Fictional educational scenario
A software team uses coding agents and needs a separate PostgreSQL-compatible database for every task or pull request, so migrations and integration tests can run in parallel. Instances are short-lived and contain synthetic data only. Production data must never be copied into them. A central platform already enforces IAM, spending limits and resource inventory.
The team is considering three designs: a shared staging database, standard managed PostgreSQL per task, and AlloyDB Express per task. Google announced AlloyDB Express on 4 November 2026 as Preview, not GA. The vendor reports provisioning in approximately two seconds through public DNS and 20 seconds through Private Service Connect; resuming an idle instance may take roughly 20 seconds. Preview limitations include a 20 GB data limit, no CMEK, no SLA and no backups.
The team has not established whether those claims hold in its own network and CI pipeline. The architect must choose a limited proof-of-concept design and define the agent's authority. This is not an account of a real deployment or an endorsement of Preview for sensitive data.
What architecture decision would you make for this short-lived PoC, and what constraints would you place on agent autonomy? Explain your choice in terms of isolation, security, cost and exit strategy.
Guiding questions
- Does fast provisioning justify giving the agent broad infrastructure administration rights?
- Who deletes the database after the pull request closes or a job terminates unexpectedly?
- Which Preview limitations would disqualify this approach for a different data classification?
- What measurements would demonstrate an advantage over standard managed PostgreSQL?