Best way to return multiple values from a function?

The canonical way to return multiple values in languages that support it is often tupling.

Option: Using a tuple

Consider this trivial example:

def f(x):
  y0 = x + 1
  y1 = x * 3
  y2 = y0 ** y3
  return (y0, y1, y2)

However, this quickly gets problematic as the number of values returned increases. What if you want to return four or five values? Sure, you could keep tupling them, but it gets easy to forget which value is where. It's also rather ugly to unpack them wherever you want to receive them.

Option: Using a dictionary

The next logical step seems to be to introduce some sort of 'record notation'. In Python, the obvious way to do this is by means of a dict.

Consider the following:

def g(x):
  y0 = x + 1
  y1 = x * 3
  y2 = y0 ** y3
  return {'y0': y0, 'y1': y1 ,'y2': y2}

(Just to be clear, y0, y1, and y2 are just meant as abstract identifiers. As pointed out, in practice you'd use meaningful identifiers.)

Now, we have a mechanism whereby we can project out a particular member of the returned object. For example,

result['y0']

Option: Using a class

However, there is another option. We could instead return a specialized structure. I've framed this in the context of Python, but I'm sure it applies to other languages as well. Indeed, if you were working in C this might very well be your only option. Here goes:

class ReturnValue:
  def __init__(self, y0, y1, y2):
     self.y0 = y0
     self.y1 = y1
     self.y2 = y2

def g(x):
  y0 = x + 1
  y1 = x * 3
  y2 = y0 ** y3
  return ReturnValue(y0, y1, y2)

In Python the previous two are perhaps very similar in terms of plumbing - after all { y0, y1, y2 } just end up being entries in the internal __dict__ of the ReturnValue.

There is one additional feature provided by Python though for tiny objects, the __slots__ attribute. The class could be expressed as:

class ReturnValue(object):
  __slots__ = ["y0", "y1", "y2"]
  def __init__(self, y0, y1, y2):
     self.y0 = y0
     self.y1 = y1
     self.y2 = y2

From the Python Reference Manual:

> The __slots__ declaration takes a sequence of instance variables and reserves just enough space in each instance to hold a value for each variable. Space is saved because __dict__ is not created for each instance.

Option: Using a dataclass (Python 3.7+)

Using Python 3.7's new dataclasses, return a class with automatically added special methods, typing and other useful tools:

@dataclass
class Returnvalue:
    y0: int
    y1: float
    y3: int

def total_cost(x):
    y0 = x + 1
    y1 = x * 3
    y2 = y0 ** y3
    return ReturnValue(y0, y1, y2)

Option: Using a list

Another suggestion which I'd overlooked comes from Bill the Lizard:

def h(x):
  result = [x + 1]
  result.append(x * 3)
  result.append(y0 ** y3)
  return result

This is my least favorite method though. I suppose I'm tainted by exposure to Haskell, but the idea of mixed-type lists has always felt uncomfortable to me. In this particular example the list is -not- mixed type, but it conceivably could be.

A list used in this way really doesn't gain anything with respect to the tuple as far as I can tell. The only real difference between lists and tuples in Python is that lists are mutable, whereas tuples are not.

I personally tend to carry over the conventions from functional programming: use lists for any number of elements of the same type, and tuples for a fixed number of elements of predetermined types.

Question

After the lengthy preamble, comes the inevitable question. Which method (do you think) is best?

Named tuples were added in 2.6 for this purpose. Also see os.stat for a similar builtin example.

>>> import collections
>>> Point = collections.namedtuple('Point', ['x', 'y'])
>>> p = Point(1, y=2)
>>> p.x, p.y
1 2
>>> p[0], p[1]
1 2

In recent versions of Python 3 (3.6+, I think), the new typing library got the NamedTuple class to make named tuples easier to create and more powerful. Inheriting from typing.NamedTuple lets you use docstrings, default values, and type annotations.

Example (From the docs):

class Employee(NamedTuple):  # inherit from typing.NamedTuple
    name: str
    id: int = 3  # default value

employee = Employee('Guido')
assert employee.id == 3

For small projects I find it easiest to work with tuples. When that gets too hard to manage (and not before) I start grouping things into logical structures, however I think your suggested use of dictionaries and ReturnValue objects is wrong (or too simplistic).

Returning a dictionary with keys "y0", "y1", "y2", etc. doesn't offer any advantage over tuples. Returning a ReturnValue instance with properties .y0, .y1, .y2, etc. doesn't offer any advantage over tuples either. You need to start naming things if you want to get anywhere, and you can do that using tuples anyway:

def get_image_data(filename):
    [snip]
    return size, (format, version, compression), (width,height)

size, type, dimensions = get_image_data(x)

IMHO, the only good technique beyond tuples is to return real objects with proper methods and properties, like you get from re.match() or open(file).

A lot of the answers suggest you need to return a collection of some sort, like a dictionary or a list. You could leave off the extra syntax and just write out the return values, comma-separated. Note: this technically returns a tuple.

def f():
    return True, False
x, y = f()
print(x)
print(y)

gives:

True
False

I vote for the dictionary.

I find that if I make a function that returns anything more than 2-3 variables I'll fold them up in a dictionary. Otherwise I tend to forget the order and content of what I'm returning.

Also, introducing a 'special' structure makes your code more difficult to follow. (Someone else will have to search through the code to find out what it is)

If your concerned about type look up, use descriptive dictionary keys, for example, 'x-values list'.

def g(x):
  y0 = x + 1
  y1 = x * 3
  y2 = y0 ** y3
  return {'y0':y0, 'y1':y1 ,'y2':y2 }

Another option would be using generators:

>>> def f(x):
	    y0 = x + 1
	    yield y0
	    yield x * 3
	    yield y0 ** 4

	
>>> a, b, c = f(5)
>>> a
6
>>> b
15
>>> c
1296

Although IMHO tuples are usually best, except in cases where the values being returned are candidates for encapsulation in a class.

I prefer:

def g(x):
  y0 = x + 1
  y1 = x * 3
  y2 = y0 ** y3
  return {'y0':y0, 'y1':y1 ,'y2':y2 }

It seems everything else is just extra code to do the same thing.

I prefer to use tuples whenever a tuple feels "natural"; coordinates are a typical example, where the separate objects can stand on their own, e.g. in one-axis only scaling calculations, and the order is important. Note: if I can sort or shuffle the items without an adverse effect to the meaning of the group, then I probably shouldn't use a tuple.

I use dictionaries as a return value only when the grouped objects aren't always the same. Think optional email headers.

For the rest of the cases, where the grouped objects have inherent meaning inside the group or a fully-fledged object with its own methods is needed, I use a class.

>>> def func():
...    return [1,2,3]
...
>>> a,b,c = func()
>>> a
1
>>> b
2
>>> c
3

Generally, the "specialized structure" actually IS a sensible current state of an object, with its own methods.

class Some3SpaceThing(object):
  def __init__(self,x):
    self.g(x)
  def g(self,x):
    self.y0 = x + 1
    self.y1 = x * 3
    self.y2 = y0 ** y3

r = Some3SpaceThing( x )
r.y0
r.y1
r.y2

I like to find names for anonymous structures where possible. Meaningful names make things more clear.

+1 on S.Lott's suggestion of a named container class.

For Python 2.6 and up, a named tuple provides a useful way of easily creating these container classes, and the results are "lightweight and require no more memory than regular tuples".

Python's tuples, dicts, and objects offer the programmer a smooth tradeoff between formality and convenience for small data structures ("things"). For me, the choice of how to represent a thing is dictated mainly by how I'm going to use the structure. In C++, it's a common convention to use struct for data-only items and class for objects with methods, even though you can legally put methods on a struct; my habit is similar in Python, with dict and tuple in place of struct.

For coordinate sets, I'll use a tuple rather than a point class or a dict (and note that you can use a tuple as a dictionary key, so dicts make great sparse multidimensional arrays).

If I'm going to be iterating over a list of things, I prefer unpacking tuples on the iteration:

for score,id,name in scoreAllTheThings():
    if score > goodScoreThreshold:
        print "%6.3f #%6d %s"%(score,id,name)

...as the object version is more cluttered to read:

for entry in scoreAllTheThings():
    if entry.score > goodScoreThreshold:
        print "%6.3f #%6d %s"%(entry.score,entry.id,entry.name)

...let alone the dict.

for entry in scoreAllTheThings():
    if entry['score'] > goodScoreThreshold:
        print "%6.3f #%6d %s"%(entry['score'],entry['id'],entry['name'])

If the thing is widely used, and you find yourself doing similar non-trivial operations on it in multiple places in the code, then it's usually worthwhile to make it a class object with appropriate methods.

Finally, if I'm going to be exchanging data with non-Python system components, I'll most often keep them in a dict because that's best suited to JSON serialization.

"Best" is a partially subjective decision. Use tuples for small return sets in the general case where an immutable is acceptable. A tuple is always preferable to a list when mutability is not a requirement.

For more complex return values, or for the case where formality is valuable (i.e. high value code) a named tuple is better. For the most complex case an object is usually best. However, it's really the situation that matters. If it makes sense to return an object because that is what you naturally have at the end of the function (e.g. Factory pattern) then return the object.

As the wise man said:

> Premature optimization is the root of all evil (or at least most of > it) in programming.

In languages like Python, I would usually use a dictionary as it involves less overhead than creating a new class.

However, if I find myself constantly returning the same set of variables, then that probably involves a new class that I'll factor out.

I would use a dict to pass and return values from a function:

Use variable form as defined in form.

form = {
    'level': 0,
    'points': 0,
    'game': {
        'name': ''
    }
}


def test(form):
    form['game']['name'] = 'My game!'
    form['level'] = 2

    return form

>>> print(test(form))
{u'game': {u'name': u'My game!'}, u'points': 0, u'level': 2}

This is the most efficient way for me and for processing unit.

You have to pass just one pointer in and return just one pointer out.

You do not have to change functions' (thousands of them) arguments whenever you make a change in your code.