Support Numpy structured arrays containing an object

requirements_dev_optional.txt

@@ -9,6 +9,7 @@ ipytree==0.2.1
 azure-storage-blob==12.8.1 # pyup: ignore
 redis==3.5.3
 types-redis
+types-setuptools
 pymongo==3.12.0
 # optional test requirements
 tox==3.24.1

zarr/storage.py

@@ -423,7 +423,7 @@ def _init_array_metadata(
         filters_config = []
 
     # deal with object encoding
-    if dtype == object:
+    if dtype.hasobject:
         if object_codec is None:
             if not filters:
                 # there are no filters so we can be sure there is no object codec

zarr/tests/test_core.py

@@ -15,6 +15,7 @@
 from numcodecs.compat import ensure_bytes, ensure_ndarray
 from numcodecs.tests.common import greetings
 from numpy.testing import assert_array_almost_equal, assert_array_equal
+from pkg_resources import parse_version
 
 from zarr.core import Array
 from zarr.meta import json_loads
@@ -1362,6 +1363,44 @@ def test_object_codec_warnings(self):
         if hasattr(z.store, 'close'):
             z.store.close()
 
+    @unittest.skipIf(parse_version(np.__version__) < parse_version('1.14.0'),
+                     "unsupported numpy version")
+    def test_structured_array_contain_object(self):
+
+        if "PartialRead" in self.__class__.__name__:
+            pytest.skip("partial reads of object arrays not supported")
+
+        # ----------- creation --------------
+
+        structured_dtype = [('c_obj', object), ('c_int', int)]
+        a = np.array([(b'aaa', 1),
+                      (b'bbb', 2)], dtype=structured_dtype)
+
+        # zarr-array with structured dtype require object codec
+        with pytest.raises(ValueError):
+            self.create_array(shape=a.shape, dtype=structured_dtype)
+
+        # create zarr-array by np-array
+        za = self.create_array(shape=a.shape, dtype=structured_dtype, object_codec=Pickle())
+        za[:] = a
+
+        # must be equal
+        assert_array_equal(a, za[:])
+
+        # ---------- indexing ---------------
+
+        assert za[0] == a[0]
+
+        za[0] = (b'ccc', 3)
+        za[1:2] = np.array([(b'ddd', 4)], dtype=structured_dtype)  # ToDo: not work with list
+        assert_array_equal(za[:], np.array([(b'ccc', 3), (b'ddd', 4)], dtype=structured_dtype))
+
+        za['c_obj'] = [b'eee', b'fff']
+        za['c_obj', 0] = b'ggg'
+        assert_array_equal(za[:], np.array([(b'ggg', 3), (b'fff', 4)], dtype=structured_dtype))
+        assert za['c_obj', 0] == b'ggg'
+        assert za[1, 'c_int'] == 4
+
     def test_iteration_exceptions(self):
         # zero d array
         a = np.array(1, dtype=int)
@@ -1893,6 +1932,10 @@ def test_object_arrays_danger(self):
         # Cannot hacking out object codec as N5 doesn't allow object codecs
         pass
 
+    def test_structured_array_contain_object(self):
+        # skip this one, this data type not supported by N5
+        pass
+
     def test_attrs_n5_keywords(self):
         z = self.create_array(shape=(1050,), chunks=100, dtype='i4')
         for k in n5_keywords:
@@ -2326,6 +2369,10 @@ def test_object_arrays_danger(self):
         # skip this one, cannot use delta with objects
         pass
 
+    def test_structured_array_contain_object(self):
+        # skip this one, cannot use delta on structured array
+        pass
+
 
 # custom store, does not support getsize()
 class CustomMapping(object):