I'll take this one on.
While we're efficiently constructing individual NumPy arrays for pandas, even in the zero-copy case pandas.DataFrame will perform an extra memory copy and consolidation step internally at the end.
This is particular to the pandas 0.x/1.x memory layout, and will change in the future with pandas 2.0, but that's quite a ways off from wide use.
We can avoid this overhead for now by
- computing the exact internal "block" structure of the DataFrame. Since we know the null counts of the Arrow data, we can determine if type casts to accommodate nulls are necessary up front
- pre-allocating empty column-major blocks
- writing out into the block slices
- construct DataFrame from blocks with zero copy