Lilith.
⌕
Editorial illustration: IBM puts time-series forecasting directly into streaming SQL
Lilith illustration · editorial remix

IBM and Confluent have integrated Granite time-series models into Apache Flink as built-in SQL functions. They are managed in Confluent Cloud and can run on premises with Confluent Platform. The models use CPUs, and IBM says the family has passed 50 million downloads.

Forecasting moves from a batch job into the event stream

The integration exposes forecasting and anomaly detection over incoming data. Flink prepares the data, invokes a model and attaches the result to the events that produced it. An application can react without waiting for the next batch calculation.

IBM describes several models and tasks. Tiny Time Mixers handle forecasting, TSPulse creates representations for similarity search and AI_DETECT_ANOMALIES compares a new observation with an expected range. A pretrained model provides a common starting point across many series without training a separate model for each one.

SQL shortens the path from a signal to an operating decision

The integration matters most where a forecast loses value quickly. A production line can evaluate quality every 10 minutes and forecast 4 hours ahead. Inventory planning can refresh a 28-day outlook as sales arrive. A trading algorithm can forecast the next 30 minutes of volume on each one-minute bar.

These examples combine a model with rules, context and an operator's decision. That is more important than model accuracy alone. The forecast appears where the application already manages an order, payment or machine, reducing the need for a separate service merely to transport model output.

Illustrative workflows are not production results

The ice cream line, oat milk replenishment and trading scenarios demonstrate possible workflows. The source explicitly labels some data as simulated and provides no production metrics for accuracy, latency or cost. Download counts also do not reveal how many organizations operate the models.

A foundation model does not remove data quality work. Sensor changes, seasonality or a new operating regime can break an expected range. In payments and trading, the output should remain one signal among rules, other models and an audit trail.

Latency, false alarms and version control will decide adoption

Real customer references should show latency across large numbers of series, CPU consumption and accuracy against simpler statistical methods. For anomaly detection, false alarm rates will be decisive because a stream of bad alerts quickly erases the value of a cheap model.

The second test is operational governance. Teams need to trace the model version, input data and decision influenced by each forecast. If Confluent preserves that trail inside the stream and supports safe rollback, a SQL function can become useful infrastructure rather than a polished demo.

Lilith's verdict

IBM did not seat a fortune teller beside the database. It put forecasting onto the conveyor carrying payments, inventory and sensor readings. Now it must show how many false alarms fall off the end.