diff --git a/pyathena/filesystem/s3.py b/pyathena/filesystem/s3.py
index eb48eccee..e05abad65 100644
--- a/pyathena/filesystem/s3.py
+++ b/pyathena/filesystem/s3.py
@@ -1367,7 +1367,7 @@ def _finish_multipart_upload(
                 upload_id=upload_id,
                 parts=parts,
             )
-        except Exception:
+        except Exception as error:
             for future in futures:
                 future.cancel()
             try:
@@ -1381,6 +1381,9 @@ def _finish_multipart_upload(
                 _logger.exception(
                     f"Failed to abort multipart upload {upload_id} to s3://{bucket}/{key}."
                 )
+                # The upload still exists, so the caller must keep its upload ID
+                # for a later discard() to retry the abort.
+                error.multipart_abort_failed = True  # type: ignore[attr-defined]
             raise
 
     def cat_file(
@@ -2391,11 +2394,13 @@ def commit(self) -> None:
                     upload_id=cast(str, self.multipart_upload.upload_id),
                     futures=self.multipart_upload_parts,
                 )
-            except Exception:
-                # The multipart upload has been aborted by the helper;
-                # prevent discard() from aborting it again.
-                self.multipart_upload = None
-                self.multipart_upload_parts = []
+            except Exception as error:
+                # Unless the abort failed, the helper has aborted the upload;
+                # clear it so discard() does not abort it again. If the abort
+                # failed, keep the upload ID so discard() can retry the abort.
+                if not getattr(error, "multipart_abort_failed", False):
+                    self.multipart_upload = None
+                    self.multipart_upload_parts = []
                 raise
 
         self.fs.invalidate_cache(self.path)
diff --git a/tests/pyathena/filesystem/test_s3.py b/tests/pyathena/filesystem/test_s3.py
index 7e4f36e5b..0515bed0c 100644
--- a/tests/pyathena/filesystem/test_s3.py
+++ b/tests/pyathena/filesystem/test_s3.py
@@ -1955,6 +1955,45 @@ def test_discard(self, multipart):
         assert file.multipart_upload is None
         assert file.multipart_upload_parts == []
 
+    @pytest.mark.parametrize(
+        ("abort_fails", "keeps_upload_id"),
+        [(True, True), (False, False)],
+    )
+    def test_commit_failed_completion_upload_id(self, abort_fails, keeps_upload_id):
+        # When CompleteMultipartUpload fails, the abort runs. If the abort
+        # also fails, the upload still exists, so commit() must keep the upload
+        # ID for discard() to retry. Otherwise the upload ID is cleared, so a
+        # later discard() does not abort an upload that is already gone.
+        file = self._make_multipart_write_file(b"x" * 8, autocommit=False)
+        part = Future()
+        part.set_result(SimpleNamespace(etag='"e1"', part_number=1))
+        file.multipart_upload_parts = [part]
+        real_fs = S3FileSystem.__new__(S3FileSystem)
+        real_fs._client = mock.MagicMock()
+        real_fs._call = mock.MagicMock(
+            side_effect=RuntimeError("abort failed") if abort_fails else None
+        )
+        real_fs._complete_multipart_upload = mock.MagicMock(
+            side_effect=RuntimeError("complete failed")
+        )
+        file.fs._finish_multipart_upload.side_effect = lambda **kw: (
+            S3FileSystem._finish_multipart_upload(real_fs, **kw)
+        )
+        file.fs.invalidate_cache = mock.MagicMock()
+
+        with pytest.raises(RuntimeError, match="complete failed"):
+            file.commit()
+
+        assert (file.multipart_upload is not None) is keeps_upload_id
+        assert (file.multipart_upload_parts != []) is keeps_upload_id
+
+        file.discard()
+        if keeps_upload_id:
+            assert file.fs._call.call_args.args[0] == "abort_multipart_upload"
+            assert file.fs._call.call_args.kwargs["UploadId"] == "uploadid"
+        else:
+            file.fs._call.assert_not_called()
+
     @pytest.mark.parametrize("autocommit", [True, False])
     def test_upload_chunk_multipart(self, autocommit):
         # Multipart upload (CompleteMultipartUpload), completed from the uploaded