Programmer Humor

There is a subtle scenario with read-modify-write transactions in MVCC where SQLite lacks some grace (in my opinion).

In MVCC, transactions work with a point-in-time (read "between atomic transactions") consistent "read snapshot" of the database.

Consider this example:

1. Transaction A begins and reads from table foo.
2. Transaction B begins and updates table foo.
3. Both transactions commit.

There is no conflict here because these transactions are isolated from each other via the snapshot mechanism. Transaction A's read snapshot is immutable and will not see any writes from transaction B, even if they are happening concurrently.

Now what happens in this example (from the OP):

1. Transaction A begins and reads from table foo.
2. Transaction B begins and updates table foo.
3. Transaction B commits.
4. Transaction A tries to update foo.

This is a true conflict because both transactions are trying to write to foo, and transaction A's writes might be based on what it just read. There is no consistent way for A to proceed, because B already wrote to foo, invalidating A's read snapshot.

So SQLite handles this by returning an error to A, effectively requiring A to restart the transaction.

There are other ways this could be handled though. The DB could optimistically retry the transaction for you. There is even a special BEGIN IMMEDIATE; statement that it could use to proactively take a write lock on foo so that the transaction doesn't get starved by other writers. But SQLite puts all of the responsibility on users to handle this.

I'm not an expert, so there could be a very good reason that SQLite works this way, but it feels a bit annoying as a user.

I don't actually know off the top of my head how PostgresQL handles this particular scenario.