Oh, I see what you mean. But then containers would depend on GADTs, which is rather counter to tradition. And I don't really see the benefit over a newtype and smart constructors around the existing Seq type. On Thu, Apr 25, 2019, 1:41 PM Zemyla <zemyla@gmail.com> wrote:
We could have a GADT for the FingerTree:
data Emptiness = E | NE
data FingerTree e a where EmptyT :: FingerTree 'E a SingleT :: a -> FingerTree e a DeepT :: !(Digit a) -> FingerTree e' (Node a) -> !(Digit a) -> FingerTree e a
And because Seq is a newtype wrapper around FingerTree (Elem a) anyway, Seq and NonEmptySeq would become newtype wrappers around FingerTree 'E (Elem a) and FingerTree 'NE (Elem a). This avoids the majority of the code duplication.
On Thu, Apr 25, 2019, 12:34 John Ericson <john.ericson@obsidian.systems> wrote:
Thanks for the clarification, David. Sounds like a good wait-and-see thing: after this lands, Zemyla's proposal and the experience with NonEmpty{Map,Set} can inform what Seq should end up looking like.
John On 4/25/19 1:28 PM, David Feuer wrote:
No, a Seq can be Empty at the bottom too. It's definitely not a mutually recursive situation like Set or Map, and it's not immediately a top-level one like IntSet or IntMap. A nonempty sequence type would need somewhat different functions for everything. In most cases, the changes seem fairly straightforward. I'm not sure about weird functions like inits and tails. Some extra care might be required for replicate.
I believe it would be possible to restructure things to make the distinction a top-level one, by using a possibly-empty type for the top and a nonempty one below. Zemyla has already identified some very solid unrelated reasons to want to separate the tops from the rest, but I'm somewhat concerned about source code duplication with that general approach. Even if we do that, it's not clear to me that we can make a non-empty/possibly-empty distinction without incurring a performance penalty.
On Thu, Apr 25, 2019, 1:17 PM John Ericson <john.ericson@obsidian.systems> <john.ericson@obsidian.systems> wrote:
I haven't looked into `Seq` in addition to `Map` and `Set`, just `IntSet` and `IntMap`. But it might be a similar thing? I take it that with `Seq` today only the root can be empty and everything else is single or deep? That means we *would* just use a single Maybe-like thing at the top level, no mutual recursion. But on the other hand to make that to make that work efficiently we would would need GHC to support unboxing sums, so 1 + 2 variants can become a flat 3.
Also, https://github.com/haskell/containers/pull/616 is now where the actual implementation is happening, not just the datatype changes as before. Feel free to comment on the concrete work in progress, everyone!
John
On 4/25/19 11:36 AM, Zemyla wrote:
A Seq has either Empty, Single, or Deep. A NonEmptySeq would have just Single or Deep.
On Thu, Apr 25, 2019, 09:55 David Feuer <david.feuer@gmail.com <mailto:david.feuer@gmail.com>> wrote:
I don't see the benefit there, unless you see a way to work it into the representation.
On Thu, Apr 25, 2019, 10:53 AM Zemyla <zemyla@gmail.com <mailto:zemyla@gmail.com>> wrote:
As long as we're doing this, can we also add NonEmptySeq as well?
On Thu, Apr 25, 2019, 09:11 Artyom Kazak <yom@artyom.me <mailto:yom@artyom.me>> wrote:
I'm -1 on any kind of |Map = NEMap|.
An ordinary map and a non-empty map are semantically different. I believe that if I non-empty maps were already in |containers|, I would pretty much always care whether a |Map| I see in code is a 0-map or 1-map.
Similarly, I prefer |Int| and |Word| instead of |Int| and |Unsigned.Int|. (Luckily that's already the case.)
We already have a precedent with |Text| and |ByteString|, where the lazy and the strict versions are only distinguished by the module prefix. In my experience, modules where both are used are pretty common, and I end up just introducing |type LByteString = Lazy.ByteString| in all my projects, because otherwise I need to scroll to the imports section whenever I need to know which flavor of bytestring is being used. (Or if I'm reading haddocks, I have to look at the link because Haddock hides module prefixes.)
"why not both" is even worse. I still can't trust the |Map|, but now I also have to learn and remember that two modules are the same. Speaking from experience again – most people seem to be surprised by the fact that |Data.Map.Lazy| and |Data.Map.Strict| export the same |Map| type. The proposed module hierarchy would move |containers| to the top of my "packages that confuse beginners" list, beating even |aeson|.
As an aside, I wish we had a proper interface for container-like structures, or at least a solution to name scoping. I really like the way Rust does it, for instance, where certain functions can be "attached" to a type – I'm hesitant to call them "methods" because Rust is not an OOP language. On Apr 25 2019, at 2:49 pm, Mario Blažević <mblazevic@stilo.com <mailto:mblazevic@stilo.com>> wrote:
On 2019-04-18 11:00 p.m., David Feuer wrote:
I'm in favor of the proposal. I find the isomorphism between Map (a,b) v and Map a (NonemptyMap b v) very pleasant. The fact that others have written less-performant implementations of this idea is rather convincing. The fact that doing this removes partial matches in the implementation is nice. And I'll take performance improvements where I can get them. The main question is the proper name of the type. Just Data.Map.Nonempty.Map, or .NonemptyMap? Should the empty be capitalized?
There seems to be a consensus for Data.Map.NonEmpty.NEMap, with the type and the functions slightly off the regular ones. This design would make it easier to use regular and non-empty containers together, but it be annoying for the use case of replacing all uses of an existing regular container with a non-empty one. I'd rather change just the import declaration than all occurrences of the type name and functions.
I don't want to derail the implementation with bikeshedding, so I'm just going to ask why not both? The library can both export the tweaked names and add a module, say Data.NonEmpty.Map.Lazy, that exports the type synonym Map = NEMap. It would also rename all the functions back to their names from Data.Map.Lazy.
On Thu, Apr 18, 2019, 7:15 PM John Cotton Ericson <John.Ericson@obsidian.systems> <John.Ericson@obsidian.systems> wrote:
In https://github.com/haskell/containers/issues/608 I proposed adding non-empty variants of Map and Set, analogous to Data.List.NonEmpty for List, to containers. semigroupoids demonstrates the many uses and structure of non-empty containers in general, and libraries such as https://github.com/mstksg/nonempty-containers and https://github.com/andrewthad/non-empty-containers demonstrate the interest in non-empty maps and sets in particular. My favorite use-case is that they're needed to "curry" containers: for example, |Map (k0, k1) v| is isomorphic not to |Map k0 (Map k1 v)| but to |Map k0 (NonEmptyMap k1 v)|. I like this use-case because it comes from the containers themselves.
Importantly, there's no good way to do this outside of containers; doing so leads to imbalancing / extra indirection, or massive code duplication. If one wraps the container was an extra value like Data.List.NonEmpty, one's left with an unavoidable extra indirection/imbalance. One can rectify this by copying and modifying the implementation of containers, but that's hardly maintainable; even as though the algorithms are the same, enough lines are touched that merging upstream containers is nigh impossible.
On the other hand, the non-empty containers can be elegantly and sufficiently implemented alongside their originals by taking the Bin constructor and breaking it out into it's own type, mutually recursive with the original. This avoids the indirection/imbalancing and code duplication problems: the algorithms work exactly as before creating the same trees (remember the UNPACK), and no code duplicated since the functions become mutually recursive matching the types.
To briefly summarize the thread:
1. I proposed the issue after performing this same refactor on the dependent-map package:
https://github.com/obsidiansystems/dependent-map/tree/non-empty,
a fork of containers. 2. I made https://github.com/haskell/containers/pull/616 which just changes the types, to make sure UNPACK preserved the importance. 3.
https://gist.github.com/Ericson2314/58709d0d99e0c0e83ad266701cd71841
the benchmarks showed rather than degrading performance, PR 616 actually /improved/ it.
If there is preliminary consensus, I'll make a second PR on top which generalizes the functions like on my dependent-map branch.
Thanks,
John
_______________________________________________ Libraries mailing list Libraries@haskell.org <mailto:Libraries@haskell.org> <mailto:Libraries@haskell.org <mailto:Libraries@haskell.org>>
http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org <mailto:
Libraries@haskell.org>
http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org <mailto:Libraries@haskell.org>
http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org <mailto:Libraries@haskell.org> http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org <mailto:Libraries@haskell.org> http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries