Should I use inplace=True in pandas?

The short version

Generally, no. inplace=True does not reliably save memory (pandas often copies internally anyway), it prevents method chaining, and it returns None, which causes a classic bug when someone assigns the result. Reassigning the returned frame is clearer, safer and the direction pandas itself is moving.

The bug it invites

Because inplace=True returns None, assigning the result destroys your data.

import pandas as pd

df = pd.DataFrame({"a": [3, 1, 2]})

df = df.sort_values("a", inplace=True)   # df is now None!

This is an easy mistake, and the error it causes appears later, somewhere else, as a baffling AttributeError: 'NoneType' object has no attribute ....

The two correct forms

Either reassign the returned frame (recommended), or call the method for its effect without assigning.

df = df.sort_values("a")          # clear, chainable, recommended
df.sort_values("a", inplace=True) # works, but returns None, do not assign

Why reassignment is better

It keeps method chaining possible, which is how readable pandas pipelines are written.

result = (
    df
    .dropna(subset=["a"])
    .sort_values("a")
    .reset_index(drop=True)
)

With inplace=True at each step this becomes three separate statements and cannot be expressed as one flow. The chained form reads as a description of the transformation, which is easier to follow and to modify.

The memory argument does not hold

People reach for inplace believing it avoids a copy. In many cases pandas still makes a copy internally and then reassigns it to the same object, so the saving is illusory. Under Copy-on-Write (the default from pandas 3.0), the internal behaviour changes again and inplace is even less meaningful as an optimisation. If memory is genuinely your constraint, reducing dtypes and reading in chunks will help far more than inplace ever will.

The direction of travel

The pandas team has discouraged inplace for years, and some methods have had it deprecated. Writing code that reassigns is future-proof; code that leans on inplace may need changing.

The one honest exception

Setting a value into an existing frame is naturally in place and does not involve inplace=True at all.

df.loc[df["a"] > 1, "flag"] = True

This is the correct way to modify a frame, and it is unrelated to the inplace argument. The advice here is specifically about the inplace=True keyword on methods like sort_values, drop, rename, fillna and reset_index: leave it out, and assign the result.