These changes are all independent of each other; I just didn’t feel
like making dozens of commits for them.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Calling `email.save()` is only needed if we altered `email.address`;
it is unnecessary if we called `email.users.add(...)` which will have
done its own INSERT.
This fixes two bugs: the most obvious is that there is a race where a
ScheduledEmail object could be observed in the window between creation
and when users are added; this is a momentary instance when the object
has no users, but one that will resolve itself.
The more subtle is that .save() will, if no records were found to be
updated, _re-create_ the object as it exists in memory, using an
INSERT[1]. Thus, there is a race with `deliver_scheduled_emails`
between when the users are added, and when `email.save()` runs:
1. Web request creates ScheduledEmail object
2. Web request creates ScheduledEmailUsers object
3. deliver_scheduled_emails locks the former, preventing updates.
4. deliver_scheduled_emails deletes both objects, commits, releasing lock
5. Web request calls `email.save()`; UPDATE finds no rows, so it
re-creates the ScheduledEmail object.
6. Future deliver_scheduled_emails runs find a ScheduledEmail with no
attending ScheduledEmailUsers objects
Wrapping the logical creation of both of these in a single transaction
avoids both of these races.
[1] https://docs.djangoproject.com/en/3.2/ref/models/instances/#how-django-knows-to-update-vs-insert
Only clear_scheduled_emails previously took a lock on the users before
removing them; make deliver_scheduled_emails do so as well, by using
prefetch_related to ensure that the table appears in the SELECT. This
is not necessary for correctness, since all accesses of
ScheduledEmailUser first access the ScheduledEmail and lock it; it is
merely for consistency.
Since SELECT ... FOR UPDATE takes an UPDATE lock on all tables
mentioned in the SELECT, merely doing the prefetch is sufficient to
lock both tables; no `on=(...)` is needed to `select_for_update`.
This also does not address the pre-existing potential deadlock from
these two use cases, where both try to lock the same ScheduledEmail
rows in opposite orders.
No codepath except tests passes in more than one user_profile -- and
doing so is what makes the deduplication necessary.
Simplify the API by making it only take one user_profile id.
This fixes a bug where email notifications were sent for wildcard
mentions even if the `enable_offline_email_notifications` setting was
turned off.
This was because the `notification_data` class incorrectly considered
`wildcard_mentions_notify` as an indeoendent setting, instead of a wrapper
around `enable_offline_email_notifications` and `enable_offline_push_notifications`.
Also add a test for this case.
Previously, the output would make it look like we sent an actual email
to the first user in the dry_run output, which is very confusing.
The `dry_run` code path already prints all the accounts that would
have been emailed at the end, so there's no reason to have this line
before the dry_run check.
Additionally, we move after the `get_connection` check because
failures at that stage shouldn't result in logging an attempt to send
an email.
This way we can stop reading as soon as we get to the body. Also,
send an Accept header, check that the request was actually successful,
use lxml.etree.iterparse instead of a broken hand-rolled state
machine, and support XHTML, all for negative 28 lines of code.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This reverts commit 1965584eec.
This syntax has a bad interaction with table syntax and needs to be
rethought.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This is more efficient than get_lexer_by_name, since we don’t need to
instantiate the class just to get its name.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
The BlockingChannel annotations in TornadoQueueClient were flat-out
wrong. BlockingChannel and Channel have no common base classes.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
