Re: Bit by "commands ignored until end of transaction block" again

Craig Ringer wrote:
> On Sun, 2009-07-26 at 19:15 -0400, Science wrote:
>
>> FWIW, the way the Rails ORM ActiveRecord (another fairly damaged ORM)
>> handles this is by allowing you to open any number of transaction
>> blocks, but only the outer transaction block commits (in Pg):
>>
>> Property.transaction { # SQL => 'BEGIN'
>> User.transaction {
>> Foo.transaction {
>> Foo.connection.execute('--some sql code') # SQL => '--some sql code'
>> }
>> }
>> } # SQL => 'COMMIT'
>
> What happens if, Foo.transaction does something that causes an error,
> though, or issues a rollback? It's not creating savepoints, so if
> Foo.transaction rolls back it throws out the work of User.transaction
> and Property.transaction too.
>
> Ugh.
>
> That design would be quite good _IF_ it used savepoints:
>
>
> Property.transaction { # SQL => 'BEGIN'
> User.transaction { # SQL => SAVEPOINT User
> Foo.transaction { # SQL => SAVEPOINT Foo
> Foo.connection.execute('--some sql code') # SQL => '--some sql code'
> } # SQL => RELEASE SAVEPOINT Foo
> } # SQL => RELEASE SAVEPOINT User
> } # SQL => 'COMMIT'
>
> ... so that inner transactions could ROLLBACK TO SAVEPOINT on error ,
> and so that asking for a rollback would give you a ROLLBACK TO SAVEPOINT
> if the transaction is a subtransaction.
>

For all I know that's how it works these days. I haven't looked at the
code underneath this in a couple of years. It should be trivial to
implement in the way you describe based on how the Ruby codebase is set
up -- hopefully all this will help OP with the Django/Python version of
the same problem.

Steve