From 9bfe855a6cfd6985073fd5e4e5b5be027628eca8 Mon Sep 17 00:00:00 2001 From: Davis Bennett Date: Tue, 14 Jul 2026 13:25:16 +0200 Subject: [PATCH 1/2] fix: byte-order handling for structured dtypes in the bytes codec (#220) * fix: byte-order handling for structured dtypes in the bytes codec The bytes codec neither byte-swapped structured-dtype fields to its configured endian on encode (numpy reports byteorder '|' for void dtypes, so the top-level byteorder comparison never detected a mismatch) nor honored its endian when decoding, silently corrupting any structured data whose field byte order differed from the stored one (e.g. virtual references to external big-endian data). Encode now detects byte-order mismatches by comparing full dtypes via newbyteorder, and decode reinterprets raw bytes in the stored byte order before converting to the data type's declared byte order, so the stored layout (codec state) and the in-memory layout (array data type) are independent. Closes #4141 Assisted-by: ClaudeCode:claude-fable-5 * test: fold structured byte-order cases into existing bytes codec tests Extend test_endian's parametrization with structured dtypes and test_bytes_codec_sync_roundtrip with endian/dtype parametrization plus stored-layout and decoded-dtype assertions, instead of adding parallel test functions for the same properties. Assisted-by: ClaudeCode:claude-fable-5 * refactor: rename stored_dtype to view_dtype in BytesCodec decode The variable is the dtype used to view the raw chunk bytes (byte order from the codec's endian configuration), not a property of the stored data or of the returned buffer, which always carries the array's declared dtype. Assisted-by: ClaudeCode:claude-fable-5 * docs: note that the decode-side byte-order conversion copies the chunk Assisted-by: ClaudeCode:claude-fable-5 --- changes/4141.bugfix.md | 1 + src/zarr/codecs/bytes.py | 33 +++++++++++----- tests/test_codecs/test_bytes.py | 69 ++++++++++++++++++++++++++------- 3 files changed, 80 insertions(+), 23 deletions(-) create mode 100644 changes/4141.bugfix.md diff --git a/changes/4141.bugfix.md b/changes/4141.bugfix.md new file mode 100644 index 0000000000..6a132da3f5 --- /dev/null +++ b/changes/4141.bugfix.md @@ -0,0 +1 @@ +Fix silent byte-order corruption for structured dtypes with the `bytes` codec: multi-byte fields are now byte-swapped to the codec's configured `endian` on write and decoded honoring it on read, so non-native-endian structured data (e.g. big-endian fields, as produced by virtual references to external data) round-trips correctly. diff --git a/src/zarr/codecs/bytes.py b/src/zarr/codecs/bytes.py index 240c077627..fae762fd08 100644 --- a/src/zarr/codecs/bytes.py +++ b/src/zarr/codecs/bytes.py @@ -100,14 +100,28 @@ def _decode_sync( chunk_spec: ArraySpec, ) -> NDBuffer: endian_str = self.endian + dtype = chunk_spec.dtype.to_native_dtype() + # The byte order of the stored data is set by this codec's `endian` + # configuration; the byte order of the decoded array is set by the array's + # data type. The two are independent: the raw bytes are viewed with a dtype + # in the stored byte order, then converted to the declared dtype if needed. if isinstance(chunk_spec.dtype, HasEndianness): - dtype = replace(chunk_spec.dtype, endianness=endian_str).to_native_dtype() # type: ignore[call-arg] + view_dtype = replace(chunk_spec.dtype, endianness=endian_str).to_native_dtype() # type: ignore[call-arg] + elif isinstance(chunk_spec.dtype, Struct) and endian_str is not None: + # Per the struct data type spec, all multi-byte fields are stored in the + # byte order configured on this codec. + view_dtype = dtype.newbyteorder(endian_str) else: - dtype = chunk_spec.dtype.to_native_dtype() + view_dtype = dtype as_array_like = chunk_bytes.as_array_like() chunk_array = chunk_spec.prototype.nd_buffer.from_ndarray_like( - as_array_like.view(dtype=dtype) # type: ignore[attr-defined] + as_array_like.view(dtype=view_dtype) # type: ignore[attr-defined] ) + if view_dtype != dtype: + # This byte-swapping conversion copies the chunk. The dtype inequality + # guard keeps the common case, where the stored and declared byte orders + # already match, on the zero-copy view path above. + chunk_array = chunk_array.astype(dtype) # ensure correct chunk shape if chunk_array.shape != chunk_spec.shape: @@ -129,13 +143,14 @@ def _encode_sync( chunk_spec: ArraySpec, ) -> Buffer | None: assert isinstance(chunk_array, NDBuffer) - if ( - chunk_array.dtype.itemsize > 1 - and self.endian is not None - and self.endian != chunk_array.byteorder - ): + if chunk_array.dtype.itemsize > 1 and self.endian is not None: + # Compare full dtypes rather than the top-level byteorder: numpy reports + # byteorder '|' for structured dtypes even when their fields are + # byte-order-sensitive, so newbyteorder is the only reliable way to + # detect (and normalize) a byte-order mismatch. new_dtype = chunk_array.dtype.newbyteorder(self.endian) - chunk_array = chunk_array.astype(new_dtype) + if new_dtype != chunk_array.dtype: + chunk_array = chunk_array.astype(new_dtype) nd_array = chunk_array.as_ndarray_like() # Flatten the nd-array (only copy if needed) and reinterpret as bytes diff --git a/tests/test_codecs/test_bytes.py b/tests/test_codecs/test_bytes.py index 03dd0b40c6..ead778f526 100644 --- a/tests/test_codecs/test_bytes.py +++ b/tests/test_codecs/test_bytes.py @@ -33,29 +33,50 @@ @pytest.mark.parametrize("store", ["local", "memory"], indirect=["store"]) -@pytest.mark.parametrize("input_dtype", [">u2", "u2", + "f4"), ("mask", ">i4")], + [("flux", "u2", " None: """ The `bytes` codec stores multi-byte data in the byte order configured on the codec, regardless of the input array's byte order, and reads it back to the - original values. The input-dtype/store-endian cross-product exercises the - encode-side byteswap (input byte order != store byte order) and the no-op - case alike. Compression is disabled so the stored chunk is the codec's raw - output and its byte layout can be asserted directly. + original values. For structured dtypes this applies to every multi-byte + field, per the `struct` data type spec; the struct cases guard against the + endianness bugs from + https://github.com/zarr-developers/zarr-python/issues/4141, where the + encode path never byte-swapped struct fields (numpy reports byteorder '|' + for void dtypes) and the decode path ignored the codec's endian entirely. + The input-dtype/store-endian cross-product exercises the encode-side + byteswap (input byte order != store byte order) and the no-op case alike. + Compression is disabled so the stored chunk is the codec's raw output and + its byte layout can be asserted directly. """ - data = np.arange(0, 256, dtype=input_dtype).reshape((16, 16)) + dtype = np.dtype(input_dtype) + if dtype.fields is None: + data = np.arange(0, 256, dtype=dtype).reshape((16, 16)) + else: + data = np.zeros((16, 16), dtype=dtype) + data["flux"] = np.arange(0, 256).reshape((16, 16)) + data["mask"] = np.arange(256, 512).reshape((16, 16)) path = "endian" spath = StorePath(store, path) a = await zarr.api.asynchronous.create_array( spath, shape=data.shape, chunks=(16, 16), - dtype="uint16", + dtype=dtype, fill_value=0, compressors=None, serializer=BytesCodec(endian=store_endian), @@ -66,8 +87,7 @@ async def test_endian( # The stored chunk is laid out in the byte order configured on the codec. stored = await store.get(f"{path}/c/0/0", prototype=default_buffer_prototype()) assert stored is not None - expected_dtype = ">u2" if store_endian == "big" else " None: assert isinstance(BytesCodec(), SupportsSyncCodec) -def test_bytes_codec_sync_roundtrip() -> None: - codec = BytesCodec() - arr = np.arange(100, dtype="float64") +@pytest.mark.parametrize("endian", ENDIAN) +@pytest.mark.parametrize( + "native_dtype", + [np.dtype("float64"), np.dtype(">u2"), np.dtype([("a", ">f4"), ("b", " None: + """ + The synchronous encode/decode path round-trips data, and the two byte + orders involved are independent: the codec's `endian` configuration governs + only the stored byte layout (every multi-byte value, including struct + fields, is laid out in the codec's byte order regardless of the input + array's byte order), while the decoded buffer's byte order is governed by + the array's data type regardless of the codec's. The mixed-endian struct + case pins that per-field byte order of the in-memory dtype survives a + roundtrip through a single stored byte order. + """ + if native_dtype.fields is None: + arr = np.arange(100, dtype=native_dtype) + else: + arr = np.array([(1.5, 2), (3.5, 4), (5.5, 6), (7.5, 8)], dtype=native_dtype) zdtype = get_data_type_from_native_dtype(arr.dtype) spec = ArraySpec( shape=arr.shape, @@ -91,11 +129,14 @@ def test_bytes_codec_sync_roundtrip() -> None: ) nd_buf: NDBuffer = default_buffer_prototype().nd_buffer.from_numpy_array(arr) - codec = codec.evolve_from_array_spec(spec) + codec = BytesCodec(endian=endian).evolve_from_array_spec(spec) encoded = codec._encode_sync(nd_buf, spec) assert encoded is not None + assert encoded.to_bytes() == arr.astype(native_dtype.newbyteorder(endian)).tobytes() + decoded = codec._decode_sync(encoded, spec) + assert decoded.dtype == zdtype.to_native_dtype() np.testing.assert_array_equal(arr, decoded.as_numpy_array()) From 87802b7eb0776f08037f3d3e4f66f7e53452349f Mon Sep 17 00:00:00 2001 From: Davis Vann Bennett Date: Wed, 22 Jul 2026 15:16:39 +0200 Subject: [PATCH 2/2] docs: amend 3.3.0 release notes with behavior-change and dependency details Add pre-release clarifications to the 3.3.0 section: - #4141: note the read-time behavior change for structured-dtype arrays written by <= 3.2.1 with an explicitly non-native bytes-codec endian. - #3417: note that V3 metadata omitting endian for multi-byte dtypes now fails at open time with ValueError instead of assuming native order. - New Misc entry: typing_extensions >= 4.14 is now required (Sentinel), and the gpu extra skips CuPy on macOS. - #3963 / #3968: clarify that enum-specific idioms on the now-string attributes raise immediately (no deprecation cycle), while equality against enum members still holds. Assisted-by: ClaudeCode:claude-fable-5 --- docs/release-notes.md | 29 +++++++++++++++++++++++++++-- 1 file changed, 27 insertions(+), 2 deletions(-) diff --git a/docs/release-notes.md b/docs/release-notes.md index 3fd8a5f360..2f5c816c9f 100644 --- a/docs/release-notes.md +++ b/docs/release-notes.md @@ -27,7 +27,12 @@ - Fixed `BytesCodec.from_dict` so that `BytesCodec` instances roundtrip to / from their dict representation. `BytesCodec.from_dict` now interprets a missing `endian` configuration as `endian=None` (matching what `BytesCodec.to_dict` - emits), instead of falling back to the system's native byte order. ([#3417](https://github.com/zarr-developers/zarr-python/issues/3417)) + emits), instead of falling back to the system's native byte order. As a + consequence, Zarr V3 array metadata that omits the `endian` configuration for + a multi-byte data type — which is spec-noncompliant, and only producible by + third-party writers, since zarr-python has always written the field — + previously opened and decoded assuming machine-native byte order, and now + fails at open time with an informative `ValueError`. ([#3417](https://github.com/zarr-developers/zarr-python/issues/3417)) - Fixed `save_array`, `Group.__setitem__`, and `load` for 0-dimensional arrays. ([#3469](https://github.com/zarr-developers/zarr-python/issues/3469)) - Fixed inner-codec spec evolution for sharded arrays. The sharding codec now threads the array spec through its inner codec chain when evolving codecs, so a codec that changes the dtype upstream of `BytesCodec` no longer leaves the inner chain evolved against the wrong spec (which previously failed at decode time). This runs on the default `BatchedCodecPipeline` as well. Standard inner chains (`[BytesCodec]`, `[BytesCodec, ZstdCodec]`, transpose + bytes) are byte-identical to before. Restores the behavior of #2179. ([#3885](https://github.com/zarr-developers/zarr-python/issues/3885)) - Make chunk normalization properly handle `-1` as a compact representation of the @@ -80,7 +85,9 @@ ([#4074](https://github.com/zarr-developers/zarr-python/issues/4074)) - Fixed writing Fortran-ordered (F-contiguous) arrays through the variable-length string and bytes codecs and through numcodecs array-array filters such as `Delta`, `FixedScaleOffset` and `PackBits`. Chunks are now passed to numcodecs as C-contiguous arrays, so elements are no longer stored in transposed order. ([#4116](https://github.com/zarr-developers/zarr-python/issues/4116)) -- Fix silent byte-order corruption for structured dtypes with the `bytes` codec: multi-byte fields are now byte-swapped to the codec's configured `endian` on write and decoded honoring it on read, so non-native-endian structured data (e.g. big-endian fields, as produced by virtual references to external data) round-trips correctly. ([#4141](https://github.com/zarr-developers/zarr-python/issues/4141)) +- Fix silent byte-order corruption for structured dtypes with the `bytes` codec: multi-byte fields are now byte-swapped to the codec's configured `endian` on write and decoded honoring it on read, so non-native-endian structured data (e.g. big-endian fields, as produced by virtual references to external data) round-trips correctly. + + **Behavior change for pre-existing data**: structured-dtype arrays written by zarr-python <= 3.2.1 with an explicitly non-native `endian` in the bytes codec configuration (e.g. `endian="big"` on a little-endian machine) were stored un-swapped, in violation of their metadata; 3.3.0 now honors the metadata on read, so such arrays read back byte-swapped relative to what 3.2.1 returned. Data written with the default (native) `endian` configuration — the overwhelmingly common case — is byte-identical and unaffected. ([#4141](https://github.com/zarr-developers/zarr-python/issues/4141)) ### Improved Documentation @@ -129,6 +136,15 @@ access, and will be removed in a future release. ``BloscCodec.cname`` and ``BloscCodec.shuffle`` are now plain strings rather than enum members. + Because these attributes are now plain strings, enum-specific idioms fail + immediately rather than going through a deprecation cycle: accessing + ``.value`` or ``.name`` on e.g. ``codec.shuffle`` raises ``AttributeError``; + calling the class (``BloscShuffle("bitshuffle")``), subscripting it + (``BloscShuffle["bitshuffle"]``), and iterating over it all raise + ``TypeError``; and ``isinstance`` checks against the enum classes now always + evaluate ``False``. Equality comparisons against enum members still evaluate + ``True``. + Additional renames in ``zarr.codecs.blosc`` from the same change: the type aliases ``Shuffle`` and ``CName`` are now ``BloscShuffleLiteral`` and ``BloscCnameLiteral``, the constant ``SHUFFLE`` is now ``BLOSC_SHUFFLE`` @@ -143,6 +159,14 @@ and ``ShardingCodec.index_location`` are now plain strings rather than enum members. + Because these attributes are now plain strings, enum-specific idioms fail + immediately rather than going through a deprecation cycle: accessing + ``.value`` or ``.name`` on e.g. ``codec.endian`` raises ``AttributeError``; + calling the class (``Endian("little")``), subscripting it, and iterating + over it all raise ``TypeError``; and ``isinstance`` checks against the enum + classes now always evaluate ``False``. Equality comparisons against enum + members still evaluate ``True``. + Two follow-on changes from this deprecation: - ``NDBuffer.byteorder`` now returns a literal string (``"little"`` or @@ -162,6 +186,7 @@ ### Misc +- The minimum supported version of `typing_extensions` is now 4.14, which is required at runtime for `typing_extensions.Sentinel`. The `gpu` optional-dependency extra no longer installs CuPy on macOS (`cupy-cuda12x` now carries the environment marker `sys_platform != 'darwin'`). ([#4046](https://github.com/zarr-developers/zarr-python/issues/4046)) - [#214](https://github.com/zarr-developers/zarr-python/issues/214), [#215](https://github.com/zarr-developers/zarr-python/issues/215), [#3908](https://github.com/zarr-developers/zarr-python/issues/3908), [#3972](https://github.com/zarr-developers/zarr-python/issues/3972), [#3975](https://github.com/zarr-developers/zarr-python/issues/3975), [#3979](https://github.com/zarr-developers/zarr-python/issues/3979), [#3990](https://github.com/zarr-developers/zarr-python/issues/3990), [#3998](https://github.com/zarr-developers/zarr-python/issues/3998), [#4000](https://github.com/zarr-developers/zarr-python/issues/4000), [#4001](https://github.com/zarr-developers/zarr-python/issues/4001), [#4046](https://github.com/zarr-developers/zarr-python/issues/4046), [#4054](https://github.com/zarr-developers/zarr-python/issues/4054), [#4073](https://github.com/zarr-developers/zarr-python/issues/4073), [#4086](https://github.com/zarr-developers/zarr-python/issues/4086), [#4138](https://github.com/zarr-developers/zarr-python/issues/4138)