Pub/Sub moves toward inline AI

Most cloud message queues just move packets from one program to another. Google Cloud’s Pub/Sub now lets you rewrite or enrich a single message inside the queue itself with Single Message Transforms, including a generally available Artificial Intelligence inference option that calls Vertex Artificial Intelligence models and appends the result to the message. (cloud.google.com, docs.cloud.google.com, docs.cloud.google.com) A message queue is the digital version of a loading dock: one system drops off boxes, another picks them up later. Pub/Sub is Google Cloud’s managed version of that dock, built for real-time streams between independent applications. (docs.cloud.google.com) The old pattern was simple but heavy. If a message needed cleanup, validation, or a quick model score, teams usually had to run a separate consumer in Dataflow or Apache Flink just to touch each event before passing it on. (docs.cloud.google.com) Single Message Transforms are Google’s shortcut for that middle step. They run lightweight changes directly on message data or message attributes, and Google says you can attach them to either a topic before storage or a subscription before delivery. (docs.cloud.google.com, docs.cloud.google.com) That placement changes what every downstream system sees. A topic transform rewrites the stored copy for everyone, while a subscription transform rewrites only the version sent to one subscriber, which is closer to giving each reader its own edited edition of the same newspaper. (docs.cloud.google.com) The new Artificial Intelligence inference transform pushes that idea one step further. Pub/Sub can call Google or partner models through Vertex Artificial Intelligence, or a custom model on a Vertex endpoint, and then attach classifications, sentiment, predictions, or embeddings to each event as it passes through. (docs.cloud.google.com) That means some pipelines no longer need a dedicated enrichment worker for every event. A retailer can tag browsing events with recommendation context inside Pub/Sub, and a downstream database or alerting system can read the enriched message without waiting for a separate processor to wake up. (docs.cloud.google.com, docs.cloud.google.com) But this is not a replacement for a real streaming engine. Google’s own Dataflow guidance still treats streaming as a world of watermarks, windows, duplicates, and late data, where the hard part is not changing one message but deciding what to do when events arrive out of order or arrive after a window should have closed. (docs.cloud.google.com, docs.cloud.google.com) A watermark is Dataflow’s estimate of how far the stream has progressed in event time. If the watermark has passed the end of a time window and an older event shows up later, Dataflow treats that record as late data, which can change counts, alerts, or machine decisions after the first result was already emitted. (docs.cloud.google.com) Duplicates are the other trap. Dataflow says exactly-once mode avoids duplicated records in outputs, but Google also says at-least-once mode can cut cost and latency, with a baseline duplicate rate that is typically under 1 percent and spikes during retries or worker failures. (docs.cloud.google.com, docs.cloud.google.com) So the line is getting clearer. Pub/Sub is moving up from “mailroom” to “mailroom plus quick label maker,” while Dataflow remains the place for joins, windows, deduplication, and anything that depends on time and state across many events instead of one event at a time. (docs.cloud.google.com, docs.cloud.google.com, docs.cloud.google.com)