Deprecate generating string dtypes with unspecified length

CONTRIBUTING.rst

@@ -270,6 +270,7 @@ their individual contributions.
 * `Tariq Khokhar <https://www.github.com/tkb>`_ (tariq@khokhar.net)
 * `Tessa Bradbury <https://www.github.com/tessereth>`_
 * `Tim Martin <https://www.github.com/timmartin>`_ (tim@asymptotic.co.uk)
+* `Thomas Kluyver <https://www.github.com/takluyver>`_ (thomas@kluyver.me.uk)
 * `Tom McDermott <https://www.github.com/sponster-au>`_ (sponster@gmail.com)
 * `Tyler Gibbons <https://www.github.com/kavec>`_ (tyler.gibbons@flexport.com)
 * `Tyler Nickerson <https://www.github.com/nmbrgts>`_

hypothesis-python/RELEASE.rst

@@ -0,0 +1,11 @@
+RELEASE_TYPE: minor
+
+This patch deprecates ``min_len`` or ``max_len`` of 0 in
+:func:`~hypothesis.extra.numpy.byte_string_dtypes` and
+:func:`~hypothesis.extra.numpy.unicode_string_dtypes`.
+The lower limit is now 1.
+
+Numpy uses a length of 0 in these dtypes to indicate an undetermined size,
+chosen from the data at array creation.
+However, as the :func:`~hypothesis.extra.numpy.arrays` strategy creates arrays
+before filling them, strings were truncated to 1 byte.

hypothesis-python/src/hypothesis/extra/numpy.py

@@ -587,20 +587,60 @@ def timedelta64_dtypes(max_period="Y", min_period="ns", endianness="?"):
 
 
 @defines_dtype_strategy
-def byte_string_dtypes(endianness="?", min_len=0, max_len=16):
+def byte_string_dtypes(endianness="?", min_len=1, max_len=16):
     # type: (str, int, int) -> st.SearchStrategy[np.dtype]
     """Return a strategy for generating bytestring dtypes, of various lengths
-    and byteorder."""
-    order_check("len", 0, min_len, max_len)
+    and byteorder.
+
+    While Hypothesis' string strategies can generate empty strings, string
+    dtypes with length 0 indicate that size is still to be determined, so
+    the minimum length for string dtypes is 1.
+    """
+    if min_len == 0:
+        note_deprecation(
+            "generating byte string dtypes for unspecified length ('S0') "
+            "is deprecated. min_len will be 1 instead.",
+            since="RELEASEDAY",
+        )
+        min_len = 1
+    if max_len == 0:
+        note_deprecation(
+            "generating byte string dtypes for unspecified length ('S0') "
+            "is deprecated. max_len will be 1 instead.",
+            since="RELEASEDAY",
+        )
+        max_len = 1
+
+    order_check("len", 1, min_len, max_len)
     return dtype_factory("S", list(range(min_len, max_len + 1)), None, endianness)
 
 
 @defines_dtype_strategy
-def unicode_string_dtypes(endianness="?", min_len=0, max_len=16):
+def unicode_string_dtypes(endianness="?", min_len=1, max_len=16):
     # type: (str, int, int) -> st.SearchStrategy[np.dtype]
     """Return a strategy for generating unicode string dtypes, of various
-    lengths and byteorder."""
-    order_check("len", 0, min_len, max_len)
+    lengths and byteorder.
+
+    While Hypothesis' string strategies can generate empty strings, string
+    dtypes with length 0 indicate that size is still to be determined, so
+    the minimum length for string dtypes is 1.
+    """
+    if min_len == 0:
+        note_deprecation(
+            "generating unicode string dtypes for unspecified length ('U0') "
+            "is deprecated. min_len will be 1 instead.",
+            since="RELEASEDAY",
+        )
+        min_len = 1
+    if max_len == 0:
+        note_deprecation(
+            "generating unicode string dtypes for unspecified length ('U0') "
+            "is deprecated. max_len will be 1 instead.",
+            since="RELEASEDAY",
+        )
+        max_len = 1
+
+    order_check("len", 1, min_len, max_len)
     return dtype_factory("U", list(range(min_len, max_len + 1)), None, endianness)
 
 

hypothesis-python/tests/numpy/test_argument_validation.py

@@ -24,6 +24,7 @@
 import hypothesis.strategies as st
 from hypothesis import given
 from hypothesis.errors import InvalidArgument
+from tests.common.utils import checks_deprecated_behaviour
 
 
 def e(a, **kwargs):
@@ -142,3 +143,17 @@ def test_bad_dtype_strategy(capsys, data):
     assert capsys.readouterr().out.startswith(
         "Got invalid dtype value=%r from strategy=just(%r), function=" % (val, val)
     )
+
+
+@checks_deprecated_behaviour
+@given(st.data())
+def test_byte_string_dtype_len_0(data):
+    s = nps.byte_string_dtypes(min_len=0, max_len=0)
+    assert data.draw(s).itemsize == 1
+
+
+@checks_deprecated_behaviour
+@given(st.data())
+def test_unicode_string_dtype_len_0(data):
+    s = nps.unicode_string_dtypes(min_len=0, max_len=0)
+    assert data.draw(s).itemsize == 4

hypothesis-python/tests/numpy/test_gen_data.py

@@ -194,11 +194,29 @@ def test_can_generate_scalar_dtypes(dtype):
     assert isinstance(dtype, np.dtype)
 
 
-@given(nps.nested_dtypes())
+@given(
+    nps.nested_dtypes(
+        subtype_strategy=st.one_of(
+            nps.scalar_dtypes(), nps.byte_string_dtypes(), nps.unicode_string_dtypes()
+        )
+    )
+)
 def test_can_generate_compound_dtypes(dtype):
     assert isinstance(dtype, np.dtype)
 
 
+@given(
+    nps.nested_dtypes(
+        subtype_strategy=st.one_of(
+            nps.scalar_dtypes(), nps.byte_string_dtypes(), nps.unicode_string_dtypes()
+        )
+    ).flatmap(lambda dt: nps.arrays(dtype=dt, shape=1))
+)
+def test_can_generate_data_compound_dtypes(arr):
+    # This is meant to catch the class of errors which prompted PR #2085
+    assert isinstance(arr, np.ndarray)
+
+
 @given(nps.nested_dtypes(max_itemsize=400), st.data())
 def test_infer_strategy_from_dtype(dtype, data):
     # Given a dtype