pandas 'SettingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame' (or ChainedAssignmentError in pandas 3): what is wrong and how do I fix it properly?

Cause: you assigned through chained indexing, for example df[df["a"] > 0]["b"] = 1 or df["b"][mask] = 1. That is two separate indexing steps; the first may return a copy or a view of the data, so the write can land on a temporary object and the original DataFrame silently stays unchanged (or sometimes does change). Older pandas tried to detect this and emitted SettingWithCopyWarning. Copy-on-Write (CoW) removes the ambiguity: any derived object behaves as a copy, so chained assignment never updates the original, and pandas warns with ChainedAssignmentError. CoW is the default and only mode in pandas 3.0. Fix (best first): 1. Do the whole assignment in one .loc call with row and column indexers: df.loc[df["a"] > 0, "b"] = 1. For a positional row and a named column use df.iloc[0, df.columns.get_loc("b")] = 1. 2. If you really want an independent subset to modify, make that explicit: sub = df[df["a"] > 0].copy(); sub["b"] = 1. Then merge back deliberately if needed. 3. Replace inplace methods on a selected column (df["b"].fillna(0, inplace=True), df["b"].replace(..., inplace=True)) with assignment: df["b"] = df["b"].fillna(0). These also stop working under CoW. 4. On pandas 2.x, find code that will break on 3.0 by setting pd.options.mode.copy_on_write = "warn" in a test run; set it to True to adopt the new behavior early. 5. Do not silence the warning with pd.options.mode.chained_assignment = None; that hides a real bug where updates are lost. Verify: the warning no longer appears when running with warnings as errors (python -W error::Warning or pytest -W error), and the target cells actually have the new values in the original DataFrame. Sources: https://pandas.pydata.org/docs/user_guide/migration.html https://pandas.pydata.org/docs/user_guide/indexing.html