Google Cloud Launches Spanner Queues: Putting Message Queues Inside Database Transactions to Make AI Agents More Reliable

What Google Cloud announced

Google Cloud announced on its official blog that Spanner queues are now generally available (general availability, meaning the feature is no longer in preview). Spanner is Google Cloud's database service. According to the company, Spanner queues are a native transactional messaging feature built directly into the Spanner database. The post is bylined by product manager Nitin Sagar and engineering manager Matthew Mucklo.

A "queue" here can be thought of as a to-do list of jobs waiting to be processed. A "transaction" is a set of data changes bundled together that either all complete or are all canceled. Google Cloud says that with Spanner queues, creating a message is simply another write within a transaction. An AI agent is an AI program that can carry out tasks on its own, such as issuing refunds or managing inventory. When such an agent updates its own state, that update and the instruction for "what to do next" are written together as a single unit. Either everything succeeds or nothing takes effect.

The problem it aims to solve: state and actions out of sync

Google Cloud notes that AI agents often need to record their internal state in a database. At the same time, they dispatch asynchronous actions (work that doesn't have to finish immediately and runs later) through a separate messaging or event-queue system. The company argues that having the two systems commit independently breaks transactional consistency.

Google Cloud gives two failure scenarios. In one, the database is updated but the action is never sent, so the agent "decided but didn't act". In the other, the action is sent but the database update is rolled back (undone), so the agent acts on invalid state. The company says developers previously had to build their own workarounds, such as the outbox pattern, idempotency layers and reconciliation programs. An idempotency layer is a mechanism that ensures running the same action repeatedly doesn't produce duplicate results. Google Cloud describes these workarounds as a heavy reliability burden.

Key features

• Deciding and acting in one step: Google Cloud says an agent can update its memory or state and add a task to the queue for another agent in the same transaction. Both are treated as one indivisible ("atomic") change, so they always complete together.
• Scheduling and delays: according to the company, delayed retries, scheduled checks or SLA (service level agreement) escalation timers can all be written alongside state updates.
• Human approval and timeouts: in Google Cloud's example, an agent waiting for a manager's approval can record a pending-approval state and schedule an escalation for 72 hours later. If the manager approves early, the escalation task can be canceled in the same transaction.
• Lease management: a "lease" is the period during which a worker temporarily claims a task. Google Cloud says Spanner manages message leases automatically. Leases can be extended for long-running LLM (large language model) reasoning or external API calls.
• SQL monitoring: the company says queues can be defined, inspected and managed with GoogleSQL. Execution records can be audited with standard SQL.

Google Cloud also draws a distinction between Spanner change streams and Spanner queues. Change streams continuously capture data changes in the database and stream them to downstream analytics or storage. Spanner queues are designed specifically for transactional task orchestration.

What it means for general readers

Ordinary users won't interact with Spanner queues directly, but they may indirectly experience the problems it aims to solve. Google Cloud uses refunds and approvals as examples. If an agent's "records" and "actual actions" fall out of sync, things that should happen may not, or reminders may still arrive after approval. Google Cloud says this feature hands exactly these consistency problems over to the database.

Google Cloud says Spanner queues can also be used beyond AI agents. Its examples are social feeds, live news updates, retail order and inventory workflows, and transaction notifications in financial services. Actual results still depend on how each company designs and uses it, and no independent source has yet verified its performance.

Frequently asked questions

Browse the latest news in this topic

• AI news