From 900f7dd7b4e6bba96ad7cde0d6e8550795333796 Mon Sep 17 00:00:00 2001
From: David Berenstein <david.m.berenstein@gmail.com>
Date: Wed, 12 Aug 2026 18:08:01 +0200
Subject: [PATCH] fix: drop dtype coercion in CSV update path

Updating a run row coerced each incoming value to the dtype pandas
inferred for the existing column. Columns empty in every row are read
back as float64, so numpy.float64("") / numpy.float64(None) raised.
Rebuild the row via concat instead, which dedupes by run_id without
touching dtypes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 codecarbon/output_methods/file.py | 15 ++++-----
 tests/output_methods/test_file.py | 56 +++++++++++++++++++++++++++++++
 2 files changed, 62 insertions(+), 9 deletions(-)

diff --git a/codecarbon/output_methods/file.py b/codecarbon/output_methods/file.py
index 6a13d5b41..76eebf417 100644
--- a/codecarbon/output_methods/file.py
+++ b/codecarbon/output_methods/file.py
@@ -107,9 +107,7 @@ def out(self, total: EmissionsData, _):
         else:
             df = pd.read_csv(self.save_file_path)
             df_run = df.loc[df.run_id == total.run_id]
-            if len(df_run) < 1:
-                df = pd.concat([df, new_df])
-            elif len(df_run) > 1:
+            if len(df_run) > 1:
                 logger.warning(
                     f"CSV contains more than 1 ({len(df_run)})"
                     + f" rows with current run ID ({total.run_id})."
@@ -117,12 +115,11 @@ def out(self, total: EmissionsData, _):
                 )
                 df = pd.concat([df, new_df])
             else:
-                update_values = {}
-                for col, val in dict(total.values).items():
-                    update_values[col] = df[col].dtype.type(val)
-                df.loc[df.run_id == total.run_id, update_values.keys()] = (
-                    update_values.values()
-                )
+                # Drop the previous row for this run (if any) and re-append it.
+                # Assigning column by column would coerce values to the dtype
+                # pandas inferred for the existing column, which breaks for
+                # columns that are empty in every row (read back as float64).
+                df = pd.concat([df.loc[df.run_id != total.run_id], new_df])
             df.to_csv(self.save_file_path, index=False)
 
     def task_out(self, data: List[TaskEmissionsData], experiment_name: str):
diff --git a/tests/output_methods/test_file.py b/tests/output_methods/test_file.py
index e8bccfdf0..0b7f301e4 100644
--- a/tests/output_methods/test_file.py
+++ b/tests/output_methods/test_file.py
@@ -169,6 +169,62 @@ def test_file_output_out_update_file_exists_one_matchingrows(self):
         df = pd.read_csv(os.path.join(self.temp_dir, "test.csv"))
         self.assertEqual(df["cpu_power"].iloc[0], 2)
 
+    def test_file_output_out_update_with_always_empty_columns(self):
+        """Regression test: updating a run must not coerce incoming values to the
+        dtype pandas inferred for the existing column.
+
+        An OfflineEmissionsTracker leaves longitude/latitude empty, and
+        gpu_count/gpu_model are empty on CPU-only machines. Such columns are read
+        back from the CSV as float64, so the previous implementation evaluated
+        numpy.float64("") / numpy.float64(None) and raised on the second write.
+        """
+        empty_columns_data = EmissionsData(
+            timestamp="2023-01-01T00:00:00",
+            project_name="test_project",
+            run_id="test_run_id",
+            experiment_id="test_experiment_id",
+            duration=10,
+            emissions=0.5,
+            emissions_rate=0.05,
+            cpu_power=20,
+            gpu_power=0,
+            ram_power=5,
+            cpu_energy=200,
+            gpu_energy=0,
+            ram_energy=50,
+            energy_consumed=250,
+            water_consumed=0.1,
+            country_name="Testland",
+            country_iso_code="TS",
+            region="Test Region",
+            cloud_provider="",
+            cloud_region="",
+            os="TestOS",
+            python_version="3.8",
+            codecarbon_version="2.0",
+            cpu_count=4,
+            cpu_model="Test CPU",
+            gpu_count=None,
+            gpu_model=None,
+            longitude="",
+            latitude="",
+            ram_total_size=16,
+            tracking_mode="machine",
+        )
+
+        file_output = FileOutput("test.csv", self.temp_dir, on_csv_write="update")
+        file_output.out(empty_columns_data, None)
+
+        empty_columns_data.cpu_power = 2
+        # This should not raise.
+        file_output.out(empty_columns_data, None)
+
+        df = pd.read_csv(os.path.join(self.temp_dir, "test.csv"))
+        self.assertEqual(len(df), 1)
+        self.assertEqual(df["cpu_power"].iloc[0], 2)
+        self.assertIn("longitude", df.columns)
+        self.assertIn("gpu_model", df.columns)
+
     # def test_file_output_out_consistent_column_ordering(self):
     #     file_output = FileOutput("test.csv", self.temp_dir, on_csv_write="append")
     #     file_output.out(self.emissions_data, None)