change in oslo.db "ping" handling is causing issues in projects that are not using transactions

[Mike Bayer]

in https://review.openstack.org/#/c/106491/, the "ping" listener which emits "SELECT 1" at connection start was moved from being a connection pool "checkout" handler to a "transaction on begin" handler. Apparently Keystone and possibly others are using the Session in "autocommit" mode, despite that this is explicitly warned against in SQLAlchemy's docs (see http://docs.sqlalchemy.org/en/rel_0_9/orm/session.html#autocommit-mode), and for these projects they are seeing failed connections not transparently recovered (see https://bugs.launchpad.net/keystone/+bug/1361378).

Alternatives include:

1. move the ping listener back to being a "checkout" handler

2. fix downstream projects to not use the session in autocommit mode

In all likelihood, the fix here should involve both. I have a longer term plan to fix EngineFacade once and for all so that the correct use patterns are explicit, but that still has to be blueprinted.

[Mike Bayer (zzzeek)]

review at https://review.openstack.org/#/c/124466/

[Morgan Fainberg (mdrnstm)]

The cleanup to not use session with autocommit will be a fix we apply to Keystone in Kilo (it is something we should do regardless). I talked with Doug Hellmann and Mike Bayer and they agreed the more immidiate fix should go into oslo.db for the Juno timeframe.

As such I've targeted this to K1 (might slip to a later K-milestone as we go) but should be resolved in Kilo release cycle.

[Doug Hellmann (doug-hellmann)]

Backport for juno: https://review.openstack.org/#/c/125079/

[Morgan Fainberg (mdrnstm)]

Moved from K1 as we'll pick this up when EngineFacade is refactored. The fix to olso.db should be sufficient for handling our current use-cases.

[Brant Knudson (blk-u)]

There's a fix in oslo.db. The work to update Keystone will be part of a spec or blueprint to use new features in oslo.db once they're ready. I don't think it's worth keeping a bug open.