A Gentle Introduction to Liquid Types
goto.ucsd.edu
goto.ucsd.edu
With Liquid Haskell, the bound checks are moved from runtime to compile time, semi-automatically handled by SMT-solvers. Static types help programmers to write correct programs faster, and the programs also run faster.
As an aside, speeding up programs with static analysis (constrained dynamism) are also present in Mojo (a variant of Python) or Swift [2].
[1]: https://github.com/Gabriella439/slides/blob/main/liquidhaske... "Scrap your Bounds Checks with Liquid Haskell"
[2]: https://github.com/modularml/mojo/discussions/466 "Mojo and Dynamism"
You could theoretically remove all the unnecessary checks by hand, but it’s tricky.
Automatically removing bounds checking where safe seems like a good goal, but in all but the most trivial cases its quite complex for the compiler to decide what's safe.
Also if you are accessing a data structure out of bounds, that's a show stopping bug that shouldn't ever happen. That should be something that is prevented ahead of time, not recovered from. Tanking a program's performance to catch bugs in released programs is a bold strategy to begin with.
Haskell can do that too, with unsafeTake and unsafeDrop [1], you can opt-out of runtime bounds checks as easily as C++, Rust, and Julia. But opting-out of machine-checked guarantees is unsafe, and Haskell can do better.
What's more is that Liquid Haskell can move such runtime bounds checks to compile time, made ergonomic with SMT solvers. This way, you get machine-checked guarantees (so programmers can write correct programs faster) with no runtime overhead (the programs also run faster).
> Also if you are accessing a data structure out of bounds, that's a show stopping bug that shouldn't ever happen. That should be something that is prevented ahead of time, not recovered from.
Yes, that's the whole point of static analysis, to move runtime checks to compile time, made ergonomic with Liquid Types. Note that this compile time verification is much stronger than testing at development time, which is what C++, Rust, and Julia offer without Liquid Types.
> That should be something that is prevented ahead of time, not recovered from.
Gabriella Gonzalez agrees with you, and calls this idea of "Prefer pushing fixes upstream over pushing problems downstream" as "The golden rule of programming", at the 1st slide in her talk [2]. Her talk shows you how such machine-checked verification is possible with Liquid Haskell, which is not possible in C++, Rust, and Julia without Liquid Types.
[1]: https://github.com/Gabriella439/slides/blob/main/liquidhaske... "Documented preconditions - Scrap your Bounds Checks with Liquid Haskell"
[2]: https://github.com/Gabriella439/slides/blob/main/liquidhaske... "Scrap your Bounds Checks with Liquid Haskell"
Move simple runtime checks to compile time. Most if not all of these can be done with iterators or simple loops. Index lookups that depend on use input and data still have to be checked somewhere and should probably be done outside of a loop anyway.
Static analysis is always welcome, but this seems like a lot to go through to solve something that isn't much of an issue in the first place. Bounds checks were a big problem in C, but with more explicit iteration and runtime checks they are almost never a problem I see anymore.
If you can afford runtime checks (if they are not much of an issue in the first place), Haskell, C++, Rust, and Julia can have runtime checks.
If you want to opt-out of runtime checks, Haskell, C++, Rust, and Julia can opt-out easily, but unsafely.
So, to be fair to Haskell or any of them, this is not a complication specific only to Haskell.
But Haskell enables you to do more. With around 30 lines of Liquid Haskell (those lines between {-@ and @-}), you can upgrade simple runtime checks to compile time verification, see the full example [1]. You get great efficiency, safely, without unnecessary runtime checks, which is impossible in C++, Rust, or Julia without Liquid Types.
[1]: https://github.com/Gabriella439/slides/blob/main/liquidhaske... "Full program - Scrap your Bounds Checks with Liquid Haskell"
Just a side-note, julia's compiler is very adept at proving situations where array accesses are guaranteed to be in-bounds and will automatically elide bounds checks in those situations. Not saying this to take anything away from Liquid Haskell though, that seems like a very cool project with some very cool capabilities.
These aren't the same though. Runtime checks can check every access, static analysis isn't going to deal with every scenario like access based on inputs. Lumping them together is pretending there is something happening that isn't.
Not true.
With proper abstraction, static analysis proves the absence of bugs under all relevant inputs, which gives stronger guarantees than testing at development time (which only checks for a number of inputs), and is more performant than runtime checks (without runtime overhead).
An example is the use of a static type system, which eliminates a whole class of bugs under all possible values of the same type (those due to type mismatch: passing a string into a function expecting an integer), which is stronger than a test suite testing certain inputs, and does not have runtime overhead (unlike runtime checks).
Going further than usual type systems, people study refining types by possible values (such as positive integers as a refinement of all integers), and the result is refinement types, which is what Liquid Types is based on. And you can see this in action in the full example [1], such as { n : Int | 0 <= n } which refines n to be a positive integer.
[1]: https://github.com/Gabriella439/slides/blob/main/liquidhaske... "Full program - Scrap your Bounds Checks with Liquid Haskell"
> Lumping them together is pretending there is something happening that isn't.
Correct. Static analysis can formally verify correctness under every access (like runtime checks), but without the runtime overhead (unlike runtime checks), with proper abstraction.
That is why compile time verification is an upgrade from runtime checks.
OK, in reality the trick is that static analysis removes redundant runtime checks (check exactly once), see the sibling discussion https://news.ycombinator.com/item?id=37376963.
PS: I figured that setting the record strict may benefit others, such as https://news.ycombinator.com/item?id=37380176, hence this reply.
Seems like back peddling. If part of the input data you have is there to reference input data in an array in C++ you would check the input data ahead of time so that you don't have to check the range in the middle of a loop.
Then you can have performance that is compared to everything else instead of just other programs written in the same language.
The toy example/implementation is chosen to teach Liquid Types/Haskell, and therefore, covers the simple case of parsing UDP headers, which is a simple bounds check of "indexing an array of length 8". In this case, there is no need to use Liquid Types in any language.
But in the real use case (not shown in the talk), when you want to dig deeper into the content of UDP packets and such, you will need more than "check the input data (length) ahead of time so that you don't have to check the range in the middle of a loop". This is the case where Liquid Types help, tremendously. But such complicated cases, where Liquid Types shines, may not fit into a talk.
Again, it is not fair to think this is a complication specific to Haskell.
It is fair, because other languages have a single keyword or setting instead of a whole 'type system' with fancy labels that introduces lots of complexity for something simple.
https://news.ycombinator.com/item?id=37382215
If you can afford runtime checks (if they are not much of an issue in the first place), Haskell, C++, Rust, and Julia can have runtime checks.
https://news.ycombinator.com/item?id=37387005
It is unfair to not respond to these points, while calling "this is a complication specific to Haskell." Because Haskell can do what others can do, but more.
> lots of complexity for something simple.
"indexing an array of length 8" for parsing UDP headers is simple, agreed. And Haskell can handle that simply with unsafe functions.
Keeping track of properties with complicated dependencies to parse arbitrary content (perhaps application-logic-aware contents, not "just checking for array length before indexing in a loop") is simple, strong disagree. See the section "Beyond Bounds Checking" of "Why Liquid Haskell matters" for example [1]. This might be used to implement the kinds of static analysis needed for Mojo or Swift, for instance [2].
[1]: https://www.tweag.io/blog/2022-01-19-why-liquid-haskell/ "Why Liquid Haskell matters"
[2]: https://github.com/modularml/mojo/discussions/466 "Mojo and Dynamism"
https://news.ycombinator.com/item?id=37376010
And this cannot be solved in the languages you mentioned, either. Unless you are claiming that all parsing checks are "just checking for array length before indexing in a loop".
https://news.ycombinator.com/item?id=37416911
We are still working on completely different scales.
The talk is using a toy example to show what is possible, and you are using that example too literally, without realizing that the example scales to handle much more complex problems that people need to deal with in the real world.
It is unfair to lump together different scales and call them the same thing.
Your reaction is like criticizing a talk which shows sorting 5 numbers with programs, with "it is simple to sort 5 numbers by hand, why bother learning how to program with all those complexity for something this simple?". And then calling this fair, without realizing that in the real world, sometimes people need to sort way more than 5 numbers.
Your reaction is like criticizing a talk which shows sorting 5 numbers with programs,
It is not, because this idea that loops and bounds checks are complex in the first place isn't true. No other language invests so much complexity in a simple problem that is barely an issue in modern languages in the first place.
In Haskell you can provide your own prelude, which uses unsafe functions without bounds checks by default. With CPP or cabal tricks, you can easily have bounds checks only for debug builds, but no bounds checks for release builds. No Liquid Types needed either, though this is only testing in development but not compile time verification. Less safe, but same effort and result as in C++.
> loops and bounds checks are complex in the first place isn't true.
Liquid Types people are not thinking about only loops and bounds checks. They think about any properties, for example:
* pointer checks, a huge security issue, see [1] or https://news.ycombinator.com/item?id=37376946.
* valid expression checks (Beyond Bounds Checking of [2]), see https://news.ycombinator.com/item?id=37440896.
[1]: https://arxiv.org/abs/2207.04034 "Flux: Liquid Types for Rust"
[2]: https://www.tweag.io/blog/2022-01-19-why-liquid-haskell/ "Why Liquid Haskell matters"
Sorting 5 numbers is complex in the first place isn't true. No other technology invests so much complexity as programming and this is barely an issue in modern problem solving in the first place.
My previous comments already said loops and bounds checks are not complex in the first place, and they are only used as a demo in a talk, as Liquid Types scale to handle any properties.
I am pretty certain that a reasonable reader can judge the merits of our arguments and make up their mind. Thanks for raising a possible misconception about Haskell and Liquid Types, and this forced me to strengthen the arguments to clear the misconception for other reasonable readers. This is my last comment on this issue.
So your solution is to rewrite the haskell standard library yourself?
With CPP or cabal tricks
I don't know what cabal tricks is supposed to mean here.
> So your solution is to rewrite the haskell standard library yourself?
No need--the simplest solution is to pass flags correctly to the dependent library (set all flags BoundsChecks, UnsafeChecks, and InternalChecks to false for the vector package) for release builds.
Explanation after the one-line rebuttal:
This exactly matches the C++ behavior: no bound checks for release builds, but with bounds checks for debug builds. This shows that Haskell can do what C++ can do, very easily, just by passing compiler/cabal/stack flags correctly.
With this rebuttal, I think the issue is completely settled, at least for reasonable readers.
If performance is an issue, why not just do this instead of dealing with "Liquid Types".
And with this rebuttal, the issue is now settled for all readers and even those differently abled that are using voice readers.
For pointer checks.
And with this rebuttal, the issue is now double settled.
Excellent question. The existing "mainstream" languages (C++, Rust, Julia, or even Haskell) are not designed to support such applications as bound checks or pointer checks in their type systems, where certain runtime values/checks can affect/refine the types (as in dependent types, or refinement types which Liquid Types are based on).
It could be done (in some newer languages: Idris, Agda, Lean), but people have not figured out an ergonomic way to do so in more "mainstream" languages (e.g., not hurting type inference, light and maintainable proof burdens). Liquid Haskell is an attempt to bring refinement type to Haskell, as an extrinsic type system.
> C++ could build in pointer checks and it would be in the same type system.
People have not figured out how to do so (and I doubt if it is possible at all). Suffice to say:
* Rust is an attempt to build in lifetime-and-ownership-checks into the type system of C++, and Rust is a new language,
* Liquid Rust is an attempt to add pointer checks (refinement types) to Rust, and even then it is not built into Rust.
> And with _this_ rebuttal, the issue is now double settled.
People are still researching solutions (see Rust and Liquid Rust over C++) to this problem (building in pointer checks in the type system of C++), and they would love it if you can settle their problems.
> I don't know what cabal tricks is supposed to mean here.
It means conditional compilation, i.e. what the C preprocessor also gives you (but somewhat more principled).
(edit: I am Australian - this is sarcasm.)
On a complicated single-process system (a monolith), that can be a big improvement. For distributed systems (microservices), there will still be plenty of runtime checks. That's what allows each service to have its own lifecycle.
I'm wondering what Elixir will do, given that Erlang servers divide up work into lots of independent, distributed components?
Libraries for TypeScript like Zod[0] allow you to define types and validate against them at runtime. For example: https://stackblitz.com/edit/typescript-pwzng4?file=index.ts
At work, we use this to validate API responses, local storage contents, or URL/router data. It solves the problem you described -- dirty data coming from outside of the type system.
This isn't a plug for Zod or TypeScript -- there are other similar libraries for TypeScript, and I would imagine other statically typed languages have something to fill this role.
[0]: https://zod.dev/
- why is there `safeParse`? Why is parsing not safe by default? Have an `unsafeParse` as an escape hatch, not the other way around.
- the imperative check for `success` seems unfortunate. Then again, I am familiar with pydantic from Python, which seems awfully similar to Zod, and there you'd get an Exception on failure which you'd have to catch. Some would call that uglier (I don't). Point being, I don't know of a better alternative, but sure hope native liquid/dependent types wouldn't suffer from this anymore
> the imperative check for `success` seems unfortunate.
Yes! But the cool thing is that `success` being `true` narrows the type and ensures that the `data` field is present. You cannot access `data` without checking for success
Here's another example showing those concepts: https://stackblitz.com/edit/typescript-wbuikz?file=index.ts
Result
.fromExecution(() => schema.parse(input))
.mapOk(parsed_input => /* */)
.mapError(parse_errors => /* */)
I would hazard a guess that `parse` is a function introduced long ago, and `safeParse` came afterwards. (Safe as in "do not throw an exception")
`const result = schema.safeParse(val)` does not throw, and if `!result.success`, `result.error` contains a `ZodError` object with all the details. It's common to further transform this error object, perhaps to make it more human-readable or suitable for an API response.
Typically I find myself using `safeParse` far more than `parse`.
Glad to not have to write seemingly thousands of lines of interfaces and type guards any more.
[1] https://lexi-lambda.github.io/blog/2019/11/05/parse-don-t-va...
Validating on load is a form of runtime type checking. It’s not relying on a type assertion, which would allow the compiler to omit the type-check and the programmer to omit the error-handling when the check fails. Instead of preserving knowledge about the type constraints on some data, we prove the constraints again.
You can totally do that if you want and you fully trust that the thing that generated the data used the right types.
Similar to what you do every time you use a dynamic library.
A type system should be designed based on what invariants you can reasonably expect in the environment where the system operates, versus what you can expect will change. For example, Swift has an interesting type system that's specifically designed to make upgrading dynamic libraries easier and safer. Apple can make assumptions about what sort of changes there will and won't be to the libraries that ship with its operating systems.
Contrast with Go, which makes the assumption is that there are no dynamic libraries; all Go source code is seen by the compiler.
It's not not verifying input. It's just not verifying the types. There are no additional security issues unless you do something spectacularly stupid.
For example you can load random bytes as a Protobuf file. It never validates that the data is in fact the right type, it just loads the bytes assuming they are. The only issue you'll see is junk data (or a parse error).
In any case I think you know that. I'm not really sure what point you were trying to make.
The way I see it, types are sometimes used to represent security guarantees, so junk data getting cast to the wrong type could be a security hole. It depends on which types we're talking about.
For example, you can declare a SafeHTML type whose underlying representation is a string, and if you somehow deserialize it without actually doing the runtime check, it could be a security bug.
Verifying types and verifying input are closely related; the type keeps track of what you verified when you constructed an object of that type.
...I'd read that article. Now I'm starting to think that I should write it.
What if instead of TS, you had plain JS as-is (no compiler needed!) then just add comment like:
/* x is an integer, x > 0 ... */
function add(x, y) { ...
This link discusses, along with other usefulness, the use of dependent types for the likes of array bounds checking, creating "sorts" of numbers satisfying various predicates, and more interestingly the utility of combining Dependent Types with PURE Linear Types (as opposed to Affine Types a la Rust lol) ...
AFAIK, DML was originally intended as a gateway to constructing typed assembly (DARPA)
Having been a researcher in this field for, this quote from the original paper on 'Liquid Types' originally seemed suspect.... "a system that combines Hindley-Milner type inference with Predicate Abstraction to automatically infer dependent types precise enough to prove a variety of safety properties" .... type inference with dependent types is undecidable!? Then again, this is one of the primary focuses of ATS3, I digress...
This area of research is rich in pipe-dreams and cerebral investigation. From my biased opinion, ATS3 (the current version of the successor3 of DML, being actively devloped by the PI Professor Hongwei Xi, is the only language in this space to be aimed at practical programming in industry... of course there is always haskell....)
What I mean is that, TEMPLATES++ Linear Types + Dependent Types + Hindley Milner Type inference ....
At the end of the day, who doesn't love hyper efficient meta-programming magic?
Now, the Liquid Haskell type checker knows a lot about integers: it can deduce whether your invocation of x-y could give a negative result or not.
If it cannot result in a negative value (e.g. because y has a particular liquid type), the type checker will accept your program, and you'll be able to use the subtraction function as you normally would.
However, if your invocation of x-y could potentially result in a negative value (e.g. because the values of x,y are not sufficiently constrained), the type checker will reject your program, and tell you to do something else (e.g. add runtime checks that rule out a negative result before invoking the subtraction).
I think that's the main disadvantage of these types. It's now no longer a decidable system with (usually) easy to understand error messages. You're in the realm of formal verification and SMT solvers where the only error message is "didn't work; try something else".
Probably still good but I think it's important to be aware of the downsides.
As long as there's a good way to map from the real world into the liquid typed world, either by just telling the liquid typed world that this value can only be a positive integer, or by the liquid typed world having functions that ensure this at the boundary and return nice errors on runtime failure of this check, then it sounds really interesting.
For subtraction it becomes fallible. You get back an option, 5 - 4 = Some(1), 4 - 5 = None. (These are probably methods rather than infix operators, your language may vary.)
See for example the methods documented here: https://doc.rust-lang.org/stable/std/num/struct.NonZeroUsize...
As for it's utility, it enables certain optimizations and it can be a useful constraint. Eg for ensuring that you don't divide or multiply by zero.
That’s not true. 5 + (-5) is 0.
One usually doesn't program airplane control software. Some bugs are okay and speed of development and correctness is a trade-off.
Certainly, the language itself won't have liquid types in the compiler any time soon. But taking Ruby's Sorbet as inspiration, I wonder if you could annotate Java classes with semantic rules and do some simple verifications as a post-compilation step?
There's too much bad code out there that can't just be replaced. Tools to help make it better would be a huge boon.
[0] https://www.thestrangeloop.com/2018/contracts-for-getting-mo...
Personally I find refinement types more promising than dependent types, but time will tell.
LSPs are a sort of middle ground where I derive a lot of the value from, and even weak type systems really help the analysis of the LSP. They make me immensely more productive. I've been writing a lot of TypeScript lately, and I may not have taken the plunge to the frontend if LSPs hadn't made it dramatically easier. So in that case the impact on my productivity is +100%.
Especially when dealing with someone else's code, it can be a ceremonial roadblock for moving data around that could easily be processed without issue but artificial constraints in order to service an orthodoxy have been erected to make things needlessly more brittle.
In the name of reducing bugs, it in practice can often impose workarounds that increase the chance the confusion and more bugs.
This isn't always the case but most of the time things are built in incompetent hands (including my own). Accommodating for this reality is probably the better move. Imposition of absolutist rules doesn't improve the behavior of incompetency, it just makes it more subversive and the defects it creates more insidious
I try not to think of it as whining, I try to think of it as the type checker helpfully pointing out, "hey, you haven't done your due diligence here." I ran into this advice when learning Rust and it made it much more pleasant. In the same way that you might chase a series of failing tests to finish a task, you can chase a series of type warnings. I have them integrated into my editor, so usually the "series" is length 1 because I'm knocking them out as they arise.
The projects I've worked with that were previously untyped (Python SaaS backends), I never actually ran the type checker on the entire source code. But if I had a megabyte of type errors that weren't really going to provide value to the business, yeah that would feel like a waste of time. Unless we were having serious quality issues and it was between annotating the entire codebase and rewriting it from scratch, I'd leave well enough alone - it's a tool, not a religion. But the new code I'm checking in, that's getting annotated for sure.
I don't really see types as arbitrary absolutist rules. They're the semantics of the language, they're the rules you're following whether you have a typechecker or not. And if you violate them you're gunnuh get bugs and crashes regardless. I too am a flawed human who frequently makes mistakes, and I like having them caught immediately, before I've even checked my code in.
It can be both true that you haven't done your due diligence in the sense that you haven't taken care to handle the entire range of all possible inputs (or at least it seems like it when you're only considering the relevant code section in isolation), but that as a practical matter it's just not a problem. So the time spent attending to what it's nagging you to take care of is time wasted. Alternatively, you're not even a program maintainer, you're just trying to build the thing so you can run it for your own one-off use, but it refuses to even build because of some issue totally irrelevant to what it is you're about to use it for.
The approach that the TypeScript compiler takes is beneficial—complain about the thing, but don't let that prevent the compiler from running to completion and still producing a program that you can run.
There's type narrowing, warnings for not awaiting Awaitables, generics, protocols/interfaces, even whole function signatures can be lifted into the type system (ParamSpec). Dictionaries can be turned into typed versions as well, which is quite powerful.
By default, it's weaker, as it's optional. Rust's isn't optional. But start a greenfield project with type annotations and you won't be disappointed, I hope. Features like ADTs will still be absent, but I feel that's an orthogonal issue to type safety etc.
"Python wasn't weak, I was" is an accurate call out, I have abandoned complex annotations in Python before because there was a bunch of cruft involved and the documentation wasn't always as straightforward as I'd have liked.
The lack of ADTs (and thus a Result type) combined with the lack of exceptions as part of the function signature (like Java) make it really difficult to develop a reliable application. Maybe that's not quite a type safety issue, but it's a safety issue for sure.
Digging in, it looks like people are working around this by returning exceptions, which is very clever and I hadn't thought of it before.
Never came across that! Where did you find that? Pretty interesting. I'd call it deeply unpythonic and very surprising. It's unfortunate to not have a Result type/idiom, but IMHO faking it isn't the way to go either.
https://stackoverflow.com/a/74192977
I haven't given it a great deal if thought, it would probably mess up stack traces if you raised it later, which would be bad. I agree it's unpythonic, I've implemented result-like types in Python before which has worked okay.
ETA: That answer links to this library, which uses a decorator technique that I think will still print a good traceback if you reraise it later (definitely will if you use `raise ... from` syntax), but I can't test it at the moment:
https://github.com/alexandermalyga/poltergeist
I probably won't be putting this into production anytime soon but it was charming food for thought.
you either create a complex program, with simple language constructs, or a simple program with complex language constructs
if you use only the simplest types, your program will be very complex, if you put all the complexity in types, your type system will be very complex
you need to balance things
i think OOP is the extreme where all the complexity is in the Object system (or types system)
functional programming with complex types, i think give a nicer balance, where some logic goes into the program flow and some goes into types
anyway balance is everything, keep things balanced , dont lean too much into any direction
Languages and libraries that are well designed and compose better end up being great at taming complexity. Functional languages are made to compose better, but their popularity is a hint that they are not that simple to use effectively.
do you have any in mind ?
I often think about a blend of functional and OO as 'protocol graphs' but it's a faint idea for now
The thing is that while array programming is amazing for some specific problems it's not going to help you make sure that you don't have memory errors, race conditions, or wrong states in your program
> it's not going to help you make sure that you don't have memory errors, race conditions
Not only does array programming abstract away those kinds of errors, it also solves the problems that I care about: code that is faster, more expressive, and easier to read and write. Unwieldy type systems do not help me solve those problems.
You needn't use your real name, of course, but for HN to be a community, users need some identity for other users to relate to. Otherwise we may as well have no usernames and no community, and that would be a different kind of forum. https://hn.algolia.com/?sort=byDate&dateRange=all&type=comme...
Also, can you please stop using trollish usernames? Those aren't allowed either, because they basically troll every thread the account posts to. https://hn.algolia.com/?sort=byDate&dateRange=all&type=comme...
Type systems also enable optimizations that do make code faster. Eg monomorphization. Consider that JITs makes untyped code faster by inferring types at runtime. See also this comment https://news.ycombinator.com/item?id=37376010
I honestly can't make heads or tails of the idea that type systems don't help make code more expressive? They allow you to express your precise intent in a way that's legible to both humans and the compiler.
Was it ever not?
As a library developer I spend substantially more time on making the types right than the actual implementation. It can be brutal at times.
I guess you can complain that Typescript forces you to spend time making sure your code meets a minimal bar of correctness...