fix: ensure that the event snapshot is available when one observer defers and the other does not

[tonyandrewmeyer]

If an event has multiple observers, then if any of the (event+observer) pairs do not defer, then we need to save the snapshot so that it's available to process that (event+observer).

#1372 introduced a bug where if any of the (event+observer) pairs was skipped because an exact match was already in the deferral queue, we would skip saving the snapshot. This is ok if there's only a single observer, or if an (event+observer) that did not defer was processed before any that did defer, but not otherwise.

Fixes #1561

[tonyandrewmeyer]

I'm attempting to reproduce the failure (and then the fix) in pgbouncer-operator and postgresql-k8s-operator, although I've had limited success in the past trying to run the integration tests of complex charms outside of the regular repo.

[tonyandrewmeyer]

postgresql-k8s-operator failure in my fork. Trying to get it to build with this branch to see the fix.

pgbouncer-operator's tox -e integration had 13 failures (25 xfailed, 5 errors) locally, but none matching the expected one - I think perhaps my local Juju version isn't the one expected. I'll try that in CI in my fork also.

[benhoyt]

Thanks for the fix and the tests.

[dimaqq]

Let's confirm the observation counts for partially deferred events at today's standup.

[dimaqq]

P.S. confirming that the change fixes my reproducer.

[tonyandrewmeyer]

Reproduced failure in pgbouncer. However, getting the integration tests to run with a branch is still proving difficult, so I'll instead do the release and then do a final verify after that.

[tonyandrewmeyer]

Reproduced failure in pgbouncer. However, getting the integration tests to run with a branch is still proving difficult, so I'll instead do the release and then do a final verify after that.

pgbouncer seems good now as does postgresql-k8s 🎉