9

In Python specifically (I don't know if this generalizes) is there a "best" way to return multiple items from a function?

def func1():
    return a,b #equivalent to (a,b)

def func2():
    return[a,b]

def func3():
    return{"valueA":a,"valueB":b}

The first is what I see most generally, but I feel like the last one creates more readable code because you have named outputs, but I might be missing some sort of overhead created by this method?

  • You're not retuning multiple arguments; in all of these examples, you're returning a single result which just happens to be a compound data type (a tuple, list, and dictionary, respectively). What data type to return depends on your function and how it's meant to be used. – gardenhead Aug 6 '16 at 18:49
  • Interesting, I never thought of it that way, if you make that an answer I can accept it – Devon Muraoka Aug 6 '16 at 19:09
12

I think the choices need to be considered strictly from the caller's point of view: what is the consumer most likely to need to do?

And what are the salient features of each collection?

  • The tuple is accessed in order and immutable
  • The list is accessed in order and mutable
  • The dict is accessed by key

The list and tuple are equivalent for access, but the list is mutable. Well, that doesn't matter to me the caller if I'm going to immediately unpack the results:

score, top_player = play_round(players)
# or
idx, record = find_longest(records)

There's no reason here for me to care if it's a list or a tuple, and the tuple is simpler on both sides.

On the other hand, if the returned collection is going to be kept whole and used as a collection:

points = calculate_vertices(shape)
points.append(another_point)
# Make a new shape

then it might make sense for the return to be mutable. Homogeneity is also an important factor here. Say you've written a function to search a sequence for repeated patterns. The information I get back is the index in the sequence of the first instance of the pattern, the number of repeats, and the pattern itself. Those aren't the same kinds of thing. Even though I might keep the pieces together, there's no reason that I would want to mutate the collection. This is not a list.

Now for the dictionary.

the last one creates more readable code because you have named outputs

Yes, having keys for the fields makes heterogenous data more explicit, but it also comes with some encumbrance. Again, for the case of "I'm just going to unpack the stuff", this

round_results = play_round(players)
score, top_player = round_results["score"], round_results["top_player"]

(even if you avoid literal strings for the keys), is unnecessary busywork compared to the tuple version.

The question here is threefold: how complex is the collection, how long is the collection going to be kept together, and are we going to need to use this same kind of collection in a bunch of different places?

I'd suggest that a keyed-access return value starts making more sense than a tuple when there are more than about three members, and especially where there is nesting:

shape["transform"]["raw_matrix"][0, 1] 
# vs.
shape[2][4][0, 1]

That leads into the next question: is the collection going to leave this scope intact, somewhere away from the call that created it? Keyed access over there will absolutely help understandability.

The third question -- reuse -- points to a simple custom datatype as a fourth option that you didn't present.

Is the structure solely owned by this one function? Or are you creating the same dictionary layout in many places? Do many other parts of the program need to operate on this structure? A repeated dictionary layout should be factored out to a class. The bonus there is that you can attach behavior: maybe some of the functions operating on the data get encapsulated as methods.

A fifth good, lightweight, option is namedtuple(). This is in essence the immutable form of the dictionary return value.

  • Is there a reason you don't mention instances of user defined classes at all? That should be the goto solution in any OO language. – Esben Skov Pedersen Aug 7 '16 at 10:00
  • 1
    That's covered in the third-to-last and second-to-last paragraphs, @EsbenSkovPedersen. "A repeated dictionary layout should be factored out to a class." – Josh Caswell Aug 7 '16 at 15:24
1

Don't think about functions returning multiple arguments. Conceptually, it is best to think of functions as both receiving and returning a single argument. A function that appears to accept multiple arguments actually receives just a single argument of tuple (formally product) type. Similarly, a function that returns multiple arguments is simply returning a tuple.

In Python:

def func(a, b, c):
  return b, c

could be rewritten as

def func(my_triple):
  return (my_triple[1], my_triple[2])

to make the comparison obvious.

The first case is merely syntactic sugar for the latter; both receive a triple as an argument, but the first pattern-matches on its argument to perform automatic destructuring into its constituent components. Thus, even languages without full-on general pattern-matching admit some form of basic pattern matching on some of their types (Python admits pattern-matching on both product and record types).


To return to the question at hand: there is no single answer to your question, because it would be like asking "what should be the return type of an arbitrary function"? It depends on the function and the use case. And, incidentally, if the "multiple return values" are really independent, then they should probably be computed by separate functions.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.