diff --git a/changes/4503.bugfix.md b/changes/4503.bugfix.md
new file mode 100644
index 0000000000..245ef25820
--- /dev/null
+++ b/changes/4503.bugfix.md
@@ -0,0 +1 @@
+Writing one field of a structured array with `fields=` no longer overwrites the other fields of each selected record. A tuple naming several fields is now rejected for writes, like a list.
diff --git a/src/zarr/core/array.py b/src/zarr/core/array.py
index 06f7338755..0ced32074c 100644
--- a/src/zarr/core/array.py
+++ b/src/zarr/core/array.py
@@ -5846,6 +5846,8 @@ async def _set_selection(
     # check fields are sensible
     check_fields(fields, dtype)
     fields = check_no_multi_fields(fields)
+    # the dtype of the values being written: a single field, or the whole record
+    value_dtype = check_fields(fields, dtype)
 
     # check value shape
     if np.isscalar(value):
@@ -5854,21 +5856,39 @@ async def _set_selection(
             # TODO: need to handle array types that don't support __array_function__
             # like PyTorch and JAX
             array_like_ = cast("np._typing._SupportsArrayFunc", array_like)
-        value = np.asanyarray(value, dtype=dtype, like=array_like_)
+        value = np.asanyarray(value, dtype=value_dtype, like=array_like_)
     else:
         if not hasattr(value, "shape"):
-            value = np.asarray(value, dtype)
+            value = np.asarray(value, value_dtype)
         # assert (
         #     value.shape == indexer.shape
         # ), f"shape of value doesn't match indexer shape. Expected {indexer.shape}, got {value.shape}"
-        if not hasattr(value, "dtype") or value.dtype.name != dtype.name:
+        if not hasattr(value, "dtype") or value.dtype.name != value_dtype.name:
             if hasattr(value, "astype"):
                 # Handle things that are already NDArrayLike more efficiently
-                value = value.astype(dtype=dtype, order="A")
+                value = value.astype(dtype=value_dtype, order="A")
             else:
-                value = np.array(value, dtype=dtype, order="A")
+                value = np.array(value, dtype=value_dtype, order="A")
     value = cast("NDArrayLike", value)
 
+    if fields:
+        # A field write must preserve the other fields of each record, so read the
+        # selected records, replace the field, and write the whole records back.
+        # Missing chunks are read as the fill value, as for any partial chunk write.
+        records = prototype.nd_buffer.empty(shape=indexer.shape, dtype=dtype)
+        await _get_selection(
+            store_path,
+            metadata,
+            codec_pipeline,
+            replace(config, read_missing_chunks=True),
+            chunk_grid,
+            indexer,
+            prototype=prototype,
+            out=records,
+        )
+        records[fields] = value
+        value = records.as_ndarray_like()
+
     # We accept any ndarray like object from the user and convert it
     # to an NDBuffer (or subclass). From this point onwards, we only pass
     # Buffer and NDBuffer between components.
diff --git a/src/zarr/core/indexing.py b/src/zarr/core/indexing.py
index 8ea59eb960..4a7553d06f 100644
--- a/src/zarr/core/indexing.py
+++ b/src/zarr/core/indexing.py
@@ -1538,12 +1538,12 @@ def check_fields(fields: Fields | None, dtype: np.dtype[Any]) -> np.dtype[Any]:
         return dtype
 
 
-def check_no_multi_fields(fields: Fields | None) -> Fields | None:
-    if isinstance(fields, list):
-        if len(fields) == 1:
-            return fields[0]
-        elif len(fields) > 1:
+def check_no_multi_fields(fields: Fields | None) -> str | None:
+    if isinstance(fields, list | tuple):
+        if len(fields) > 1:
             raise IndexError("multiple fields are not supported for this operation")
+        # an empty sequence selects no fields, i.e. the whole record
+        return fields[0] if fields else None
     return fields
 
 
diff --git a/tests/test_indexing.py b/tests/test_indexing.py
index 785f089513..0ca9fde64e 100644
--- a/tests/test_indexing.py
+++ b/tests/test_indexing.py
@@ -2044,9 +2044,9 @@ def test_get_selections_with_fields(store: StorePath) -> None:
         z.get_basic_selection(Ellipsis, fields=slice(None))  # type: ignore[arg-type]
 
 
-@pytest.mark.xfail(reason="fields are not supported in v3")
+@pytest.mark.filterwarnings("ignore::zarr.errors.UnstableSpecificationWarning")
 def test_set_selections_with_fields(store: StorePath) -> None:
-    """Would verify that basic, orthogonal, coordinate, and mask set-selections with structured-array `fields` correctly write individual fields and reject multi-field assignment (xfail: fields unsupported in v3)."""
+    """Basic, orthogonal, coordinate, and mask set-selections with structured-array `fields` write only the named field and reject multi-field assignment."""
     v = np.array(
         [("aaa", 1, 4.2), ("bbb", 2, 8.4), ("ccc", 3, 12.6)],
         dtype=[("foo", "S3"), ("bar", "i4"), ("baz", "f8")],
@@ -2073,11 +2073,11 @@ def test_set_selections_with_fields(store: StorePath) -> None:
             with pytest.raises(IndexError):
                 z.set_basic_selection(Ellipsis, v, fields=fields)
             with pytest.raises(IndexError):
-                z.set_orthogonal_selection([0, 2], v, fields=fields)  # type: ignore[arg-type]
+                z.set_orthogonal_selection([0, 2], v[[0, 2]], fields=fields)  # type: ignore[arg-type]
             with pytest.raises(IndexError):
-                z.set_coordinate_selection([0, 2], v, fields=fields)
+                z.set_coordinate_selection([0, 2], v[[0, 2]], fields=fields)
             with pytest.raises(IndexError):
-                z.set_mask_selection([True, False, True], v, fields=fields)  # type: ignore[arg-type]
+                z.set_mask_selection([True, False, True], v[[0, 2]], fields=fields)  # type: ignore[arg-type]
 
         else:
             if isinstance(fields, list) and len(fields) == 1:
@@ -2131,6 +2131,55 @@ def test_set_selections_with_fields(store: StorePath) -> None:
             assert_array_equal(a, z[:])
 
 
+@pytest.mark.parametrize(
+    ("zarr_format", "shards"), [(2, None), (3, None), (3, (6,))], ids=["v2", "v3", "v3-sharded"]
+)
+@pytest.mark.parametrize("read_missing_chunks", [True, False])
+def test_set_field_preserves_other_fields(
+    zarr_format: ZarrFormat, shards: tuple[int, ...] | None, read_missing_chunks: bool
+) -> None:
+    """Writing one field of a structured array leaves the other fields of each record unchanged, including in chunks that were never written."""
+    dtype = np.dtype([("a", "i4"), ("b", "f8"), ("c", "i2")])
+    data = np.array([(i, i * 1.5, -i) for i in range(10)], dtype=dtype)
+    z = zarr.create_array(
+        {},
+        shape=(12,),
+        chunks=(3,),
+        shards=shards,
+        dtype=dtype,
+        zarr_format=zarr_format,
+        fill_value=None if zarr_format == 2 else 0,
+        config={"read_missing_chunks": read_missing_chunks},
+    )
+    # the last chunk (elements 9-11) is only partly written, the last shard is missing
+    z[:10] = data
+    expected = np.zeros(12, dtype=dtype)
+    expected[:10] = data
+
+    z.set_basic_selection(slice(2, 8), np.full(6, 7.0), fields="b")
+    expected["b"][2:8] = 7.0
+    z.set_basic_selection(4, 9, fields=["a"])
+    expected["a"][4] = 9
+    z.set_orthogonal_selection([0, 11], np.array([-5, -6], dtype="i2"), fields="c")
+    expected["c"][[0, 11]] = [-5, -6]
+    z.set_coordinate_selection([10, 1], 3.25, fields="b")
+    expected["b"][[10, 1]] = 3.25
+    mask = np.zeros(12, dtype=bool)
+    mask[[5, 9]] = True
+    z.set_mask_selection(mask, 8, fields="a")
+    expected["a"][mask] = 8
+
+    assert_array_equal(z[:], expected)
+
+
+def test_set_multiple_fields_as_tuple_raises() -> None:
+    """A tuple naming several fields is rejected for writes, like a list."""
+    dtype = np.dtype([("a", "i4"), ("b", "f8")])
+    z = zarr.create_array({}, shape=(4,), chunks=(2,), dtype=dtype, fill_value=0)
+    with pytest.raises(IndexError, match="multiple fields are not supported"):
+        z.set_basic_selection(slice(None), 1, fields=("a", "b"))
+
+
 def test_slice_selection_uints() -> None:
     """make_slice_selection accepts unsigned integer indices without error and produces correct shape."""
     arr = np.arange(24).reshape((4, 6))