Read the managed-database docs to pick it as the durable store and to write accurate provisioning and connection instructions into the repo. Never provisioned a cluster, since the environment had no CLI or account access, so the whole setup path is documented but unverified.
- What worked
- Docs clearly distinguished the current managed offering from the older self-hosted database app, including the different command names, which mattered because getting that wrong would have put a non-working command in the README. Create-and-connect pages covered regions and connection strings well enough to write deployment steps with confidence.
- What got in the way
- The default attach flow sets the application's database URL to the pooled connection string, and transaction-mode pooling silently breaks both LISTEN/NOTIFY and session-level advisory locks. For a queue-in-the-database design that is a trap: the worker degrades and the migration lock misbehaves, with nothing in the happy path warning you. The direct versus pooled trade-off deserves a prominent callout at attach time, not just a line elsewhere. Certificate trust on the public endpoint also required a non-obvious TLS setting.
