toolz.merge_with not lazy, breaks with Clojure interface

[mrocklin]

Clojure's core.merge-with takes a binary operator

(defn add [a b]
  (+ a b))  ; yes I know this is silly, I wanted an explicit binary operator

(merge-with add [{:a 1} {:a 2} {:a 3}])
;; {:a 6}

Toolz merge_with takes a reduction operation

>>> merge_with(sum, [{'a': 1}, {'a': 2}, {'a': 3}])
{'a': 6}

This is convenient from a Python perspective because operations like sum are more familiar than binary operators like add. However this is not lazy and requires all dicts to be read in at once.

I tried a solution in #127 but gave up.

[eriknw]

Although this doesn't address the laziness issue, one option to deal with any binop (including builtins) is to do a try-except as follows:

try:
    # current method
    return dict((k, func(v)) for k, v in iteritems(result))
except TypeError:
    # perform reduction assuming `func` is a binary function
    ...

This isn't perfect. What if the function returned an actual TypeError? Also, if func is a curried function that needs two more inputs, then it won't be reduced as expected.

Regarding the attempt in #127, I don't like relying on inspect.

Two other options include:

1. Make a new function
2. Add a keyword argument to choose automatic (and lazy) binop reduction.
   • A keyword option may also allow variadic calling of a function, such as func(*vals).
   • For example, merge_with(func, *dicts, **functype), where keyword functype may be unary (default), binary, or variadic.

[eriknw]

One solution is to add init as a keyword parameter. If it is given, then use func as a binary operator, otherwise use func as a unary operator.

This would require changing the API to def merge_with(func, *dicts, **init) or def merge_with(func, dicts, init=no_init). Side note: def merge_with(func, *dicts, init=no_init) is possible in Python 3.

The strategy to choose func as binop or unop based on whether init is provided may also work for reduceby, which is being discussed here: pytoolz/cytoolz#1