Code that edited a DataFrame through chained indexing stopped working after upgrading to pandas 3. Why?

Short answer

pandas 3.0 made Copy-on-Write the only behavior, so any subset behaves as its own object and writing to it never changes the parent. Chained assignment such as df["dose"][mask] = 0 now does nothing to df, so the edit must be one .loc call.

Why

Before pandas 3.0, whether df["dose"] returned a view or a copy depended on internal details, which is why SettingWithCopyWarning existed. Sometimes a chained edit changed the original and sometimes it did not. pandas 3.0, released in January 2026, made Copy-on-Write mandatory. Every indexing result now acts as a separate object, and data is copied lazily only when one side is written to.

The migration guide says chained assignment now consistently never works and raises a ChainedAssignmentError warning, and the release notes say SettingWithCopyWarning was removed. A cleaning script that zeroed negative doses with a chained edit therefore keeps the negative values after the upgrade, and every downstream summary changes. The fix is a single call, df.loc[df["dose"] < 0, "dose"] = 0, which states the rows and the column at once.

Defensive .copy() calls added only to silence the old warning can be removed. Two other pandas 3.0 changes catch scientific code. Text columns now default to the str dtype, and parsed datetimes default to microsecond resolution rather than nanoseconds.

Pin the pandas version and run the full pipeline on a reference dataset before and after any major upgrade, because a warning is easy to miss in a long log.

A strong answer covers

Copy-on-Write is mandatory from pandas 3.0
Chained assignment never updates the parent
Use one .loc call with row mask and column
SettingWithCopyWarning was removed
Compare outputs on a reference dataset across the upgrade

Common mistakes

Adding .copy() everywhere in pandas 3 code, which no longer serves any purpose for this problem
Assuming that a script which runs without an exception after an upgrade still produces the same numbers

The interviewer might ask next

What other pandas 3.0 default changes would you check in a pipeline that parses instrument timestamps?