PragTob
Library API Design error returns: nil vs error tuple vs exception
![]()
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 returnnilvalues 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/2andEnum.min/2is to raise an exception (my min/max functions basically just wrap these
) . However, Enum.sum/1returns0. 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 somenilvalue 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
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.
sasajuric
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
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! 
Thanks for being such a helpful bunch! 
@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 
@Qqwy It’s hard to say how often they’ll do it as a library can be used in all sorts of ways imo
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 
sashman
As a library user my vote would be a tuple return in general.
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").








