Avoiding space leaks at all costs(kodimensional.dev)
kodimensional.dev
Avoiding space leaks at all costs
https://kodimensional.dev/space-leak
46 comments
It is Haskell, just not Haskell2010. It is included in GHC2021.
Would you say that a company that uses a subset of C++ as a policy is not using C++? Their definition of the subset is as arbitrary as another’s that uses BangPatterns.
Would you say that a company that uses a subset of C++ as a policy is not using C++? Their definition of the subset is as arbitrary as another’s that uses BangPatterns.
I believe one could avoid BangPatterns by using seq.
I think his point is that a naive implementation of an extremely simple function uses 35x more memory than the 'optimal' implementation, which is not at all obvious without knowing internal details of the compiler.
Sure, this language feature is mainstream now. But most real-world haskell codbases vary greatly in which language extensions they've chosen to use.
Is this a C++ benefit, that the competing implementations GCC and LLVM give each other's extensions to the language more legitimacy by having a second implementation? It doesn't seem like language extensions get so much attention in C++ codebases (but maybe I'm mistaken)?
C and C++ programmers who write only or mostly for a single platform often have some extensions they insist on, and may even regard the standard language (lacking these features) as defective.
Linux won't compile in MSVC. Obviously it needn't, it's a kernel and not a Windows program, but it also can't, because it needs GNU extensions, which are now also implemented for Clang.
If you write broadly cross platform software for C++ you are committed to supporting all three compilers and so that's effectively just the standard. Indeed if all three do or don't do something de facto that is the standard. The ISO document used to say C++ has optional garbage collection, it doesn't because none of the three compilers have that. The ISO standard still doesn't say pointer provenance is a thing, but all three compilers require it, so in fact it's a thing.
Linux won't compile in MSVC. Obviously it needn't, it's a kernel and not a Windows program, but it also can't, because it needs GNU extensions, which are now also implemented for Clang.
If you write broadly cross platform software for C++ you are committed to supporting all three compilers and so that's effectively just the standard. Indeed if all three do or don't do something de facto that is the standard. The ISO document used to say C++ has optional garbage collection, it doesn't because none of the three compilers have that. The ISO standard still doesn't say pointer provenance is a thing, but all three compilers require it, so in fact it's a thing.
> The ISO document used to say C++ has optional garbage collection
I'm intrigued; what was this feature supposed to look like?
I'm intrigued; what was this feature supposed to look like?
https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n26...
The standard basically said, if your compiler supports this new feature, here's how you use it in your programs. And no compilers have ever supported the feature. So C++ 20 rips it back out.
The standard basically said, if your compiler supports this new feature, here's how you use it in your programs. And no compilers have ever supported the feature. So C++ 20 rips it back out.
GHC language extensions are like C++ Technical Reports. They are a way to evolve the language without excessive red-tape on an opt-in basis, and periodically the ones that worked out make it into a new “standard”.
It’s not really all the different from Python PEPs or Rust’s various proposals / nightly stuff.
Just about every language has some form of this.
It’s not really all the different from Python PEPs or Rust’s various proposals / nightly stuff.
Just about every language has some form of this.
That's true. And when somebody says "oh you can just use GCC's special feature to make C++ tolerable" they rightly get pushback because only a portion of these features ever make it into a standard. This means that over time you end up with a codebase that is forever tied to these nonstandard features and might even end up being incompatible with major language changes. Or maybe the feature you adopted does get standardized but in a slightly different and incompatible way and now you've got a big problem on your hands.
It is good for people to have ways of playing with a language so it evolves. It is (IMO) bad to suggest the adoption of these things as a means of making a language production-ready.
It is good for people to have ways of playing with a language so it evolves. It is (IMO) bad to suggest the adoption of these things as a means of making a language production-ready.
Just about every successful C/C++ project uses flags or intrinsics or whatever to some degree. It’s not unusual or particularly problematic. Linux famously uses a bunch of GCC stuff (though I gather that’s in decline) and non-standard defaults like ‘-fno-strict-aliasing’.
And the extensions in the article are for the most part in the GHC2021 standard. So it’s more like, this is in C++20.
Younger languages like Rust or Kotlin or something have less pressure for mechanisms like this because you have to get to a certain point in the language lifecycle before the tradeoff tilts from “let’s do a breaking change at a point release” to “let’s flag in the new behavior”, but if (as I expect) e.g. Rust lasts as long as C++ or Haskell, it will be be bedazzled with all kinds of flags and extensions and stuff too, how else do you grow a language for 40 years?
And the extensions in the article are for the most part in the GHC2021 standard. So it’s more like, this is in C++20.
Younger languages like Rust or Kotlin or something have less pressure for mechanisms like this because you have to get to a certain point in the language lifecycle before the tradeoff tilts from “let’s do a breaking change at a point release” to “let’s flag in the new behavior”, but if (as I expect) e.g. Rust lasts as long as C++ or Haskell, it will be be bedazzled with all kinds of flags and extensions and stuff too, how else do you grow a language for 40 years?
> how else do you grow a language for 40 years?
You could introduce things in one version of the standard, and then remove them in the next release if people didn't like it, such that to get the feature you have to specifically use --std=2022a, no earlier, no later; and if you want your project to track the evolution of the language rather than getting stuck on a particular version, you'd have to rip use of such features back out.
You could sort of think of it as if you had a dependencies lockfile which includes a version-constraint on the major version of the compiler; where every time there's a change in what code is accepted, that major version of the compiler goes up. Any change other than changing what code is accepted as valid goes into the lower semver fields; and gets replicated across all major versions of the compiler, where relevant. (Except that you don't have to do this by actually maintaining separate major-version branches of the compiler for every standards release; you can do it the way we currently do feature flags, only the only flag you get is which standard-version your project is written for.)
I don't think I've ever seen anyone actually do this for a language/compiler; but it's a common approach used for protocol and API-client libraries, and seems to work well there.
You could introduce things in one version of the standard, and then remove them in the next release if people didn't like it, such that to get the feature you have to specifically use --std=2022a, no earlier, no later; and if you want your project to track the evolution of the language rather than getting stuck on a particular version, you'd have to rip use of such features back out.
You could sort of think of it as if you had a dependencies lockfile which includes a version-constraint on the major version of the compiler; where every time there's a change in what code is accepted, that major version of the compiler goes up. Any change other than changing what code is accepted as valid goes into the lower semver fields; and gets replicated across all major versions of the compiler, where relevant. (Except that you don't have to do this by actually maintaining separate major-version branches of the compiler for every standards release; you can do it the way we currently do feature flags, only the only flag you get is which standard-version your project is written for.)
I don't think I've ever seen anyone actually do this for a language/compiler; but it's a common approach used for protocol and API-client libraries, and seems to work well there.
What's insufficient about the way Rust's already growing?
> Or maybe the feature you adopted does get standardized but in a slightly different and incompatible way and now you've got a big problem on your hands.
Haskell is de facto a single-implementation language: an extension is not going to be made incompatible with some future Haskell standard without having been long since deprecated in GHC.
Haskell is de facto a single-implementation language: an extension is not going to be made incompatible with some future Haskell standard without having been long since deprecated in GHC.
If our arguments can be so pedantic, then I can say: sure, Haskell is forever a research language. But that's irrelevant to me, because I use no-longer-really-Haskell.
I agree with the original critique more than this defense, but I had to upvote the perfection of the expression of the point I don't agree with.
> I think Haskell is a fine language, but given the amount of various language extensions that are needed for it to be usable in production without friction, it seems it will forever be a research language - and there is nothing wrong with that.
This isn't really a problem in the real world, only in theory.
This isn't really a problem in the real world, only in theory.
I have tried and failed to comprehend Haskell for my own uses. Certainly most of the examples in this article are meaningless gibberish to my smooth, practical brain. So kudos to those who grok the underlying math and find it a useful language.
That said, I am a fan of lazy evaluation in the right circumstances, and often find myself implementing lazy data structures (particularly in my Ruby programs).
Ultimately, though, I’m confused by this article which seems to be all about removing laziness from Haskell in ways that seem entirely counter to what I take to be its purpose. If you have to bend over backwards and rewrite your code in awkward ways to get the language to behave the way you want, maybe the project itself isn’t right for the language? It’s one thing to hack around a couple of isolated space issues in a larger context, but if you are enabling language extensions across the board by default to prefer eager evaluation, why are you using Haskell at all?
That said, I am a fan of lazy evaluation in the right circumstances, and often find myself implementing lazy data structures (particularly in my Ruby programs).
Ultimately, though, I’m confused by this article which seems to be all about removing laziness from Haskell in ways that seem entirely counter to what I take to be its purpose. If you have to bend over backwards and rewrite your code in awkward ways to get the language to behave the way you want, maybe the project itself isn’t right for the language? It’s one thing to hack around a couple of isolated space issues in a larger context, but if you are enabling language extensions across the board by default to prefer eager evaluation, why are you using Haskell at all?
Some would argue that in retrospect strict should have been the default, I sympathize with that view personally somewhat.
But every language that beats Darwin for long enough does so by iterating the stuff that can be “fixed” and developing conventions and tools for the gotchas that can’t (practically) be.
But every language that beats Darwin for long enough does so by iterating the stuff that can be “fixed” and developing conventions and tools for the gotchas that can’t (practically) be.
I would love a strict Haskell, but I don't think functional programming would have been the same today. Laziness was the big motivator for keeping the language uncompromisingly pure, and we may not have developed monads and similar without it.
I've never heard of or met a circumstance that actually warranted Haskell. Not even in finance, those guys are totally wasting their time.
It's an attractive (I'm going to say "it seems" though you ought to take that for granted) intellectual rathole. Like some forms of philosophy, like many things. Thank GOD (despite being unbelieving at the time) I went with Lisp instead. And I got real speedups.
It's an attractive (I'm going to say "it seems" though you ought to take that for granted) intellectual rathole. Like some forms of philosophy, like many things. Thank GOD (despite being unbelieving at the time) I went with Lisp instead. And I got real speedups.
Algorithmic finance might be the most ruthlessly Darwinian area of industrial software, it’s way up there certainly.
It’s also notoriously opaque, which makes it especially hard to know that so-and-so is wasting their time.
How about “I don’t like Haskell and/or OCaml because X, I recommend Lisp instead because Y” as an alternative formulation?
That one I’d be interested to read!
It’s also notoriously opaque, which makes it especially hard to know that so-and-so is wasting their time.
How about “I don’t like Haskell and/or OCaml because X, I recommend Lisp instead because Y” as an alternative formulation?
That one I’d be interested to read!
> I've never heard of or met a circumstance that actually warranted Haskell.
Sounds like you are biased to conclude it has no purpose to begin with.
Sounds like you are biased to conclude it has no purpose to begin with.
> Certainly most of the examples in this article are meaningless gibberish to my smooth, practical brain
My smooth practical brain struggles in languages that aren't Haskell.
My smooth practical brain struggles in languages that aren't Haskell.
Wait, the first suggested fix for "lazy evaluation consumes too much memory" is "make it eager"? I'm sure I'm missing some nuance here, but that seems backwards.
Lazily evaluating 2+2 requires storing two numbers and an operator for a while, eagerly evaluating requires storing just one number (4, the result).
This can be a real problem, for example if you read potentially large network input eagerly and use it to compute a small result lazily. IMO it's always a mistake and IMNSHO the proper solution is always something more nuanced than just "make it eager" or "make it properly lazy".
This can be a real problem, for example if you read potentially large network input eagerly and use it to compute a small result lazily. IMO it's always a mistake and IMNSHO the proper solution is always something more nuanced than just "make it eager" or "make it properly lazy".
In all languages, sometimes you want eager evaluation and sometimes you want lazy evaluation.
If things are lazy by default, you simply need to put a "!" before the term to evaluate in order to get eagerness.
If things are eager by default, you need to rewrite the algorithm as a streaming algorithm.
If things are lazy by default, you simply need to put a "!" before the term to evaluate in order to get eagerness.
If things are eager by default, you need to rewrite the algorithm as a streaming algorithm.
> you simply need to put a "!" before the term to evaluate in order to get eagerness
That's wrong. With the bang pattern, evaluation only goes up to WHNF, not the whole way — it's just a syntax sugar around "seq" which has been around since always. If you need to fully evaluate the term, look at the deepseq package: it's implementation is non-trivial.
That's wrong. With the bang pattern, evaluation only goes up to WHNF, not the whole way — it's just a syntax sugar around "seq" which has been around since always. If you need to fully evaluate the term, look at the deepseq package: it's implementation is non-trivial.
There's no inherent reason why doing lazy evaluation in an eager-by-default language should be any harder than the reverse. I can imagine a language that is eager by default but has a '#' operator that holds expressions in unevaluated form until terms are requested.
The combination of lazy evaluation and state mutation/side effects can be pretty difficult to reason about. For example, if you have a function that changes a global variable as a part of a lazy computation, once that function could have been called you have no way of knowing if or when that global variable will change in the future. If you have other functions that depend on the value of that variable, their future behavior is now much more challenging to reason about than in a strict language. You can also imagine something akin to a race condition in which there are multiple lazy computations which could eventually set that variable to different values and the actual sequence of state transitions depends entirely on the dependency order of a possibly unrelated piece of code. In practice, this means that in languages that are strict by default, lazy computations are often forced to run in order to reason about the code, rather than because the actual results of the computation are required.
Since pure functions compute the same results under lazy or strict evaluation and require that any data dependencies they have are explicitly provided as inputs, they interact with lazy computations in a much more tractable way. This means that adding a strictness operator to a lazy language is much easier than adding a laziness operator a a strict language.
An alternate approach is what python did with generators where there is a data type for lazy computation, but it lives apart from the rest of the language, so it is mostly used for e.g. stream processing where a default-lazy approach is conceptually straightforward and is less likely to lead to extremely non-trivial control flow. This approach does, however, basically give up on having a laziness operator that will turn a strict computation into a lazy one.
Since pure functions compute the same results under lazy or strict evaluation and require that any data dependencies they have are explicitly provided as inputs, they interact with lazy computations in a much more tractable way. This means that adding a strictness operator to a lazy language is much easier than adding a laziness operator a a strict language.
An alternate approach is what python did with generators where there is a data type for lazy computation, but it lives apart from the rest of the language, so it is mostly used for e.g. stream processing where a default-lazy approach is conceptually straightforward and is less likely to lead to extremely non-trivial control flow. This approach does, however, basically give up on having a laziness operator that will turn a strict computation into a lazy one.
Sadly even though what you say is true, I do not know of any languages that get that right. Even those that do care enough to give you a way to make lazy values, they require you to explicitly wrap them and force them all over the place, making lazy programming effectively so noisy it's unusable. Which is the main reason why I still find myself coming back to haskell (stuff like parser combinators is just so unwieldy without laziness and do notation...)
Counterpoint: C++ expression templates have existed for a long time, and require no explicit work on the user side. They are incredibly unwieldy to write and debug though.
I am gonna be completely honest, I have never quite been able to understand expression templates properly. is it possible to use them to get hassle-free laziness? is there any example of a library that does this (just to look at it and see how it's done)?
Define hassle-free. In readability of user code? Yes. In getting clean compiler errors? No.
One example would be Eigen: https://eigen.tuxfamily.org/index.php?title=Main_Page
See also their page on lazy evaluation: https://eigen.tuxfamily.org/dox/TopicLazyEvaluation.html
Of course, C++ is powerful enough that using specific tools can make the expression template abstraction fall apart (e.g. auto), but fundamentally you can write matrix code very cleanly and get great lazy-evaluated performance.
One example would be Eigen: https://eigen.tuxfamily.org/index.php?title=Main_Page
See also their page on lazy evaluation: https://eigen.tuxfamily.org/dox/TopicLazyEvaluation.html
Of course, C++ is powerful enough that using specific tools can make the expression template abstraction fall apart (e.g. auto), but fundamentally you can write matrix code very cleanly and get great lazy-evaluated performance.
> There's no inherent reason why doing lazy evaluation in an eager-by-default language should be any harder than the reverse.
lazy by default enables better composition and local reasoning.
https://publish.obsidian.md/paretooptimaldev/Advantages+of+l...
lazy by default enables better composition and local reasoning.
https://publish.obsidian.md/paretooptimaldev/Advantages+of+l...
While you're technically right, it's only in the same sense that any TC language can do anything any other TC language can -- you can theoretically always delay any computation behind a lambda, but the ergonomics have wide-ranging consequences. You also don't get any memoization that way -- without further runtime support.
There was a good Reddit thread recently where a lot of experts weighed in: https://www.reddit.com/r/haskell/comments/wpbs4z/what_things...
I recommend reading the thread -- lot's of great stuff in there, but...
TL;DR: It's a lot more complicated than it seems at first and it's far from obvious that strict-by-default is worth the cost.
There was a good Reddit thread recently where a lot of experts weighed in: https://www.reddit.com/r/haskell/comments/wpbs4z/what_things...
I recommend reading the thread -- lot's of great stuff in there, but...
TL;DR: It's a lot more complicated than it seems at first and it's far from obvious that strict-by-default is worth the cost.
Yes. Which is my main criticism of Haskell: sure, sometimes laziness is very handy, sometimes it's even indispensable... but most of the time, eager is what you want, with judicious sprinkles of explicitly lazy structures (btw, OCaml has "lazy" keyword exactly for that).
Sure, GHC goes to heroic lengths and turns as much lazy computations as it can into eager ones, but there is always a limit to what it can do on its own, and you inevitably end with the programmer having to slap strict patterns and seqs and deepseqs wherever they can reach.
Sure, GHC goes to heroic lengths and turns as much lazy computations as it can into eager ones, but there is always a limit to what it can do on its own, and you inevitably end with the programmer having to slap strict patterns and seqs and deepseqs wherever they can reach.
In OCaml lazy data is fundamentally incompatible with strict data -- you have to explicitly evaluate it, e.g. when passing to a function which takes strict data. This means you end up with two disjoint 'worlds' where you need to write algorithms twice, etc. (In fact you'd have to do that for every possible variant of where exactly the laziness lies.)
The ergonomics become incredibly bad, unfortunately.
(Also see my link to a Reddit thread elsewhere in this thread further ideas on why strict-by-default is not a simple win.)
The ergonomics become incredibly bad, unfortunately.
(Also see my link to a Reddit thread elsewhere in this thread further ideas on why strict-by-default is not a simple win.)
The same split exists in Haskell as well: Data.Text.Text vs Data.Text.Lazy.Text, etc. And in LISP IIRC lazy data are type-compatible with strict data.
Yes, it's true, and the split is also painful in Haskell. The split exists in Haskell for a choice few data structures where it truly matters for performance -- it also exists for Set/Map. (And I'd argue it is actually a bad thing from a purely language perspective, but sometimes practicality does win out, even in Haskell.)
I think it might actually have been a mistake to have the lazy variants of these data structures, esp. Text/ByteString. IME it's extremely rare to want laziness for these... and it's usually just when building output strings (where there are better solutions like Builders). Anyway, I digress...
IME the same applies to lazy Map/Set, but I'm less familiar with uses for the lazy versions of these. (IIRC they're only spine-lazy, which is rarely what you want, tho.)
> And in LISP IIRC lazy data are type-compatible with strict data.
LISP is dynamically typed -- it's trivial to do there. Also not the directionality: You can go from lazy to strict (that's just evaluation which the runtime will do for you), but you can NOT go the other way automatically. Recovering laziness from strictness is impossible.
The problem with the split happens when you have mismatching types that your compiler will shout at you for.
I think it might actually have been a mistake to have the lazy variants of these data structures, esp. Text/ByteString. IME it's extremely rare to want laziness for these... and it's usually just when building output strings (where there are better solutions like Builders). Anyway, I digress...
IME the same applies to lazy Map/Set, but I'm less familiar with uses for the lazy versions of these. (IIRC they're only spine-lazy, which is rarely what you want, tho.)
> And in LISP IIRC lazy data are type-compatible with strict data.
LISP is dynamically typed -- it's trivial to do there. Also not the directionality: You can go from lazy to strict (that's just evaluation which the runtime will do for you), but you can NOT go the other way automatically. Recovering laziness from strictness is impossible.
The problem with the split happens when you have mismatching types that your compiler will shout at you for.
The phrase "space leak" seemed unusual and a bit of a "tortured phrase", since I'm more familiar with this concept as "memory bloat" or just inefficiency in general. Upon a little more searching it appears to be mainly Haskell jargon.
It's the equivalent of "memory leak" for a garbage-collected language
> For now, this has nothing to do with lazy evaluation. Such implementation will be slow in every language. It happens because add doesn’t use tail-call recursion.
This is why I'll never really trust Haskell to be usable in a serious context. Haskellers will talk your ear off about how such restricted abstractions enable the compiler to optimize it to hell and back, and perfectly idiomatic Haskell can be compiled to almost-perfect machine code, but in point of fact GHC basically never does, and you still have to care about things like tail calls and design APIs to do things like continuation passing.
This is why I'll never really trust Haskell to be usable in a serious context. Haskellers will talk your ear off about how such restricted abstractions enable the compiler to optimize it to hell and back, and perfectly idiomatic Haskell can be compiled to almost-perfect machine code, but in point of fact GHC basically never does, and you still have to care about things like tail calls and design APIs to do things like continuation passing.
> all mainstream languages have eager evaluation models
That is incorrect, e.g. Java Streams (Javas built in library for functional processing of data) are lazy, which can surprise beginners. You say it's just an std lib and not a lazy programming language? Ok, the much older (but just as lazy) C# linq is directly integrated in the C# language, making the C# language both eager and lazy by default, depending on which parts you use.
That is incorrect, e.g. Java Streams (Javas built in library for functional processing of data) are lazy, which can surprise beginners. You say it's just an std lib and not a lazy programming language? Ok, the much older (but just as lazy) C# linq is directly integrated in the C# language, making the C# language both eager and lazy by default, depending on which parts you use.
Arguably Java streams aren't really lazy evaluation at the language level, because the programmer must use explicit lambdas around the lazy bits. Java itself is still eager.
Of course, this is simple - you just enable a language extension so that the language you are referring to Haskell is no longer really Haskell. I think Haskell is a fine language, but given the amount of various language extensions that are needed for it to be usable in production without friction, it seems it will forever be a research language - and there is nothing wrong with that.