executemany (one transaction) is ~1.25× faster than execute() per row

Batching the calls on the Python side is ~25% faster even with a single commit in both cases — a one-line change.

Result

Run	executemany (one transaction)	execute() per row	Ratio
Run 1	15.737	19.655	0.80×
Run 2	14.310	17.495	0.82×
Run 3	14.122	18.150	0.78×

Spread over 3 runs: executemany (one transaction) CV 5.0% / execute() per row CV 5.0% / Ratio CV 2.0%

con.executemany('INSERT INTO items VALUES (?,?,?,?,?)', rows); con.commit()

for r in rows: con.execute('INSERT INTO items VALUES (?,?,?,?,?)', r)
con.commit()

5,000 rows inserted into an 8,000-row table, no indexes, `synchronous=NORMAL`.

With very large rows, or when you must commit mid-way, the gap narrows.

cd bench && python sqbench.py --samples 9 --out ../data/run.json