Python code (with IPython timing) for comparing rle implementations.

rle-implementations.py

#
# Python code (with IPython timing) for comparing rle implementations.
# See: https://twitter.com/double_thunk/status/1358509387044827138?s=20
#
import random
import string
import platform
from itertools import groupby

import numpy as np
import pandas as pd

def rle_numpy(text):
    a = np.fromiter(text, '<U1')
    starts = 1 + (a[1:] != a[:-1]).nonzero()[0]
    letters = [a[0], *a[starts].tolist()]
    counts = np.diff(starts, 1, 0, 0, len(a))
    return [*zip(letters, counts.tolist())]

def rle_itertools(text):
    return [(c, sum(1 for _ in g)) for c, g in groupby(text)]

def rle_beazley(items):
    if len(items) == 0:
        return []
    result = []
    prev_item = items[0]
    count = 0
    for item in items:
        if item == prev_item:
            count += 1
        else:
            result.append((prev_item, count))
            prev_item = item
            count = 1
    if count > 0:
        result.append((prev_item, count))
    return result

def rle_pandas(items):
    s = pd.Series(list(items))
    s = s[s.ne(s.shift(-1))]
    c = s.index.to_series() + 1
    c -= c.shift(fill_value=0)
    return list(zip(s.tolist(), c.tolist()))

counts = np.random.randint(10, size=1_000_000)
letters = [random.choice(string.ascii_lowercase) for _ in counts]
text = ''.join(l*c for l,c in zip(letters, counts))[:1_000_000]
print(f"Input head: {text[:50]}")

assert rle_itertools(text) == rle_numpy(text) == rle_beazley(text)

print("-------------------")
print(f"Testing on {platform.python_implementation()}")
print("-------------------")

print("\nrle_itertools:")
%timeit r = rle_itertools(text)

print("\nrle_numpy:")
%timeit r = rle_numpy(text)

print("\nrle_beazley:")
%timeit r = rle_beazley(text)

print("\nrle_pandas:")
%timeit r = rle_pandas(text)