AI-Powered Streaming: Pub/Sub Gets Inference Capabilities

Google has announced a wave of updates to its Data Cloud platform covering the week of September 7–10, with the headline feature being the general availability of AI Inference Source Message Transforms (SMTs) for Pub/Sub. This capability lets teams apply machine learning inference directly on an incoming stream of events using models hosted in Gemini Enterprise Agent Platform, with predictions appended to each event before it reaches downstream systems such as BigQuery or Bigtable. For organisations running real-time anomaly detection or fraud scoring pipelines, this removes the need for a separate inference microservice sitting between the message bus and the data warehouse.

The practical implication is significant: instead of building and maintaining a consumer service that pulls from Pub/Sub, calls an inference endpoint, and re-publishes enriched events, teams can now configure an SMT in Pub/Sub itself. Google frames this as a simplification of anomaly detection architectures, and that framing is accurate — fewer moving parts means fewer failure points and easier operational observability.

PostgreSQL Connectivity Reaches General Availability

Also reaching general availability this week is the PostgreSQL Source Connector for Datastream, Google's change-data-capture service. This connector enables continuous replication of PostgreSQL databases into BigQuery, enabling near-real-time analytics on transactional data without requiring manual ETL jobs. The GA announcement signals that the connector has cleared Google's reliability and support bar, making it suitable for production workloads beyond early-adopter experiments.

PostgreSQL remains one of the most widely deployed open-source databases, and first-class CDC support from Google is a direct competitive response to similar capabilities in AWS Database Migration Service and Azure Database Migration Service. Organisations already using Datastream for MySQL or Oracle replication can now extend the same pattern to Postgres workloads with a supported, maintained connector.

BigQuery and Looker Updates

The update batch also includes enhancements to BigQuery and Looker, though specific details for these components were part of the broader announcement cycle. Google has been steadily integrating Looker more tightly with BigQuery since the acquisition, and recent releases have focused on reducing the friction between semantic modelling in Looker and query execution in BigQuery. Teams that have adopted the Looker-BigQuery stack benefit from these incremental improvements without requiring architectural changes.

Google's data cloud announcements follow a consistent weekly cadence, which itself reflects a strategic choice: rather than batching features into large quarterly releases, the team ships incrementally and communicates the changes through a structured update series. This approach keeps the platform feeling active and responsive to customer feedback, though it also places a burden on platform teams who must evaluate whether each update requires configuration changes in their environments.

What These Updates Mean for Data Teams

The Pub/Sub AI Inference SMT is the most architecturally significant change in this batch. Teams currently using Pub/Sub as the backbone of their event-driven pipelines can now route inference workloads through the message layer itself rather than building bespoke consumers. The trade-off is that this tightens the coupling between the streaming infrastructure and Gemini Enterprise Agent Platform — teams using third-party model hosting will not benefit from the same native integration.

The PostgreSQL CDC connector reaching GA is less dramatic but practically important. CDC-based replication is increasingly the standard pattern for keeping analytical systems in sync with operational databases, and removing the experimental caveat from the PostgreSQL connector makes it a credible choice for teams evaluating data replication strategies on GCP.

Key Takeaways

  • Pub/Sub AI Inference SMTs are now generally available, enabling in-stream model inference without a separate consumer service.
  • The PostgreSQL Source Connector for Datastream has reached GA, supporting production-grade CDC replication into BigQuery.
  • Google's weekly Data Cloud update cadence reflects an incremental shipping strategy rather than large quarterly releases.
  • The tightest benefits of the new Pub/Sub SMT apply to teams already using Gemini Enterprise Agent Platform for model hosting.