Hash collisions in memoization result in invalid values

[brandonwillard]

Our custom memoize function is mapping inputs to their hashes, then those hashes are mapped to cached values, and, because the inputs and the hashes' original values aren't compared, a simple hash collision will cause memoize to return the wrong values.

Here's a simple example:

import pymc3 as pm


class Bad1:
    def __hash__(self):
        return 329


class Bad2:
    def __hash__(self):
        return 329


b1 = Bad1()
b2 = Bad2()


# Here's what we expect:
d = {}

d[b1] = 1
d[b2] = 2

assert hash(b1) == hash(b2)
assert d[b1] != d[b2]

# Because `memoize` doesn't use the inputs as keys in the `dict`-based
# cache, we don't get the expected post-hash-lookup equality checks that
# prevent errors when faced with hash collisions

@pm.memoize.memoize
def test_func(x):
    return x


assert b1 != b2
assert test_func(b1) != test_func(b2)

The last assert will currently fail, although it should be equivalent to the assert preceding it.