Copy-on-Write Enabled by Default

Pandas 3.0 enables Copy-on-Write (CoW) by default. This mechanism fundamentally changes copy semantics: a DataFrame derived from another shares the underlying memory until a modification is made, at which point a real copy is created.

Before / After

The old behaviour could cause unexpected modifications through shared views. With CoW, every modification is isolated.

python 
import pandas as pd

df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})

# Create a subset
subset = df[['a']]

# BEFORE Pandas 3.0 (without CoW):
# subset['a'] = 99  # ALSO modifies df! (SettingWithCopyWarning)

# WITH Pandas 3.0 (CoW enabled):
subset['a'] = 99
print(subset)
#    a
# 0  99
# 1  99
# 2  99

print(df)
#    a  b
# 0  1  4  <- df is NOT modified
# 1  2  5
# 2  3  6

Performance Impact and Migration

CoW reduces unnecessary memory copies: as long as you don't modify, memory is shared. Operations like df.rename(), df.set_index(), or df.reset_index() no longer copy data. To migrate, remove defensive .copy() calls that are no longer needed.

python 
import pandas as pd

df = pd.DataFrame({'x': range(1_000_000), 'y': range(1_000_000)})

# These operations DO NOT copy data (CoW)
df2 = df.rename(columns={'x': 'col_x'})
df3 = df.reset_index()
df4 = df[['x']]  # no copy as long as we don't modify

# Copy only happens WHEN you modify
df4['x'] = 0  # here, a copy is triggered for df4

# Migration: remove defensive .copy() calls
# Before: subset = df[['a']].copy()  # necessary
# After:  subset = df[['a']]          # sufficient with CoW

Sources