In-memory full outer join for lists

Question

How do I go about writing a SQL-like full join operation on lists with the below signature?

fullJoin :: [a] -> [b] -> (a -> b -> Bool) -> [(Maybe a, Maybe b)]
fullJoin xs [] _ = map (\x -> (Just x, Nothing)) xs
fullJoin [] ys _ = map (\y -> (Nothing, Just y)) ys
fullJoin xs@(xh:xt) ys@(yh:yt) on = 
  if on xh yh
    then (Just xh, Just yh) : fullJoin xt yt
    else ???

The condition as written is provided by the a -> b -> Bool. here in the example it's set as (\n i -> length n == i), meaning the records from names which have the same length as the numbers in nums.

Example:

names = ["Foo","Bar", "John" , "Emily", "Connor"]
nums = [1,2,3,4,5]

fullJoin names nums (\n i -> length n == i)
  == [ (Just "Foo", Just 3)
     , (Just "Bar", Just 3)
     , (Just "John", Just 4)
     , (Just "Emily", Just 5)
     , (Just "Connor", Nothing)
     , (Nothing, Just 1)
     , (Nothing, Just 2)
     ]

To clarify the exact SQL equivalent of the said function, here's how it would be written in PostgreSQL:

create table names as
select * from (values 
               ('Foo'),
               ('Bar'),
               ('John'),
               ('Emily'),
               ('Connor')
              ) 
              as z(name);

create table nums as
select * from (values 
               (1),
               (2),
               (3),
               (4),
               (5)
              ) 
              as z(num);

select * from names
full join nums
on char_length(name) = num

And running this will yield:

name    num
(null)  1
(null)  2
Foo     3
Bar     3
John    4
Emily   5
Connor  (null)

Fiddle link : sqlfiddle

Answer 1

Just dully implement what it is you want it to do. Super inefficient:

fullJoin :: [a] -> [b] -> (a -> b -> Bool) 
         -> [(Maybe a, Maybe b)]
fullJoin xs ys p = concatMap (map (Just *** Just) . snd) a
                    ++ [(Just x, Nothing) | (x, []) <- a]
                    ++ [(Nothing, Just y) | (y, []) <- b]
  where
    a = [(x, [(x, y) | y <- ys, p x y]) | x <- xs]
    b = [(y, [(x, y) | x <- xs, p x y]) | y <- ys]

If we were allowed to add the Eq constraints on the a and b types, we could use Data.List.\\ to find the unmatched elements instead of making the second sweep.

Testing:

> mapM_ print $ fullJoin names nums (\n i -> length n == i)
(Just "Foo",Just 3)
(Just "Bar",Just 3)
(Just "John",Just 4)
(Just "Emily",Just 5)
(Just "Connor",Nothing)
(Nothing,Just 1)
(Nothing,Just 2)

Answer 2

Here is a naive Python implementation. It is O(n^2).

#! /usr/bin/env python

def full_join_cond (list1, list2, condition_fn):
    answer = []
    for x in list1:
        matches = []
        for y in list2:
            if condition_fn(x, y):
                matches.append(y)
        if len(matches):
            answer.extend([[x, y] for y in matches])
        else:
            answer.append([x, None])

    for y in list2:
        matched = False
        for x in list1:
            if condition_fn(x, y):
                matched = True
                break
        if not matched:
            answer.append([None, y])

    return answer


names = ["Foo","Bar", "John" , "Emily", "Connor"]
nums = [1,2,3,4,5]
cond = lambda x, y: len(x) == y

print(full_join_cond(names, nums, cond))

And here is an implementation that more closely matches how a database would perform this join. It is O(size_of_output) which is frequently O(n).

def list_map_to_dict (l, m):
    d = {}
    for x in l:
        mx = m(x)
        if mx in d:
            d[mx].append(x)
        else:
            d[mx] = [x]
    return d

def full_join_map (list1, map1, list2, map2):
    dict1 = list_map_to_dict(list1, map1)
    dict2 = list_map_to_dict(list2, map2)

    answer = []
    for mx, xs in dict1.iteritems():
        if mx in dict2:
            for y in dict2[mx]:
                answer.extend([[x, y] for x in xs])
            dict2.pop(mx)
        else:
            answer.extend([[x, None] for x in xs])

    for my, ys in dict2.iteritems():
        answer.extend([[None, y] for y in ys])
    return answer

print(full_join_map(names, lambda x: len(x), nums, lambda y: y))

And for giggles and grins the two can be combined in a function that both is general and can run fast. (I have not tried to make the function signature reasonable - just trying to show the idea.)

def full_join (list1, map1, list2, map2, cond=None):
    if map1 is None:
        map1 = lambda x: None

    if map2 is None:
        map2 = lambda y: None

    if cond is None:
        cond = lambda x, y: True

    dict1 = list_map_to_dict(list1, map1)
    dict2 = list_map_to_dict(list2, map2)

    answer = []
    for mx, xs in dict1.iteritems():
        if mx in dict2:
            answer.extend(full_join_cond(xs, dict2[mx], cond))
            dict2.pop(mx)
        else:
            answer.extend([[x, None] for x in xs])

    for my, ys in dict2.iteritems():
        answer.extend([[None, y] for y in ys])
    return answer

Answer 3

A full outer join between tables X and Y on a condition rel(x, y) can be thought of as the union of three (disjoint) parts:

• The set of pairs (x, y) where rel(x,y)
• The set of pairs (x0, null) where for all y in Y, not (rel(x0, y))
• The set of pairs (null, y0) where for all x in X, not (rel(x, y0))

We can structure our Haskell program in a similar way:

fullJoin :: [a] -> [b] -> (a -> b -> Bool) -> [(Maybe a, Maybe b)]
fullJoin xs ys rel = concat $
  [[(Just x, Just y) | y <- ys, x `rel` y] | x <- xs] -- for each x, find all the related ys
  <> [[(Just x, Nothing)] | x <- xs, all (\y -> not (x `rel` y)) ys] -- find all xs that have no related ys
  <> [[(Nothing, Just y)] | y <- ys, all (\x -> not (x `rel` y)) xs] -- find all ys that have no related xs

The way the problem is posed, you can't get this faster than O(n^2). That said, my solution, while being O(n^2), it not optimal: it traverses xs twice. You can think about what you'd have to do in order to only traverse xs once. It has to do with finding a way to keep track of which xs you've already seen.