I would probably invert this whole process. Don’t delete the record from the database first. Mark the database record as deleted_at pending_deletion or something, and then enqueue a job in something like Oban which can try to hit AWS to delete the record. Only after that has succeeded do you delete the record in the DB. The flag lets you filter it out from UI lists immediately, but you want the record around to reason about the file until it’s actually gone.

The job part here is critical. AWS S3 has good uptime, but not perfect up time. Requests to delete files will fail periodically and will need to be retried.

If you like Arc, you might take a look at Waffle (I've forked Arc to Waffle and merged all open PRs). Arc seems to have gone unmaintained and Waffle is a fork with fixes. I prefer the transparency and control of working directly with ex_aws_s3 and orchestrating uploads and deletions via Ecto.Multi, but understand the allure of an opinionated package.

Hm, I see. In this case I recommend you switch approaches. You’re better off putting a “please delete attachmend with ID 123” command on a message queue (RabbitMQ, Kafka, you can also use your DB as well) and then have a background worker that periodically works through those and just have that delete the files from AWS.