I've seen multiple requests for this and it isn't obvious how to do it with the existing tooling.  

The example currently given in the recipes section isn't scalable because it involves expanding the population into a big list with repeated elements:

      sample(['x', 'x', 'x', 'x', 'y', 'y', 'z'], k=5)

Example task:  Given an urn with 8 red balls, 2 white balls, and 3 green balls, choose ten without replacement:

    >>> population = ['red', 'blue', 'green']
    >>> weights =    [  8,      5,      3   ]
    >>> sample(population, weights=weights, k=10)
    ['red', 'green', 'blue', 'red', 'green', 'blue', 'red', 'blue', 'red', 'blue']

I could also add *cum_weights* as an optional optimization but think it best to wait until someone asks for it ;-)

+1 for the functionality.

How about "counts" instead of "weights"?

I found the name "weights" misleading - my first thought was that this would be doing a weighted sampling without replacement (like NumPy's `random.choice(..., replace=False, p=weights)`).

Of course, now I've read the docs, I know better.

> How about "counts" instead of "weights"?

That makes sense.

New changeset 81a5fc38e81b424869f4710f48e9371dfa2d3b77 by Raymond Hettinger in branch 'master':
bpo-40541: Add optional *counts* parameter to random.sample() (GH-19970)
https://github.com/python/cpython/commit/81a5fc38e81b424869f4710f48e9371dfa2d3b77