on_csv_write="update" crashes on second write when a column is empty in all existing rows

[davidberenstein1957]

Problem

With on_csv_write="update", the second write for a given run_id crashes when any column is empty in every existing CSV row. This is easy to hit in practice: an OfflineEmissionsTracker leaves longitude/latitude empty, and gpu_count/gpu_model are empty on CPU-only machines. Any run that writes twice — a flush() followed by stop(), or a tracker with measure_power_secs scheduling — crashes on the second write.

Reproduction

import tempfile
from codecarbon.output_methods.emissions_data import EmissionsData
from codecarbon.output_methods.file import FileOutput

d = EmissionsData(
    timestamp="2023-01-01T00:00:00", project_name="p", run_id="r", experiment_id="e",
    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="R",
    cloud_provider="", cloud_region="", os="TestOS", python_version="3.8",
    codecarbon_version="2.0", cpu_count=4, cpu_model="CPU", gpu_count=None,
    gpu_model=None, longitude="", latitude="", ram_total_size=16, tracking_mode="machine",
)

f = FileOutput("test.csv", tempfile.mkdtemp(), on_csv_write="update")
f.out(d, None)
f.out(d, None)  # raises

Root cause

codecarbon/output_methods/file.py:122, in the branch handling exactly one existing row with the current run_id:

update_values = {}
for col, val in dict(total.values).items():
    update_values[col] = df[col].dtype.type(val)

df[col].dtype.type(val) coerces the incoming value to whatever dtype pandas inferred when reading the existing CSV back. A column that is empty in every existing row is read back as float64, so this evaluates numpy.float64("") or numpy.float64(None).

Expected vs actual

Expected: the existing row for the run is replaced by the new values, empty columns included.

Actual:

File "codecarbon/output_methods/file.py", line 122, in out
    update_values[col] = df[col].dtype.type(val)
ValueError: could not convert string to float: ''

(None instead of "" gives TypeError: float() argument must be a string or a real number, not 'NoneType'.)

The coercion serves no purpose here — pandas can assign the values directly, or the row can simply be rebuilt.