diff --git a/tests/test_csv_ingestor.py b/tests/test_csv_ingestor.py
index a542e17..1353487 100644
--- a/tests/test_csv_ingestor.py
+++ b/tests/test_csv_ingestor.py
@@ -96,11 +96,24 @@ def test_read_data_unique_id_column_missing_raises(make_csv):
         list(ing.read_data(str(path)))
 
 
-def test_read_data_empty_file_returns_nothing(tmp_path):
+def test_read_data_empty_file_raises_with_clear_message(tmp_path):
+    """An empty (zero-byte) CSV is a hard input error, not a successful
+    '0 rows' run. The old code logged a WARNING and silently returned an
+    empty generator — the ingestor then created an empty MySQL table and
+    called `send_generate_edge_label_meta`, which 400'd with the misleading
+    'No data found for table X' message (same cascade #213 traced for
+    self-supervised + label mismatch). The user blamed the backend.
+
+    Fail fast at the read layer with a clear, source-truthful message
+    naming the path and pointing at staging as the likely cause —
+    DataValidator's existing 'No data found to validate' path catches it
+    before any backend round-trip.
+    """
     p = tmp_path / "empty.csv"
     p.write_text("")
     ing = make_csv_ingestor(schema={})
-    assert list(ing.read_data(str(p))) == []
+    with pytest.raises(ValueError, match="Empty CSV file"):
+        list(ing.read_data(str(p)))
 
 
 def test_validate_csv_type_coercion():
diff --git a/tracebloc_ingestor/ingestors/csv_ingestor.py b/tracebloc_ingestor/ingestors/csv_ingestor.py
index a8ff312..46c6f93 100644
--- a/tracebloc_ingestor/ingestors/csv_ingestor.py
+++ b/tracebloc_ingestor/ingestors/csv_ingestor.py
@@ -490,8 +490,24 @@ def read_data(self, file_path: str) -> Generator[Dict[str, Any], None, None]:
                     yield record
 
         except pd.errors.EmptyDataError:
-            logger.warning(f"{YELLOW}Empty CSV file: {file_path}{RESET}")
-            return
+            # An empty (zero-byte) CSV is a hard input error, not a
+            # successful "0 rows" run. Previously this branch logged a
+            # WARNING and silently returned an empty generator — the
+            # ingestor then proceeded to create an empty MySQL table and
+            # called `send_generate_edge_label_meta`, which 400'd with
+            # the misleading "No data found for table X" message that
+            # blamed the BACKEND instead of the input. (Same misleading
+            # cascade #213 traced for self-supervised + label mismatch.)
+            # Raise here so DataValidator's existing "No data found to
+            # validate" error path surfaces the empty-input cause with a
+            # clear, source-truthful message — and no backend round-trip.
+            raise ValueError(
+                f"{RED}Empty CSV file: {file_path}. The file has no "
+                f"header and no rows. Either stage a non-empty CSV at "
+                f"this path, or check the cluster-side path (the chart "
+                f"mounts your PVC at /data/shared/ — confirm staging "
+                f"completed before helm install).{RESET}"
+            )
 
         except (pd.errors.ParserError, Exception):
             raise