Sequelize Nested Transactions: Unclear Rollback Behavior in Deployment Logs

Answer: ROLLBACK TO SAVEPOINT does not abort the outer transaction

A logged “ROLLBACK TO SAVEPOINT” only undoes changes made after that savepoint; the parent transaction remains active and can still be committed unless it is later rolled back explicitly.

Confirmed facts

• Sequelize implements nested transactions as savepoints for MySQL and PostgreSQL; SQLite emulates savepoints with limited durability.
• The outer transaction inherits the isolation level of the parent and stays open after a savepoint rollback.
• A child transaction that commits before its parent is not durable if the parent later rolls back.

Likely explanation for the log ambiguity

Deployment pipelines see the savepoint rollback entry and assume the whole transaction failed, but the log does not contain a subsequent full ROLLBACK statement that would indicate the outer transaction was aborted.

Steps to detect whether the outer transaction was aborted

1. Enable Sequelize logging (or your DB’s general log) and look for a ROLLBACK statement that is not prefixed with TO SAVEPOINT after the savepoint rollback.
2. In code, inspect the transaction object after the nested block: transaction.finished will be true only if the outer transaction was committed or rolled back.
3. Alternatively, listen for the afterRollback hook on the Sequelize instance; it fires only for a full transaction rollback.

Best‑practice configuration to avoid orphaned changes

• Always wrap the outer transaction in a try/catch and call await transaction.rollback() in the catch block; do not rely on savepoint rollbacks to abort the parent.
• Set transaction.autocommit = false (default) and avoid manually committing a child transaction before the parent unless you explicitly handle parent‑level rollback logic.
• For SQLite, consider disabling nested transactions or using a single transaction per operation, because emulated savepoints may not survive process crashes.

Missing diagnostic detail

Which SQL dialect are you using (MySQL, PostgreSQL, or SQLite)? The answer changes only for SQLite, where savepoint emulation may lose state on process exit and may require a different verification approach.