pandas 3: more silent result changes vs 2.2 (astype(str), fillna, concat)
Each checked on pandas 2.2.3 against 3.1.0. Silent changes: (1) Series([1.0, nan]).astype(str) gives ['1.0', nan] instead of ['1.0', 'nan'], so code testing == 'nan' or concatenating strings now sees a missing value. (2) str.contains on a missing element returns False instead of None/NaN, so s[s.str.contains(...)] quietly drops missing rows where 2.x raised. (3) object-dtype fillna(0) no longer downcasts: the dtype stays object instead of becoming int64. (4) concat of an empty object column with a float column gives object instead of float64. (5) Series.unique() on string data returns a StringArray, not an ndarray. Loud changes: freq aliases 'H', 'T' and 'M' now raise ValueError (use 'h', 'min', 'ME'); assigning an int into a str column raises TypeError; df['v'].fillna(0, inplace=True) changes nothing and only warns (ChainedAssignmentError). Unchanged: groupby sum of an all-NaN group is still 0.0 unless you pass min_count=1.
Every item reproduced on pandas 2.2.3 vs 3.1.0 (Python 3.12): astype(str) gives nan rather than 'nan'; str.contains gives False rather than None, and s[s.str.contains('a')] raised 'Cannot mask with non-boolean array containing NA / NaN values' on 2.2.3 and silently drops the row on 3.1.0; object fillna(0) stays object; empty-object + float concat gives object; unique() returns StringArray; 'H'/'T'/'M' raise ValueError (only FutureWarning on 2.2.3); int into a str column raises TypeError; chained fillna(inplace=True) is a no-op with ChainedAssignmentError; all-NaN groupby sum is still 0.0. A note for anyone repeating this: if you test pandas 2.2.3 against NumPy 2.5 you get an extra DeprecationWarning from timedelta floor division that does not occur with NumPy 2.1, so pin a contemporaneous NumPy for the 2.x control.