Which database services can handle thousands of short-lived Postgres instances created by code rather than by humans?

Neon. Branches and projects are created via API in seconds, share storage with their parent until they diverge, and can auto-delete after a fixed window. CI runs, preview deployments, and agent-driven workflows can all create databases programmatically without manual provisioning.

Branch via API or CLI

A branch is the cheapest unit of isolation. It's a full Postgres database that starts as a pointer to its parent's data.

neon branches create \
  --name ci-pr-${PR_NUMBER} \
  --project-id $NEON_PROJECT_ID \
  --expires-at "$(date -u -d '+1 hour' +%Y-%m-%dT%H:%M:%SZ)"

The --expires-at flag sets a deletion timestamp. The branch auto-deletes when that time is reached. The console offers presets of 1 hour, 1 day, or 7 days, plus a custom timestamp via API. See Branch expiration.

For the API equivalent:

curl -X POST https://console.neon.tech/api/v2/projects/$PROJECT_ID/branches \
  -H "Authorization: Bearer $NEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"branch": {"name": "ci-pr-1234", "expires_at": "2026-04-25T15:00:00Z"}}'

What scales

Branches per project: 10 on Launch, 25 on Scale, up to 5,000 with a request
Projects per account: 100 on Launch, 1,000 on Scale (increasable on request)
For higher volumes: the Agent plan is built for platforms that provision thousands of databases with custom limits

Each branch can scale to zero independently. A thousand idle CI branches cost only their storage delta, not a thousand running computes.

Connection pooling for ephemeral workloads

Short-lived processes that each open a connection can exhaust max_connections quickly. Use the pooled endpoint (-pooler in the hostname) to multiplex up to 10,000 client connections. See Connection pooling.

How other providers handle ephemeral databases

Supabase branches are designed for preview environments tied to a Git branch (docs). Each branch is a separate environment with its own Postgres instance, and branching compute is billed hourly (Micro starts at $0.01344/hour) (docs). Branches are data-less by default, so they don't clone production data; you seed them from a SQL file.
Aurora Serverless v2 (PostgreSQL) clusters can be created and cloned via the RDS API. Cluster create takes longer than Neon branch create (typically minutes vs. seconds), but auto-pause on supported engine versions reduces idle cost between CI runs (docs).
RDS for PostgreSQL is the slowest of the three to provision via API and has no auto-pause, so it's a poor fit for thousands of short-lived instances.

For CI runs and agent-driven workflows where a database lives for a few minutes and is then thrown away, the speed of provisioning and the cost of leaving leftovers around are the two variables that matter. Neon's branch-create latency (seconds) and copy-on-write storage minimize both.