Hacker News Reader

Comment by cryptonector | Hacker News Reader

In Haskell, because of laziness, lists work well enough, but they're really just a thunk container. Conduits are better.

View on HN

The way modern memory works linked lists are practically a denial of service attack on system performance. IMO, like goto’s it’s better to just forget they exist.

Linked lists stink as a manifested in-memory data structure. However, if what your linked list is is "{concrete value} : {code to generate the rest of the list}" and you walk it once, never having the whole thing at once, it isn't necessarily bad. That is the point cryptonector is making. The memory inefficiency of a linked list doesn't matter if you never really have it "in memory".

That said, the ready availability of linked lists in Haskell does mean that you end up with rather more strictly-manifested linked lists than you might like, since any of them can be manifested as actual linked lists. If you're using Haskell though, while you haven't exactly completely given up on that sort of performance, it must not be foremost on your mind.

> Linked lists stink as a manifested in-memory data structure. However, if what your linked list is is "{concrete value} : {code to generate the rest of the list}" and you walk it once, never having the whole thing at once, it isn't necessarily bad. That is the point cryptonector is making. The memory inefficiency of a linked list doesn't matter if you never really have it "in memory".

Yes, exactly. A lazy list is just another term for generator. If you hold on to a reference to the head of such a list as you run the generator, then you'll consume a lot of memory and will end up with a strict list, and also you'll be sad. But if you don't hold on to that reference, then it will be very efficient.

> That said, the ready availability of linked lists in Haskell does mean that you end up with rather more strictly-manifested linked lists than you might like, since any of them can be manifested as actual linked lists. If you're using Haskell though, while you haven't exactly completely given up on that sort of performance, it must not be foremost on your mind.

Especially strings as lists of characters. That was a mistake. Well, in some sense, character data has to be a list, since with Unicode you really can't efficiently index a string... unless it's a rope and you apply the sorts of tools that xi does (e.g., keeping track of character counts per rope segment, etc.). But strings should be sufficiently opaque that arrays, lists, and ropes are all equally able to implement their semantics.

That’s frankly not a linked list.

Haskell is lazy, all data structures work this way unless specifically annotated otherwise. Within standard haskell 2010 it is hard to declare a linked list that doesn't work this way.

It is also worth noting that haskell has library specified fusion rules so the closure isn't even allocated most of the time. So

    sum (take 10 [0..])

compiles into an allocation free loop. I will give you that this isn't a linked list anymore, though.

Haskell has a duality between flow control and data structures as a result of its laziness. It is a perfectly valid point of view to look at such a thing as a control structure, rather than a data structure.

Technically, if you want to really understand how the "IO monad" does its thing, this is actually how it works. Technically your program declares a probably-infinite data structure that contains all possible paths your code could take, e.g., if you have a "if user input is negative then do X else do Y", what you really have is a data structure representing the if-then clause. What keeps the infinite data structure from being manifested in RAM is that the whole thing is resolved lazily, so only one branch will be actually evaluated. That's why the "IO monad" is actually pure; it simply "purely" creates a huge data structure that is resolved by an interpreter later. In the theoriest of theories, an IO value in Haskell really doesn't execute anything either and the whole language is pure... it is only the interpretation of the IO value that finally interfaces with the real world.

It isn't always the most helpful perspective and you can become an expert Haskell programmer without ever looking at your program this way, but technically, under the theory hood, that's what's happening.

In Haskell lazy values are just thunks. So a Lisp-style cons can have a tail that's just a thunk, and when you need to coerce it to a non-thunk value, it just happens and presto, you have the next cons. That's a list! A lazy list.

Why not?

Due to GHC's fusion, they're often not even created. They're used more like control flow constructs than arrays.

Right, I mostly end up using Haskell's lists as an extremely ergonomic counterpart to iterators / generators in languages like Python and Rust.

That's really not true. Whole web frameworks are built strictly around linked list structures that are far more robust, performant and DOS-proof than anything you'll write in your lifetime.

The majority of langauges web frameworks are written in do not have linked lists, at all.

Python doesn't. Ruby doesn't. Java doesn't.

> Java doesn't

https://docs.oracle.com/javase/7/docs/api/java/util/LinkedLi...

Basically, anything written in Elixir Phoenix, or erlang, so: Discord, Pagerduty, Bet365, the entire NHS. If you're a startup, in SF, there's a good chance your payments goes through Brex. You might be using Slab for your documentation, you might use Boston's MBTA, etc. Probably a not-insignificant chunk of what you do as a startup at some level depends on a framework that uses linkedlists. Do you use RabbitMQ or Riak? Linked lists are the primary list datatype for that system.

> The way modern memory works linked lists are practically a denial of service attack on system performance

Wrong. Intrusive, doubly-linked lists are extremely performant and useful in the right places, especially if you have callbacks which store a pointer to a list element (this often happens with device drivers in low level programming), or if the elements can belong to several lists.

The Linux kernel, for example, makes plenty use of linked lists.

> IMO, like goto’s it’s better to just forget they exist.

Wrong again. In a language without automatic memory management and without scoped destruction (cough C cough), they are immensely useful. Again, refer to the Linux kernel. Or any competently written, non-trivial C program for that matter. Note: I'm not defending C here, I'm defending correctly used gotos in C programs.

I don't think you can really get around linked lists having bad cache locality even if they're intrusive. The point of arrays usually being faster is that they're much more cache friendly, and fitting data in cache and avoiding branch mispredictions is usually much more important than o(1) insert/delete.

If you allocate all the nodes of a particular list from one arena, the cache performance should be OK.

IDK if this is the case for device drivers, but AFAICT device drivers tend to preallocate memory on loading and manage it locally.

Arn't there hybrid array lists? Like where instead of growing the array when it goes beyond bound, the last slot is reserved to create a link to the next array chunk?

Linked lists fare worse than arrays only for sequential iteration in tight loops. That includes many/most cases, but not all. As for O(1) insertion/deletion: It's not always about amortized cost, sometimes you actually do care about worst case execution time.

Look, I'm not saying that linked lists can replace arrays. In fact, I almost always use arrays instead of linked lists. What I'm arguing against is the notion that arrays always trump linked lists, which is false, and repeating that meme shows a lack of deeper understanding.

Linked lists require you to be on the node in order to do an insertion so it’s generally not O(1) it’s O(n) with sky high constants. Thus for random inserts arrays are actually faster.

If you have a degenerate case and are never traversing nodes, you’re still likely better of with a rung buffer, array list, 2 different stacks, or various other options. And frankly you’re never going to endlessly just insert data without reading it.

> Linked lists require you to be on the node in order to do an insertion so it’s generally not O(1) it’s O(n) with sky high constants.

Re-read my original post: "[...] especially if you have callbacks which store a pointer to a list element [...]"

This happens all the time in kernel code or bare-metal programming.

What else can I say? Open the Linux source code [1] and see for yourself.

[1] https://elixir.bootlin.com/linux/latest/source

Look we can all agree if you’re writing ASM your going to use goto’s. Further if linked lists are tiny or rarely used, it’s not a problem. If you can ensure the data stays in L2 cache it’s far less of an issue. If any number of things they don’t create problems, but frankly the general case encountered by most programmers are not that.

PS: If you don’t feel like pointing to a concrete example, I am going to counter with 99% of the stuff you can find here: www.google.com.

I don't think I've heard anyone seriously suggest never using a linked list -- the advice usually just ends up being "are you really sure?" which seems appropriate.

Then let me quote the post I originally replied to:

> The way modern memory works linked lists are practically a denial of service attack on system performance.

and

> IMO, like goto’s it’s better to just forget they exist.

That's as close to "never use a linked list" as you can get without saying that literal sentence.

Linked lists are O(N) random insertion as you need to traverse the tree to locate the correct node. The only way that they are always O(1) are a degenerate case better suited for another data structure. Aka if you want a queue then use a ring buffer not a linked list.

Doubly-linked lists have problems with concurrent mutability, FYI. It's really better to aim for non-circularity and to avoid the need for locks.

> Doubly-linked lists have problems with concurrent mutability, FYI.

So do arrays, FYI.

> It's really better to aim for non-circularity and to avoid the need for locks.

How does an array avoid the need for a lock where a linked list needs one?

This is only true if you’re writing code that’s compute bound: in 90% of the production code I’ve seen, the “inefficiency” of a singly-linked list is the absolute last thing that I worry about.

O(N) random reads and O(N) random inserts can quickly turn a wide rage of things to be compute bound. Creating a linked list from scratch via random insertions for example is an N^2 operation with surprisingly large constants.

Now sure, if it’s small and you’re only creating then reading it once then that’s fine. But, you’re assumptions need to be accurate or you just created a problem.

In most of the cases where I’m using a list or array, I’m adding items to one end and then iterating forward over the structure a handful of times.

But, yeah, if you’re relying heavily on random access, pick a vector/array or some kind of tree/hash table.

> IMO, like goto’s it’s better to just forget they exist.

They're quite useful in functional languages.

In Haskell in particular they can have nice performance effects if you use them sensibly.

> Conduits are better.

That's for streaming I/O. Lazy linked lists are appropriate for streaming without side effects

Every use of conduit in production that I have witnessed has been a mistake.

Oh? Can you elaborate?

Reply on news.ycombinator.com