database adapter serialisation tests disabled

[James Henstridge]

The database adapter serialisation tests started failing again when mthaddon was trying to merge the new Bazaar release.

I have moved them out of the test_adapter.txt page test into a separate adapter-serialisation.txt doctest, and disabled that test (so the remainder of test_adapter.txt is still in force). We should re-enable this test when we work out a real fix for the problem.

The bug is marked critical at the request of SteveA.

[James Henstridge (jamesh)]

To re-enable the test, uncomment the appropriate lines in lib/canonical/launchpad/webapp/ftests/test_adapter.py.

This appears to be the same errors we were seeing previously in bug 32231.

[Stuart Bishop (stub)]

We need the output of the failure as we probably won't be able to reproduce this locally.

[James Henstridge (jamesh)]

The output of Tom's test suite run is:
    https://devpad.canonical.com/~mthaddon/2007-07-31-bzr-upgrade-failure.txt

It appears to be the same exception as in bug 32231 (as mentioned in my previous comment).

[James Henstridge (jamesh)]

Steve asked for an explanation of the problem, so here goes:

the commit() and rollback() implementations in psycopg1 have a weird error handling: if the statement can't be executed or the result status is not PGRES_COMMAND_OK, then it calls pgconn_set_critical(). This function stores the error message in the cursor.

Most psycopg cursor methods call the EXC_IFCRITICAL() macro which calls pgconn_resolve_critical() if an error message has previously been set with pgconn_set_critical(). The pgconn_resolve_critical() closes the connection and raises a Python exception.

So a failed commit() or abort() can end up closing the connection. I believe this is the root cause of the problem we've seen, and I guess it is down to timing on which statement in the test fails. If it is the commit, then the connection is screwed.

There is similar code in psycopg2, and I submitted a bug report with a test case that exercises the bug.

[Joey Stanford (joey)]

Per conversation with Steve, demoting this to high.

[James Henstridge (jamesh)]

As part of the work on improving psycopg2 for use with Storm and Launchpad, I've greatly reduced the use of the set_critical()/resolve_critical() code. So this test should not hit the bug when we move over to that version.

[James Henstridge (jamesh)]

As these tests have been replicated in the psycopg2 test suite, we can probably remove them completely from the LP test suite when switching over to Storm.