Google's AlloyDB Express preview: a small Postgres 18 database that pauses when idle

Google has put AlloyDB Express into Preview. It is a small configuration of AlloyDB for PostgreSQL, not a separate product, aimed at developers, test pipelines and AI coding agents that need a throwaway database fast. The release notes list it on 4 November for PostgreSQL 18, and the docs place it under Google’s Pre-GA Offerings Terms, which means “as is” and possibly limited support.

This is a different thing from the AlloyDB agent query compute we covered in October, which adds isolated compute for agents on an existing production cluster. Express is a standalone starter database.

What you get

• Fast creation. The docs say a database is ready in about 2 seconds with a public DNS endpoint, or within 20 seconds with Private Service Connect. Google’s blog says “under two seconds.”
• Scale to zero. When an instance is idle, compute drops to zero and you stop paying for it. The blog says the pause comes after 15 minutes without activity; the docs say only “when idle.” On the next connection or query, the docs say it resumes in about 20 seconds, a slower figure than the blog’s “wakes up automatically in seconds.” Plan around the docs number.
• Small, fixed sizes. One zonal primary instance with 1 vCPU and 4 GB of RAM or 2 vCPUs and 8 GB. Storage grows automatically to 20 GB, and at 20 GB further writes are refused. max_connections is fixed at 200.
• AI features included. The docs say Express keeps full AlloyDB AI: ScaNN vector search, the vector (pgvector) extension, natural language queries and SQL functions that call models. The blog names ai.if, ai.rank and ai.generate for Gemini calls inside SELECT and INSERT.
• One-way upgrade. One click in the console turns an Express cluster into a full AlloyDB cluster, keeping data, schemas, users, network configuration, connection string and DNS name. You can’t downgrade back.

What it leaves out

Per the docs: no high availability, read pools or cross-region replication, no automated or on-demand backups and no point-in-time recovery, no SLA, no customer-managed encryption keys, PostgreSQL 18 only, and no committed-use discounts. Connections go through a public DNS endpoint or Private Service Connect; public IP and private services access are listed only for full AlloyDB. Google patches Express itself, typically when an instance resumes or during low-usage hours.

One conflict to know about: the blog lists Terraform among the ways to create a database, but the docs comparison table gives Express only the console, gcloud CLI and REST API, with Terraform and Config Connector listed for full AlloyDB. If your test harness is Terraform, check before you build on it. The blog also mentions creation over MCP.

Pricing and regions

From the pricing page: compute is billed per second by compute unit (1 vCPU with 4 GB), from $0.115 per hour in us-west1 up to $0.161 in asia-east2, and storage from $0.30 per GiB per month in us-west1. Paused compute costs nothing, but storage is billed while paused. The free tier “includes 50 compute units and 2 GB of storage per billing account” each month for the first 12 months, and the page says “Billing starts from December 2026,” with no exact day.

Express runs in seven regions: asia-east1, asia-east2, asia-northeast1, asia-southeast1, australia-southeast1, europe-west2 and us-west1.

Why this matters

If your agents or CI jobs create a fresh Postgres for every task, Express gives each one an isolated database with pgvector and ScaNN, and you pay compute only while it runs. The limits put it firmly in the test-and-prototype lane: no backups, no SLA, a 20 GB ceiling and a wake-up of about 20 seconds that a CI step might notice. Its cleanest selling point is the in-place upgrade, which keeps the connection string, while the Postgres-for-AI case we made in September still holds for when that prototype becomes the real app.