fix: S3Downloader - improve concurrent behavior

integrations/amazon_bedrock/src/haystack_integrations/components/downloaders/s3/s3_downloader.py

@@ -3,6 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import os
+import uuid
 from collections.abc import Callable
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
@@ -200,14 +201,19 @@ def _download_file(self, document: Document) -> Document | None:
 
         file_path = self.file_root_path / Path(file_name)
 
-        if file_path.is_file():
-            # set access and modification time to now without redownloading the file
-            file_path.touch()
-
-        else:
+        # if the file exists, avoid downloading it and just update the timestamp
+        try:
+            os.utime(file_path, None)
+        except FileNotFoundError:
             s3_key = self.s3_key_generation_function(document) if self.s3_key_generation_function else file_name
-            # we know that _storage is not None after warm_up() is called, but mypy does not know that
-            self._storage.download(key=s3_key, local_file_path=file_path)  # type: ignore[union-attr]
+            # download to a temp path to prevent other downloaders running concurrently to see a partially-written file
+            tmp_path = file_path.with_name(f"{file_path.name}.tmp-{uuid.uuid4().hex}")
+            try:
+                # we know that _storage is not None after warm_up() is called, but mypy does not know that
+                self._storage.download(key=s3_key, local_file_path=tmp_path)  # type: ignore[union-attr]
+                os.replace(tmp_path, file_path)
+            finally:
+                tmp_path.unlink(missing_ok=True)
 
         document.meta["file_path"] = str(file_path)
         return document

integrations/amazon_bedrock/tests/test_s3_downloader.py

@@ -308,6 +308,29 @@ def test_run_returns_empty_when_all_filtered(self, tmp_path, mock_s3_storage, mo
         assert out["documents"] == []
         mock_s3_storage.download.assert_not_called()
 
+    def test_download_writes_to_temp_path_then_renames(self, tmp_path, mock_boto3_session):
+        d = S3Downloader(file_root_path=str(tmp_path))
+
+        final_path = tmp_path / "test.pdf"
+        captured_paths = []
+
+        def fake_download(key, local_file_path: Path):
+            captured_paths.append(Path(local_file_path))
+            assert not final_path.exists(), "final path must not exist while download is in progress"
+            Path(local_file_path).write_bytes(b"complete content")
+
+        mock_storage = MagicMock(spec=S3Storage)
+        mock_storage.download.side_effect = fake_download
+        d._storage = mock_storage
+
+        d.run(documents=[Document(meta={"file_name": "test.pdf"})])
+
+        assert len(captured_paths) == 1
+        assert captured_paths[0] != final_path
+        assert captured_paths[0].name.startswith("test.pdf.tmp-")
+        assert final_path.exists()
+        assert final_path.read_bytes() == b"complete content"
+
     def test_cleanup_cache_evicts_old_files(self, tmp_path, mock_s3_storage, mock_boto3_session):
         d = S3Downloader(file_root_path=str(tmp_path), max_cache_size=1)
         d._storage = mock_s3_storage