D1 batch() wraps all statements in one transaction that rolls back entirely on failure

db.batch([stmt1, stmt2]) sends the statements to D1 as a single unit inside an implicit transaction. If any statement fails, none of them commit and the returned promise rejects. Statements run in order and the results array matches the input order, so meta.last_row_id can be read from the specific insert that produced it.

Statements sent individually autocommit: two separate run() calls are two transactions and can interleave with other writers. That is why multi-table writes which must stay consistent, such as inserting a row and updating a counter or a search index, belong in one batch() call. Prepared statements are reusable, so build the statement once outside the loop and call bind() per row rather than preparing inside the loop, which also keeps the statement cache effective.