Thanks for the report, confirmed on main. PRs to fix are welcome!

It's because the dtype changed : int64->float64->int64

To combine with 'b', NaN need to be added to 'a',
so to include NaN, 'int64' was promoted to 'float64' by below flow:

combine_first core/frame.py:8785
combine core/frame.py:8644
align core/generic.py:9447
_align_frame core/generic.py:9524
_reindex_with_indexers core/generic.py:5423
reindex_indexer core/internals/managers.py:832
take_nd core/internals/blocks.py:1020
take_nd core/array_algos/take.py:115
_take_nd_ndarray core/array_algos/take.py:131
_take_preprocess_indexer_and_fill_value core/array_algos/take.py:531
maybe_promote core/dtypes/cast.py:589

IMHO, it's a genaral practice, and converting this 'float64' back to 'int64' seems not natural.

So, I think making the result 'float64' could be a solution. WDYT?

Just for reference> if NaN exists from the first, it could be handled as 'int' :

a = pd.Series([1666880195890293744, 5,pd.NA]).to_frame()
b = pd.Series([6, 7, 8]).to_frame()
a.combine_first(b)

So, I think making the result 'float64' could be a solution. WDYT?

It seems to me we should be able to carry out this operation without passing through floats.

The cause of 'passing through floats' is, it tries to insert NaN , while converting 'a' from len:2 to len:3. In this case, should we insert other values (like 0 ) to keep the int64 type ?

I am merely a user of Pandas, and the underlying code is far over my head, so maybe I should not be commenting here, but I wonder would it be possible to use here whatever solution was used to solve issue #39051 ?

The above patch could solve this issue ( when all the columns are 'int64' ), but could not cover mixed case like below:

a = pd.DataFrame({0:pd.Series([1666880195890293744, 5]),1:pd.Series([1.0,2.0])})
b = pd.Series([6, 7, 8]).to_frame()
a.combine_first(b)

Currently, NDFrame::align simply get just 1 fill_value as a parameter.
IMHO, to solve above case, we need to pass more context as parameters. Is this a right way?