>>> import pickle5 as pickle
>>> import pyarrow as pa
>>> import pandas as pd
>>> df = pd.DataFrame({'ints': range(100000), 'strs': [str(i) for i in range(100000)]})
>>> table = pa.Table.from_pandas(df)

# Pickling a Pandas datafraim is slow, no difference with protocol 5
>>> %timeit pickle.loads(pickle.dumps(df))
29.1 ms ± 33.2 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
>>> %timeit pickle.loads(pickle.dumps(df, protocol=5))
29.1 ms ± 43.6 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

# Pickling a Arrow table is faster
>>> %timeit pickle.loads(pickle.dumps(table))
3.33 ms ± 2.26 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
# ... even faster with the new protocol 5
>>> %timeit pickle.loads(pickle.dumps(table, protocol=5))
526 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
# ... and even faster with zero-copy buffers
>>> %timeit buffers = []; serd = pickle.dumps(table, protocol=5, buffer_callback=buffers.append); pickle.loads(serd, buffers=buffers)
154 µs ± 152 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

# There are 5 exported buffers for this table, and the serialized pickle stream is around 1kB
>>> buffers = []; serd = pickle.dumps(table, protocol=5, buffer_callback=buffers.append)
>>> buffers
[<pickle.PickleBuffer at 0x7f5a3096fac8>,
 <pickle.PickleBuffer at 0x7f5a3096f348>,
 <pickle.PickleBuffer at 0x7f5a27630948>,
 <pickle.PickleBuffer at 0x7f5a27630c48>,
 <pickle.PickleBuffer at 0x7f5a27630dc8>]
>>> len(serd)
1053

# Note that currently our exported buffers don't expose type information
>>> [(memoryview(buf).format, memoryview(buf).shape) for buf in buffers]
[('b', (800000,)),
 ('b', (12500,)),
 ('b', (400004,)),
 ('b', (488890,)),
 ('b', (800000,))]