fix(codecs): reshape flattened numcodecs filter output on encode

changes/4560.bugfix.md

@@ -0,0 +1,8 @@
+Encoding a chunk through a numcodecs filter no longer passes the
+filter's flattened output to the next codec: `_encode_sync` reshapes the
+flattened array back to the chunk shape, so downstream codecs
+(e.g. `TransposeCodec` or an `imagecodecs` serializer chained after a
+`Delta` or `FixedScaleOffset` filter) receive a chunk of the expected
+shape instead of a flat 1-D array that could make writing fail or
+silently corrupt data. Shape-changing codecs such as `PackBits` are
+unaffected.

src/zarr/codecs/numcodecs/_codecs.py

@@ -164,6 +164,11 @@ def _encode_sync(self, chunk_data: NDBuffer, chunk_spec: ArraySpec) -> NDBuffer:
         # would be encoded in transposed element order (gh-3558)
         chunk_ndarray = np.ascontiguousarray(chunk_data.as_ndarray_like())
         out = self._codec.encode(chunk_ndarray)
+        if out.ndim == 1 and out.size == np.prod(chunk_spec.shape):
+            # most numcodecs filters return a flattened array; reshape it so
+            # downstream codecs receive a chunk of the shape the pipeline
+            # promised them. Shape-changing codecs (e.g. PackBits) keep out.
+            out = out.reshape(chunk_spec.shape)
         return chunk_spec.prototype.nd_buffer.from_ndarray_like(out)
 
     async def _encode_single(self, chunk_data: NDBuffer, chunk_spec: ArraySpec) -> NDBuffer:

tests/test_codecs/test_numcodecs.py

@@ -11,6 +11,7 @@
 from zarr import config, create_array, open_array
 from zarr.abc.numcodec import _is_numcodec, _is_numcodec_cls
 from zarr.codecs import numcodecs as _numcodecs
+from zarr.codecs.transpose import TransposeCodec
 from zarr.errors import UnknownCodecError
 from zarr.registry import get_codec_class, get_numcodec
 
@@ -192,6 +193,45 @@ def test_generic_filter_f_contiguous(
     np.testing.assert_array_equal(data, a[:, :])
 
 
+def test_generic_filter_flattened_output_reshaped() -> None:
+    # gh-4553: numcodecs filters return a flattened array; `_encode_sync` must
+    # reshape it back to the chunk shape, so downstream codecs receive chunks
+    # with the shape the pipeline promised them
+    data = np.arange(256, dtype="int64").reshape(16, 16)
+
+    a = create_array(
+        {},
+        shape=data.shape,
+        chunks=(16, 16),
+        dtype=data.dtype,
+        fill_value=0,
+        filters=[_numcodecs.Delta(dtype="int64"), TransposeCodec(order=(1, 0))],
+    )
+
+    a[:, :] = data
+    with codec_conf():
+        b = open_array(a.store, mode="r")
+    np.testing.assert_array_equal(data, b[:, :])
+
+
+def test_generic_filter_packedbits_output_not_reshaped() -> None:
+    # PackBits drives the size of the chunk, so `_encode_sync` must not reshape
+    # its output
+    data = np.tril(np.ones((16, 16), dtype=bool))
+
+    a = create_array(
+        {},
+        shape=data.shape,
+        chunks=(16, 16),
+        dtype=data.dtype,
+        fill_value=0,
+        filters=[_numcodecs.PackBits()],
+    )
+
+    a[:, :] = data
+    np.testing.assert_array_equal(data, a[:, :])
+
+
 def test_generic_filter_bitround() -> None:
     data = np.linspace(0, 1, 256, dtype="float32").reshape((16, 16))