Using the clever concurrency testing borrowed from SubscriptionPlacementJob, but I thought a shorter pause time (just 100ms) would be sufficient.
I considered doing this with a new 'state' field (upcoming/open/close), but decided to keep it simple.
Although we won't be allowing multiple in the this PR, we certainly plan to in the future.
The migration helper add_reference couldn't handle the custom column name, so I had to put it together manually.
This will store the URL for each user that wants a notification.
We probably don't need URL validation (it's not done on Enterprise for example). It could be validated by browser input, and anyway will be validated if the webhook actually works or not.
Inspired by Keygen: https://keygen.sh/blog/how-to-build-a-webhook-system-in-rails-using-sidekiq/
Best reviewed with whitespace hidden.
Unfortunately the spec isn't allowed in CI. But it worked on my environment, I promise.
I chose `xit` so that it doesn't run unnecessarily. Perhaps we could use `pending` instead, which would execute, and notify us if it suddenly started working one day. But I doubt it.
This job is responsible for delivering a payload for one webhook event only. It allows the action to run asynchronously (and not slow down the calling process).
The route to update a return authorization must include the order number
as id:
/admin/orders/:order_id/return_authorizations/:id(.:format)
The spec only worked because the controller's ivars were not reset
between requests and the order was already set. But Rails 7 will reset
the ivars and it failed finding the order without a given id.
When calling `save!` without changing any attributes then Rails doesn't
always touch other records because nothing changed. So I changed the
spec to `touch` explicitely and it turns out that everything passes.
Tada, our code seems correct and it was only the spec which seemed
broken in Rails 7.
I didn't observe it but if the spec code would run within the same
millisecond then we wouldn't be able to observe a change to
`updated_at`. Time travel solves this potential problem.
I found this because Rails 7 converts timestamps to database precision
straight away. While we may have some broken logic in the code, most of
these cases may just be broken spec code. Watch this space.
Storing a timestamp to the database has less accuracy than a Ruby Time
object. So `updated_at` changes after being written and loaded from the
database. Rails 7 accounts for that by rounding it in the model already
before it's written to the database. That made one spec fail.
Before it said:
```
WARNING: Using the `raise_error` matcher without providing a specific error or message risks false positives, since `raise_error` will match when Ruby raises a `NoMethodError`, `NameError` or `ArgumentError`, potentially allowing the expectation to pass without even executing the method you are intending to call. Actual error raised was #<RuntimeError: Cannot use t(".legacy_translation") shortcut because path is not available>. Instead consider providing a specific error class or message. This message can be suppressed by setting: `RSpec::Expectations.configuration.on_potential_false_positives = :nothing`.
```
Instead of knowing the input id generated by select2, we now rely on the
fact that the owner is in the fifth column. Both could change but this
is less code.
