Hi,
the connection.autocommit feature has created the problem shown here:
https://code.djangoproject.com/ticket/16250
I've taken a look and they have a set_autocommit method
<https://code.djangoproject.com/browser/django/trunk/django/db/backends/creation.py#L347>
implemented in a painful way: instead of being driver-specific they
just invoke some random method on the connection, and the thing is
compounds with the fact they *don't even know* there is a transaction
somehow already open. As per discussion
<http://archives.postgresql.org/psycopg/2011-05/msg00033.php>,
psycopg's set_session gives an error if invoked in a transaction.
Now, I would love to argue that Django's set_autocommit is written
with the wrong anatomical part, the bug is theirs and it's all their
problem. However, knowing the painful process they use to fix a bug I
wouldn't be surprised it would take months (see how they handled the
idle in transaction mess) and this would only be a problem for django,
postgres and psycopg users, as 2.4.2 is the version installed by
default by easy_install and friends. So I'm positive to change the
semantics of set_session/autocommit and issue an implicit rollback if
in transaction instead of raising an exception, as set_isolation_level
does.
Thoughts?
-- Daniele
