diff --git a/docs/release-notes/4142.fix.md b/docs/release-notes/4142.fix.md
new file mode 100644
index 0000000000..9305dcfed4
--- /dev/null
+++ b/docs/release-notes/4142.fix.md
@@ -0,0 +1 @@
+Download cached dataset files atomically so that parallel processes sharing a dataset directory (e.g. `pytest-xdist` workers) can no longer read a partially-written file {smaller}`gaoflow`
diff --git a/src/scanpy/readwrite.py b/src/scanpy/readwrite.py
index 9a3d3748fe..fc10b07bee 100644
--- a/src/scanpy/readwrite.py
+++ b/src/scanpy/readwrite.py
@@ -1084,6 +1084,7 @@ def _get_filename_from_key(key, ext=None) -> Path:
 
 def _download(url: str, path: Path):
     from ssl import create_default_context
+    from tempfile import NamedTemporaryFile
     from urllib.request import Request, urlopen
 
     from certifi import contents
@@ -1092,6 +1093,8 @@ def _download(url: str, path: Path):
     blocksize = 1024 * 8
     blocknum = 0
 
+    # Write to a temp file and rename so readers never see a partial file (#4097).
+    tmp_path: Path | None = None
     try:
         req = Request(url, headers={"User-agent": "scanpy-user"})
 
@@ -1105,8 +1108,14 @@ def _download(url: str, path: Path):
                     unit_divisor=1024,
                     total=total if total is None else int(total),
                 ) as t,
-                path.open("wb") as f,
+                NamedTemporaryFile(
+                    dir=path.parent,
+                    prefix=f"{path.name}.",
+                    suffix=".part",
+                    delete=False,
+                ) as f,
             ):
+                tmp_path = Path(f.name)
                 block = resp.read(blocksize)
                 while block:
                     f.write(block)
@@ -1114,10 +1123,13 @@ def _download(url: str, path: Path):
                     t.update(len(block))
                     block = resp.read(blocksize)
 
+        tmp_path.replace(path)
+        tmp_path = None
+
     except (KeyboardInterrupt, Exception):
-        # Make sure file doesn’t exist half-downloaded
-        if path.is_file():
-            path.unlink()
+        # Only remove our own temp file; leave path, which may be another process's.
+        if tmp_path is not None:
+            tmp_path.unlink(missing_ok=True)
         raise
 
 
diff --git a/tests/test_datasets.py b/tests/test_datasets.py
index e97d12a65a..7d72ddd236 100644
--- a/tests/test_datasets.py
+++ b/tests/test_datasets.py
@@ -21,6 +21,7 @@
 
 if TYPE_CHECKING:
     from collections.abc import Callable
+    from typing import Self
 
     from anndata import AnnData
 
@@ -168,6 +169,82 @@ def test_download_failure() -> None:
     excinfo.value.close()
 
 
+def test_download_atomic(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    """The destination must not appear until the download finished (#4097)."""
+    import io
+    import urllib.request
+
+    from scanpy.readwrite import _download
+
+    content = b"0123456789" * 5_000
+    dest = tmp_path / "cache" / "data.bin"
+    dest.parent.mkdir()
+    dest_present_during_download: list[bool] = []
+
+    class FakeResponse:
+        def __init__(self) -> None:
+            self._buf = io.BytesIO(content)
+
+        def info(self) -> dict[str, str]:
+            return {"content-length": str(len(content))}
+
+        def read(self, size: int) -> bytes:
+            chunk = self._buf.read(size)
+            if chunk:
+                dest_present_during_download.append(dest.exists())
+            return chunk
+
+        def __enter__(self) -> Self:
+            return self
+
+        def __exit__(self, *exc: object) -> bool:
+            return False
+
+    monkeypatch.setattr(urllib.request, "urlopen", lambda *a, **k: FakeResponse())
+
+    _download("http://example.invalid/data.bin", dest)
+
+    assert dest.read_bytes() == content
+    assert len(dest_present_during_download) > 1
+    assert not any(dest_present_during_download)
+    assert list(dest.parent.iterdir()) == [dest]
+
+
+def test_download_failure_keeps_existing_file(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """A failed download must not delete an already-present destination (#4097)."""
+    import urllib.request
+
+    from scanpy.readwrite import _download
+
+    dest = tmp_path / "cache" / "data.bin"
+    dest.parent.mkdir()
+    dest.write_bytes(b"complete")
+
+    class FailingResponse:
+        def info(self) -> dict[str, str]:
+            return {"content-length": "100"}
+
+        def read(self, size: int) -> bytes:
+            msg = "connection reset"
+            raise OSError(msg)
+
+        def __enter__(self) -> Self:
+            return self
+
+        def __exit__(self, *exc: object) -> bool:
+            return False
+
+    monkeypatch.setattr(urllib.request, "urlopen", lambda *a, **k: FailingResponse())
+
+    with pytest.raises(OSError, match="connection reset"):
+        _download("http://example.invalid/data.bin", dest)
+
+    assert dest.read_bytes() == b"complete"
+    assert list(dest.parent.iterdir()) == [dest]
+
+
 # These are tested via doctest
 DS_INCLUDED = frozenset({"krumsiek11", "toggleswitch", "pbmc68k_reduced"})
 # These have parameters that affect shape and so on