The timestamps available in the SES receipt don't always correspond
to the time the notification has been sent. We've seen callbacks with
a current timestamp in both 'mail' and 'bounce' objects that referenced
a notification sent a week ago, which means we can't rely on it to skip
archived notifications.
One possible approach would be to look up the notification reference in
the notification_history table, but this goes against our plans to stop
relying on it in the future.
This changes the SES receipts logic to retry missing notifications once
(if the callback timestamp is within the last 5 minutes the task will
retry after a 5 minute delay) to capture callbacks arriving before the
notification reference has been persisted to the DB. Otherwise, we log
the missing notification as a warning instead of error.
Celery `self.retry` raises an exception to communicate that the task
needs to be retried. Since our ses task is wrapped in a catch-all
except block it logs that exception as an error before retrying.
Handling Retry class separately allows us to raise it without logging
the traceback.
We've seen errors caused by what we suspect is a race condition when
SES callback processing tries to look up the notification before the
sender worker has saved notification reference from the SES POST
response to the database.
This adds a retry for SES callback task if the notification was not
found and the message is less than 10 minutes old and removes the
error log message for notifications older than 3 days (since they
might no longer exist in the notifications table and would've been
marked as failure by then either way).
In order to be able to call retry and silence the error log based on
notification time this change inlines `process_ses_response` and
`update_notification_by_reference` functions into the celery task.
It also removes a lot of defensive error-handling that doesn't appear
to have been triggered in the last few months (for things like missing
keys in SES callback data).
