pandas

Warnings

• breaking Copy-on-Write (CoW) is now the only mode in pandas 3.0. Chained assignment df['col'][mask] = value silently does nothing — no error, no warning, no modification. This is the most common invisible bug when upgrading.
  Fix: Use df.loc[mask, 'col'] = value for conditional assignment. Use df = df.assign(col=...) for derived columns. Remove all defensive .copy() calls added to silence old SettingWithCopyWarning.
• breaking String columns now default to str dtype instead of numpy object dtype. Code checking dtype == object or dtype == 'O' to detect string columns will fail silently in pandas 3.0.
  Fix: Replace dtype == object checks with pd.api.types.is_string_dtype(col) or dtype == 'str'. For library code: handle both 'object' and 'str' dtypes during transition.
• breaking Python 3.10 and below dropped. pandas 3.0 requires Python >=3.11.
  Fix: Pin pandas<3.0 for Python <=3.10 environments. Upgrade Python to 3.11+ to use pandas 3.0.
• breaking Datetime default resolution changed from nanoseconds to microseconds (or input resolution). pd.Timestamp arithmetic and comparisons with nanosecond precision may produce different results.
  Fix: Explicitly pass unit='ns' where nanosecond precision is required: pd.to_datetime(arr, unit='ns').
• breaking DataFrame.groupby() observed parameter default changed to True for Categorical columns. Previously unobserved categories were included by default, causing silent behavior changes on groupby aggregations.
  Fix: Pass observed=False explicitly to restore old behavior if unobserved categories are needed.
• deprecated mode.copy_on_write option deprecated — setting it has no effect in pandas 3.0 and will be removed in 4.0.
  Fix: Remove pd.options.mode.copy_on_write = True/False from your code — CoW is always on.
• gotcha pyarrow is not required but strongly recommended for pandas 3.0. Without pyarrow, the new str dtype falls back to numpy object-backed storage, losing most performance benefits.
  Fix: pip install pyarrow alongside pandas. Verified by: pd.Series(['a']).dtype shows 'str' regardless, but performance differs significantly.
• gotcha Many third-party libraries (scikit-learn, seaborn, statsmodels, SHAP) had pandas 3.0 compatibility issues at release. Check library versions when upgrading.
  Fix: Test your full dependency stack against pandas 3.0 before upgrading. Use pandas 2.3.x as a stepping stone to surface deprecation warnings first.

Install

• pip install pandas Core
• pip install pandas[performance] With performance extras (bottleneck, numexpr, numba)
• pip install pandas pyarrow Recommended: with PyArrow for faster string dtype
• pip install pandas[all] All optional dependencies

Imports

• pandas

  import pandas as pd
  
  df = pd.DataFrame({'a': [1, 2], 'b': ['x', 'y']})
  # Modify using loc in one step
  df.loc[df['a'] > 1, 'b'] = 'z'

  Copy-on-Write is now mandatory in 3.0. Chained assignment (df['col'][mask] = val) no longer modifies the DataFrame — it silently does nothing.
• string dtype

  ser = pd.Series(['a', 'b'])
  ser.dtype  # dtype('str') in pandas 3.0
  
  # Check for string dtype in 3.0-compatible way:
  if pd.api.types.is_string_dtype(ser):
      ...

  String columns now infer to str dtype instead of numpy object. Code checking dtype == object or dtype == 'O' for string detection will silently miss string columns.

Quickstart

pandas 3.0 patterns. Use loc for in-place modification. Strings are str dtype not object.

import pandas as pd
import numpy as np

# Create DataFrame
df = pd.DataFrame({
    'name': ['Alice', 'Bob', 'Charlie'],
    'score': [85, 92, 78],
    'dept': ['eng', 'eng', 'mkt']
})

# Correct modification in pandas 3.0 (CoW)
df.loc[df['score'] > 80, 'grade'] = 'pass'

# Or use assign() for derived columns (returns new DataFrame)
df = df.assign(grade=lambda x: np.where(x['score'] > 80, 'pass', 'fail'))

# Check dtypes — strings are now 'str', not 'object'
print(df.dtypes)
# name     str
# score    int64
# dept     str