Can you create an unbiased sample of size k from a large stream in constant memory?

stream_sample_experiment.py

import random
import matplotlib.pyplot as plt

k = 1000
array = []
for n, x in enumerate([range(k)[random.randrange(k)] for x in range(100000)]):
    if n < k:
        array.append(x)
    else:
        if random.random() < k/float(n):
            array[random.randrange(k)] = x

plt.hist(array)