BUG: sometimes when using ~ and & operators for indexing it evaluated incorrectly

[adamszelestey-se]

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

Edit [rhshadrach]: The code below does not reproduce the issue.

# No idea how to exactly reproduce it, but it occurs sometimes. Logic is this:
import pandas as pd
bool_df = pd.DataFrame([
    {"first": True, "second": False, "third": True},
    {"first": True, "second": True, "third": True},
    {"first": True, "second": False, "third": True},
    {"first": True, "second": True, "third": True},
    {"first": True, "second": True, "third": True},
    {"first": True, "second": False, "third": True},
    {"first": True, "second": True, "third": True},
    {"first": False, "second": False, "third": True},
    {"first": True, "second": True, "third": True},
    {"first": True, "second": False, "third": True},
])
bool_df = bool_df[bool_df["third"]][["first", "second"]]
# In some cases, this line prints the length of the DataFrame (10)
print(len(bool_df[(~bool_df["first"]) & (~bool_df["second"])])) # Sometimes prints 10

# This line prints the expected output (1)
print(len(bool_df[(bool_df["first"] == False) & (bool_df["second"] == False)])) # Prints 1

# Using De Morgan's law also returned with the expected output
print(len(bool_df[~((bool_df["first"]) | (bool_df["second"]))])) # Prints: 1

Issue Description

We don't know when and why this occurs. We werre looking for any rational explanation for hours. Anyone else experienced similar? How could this be possible?
(Environment: MacBook Pro 2023, Sequoia 15.3)

Expected Behavior

print(len(bool_df[(~bool_df["first"]) & (~bool_df["second"])])) # Print 1

Installed Versions

INSTALLED VERSIONS

commit : 0691c5c
python : 3.13.1
python-bits : 64
OS : Darwin
OS-release : 24.3.0
Version : Darwin Kernel Version 24.3.0: Thu Jan 2 20:24:23 PST 2025; root:xnu-11215.81.4~3/RELEASE_ARM64_T6031
machine : arm64
processor : arm
byteorder : little
LC_ALL : None
LANG : None
LOCALE : None.UTF-8

pandas : 2.2.3
numpy : 2.2.3
pytz : 2025.1
dateutil : 2.9.0.post0
pip : 24.3.1
Cython : None
sphinx : None
IPython : 9.0.1
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : 4.12.3
blosc : None
bottleneck : None
dataframe-api-compat : None
fastparquet : None
fsspec : None
html5lib : None
hypothesis : None
gcsfs : None
jinja2 : 3.1.5
lxml.etree : 5.3.1
matplotlib : 3.10.1
numba : None
numexpr : None
odfpy : None
openpyxl : 3.1.5
pandas_gbq : None
psycopg2 : None
pymysql : None
pyarrow : 19.0.1
pyreadstat : None
pytest : 8.3.5
python-calamine : None
pyxlsb : None
s3fs : None
scipy : 1.15.2
sqlalchemy : 2.0.38
tables : None
tabulate : 0.9.0
xarray : None
xlrd : None
xlsxwriter : None
zstandard : 0.23.0
tzdata : 2025.1
qtpy : None
pyqt5 : None

[snitish]

I'm unable to reproduce this behavior both on 2.2.3 as well as dev. @adamszelestey-se do you have any stats on how often this line prints 10?

print(len(bool_df[(~bool_df["first"]) & (~bool_df["second"])])) # Sometimes prints 10

[rhshadrach]

• How are you running this? A Python script, a notebook like Jupyter, IPython? If a notebook/IPython, are you restarting the kernel between runs?
• Can you add print(bool_df.dtypes) on the line prior to the first print. What is the output of this when the unexpected result occurs?

[adamszelestey-se]

I'm unable to reproduce this behavior both on 2.2.3 as well as dev. @adamszelestey-se do you have any stats on how often this line prints 10?

print(len(bool_df[(~bool_df["first"]) & (~bool_df["second"])])) # Sometimes prints 10

Yeah unfortunately it is really unpredicable for us too, but it occured pretty often, I would say 1 out of 2 cases.
The code is not exactly the one I provided, but the logic shouldn't differ.

• How are you running this? A Python script, a notebook like Jupyter, IPython? If a notebook/IPython, are you restarting the kernel between runs?
• Can you add print(bool_df.dtypes) on the line prior to the first print. What is the output of this when the unexpected result occurs?

We were running it as a regular python module in a poetry environment.

I wanted to create the ticket if someone else also experience something like this, they could add more context, I can understand that you can not do anything with this until we figure out how to reproduce it.

[rhshadrach]

The code is not exactly the one I provided, but the logic shouldn't differ.

Am I correct in saying that you can never reproduce the bug with the code example that's in the OP?

[adamszelestey-se]

The code is not exactly the one I provided, but the logic shouldn't differ.

Am I correct in saying that you can never reproduce the bug with the code example that's in the OP?

Unfortunately, you are correct meant to showcase that as the pseudo code if someone comes across the same issue

This is the best I can give you, but can't reproduce the error at the moment:
`

from concurrent.futures import as_completed, ThreadPoolExecutor
import pandas as pd
import time
import random

random.seed(42)

def get_result(_id: str):
    result = {}
    sleep = random.random() * 5
    time.sleep(sleep)
    result["id"] = _id
    return result


def get_single_metric(result, expected):
    sleep = random.random() * 5
    time.sleep(sleep)
    result["internal_success"] = random.random() > 0.5
    result["perfect_extraction"] = random.random() > 0.5
    return result


def create_global_metrc(metric_df):
    print(len(metric_df)) # 50
    print(metric_df[["internal_success", "perfect_extraction"]].dtypes)  # bool, bool
    
    confusion_matrix = {
        "Correct": {
            "TP": len(metric_df[metric_df["internal_success"] & metric_df["perfect_extraction"]]),
            "FN": len(metric_df[~metric_df["internal_success"] & metric_df["perfect_extraction"]]),
        },
        "Incorrect": {
            "FN": len(metric_df[metric_df["internal_success"] & ~metric_df["perfect_extraction"]]),
            "TN": len(
                metric_df[~metric_df["internal_success"] & ~metric_df["perfect_extraction"]]
            ),
        },
    }
    print(confusion_matrix)  # {'Correct': {'TP': 11, 'FN': 10}, 'Incorrect': {'FN': 17, 'TN': 12}}, but it printed {'Correct': {'TP': 11, 'FN': 10}, 'Incorrect': {'FN': 17, 'TN': 50}}, where 50 is incorrect

ids = [f"invoice_{i}" for i in range(1, 51)]
expected = [True] * 50

metrics = []
results = []


with ThreadPoolExecutor(max_workers=25) as extraction_executor, ThreadPoolExecutor(
    max_workers=25
) as metrics_executor: # Create two separate thread pools - one for extractions and one for metrics

    extraction_futures = {
        extraction_executor.submit(get_result, id): id for id in ids     # First submit all extraction tasks
    }


    metrics_futures = []
    for future in as_completed(extraction_futures):    # Then submit metrics tasks as extractions complete
        result = future.result()
        invoice_id = extraction_futures[future]
        results.append(result)

        metrics_futures.append(
            metrics_executor.submit(
                get_single_metric,  # fill metric fields of result
                result,
                expected
            )
        )


    for future in as_completed(metrics_futures):     # Collect all metrics results
        metrics.append(future.result())

metric_df = pd.DataFrame.from_dict(metrics)
results_df = pd.DataFrame.from_dict(results)
global_metric = create_global_metrc(metric_df)

`

[rhshadrach]

Unfortunately, you are correct meant to showcase that as the pseudo code if someone comes across the same issue

Thanks for clarifying. In the future, please make this more clear when you open an issue as maintainers spend time reproducing bugs that are reported.

I would suggest to try to capture the output of print(bool_df.dtypes) on your real data if possible, it may shed some light on what's going on.

I understand that it can be difficult to find a reproducer, but unfortunately we cannot do anything until one is produced. So closing for now. If you are able to come up with a reproducible example, post it here and we'll be glad to reopen!