Why this change?
We have been getting flaky test failures from these specs and the
failure screenshot shows that the user is not logged in when it is
supposed to be. Futher investigation shows that when the test flakes, it
is because the request to view the topic is using an auth token that is
different from the one which was created when the user was signed in.
What does this change do?
1. Add the `capture_log` metadata to all the tests in this file.
2. Enables the `verbose_auth_token_logging` site setting to give us more
debugging information in the logs.
Why this change?
The spec is flaky and it seems to be DB transaction related where after
signing in as a user, the failure screenshot shows that the user has not
been signed in.
What does this change do?
Set `capture_log: true` which will log all ActiveRecord DB statements
execute while running the spec.
Currently, when only one user is available for the random auto-assign
and if that user was already assigned, then the assign will fail.
This patch addresses this issue by allowing to reassign a user who’s
already assigned. A new parameter (`allow_self_reassign`) has been added
to `Assigner#assign` with a default value of `false`.
Sometimes, the topic id actually exists for that spec (id being 1) and
it will raise about not finding the group id instead of not finding the
provided topic because the assignees group wasn’t provided in the
`fields` parameter.
Currently, when the auto-assign logic can’t find a user to assign, it
will fail saying there was no one to assign. The current logic is this
one:
- Don’t pick anyone who’s been picked in the last 180 days
- If no one has been found, then try the same thing but only for the
last 14 days.
While this is working relatively well for large enough groups, it
doesn’t work at all with very small groups (like 2 people) and it
creates unnecessary noise.
This patch addresses this issue by adding a fallback to the current
logic. Now, if the two first rules fail, instead of saying that no one
was assigned, we assign the least recently assigned person. This way,
the logic will continue to work with large groups but will also work
nicely with small groups.
The recent changes to the user menu changed how we were ordering the
items in it. A small bug was noticed though, as we’re not displaying the
unread assignments at the very top.
This patch fixes the ordering like this:
- Unread individual assignment notifications
- Unread group assignment notifications
- Read individual assignments
- Read group assignments
In each group of items, they are sorted by date, the most recent ones
being at the top.
Currently, we display a mix of topics and notifications in the user menu
assignments tab. This has a number of issues like having to maintain
hard to understand code instead of simply relying on the notifications
system we have, we can’t display more than one assignment per topic and
it’s not clear if an assignment is for a topic or a post nor if it’s a
group assignment or an individual one.
This patch addresses those issues by relying on the notifications system
we’re using for most of the other user menu tabs instead of a custom
implementation. This led to some heavy refactoring but it was
worthwhile as things are a bit more normalized and easier to reason
about. Instead of serializing topics with different attributes to access
various assignments, we now simply return a notification for each
existing assignment.
The UI changed a bit: tooltips are now explaining if the assignment is
for a topic or a post and if it’s for an individual or a group. Icons
are also properly set (so no more individual icon for group assignments)
and the assigned group name is displayed.
The background jobs signatures changed a bit (`assignment_id` is now
needed for the unassign job) so when deploying this patch, it’s expected
to have some jobs failing. That’s why a post-migration has been written
to handle the creation of missing notifications and the deletion of
extra notifications too.
Why this change?
We have been seeing tests failures due to errors like
```
Capybara::ElementNotFound:
Unable to find css "#topic-footer-button-assign"
```
when running our system tests in CI. However, I can't quite figure out
why that is happening and have invested way too much time to do so.
Therefore, I'm trying to fix this by closing a topic via the topic
timeline controls instead of the footer button instead.
This was causing the tests to be flaky with the following errors on CI:
```
PG::UniqueViolation: ERROR: duplicate key value violates unique constraint "index_assignments_on_target_id_and_target_type"
DETAIL: Key (target_id, target_type)=(39, Topic) already exists.
```
Otherwise, when an allowed group member is assigned to the PM it invites the user (which is unnecessary) and creates a public small post for the invite.
The goal is to add a class so when assigns are not public, their descriptions can be styled similar to whispers... this is a light way to reassure admins of assign message visibility.
When assignment data is preloaded, cache variable should be set when topic is assigned but also when topic is not assigned to avoid additional queries.
The previous behavior was:
1. Assign a topic to someone
2. Close the topic with the setting unassign_on_close enabled
3. Assign the topic to the same user as before
This caused small broken or empty posts to be displayed.
This commit also introduces system specs
This spec has been detected to be flaky by
https://github.com/discourse/flaky-tests-detective which we run
internally. However, I could not reproduce the flakiness after multiple
attempts so I rewrote the assertions such that we set up the
expectations of what `Assignment#assigned_to_id` should be before and
after the triggering of the DiscourseEvent.
If a topic is initially assigned to user A then reassigned to user B,
this change will make it so that notifications sent to user A are
removed. This matches the plugin's behaviour when unassigning a topic.
There is almost no reason for us to use `Fabricate.build` which does not persist the record to the database and is not what we encounter in production.
See https://github.com/discourse/discourse/pull/18437,
we are removing any references to enable_personal_messages
in core and using only personal_message_enabled_groups,
which requires auto groups to be assigned in certain specs
for them to keep working.
The basic group serializer complains that the flair_upload relation has to be eager loaded when attempting to included assigned groups in search results, which is a consequence of using strict loading during search's preload to
avoid N+1.
Eager loading the association during this would definitely fix the problem, but is it the right solution? It seems to me that this serializer does too much, when we only need a small number of attributes to display the assigned
widget. Instead, this commit adds a lightweight custom serializer which grabs only the essentials for it to work.
Only 5 assignments can exist simultaneously for a topic or its post.
The code that counted the assignments did not ignore the inactive
assignments which made the limit check more strict than it should have
been.
There were two problems with the way current automation script for automatic
assignment works:
- it tried to assign users that were not allowed to be assigned because they
were not part of groups that are allowed to use discourse-assign - fixed by
skipping users that are not a part of assign_allowed_on_groups groups
- it assigned new users - fixed by adding a new user that can skip new users
Alan Guo Xiang Tan
Loïc Guitaut
David Taylor
Natalie Tay
Jarek Radosz
Krzysztof Kotlarek
Joffrey JAFFEUX
Martin Brennan
Vinoth Kannan
Ted Johansson
Arpit Jalan
Bianca Nenciu
Kris
Jan Cernik
Gerhard Schlager
Andrei Prigorshnev
Penar Musaraj
Roman Rizzi
Keegan George