BUG: inconsistent comparison results when using PyArrow backend.

[lenardkoomen-fin]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

for do_parq_roundtrip in [False, True]:
    for do_fillna in [False, True]:
        df = df_comparison.copy().reset_index(drop=True)

        print("=" * 100)
        parq_str = "with" if do_parq_roundtrip else "without"
        fillna_str = "with" if do_fillna else "without"
        print(f"{parq_str} parquet roundtrip, {fillna_str} fillna")
        print("=" * 100)

        # Do parquet roundtrip
        if do_parq_roundtrip:
            df.to_parquet("df.parquet", index=True)
            df = pd.read_parquet("df.parquet")

        print("Column datatypes:")
        print(df.dtypes)
        print()

        # Detect mismatch
        if do_fillna:
            is_mismatch = df.col_a.fillna(2) != df.col_b.fillna(2)
        else:
            is_mismatch = df.col_a != df.col_b

        print("Mismatch detected @:")
        print(np.argwhere(is_mismatch.fillna(False)))
        print()

        df["is_mismatch"] = is_mismatch

        # Print rows where there is a detected mismatch
        print("Detected mismatch rows:")
        print(df[df.is_mismatch])
        print("=" * 100)
        print()

Issue Description

I get inconsistent results when comparing two int64[pyarrow] columns, depending on whether I use fillna or first store the dataframe in parquet and read it again.

The input data is the result of merging two dataframes resulting from df = pd.read_sql(f"select * from some.Table", some_connection, dtype_backend="pyarrow"). I've redacted the results by selecting only certain columns and renaming them. The dataframe contains two columns (col_a and col_b) that are compared. These columns are of dtype int64[pyarrow] and contain either 0, 1, or NA. There is a third column ManualIndex which I've added to make sure nothing cheeky is happening with the index, but which is probably useless.

I've attached a .parquet and .feather export of the data. But keep in mind storing and then re-loading the data apparently has an effect on the output of the comparison, so loading this data and running the script probably gives a different output then the one I've posted below.

df.zip

The output of the script gives me the following results:

====================================================================================================
without parquet roundtrip, without fillna
====================================================================================================
Column datatypes:
ManualIndex             int64
col_a          int64[pyarrow]
col_b          int64[pyarrow]
dtype: object

Mismatch detected @:
[[252518]
 [252519]]

Detected mismatch rows:
        ManualIndex  col_a  col_b is_mismatch
252518       252518      1   <NA>        <NA>
252519       252519      1   <NA>        <NA>
====================================================================================================

====================================================================================================
without parquet roundtrip, with fillna
====================================================================================================
Column datatypes:
ManualIndex             int64
col_a          int64[pyarrow]
col_b          int64[pyarrow]
dtype: object

Mismatch detected @:
[[252512]
 [252513]
 [252518]
 [252519]]

Detected mismatch rows:
        ManualIndex  col_a  col_b is_mismatch
252512       252512      1      1        True
252513       252513      1      1        True
252518       252518      1   <NA>        True
252519       252519      1   <NA>        True
====================================================================================================

====================================================================================================
with parquet roundtrip, without fillna
====================================================================================================
Column datatypes:
ManualIndex             int64
col_a          int64[pyarrow]
col_b          int64[pyarrow]
dtype: object

Mismatch detected @:
[]

Detected mismatch rows:
Empty DataFrame
Columns: [ManualIndex, col_a, col_b, is_mismatch]
Index: []
====================================================================================================

====================================================================================================
with parquet roundtrip, with fillna
====================================================================================================
Column datatypes:
ManualIndex             int64
col_a          int64[pyarrow]
col_b          int64[pyarrow]
dtype: object

Mismatch detected @:
[[252505]
 [252506]]

Detected mismatch rows:
        ManualIndex  col_a  col_b is_mismatch
252505       252505      1      1        True
252506       252506      1      1        True
====================================================================================================

As you can see the results are different for each combination of settings.

Expected Behavior

My expectations:

• I expect the roundtrip to a parquet file and back to have no bearing on the comparison at all.
• I expect is_mismatch to be True only at ManualIndex 252518 and 252519 when using fillna(2) before comparing.
• I expect is_mismatch to be NA at ManualIndex 252518 and 252519 when not using fillna before comparing.
• I expect is_mismatch to be False everywhere else.
• I expect that when is_mismatch is NA that it does not get found by np.argwhere or result in returned rows when using it to index into a dataframe.

I don't understand how the rows where col_a and col_b are both 1 can ever result in them not being equal according to the comparison. When I filter on these rows and do the comparison manually suddenly the comparison evalues to them being equal.

In short I am very confused, am I doing something wrong here?

Installed Versions

INSTALLED VERSIONS

commit : 4665c10
python : 3.13.4
python-bits : 64
OS : Windows
OS-release : 11
Version : 10.0.26100
machine : AMD64
processor : Intel64 Family 6 Model 106 Stepping 6, GenuineIntel
byteorder : little
LC_ALL : None
LANG : en_US.UTF-8
LOCALE : English_Netherlands.1252

pandas : 2.3.2
numpy : 2.2.6
pytz : 2025.2
dateutil : 2.9.0.post0
pip : None
Cython : None
sphinx : None
IPython : 9.2.0
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : 4.13.4
blosc : None
bottleneck : None
dataframe-api-compat : None
fastparquet : None
fsspec : None
html5lib : 1.1
hypothesis : None
gcsfs : None
jinja2 : 3.1.6
lxml.etree : 5.4.0
matplotlib : 3.10.3
numba : None
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
psycopg2 : None
pymysql : None
pyarrow : 21.0.0
pyreadstat : None
pytest : 8.3.5
python-calamine : None
pyxlsb : None
s3fs : None
scipy : None
sqlalchemy : 2.0.42
tables : None
tabulate : None
xarray : None
xlrd : None
xlsxwriter : None
zstandard : None
tzdata : 2025.2
qtpy : None
pyqt5 : None

[rhshadrach]

In short I am very confused, am I doing something wrong here?

I don't believe so. However I cannot reproduce your issues, as you indicated. Unfortunately there is nothing we can do to debug without a reproducible example and for this reason do not allow issues to remain open without a reproducer. Closing for now, happy to reopen if one is provided.