PragTob

PragTob

Library API Design error returns: nil vs error tuple vs exception

:wave:

I’m currently extracting the statistics calculation part from benchee and stumbled upon how to present error conditions. In the concrete example statistics of an empty list [] don’t make much sense. There’s no average, no total, no nothing.

I see three possible ways to represent this error condition right now:

  • return nil - feels kind of right as in this specific case an empty list just doesn’t have an average, a total or what not. So it’s nothing. Feels kind of wrong because some of these also return nil values normally (there doesn’t have to be a mode). Also nil tends to leak and be a bit hard to debug.
  • return {:error, reason} - this seems to be the most conventional wisdom in elixir looking around. Mostly it’s paired with an {:ok, value} return value which for my concrete use case I’m not the biggest fan of. It’s easy to match against etc. and should one decide to want to introduce an exception raising API it’d be easy to implement !-bang methods that raise instead of the tuple. What I also like about this, is that it’s easy to represent in the type spec.
  • raise a (custom) exception - this is mostly discouraged in elixir, the main argument being that it should only be used if it would leave the application in an undefined state. Here’s [a good blog post] (https://michal.muskala.eu/post/error-handling-in-elixir-libraries/) by @michalmuskala going over error tuples vs. exceptions. So why am I even considering this? I find it curious that the default behavior of Enum.max/2 and Enum.min/2 is to raise an exception (my min/max functions basically just wrap these :angel:) . However, Enum.sum/1 returns 0. So at least part of the elixir stdlib does this (to an extent, you can pass your own empty element function).

I think it’s important to consider what kind of usage each pattern promotes for the users of a library:

  • nil - get some error that they had some nil value somewhere where they were expecting something else having to debug it back to the library call and ultimately the bad value they provided. Not great.
  • {:error, reason} - pattern match on calls either in case or directly assertive through {:ok, result} = Lib.call(input) which I think is ok. However, if users didn’t read the docs (bad!) tracing back where that tuple originated instead of a real value came from might still be a bit of a bother.
  • exception - never ever pass me that value i.e. the value should be caught by users beforehand. If they don’t the stack trace points them right where they need to look and fix the input (hopefully with a very good error message). I kind of like this for my specific case as handing an empty list to a statistics function seems like something that mostly shouldn’t be done and easy to catch on the user side. However, it’s not an “application critical state”, definitely recoverable and might be considered a valid input.

Right now I use value return but an {:error, reason} error return type - which I realize is sub optimal as it’s hard to match without an {:ok, value}. I kind of like just raising exceptions for my specific case but it feels wrong to do so. After having written all this it feels like implementing both (with !-bang methods) seems like the safest choice without too much extra work.

What’s your take? What do you think about the different options in general? What do you think about the different options in the concrete statistics calculation use case?

Most Liked

michalmuskala

michalmuskala

I agree with @sasajuric - passing in an empty list to a function that doesn’t make sense on empty lists is most probably a programmer error and should just fail. I’d say that the case here is almost the definition of the ArgumentError exception.

11
Post #5
sasajuric

sasajuric

Author of Elixir In Action

In this case I’d raise an exception for two reasons:

  • according to your description, an empty list is a non-domain input (e.g. because avg([]) is undefined)
  • library users can easily handle empty list anywhere up in the call stack

In addition, I’d communicate this through typespec, by using [...] or [element_type, ...].

PragTob

PragTob

Thank you everyone for your input! Before the post my hunch was to go with an exception but it somehow also felt wrong. I’ll do that now and raise an ArgumentError as suggested by many of you! :+1:

Thanks for being such a helpful bunch! :tada:

@xlphs I’d rather avoid that the users of my lib see errors involved with what the library does but rather what they did “wrong” so I think ArgumentError is better :slight_smile:

@Qqwy It’s hard to say how often they’ll do it as a library can be used in all sorts of ways imo :slight_smile: I think they should never do it/if it can be empty catch it before ideally though. I like your thoughts on the ways to generate “useless values” - thank you!

@sasajuric yup thank you, type spec will definitely be [number, ...], I remember the dialyzer on benchee complaining when I did that though but that just means I need to tune something :grin:

sashman

sashman

As a library user my vote would be a tuple return in general.

dimitarvp

dimitarvp

It doesn’t make sense to return anything in your case. I’d just do raise(ArgumentError, "cannot calculate average of an empty list").

Where Next?

Popular in Discussions Top

New
bartblast
With the core component system and HTTP/WebSocket infrastructure solid, it’s time to tackle Pub/Sub support. What We Have vs What’s Miss...
New
marcandre
Compilation time for our project degraded from ~26 seconds to ~32 seconds (~23% slower) :disappointed_face:. Did anyone else compare com...
New
rump13
Hi everyone, I’ve been following Elixir since around 2016 and tinkering with it on and off, but I haven’t had the opportunity to use it ...
New
josefrichter
Has anyone tried Github Speckit or similar projects like AgentOS, BMAD, etc.? They basically offer a bit more structure in context windo...
New
chulkilee
I’ve heard of several companies using Elixir + gRPC successfully. Tubi is one example - which sponsors grpc However, grpc still depends ...
New
hauleth
With OTP 21 and the creation of logger Elixir Logger application became quite limiting (no structured logging). I would like to discuss w...
New
AstonJ
There seems to be a lot of buzz around DeepSeek at the moment, with some saying it’s a ChatGPT killer. The most remarkable thing (if they...
New
ScriptyScott
Hey Folks, I just spent the last couple of days doing a deep dive into using Swoosh with Amazon’s simple email service, check out this t...
New
Nvim
Elixir appears to be a superior language to Python. I don’t see any advantage of Python over Elixir. Are there any?
New

Other popular topics Top

JorisKok
I have a server on AWS, and was running a load test using artillery. When looking at the Phoenix dashboard I see the Ports going to 100% ...
New
New
nsuchy
Hi. I’ve noticed that Windows Powershell has it’s own IEX command and you cannot access Elixir’s IEX due to the conflict. This isn’t a cr...
New
vertexbuffer
Hello, can anybody help here..? I have a list of players and I what to delete an element, but every for loop the list is reverting to ori...
New
lastday4you
I wanted to check elixir version in phoenix because i found that my elixir is 1.5 but when i use Enum.chunk_by it said the function is un...
New
myronmarston
The Elixir Typespec docs show the following syntax for keyword lists in typespecs: # ... | [key: type] # keyword lis...
New
ashish173
I am using Ecto timestamps with postgres, I can see the timestamps() use the :naive_dateime but for my use case I wanted to store the ti...
New
josevalim
Hi everyone, One of the features added to Elixir early on to help integration with Erlang code was the idea of overridable function defi...
New
romenigld
I am trying to run a deploy with docker and I successfully runned with this command: docker build -t romenigld/blog-prod . but when I t...
New
Nvim
Elixir appears to be a superior language to Python. I don’t see any advantage of Python over Elixir. Are there any?
New

We're in Beta

About us Mission Statement