ecosystem more than makes up for the lack of compile time checks.

I've never been able to make heads or tails of Girard's paper but http://pauillac.inria.fr/~fpottier/slides/fpottier-2007-05-l... is an introductory divulgation of some of the pre-Rust history of this stuff. It makes me think I should read Wadler's 01990 paper.

So I think functional languages kind of had a head start, but also in a sense they have an easier problem to solve. ML's type system doesn't even have subtyping, much less mutation and resource leak safety.

OCaml and Rust would fit the bill. You can do procedural and OO in both. You could add Scala to that. I think the conclusion is not that "functional is better than OO" but "they both have their strength and are not mutually exclusive".

[1] https://nim-lang.org

To give an example from Haskell:

You can stick everything into IO, or you can be more careful with your effects.

Or to give a more main stream example:

Even in a language like Go, you can do lots of runtime casting (which is essentially dynamic typing), or you can try to cajole the type system into recognizing what your are doing statically. You can do the latter effort incrementally.

What about the decision whether a function argument is an integer or a string? A floating-point value or an optional floating-point value? How could the compiler ever decide that?

The size of an integer is only a very, very small part of a modern static type system.

If a number is an int, float, large or what have you, the question I would ask is how many decimal points should be printed or if you divide 2 with 2, do you get 1 or 0.99999. A good language/compile should have sane defaults that works almost every time, while giving me options to override it if in testing it get an unexpected result where a human would disagree with the computer.

The source of tests should be the customer requirements, and types is just an set of automatic tests being applied on the code which has no direct connection to the customer requirements.

I have no idea what you're going on about. The customer doesn't care about the particulars of your unit or integration tests, instead they care about whether your software works properly. A modern, powerful static type system is one tool that helps a developer achieve that goal of properly working software, and the developer absolutely should care about whether that function argument is an integer or a string.