Re: [SQLObject] Python 2.7.12 + tox + PostgreSQL

[Neil M. <drn...@gm...>]

On 23 September 2016 at 00:03, Oleg Broytman <ph...@ph...> wrote:
> Hi, all!
>
> On Thu, Aug 25, 2016 at 11:18:45PM +0200, Oleg Broytman <ph...@ph...> wrote:
>>    Recently tests started to fail both at Travis and Circle, but only
>> with Postgres. Initially I couldn't reproduce the problem locally but
>> after a dozen of experiments I got it: the problem manifests itself
>> only with Python 2.7.12 + tox + PostgreSQL. I have to install Python
>> 2.7.12 and tox to reproduce it locally (I have Debian with Python 2.7.9
>> and I usually don't use tox - I run py.test directly).
>>
>>    Somehow I managed to fix tests by changing test order - I explicitly
>> listed tests directories with 'tests' at the top of the other dirs.
>
>    After a few hundreds successful and failed test runs I narrowed the
> problem a bit and reproduced it with any python version and without tox.
> So the problem is clearly in SQLObject tests.
>    The problem is manifested if tests from sqlobject/tests are run after
> sqlobject/inheritance/tests:
>
> createdb test
> py.test sqlobject/inheritance/tests tests/test_schema.py
> dropdb test
>

I spent some time today poking at this. While I haven't isolated the
cause, I have some more information on what's happening.

When the test fails, the connection pool has two entries. Because of
how getConnection & releaseConnection are implemented, the connection
returned by getConnection will cycle between these two entries. In
this case, the call to CREATE TABLE and the query end up happening in
different connections, which have different values for search_path
(since that's local to the underlying connection), which causes the
failure.

I haven't worked out where the extra connection in the pool gets
created or why this only happens some of the time.

-- 
Neil Muller
drn...@gm...

I've got a gmail account. Why haven't I become cool?