Numpy array default hasher

[filyp]

I cache a function which is called with large numpy arrays. There is an efficient way to hash them with https://github.com/ifduyue/python-xxhash

I think it's a common use case, so I can merge my method to the default hasher in this library.

The only problem is that hash collisions are possible, and the cache would then work incorrectly, but it's very rare, and also the existing functionality isn't broken. I can also add an option to not hash them but just convert to string with array.tobytes(), which prevents collisions but is only is suited for small arrays.

The same can be done for regular python lists (with xxhash or str(list)).

If you want, I can make a PR with this.

[shaypal5]

Sounds great! Please find a way to benchmark the improvement, though!

Thank you for suggesting a PR!

[Borda]

How about converting the list to a tuple? #136

[Borda]

I think this is resolved with #183; see:

import numpy as np
import cachier

count = 0

@cachier.cachier()
def dummy_func(a, b=np.array([2])):
    global count
    count += 1
    return a + b

assert count == 0
dummy_func(np.array([1]))
dummy_func(np.array([1]), np.array([2]))
dummy_func(np.array([1]), b=np.array([2]))
dummy_func(a=np.array([1]), b=np.array([2]))
assert count == 1

[shaypal5]

I think we need to have to things to close this confidently:

1. A test covering this case.
2. A benchmark making sure performance, time-wise, for very large numpy array using the current solution are comparable to using hpython-xxhash.

Is 1 already in place?
If so, a contribution of 2 will enable me to feel comfortable with closing this issue.