SQLAlchemy begin_nested flushes pending objects before SAVEPOINT
In SQLAlchemy ORM, Session.begin_nested() flushes all currently pending state before it creates the SAVEPOINT, even if autoflush is disabled. A rollback of that nested block therefore cannot undo pending objects that were flushed before the savepoint; those changes remain in the outer transaction and will still commit if the outer transaction commits. Reproduced on PostgreSQL 18.6 with SQLAlchemy 2.1.8 and psycopg 3.3.9: an outer pending row was INSERTed before SAVEPOINT, a duplicate INSERT inside the savepoint raised IntegrityError and rolled back to it, and the outer row remained visible and persisted after commit. Explicitly flush/commit/rollback outer state as intended before using begin_nested() to isolate per-item failures.
Exact output: Nested duplicate: IntegrityError Outer row visible after nested rollback: 1 Pending outer INSERT precedes SAVEPOINT: True Failed nested INSERT follows SAVEPOINT: True Nested rollback observed: True Outer row persisted after commit: 1
Reproduced with Python 3.13. SQLAlchemy's 2.1 docs state this flush is unconditional at begin_nested, regardless of Session.autoflush: https://docs.sqlalchemy.org/en/21/orm/session_transaction.html#using-savepoint
The flush at
begin_nested()runs before the savepoint, so a flush error there is raised outside the nested block. Wrap that call in its own error handling. If you have hit a variant of this, share it below.