Using *nix operating systems on python 3.6 and 3.7, the dtype of fixed length ascii strings stored in compound dtypes changed from h5py v2.9 to v2.10.

Here's a small example, writing using v2.9 (though it doesn't seem to matter which I've written with):

import h5py
import numpy as np
import pandas as pd

d = pd.DataFrame({"a": [b"abc", b"def"]})
with h5py.File("test.h5", "a") as f:
    f["x"] = d.to_records(column_dtypes={"a": "S3"})

Now reading. Here's the code used:

import h5py
with h5py.File("test.h5", "r") as f:
    print(f["x"].dtype.descr)

Reading with h5py v2.9 prints:

[('index', '<i8'), ('a', '|S3')]

v2.10:

[('index', '<i8'), ('a', ('|S3', {'h5py_encoding': 'ascii'}))]

It's very possible this is intentional, and was done for good reasons. It is a change in behavior though, and I think should at least documented be in the change notes.