Use protocol=5 for dumping arrays with dtype=object - #1682

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.53%. Comparing base (863994d) to head (cd469f5).
⚠️ Report is 11 commits behind head on main.

@@            Coverage Diff             @@
##             main    #1682      +/-   ##
==========================================
- Coverage   95.57%   95.53%   -0.04%     
==========================================
  Files          46       46              
  Lines        7865     7865              
==========================================
- Hits         7517     7514       -3     
- Misses        348      351       +3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

Not quite sure why it was protocol=2, maybe because at the time compatibility between Python 2 and Python 3 was important?

I think this is the reason.

For the record, here are benchmark results when using an object dtype instead of an unboxed fixed-size dtype:

%load_ext memory_profiler
import pickle
import numpy as np

arr = np.array(['a'*1000 for _ in range(100_000)], dtype=object)

for protocol in [2, 4, 5]:
    print(f"{protocol=}")
    %timeit pickle.dumps(arr, protocol=protocol)
    %memit pickle.dumps(arr, protocol=protocol)

protocol=2
1.35 ms ± 65.3 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
peak memory: 89.78 MiB, increment: 0.00 MiB
protocol=4
1.88 ms ± 229 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
peak memory: 88.08 MiB, increment: 0.27 MiB
protocol=5
1.53 ms ± 64.7 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
peak memory: 88.23 MiB, increment: -0.80 MiB

So the change of protocol does not seem to matter in this case.