Passing in options: Maps vs. Keyword lists

Chat / Discussions
,
1

Hey everyone, this has been on my mind for some time and I’d love your input on it!

TLDR: I feel like maps are superioer for storing and working with configuration options as compared to keyword lists and am wondering what I might be missing :slight_smile:

So, in Elixir it seems like Keyword lists are the de facto standard for passing options around. It’s what Mix.Config does and sort of advocates, to my understanding. Also it is advocated in the guides

These characteristics are what prompted keyword lists to be the default mechanism for passing options to functions in Elixir

The special characteristics mentioned are:

  • Keys must be atoms
  • Keys are ordered, as specified by the developer.
  • Keys can be given more than once

While I agree that the first one is cool/important for options, I don’t see the value of the other two for options as I see them. Given options there is only ever one option and there should be one key/value pair for it (e.g. should this print a warning? How long should this run?) and I don’t care about the ordering of the keys as I just want to get my specific value.

I can clearly see the value of these properties (and therefore Keyword lists) for DSLs like Ecto (as mentioned in the docs as well) but not for options.

In practice I found keyword lists awkward to deal with when used as options. My two main pain points are:

  • keys might occur twice while I want exactly one (when I override the value of that one key I probably want it)
  • pattern matching is hard (you gotta match on the length of the list and the order of keys) - while I love to pattern match on configuration options

To elaborate on the last point, I found it to be quite nice to pattern match on configuration options to handle them which reads very nice imo:

  defp print_configuration_information(_, %{print: %{configuration: false}}) do
    nil
  end
  defp print_configuration_information(jobs, config) do
    # do printing magic
  end

This is also why I use maps for configuration in benchee.

So in short, I feel like maps are superior for storing and working with configuration options but at the same time I somehow feel like I’m not writing idiomatic elixir and that I’m missing some great advantage of keyword lists.

Opinions/Advice/Insights? :slight_smile:

Thanks!
Tobi

15 Likes
2

Have you seen how, for example, System.cmd/3 does it?

7 Likes
3

There are situations where the latter two properties are very handy. Ecto’s query syntax and Elixir’s OptionParser are situations where you need both. The import special form also leverages the last property (keys can be given more than once). I am pretty sure there are more examples but those are three from the top of my mind.

One of the reasons I also like to use keywords list for options is because we have a well defined dichotomy where keyword lists are used for incomplete sets (you do not need to pass all keys) while maps expect all keys to be there. This allows me to use the assertive map.key syntax when using maps. It doesn’t need to be an OR choice though. I agree pattern matching on maps is really handy and for such cases I would likely accept a keyword list but convert it to a map with defaults:

map = Enum.into(user_options, %{key: "default", other_key: false, ...})

This way the API is consistent with Elixir and you can still leverage the map properties when you don’t care about duplicates nor ordering.

24 Likes
4

This will sound terrible, but sometimes the biggest advantage of doing something is just because everybody else is doing it that way.

In particular, if you are supporting a public library going against the flow is going to lead to an endless stream of questions about keywords not working as options. You’d think programmers would be slightly better than most at reading documentation, but sadly programmers are just people. People as a rule don’t read documentation[1].

While this may seem like a little thing, it’s annoyances like this that sap your energy for supporting open software. It’s one thing to be paid to read man pages to people, it’s a whole 'nuther thing to do it for free in your spare time. Matz’s “principle of least surprise” that drove the design of Ruby is a valuable lesson.

Compared to Elixir, Ruby had terrible documentation for a large part of it’s early history, yet it thrived despite that. One of the reasons for this is that you’d just try the obvious thing and 90% of the time, it would just work.

[1]- Actually this is not true, there’s just always a significant percentage that don’t and seem much a much larger percentage than they really are.

5

Yeah I agree. Phoenix and Elixir both use Keyword Lists so I do as well.

6

Thank you all for your answers! I think I’ll play with converting keyword lists to maps for internal use, that seems like a good solution :slight_smile: Also thanks for arguing for using the Elixir standard against what I might prefer and the way System.cmd/3 deals with the problem - you’ve all been very helpful (as always!) :smiley:

1 Like
7

The Erlang Team has changed the default in favor of maps. Not sure if Elixir will follow.

10 Likes
8

bit late but thanks for passing this along, very interesting and good to know!

9

In his talk at the virtual DevsForUkraine conference, José Valim addressed the question.

TLDL; he still favors lists

But when we want him to reconsider, all we need to say is: José, I think there is a performance issue with using lists for config. He will then benchmark it extensively, re-read every line involved, find 4 unrelated bugs, write fixes, push them to master and finally answer. All within 15 minutes.

5 Likes
10

Thanks for summarizing it!

I was actually the one asking José that question there, thanks to this thread :grin:

Link to the moment in the Q&A for future reference: Devs For Ukraine - Q&A with José Valim - YouTube

As a bit of input here, I don’t think performance is or will be an issue. Even erlang maps are internally represented as lists until there is a map size of 32 iirc. Reason being, hashing is expensive and so for small maps a linear search is actually faster. Since I don’t hope many people pass along more than 30 (I hope more than 5) options, performance should be the same/comparable.

3 Likes