It is just a slow implementation, the usage of a dictionnary rather than a list is weird and the setup will be long, `get_sentinel` scan the whole list while it should just continue so it multiplies the complexity of the algorithm by n, line 22 scan the remainder of the list while it should just go for the multiples.

Here's a naive implementation I just wrote:

    def crible(n):
        nums = [True] * n
        nums[:1] = [False, False]

        for i, isPrime in enumerate(nums):
            if not isPrime:
                continue
            yield i
            j = 2
            while i*j < len(nums):
                nums[i*j] = False
                j += 1

It is much faster (and is about 10 lines of Python):

    timeit.timeit("list(crible(1000000))", number=1, globals={'crible': crible})
    0.9441672329999165