Update thanks to @shoyer

Author: znicholls

xarray/coding/strings.py

@@ -18,7 +18,7 @@
 
 def create_vlen_dtype(element_type):
     if element_type not in (str, bytes):
-        raise TypeError("Unsupported type for vlen_dtype: `{}`".format(element_type))
+        raise TypeError("unsupported type for vlen_dtype: {!r}".format(element_type))
     # based on h5py.special_dtype
     return np.dtype("O", metadata={"element_type": element_type})
 

xarray/conventions.py

@@ -157,8 +157,10 @@ def _infer_dtype(array, name=None):
         return np.dtype(float)
 
     element = array[(0,) * array.ndim]
+    # We use the base types to avoid subclasses of bytes and str (which might
+    # not play nice with e.g. hdf5 datatypes), such as those from numpy
     if isinstance(element, bytes):
-        return strings.create_vlen_dtype(type(element))
+        return strings.create_vlen_dtype(bytes)
     elif isinstance(element, str):
         return strings.create_vlen_dtype(str)
 

xarray/tests/test_coding_strings.py

@@ -36,13 +36,10 @@ def test_vlen_dtype():
     assert strings.check_vlen_dtype(np.dtype(object)) is None
 
 
-@pytest.mark.parametrize("str_type", (str, np.str_))
-def test_numpy_str_handling(str_type):
-    dtype = strings.create_vlen_dtype(str_type)
-    assert dtype.metadata["element_type"] == str_type
-    assert strings.is_unicode_dtype(dtype)
-    assert not strings.is_bytes_dtype(dtype)
-    assert strings.check_vlen_dtype(dtype) is str_type
+@pytest.mark.parametrize("numpy_str_type", (np.str_, np.bytes_))
+def test_numpy_subclass_handling(numpy_str_type):
+    with pytest.raises(TypeError, match="unsupported type for vlen_dtype"):
+        strings.create_vlen_dtype(numpy_str_type)
 
 
 @requires_netCDF4