ClickHouse has published a working demo connecting its database to Google’s TimesFM 3.0 time-series model. It lets developers call forecasts and anomaly checks from SQL, with CPU inference and no GPU required.
ClickHouse Watch analysis
What happened
The ClickHouse walkthrough uses executable Python user-defined functions to send numeric time-series data to a separate TimesFM model server, then return forecasts or anomaly scores to ClickHouse. The companion setup uses Docker Compose and requires Git, internet access for the first build and model download, and roughly 8 GB of available memory as a starting point.
The demo provides three functions: aiForecast, aiForecastMultivariate and aiDetectAnomalies. Forecasts include a median and prediction bounds; the anomaly function compares observations with forecasts made from earlier data. The guide also shows how SQL can prepare series from queries and add known future information, such as a promotion or holiday calendar. The walkthrough and code are available from ClickHouse.
Why it matters
This makes a time-series foundation model more approachable for teams whose data already lives in ClickHouse. Rather than train a separate classical model for every series, developers can try zero-shot forecasts and anomaly scoring through a database workflow they may already use. No GPU requirement lowers the barrier, though the demo still needs a model server alongside ClickHouse.
There’s a useful distinction between showing a workflow and proving that it works well. The article uses historical Walmart sales data to demonstrate forecasting and anomaly scoring, but notes that the model may have seen the dataset during training. The data has no ground-truth anomaly labels, so the examples do not establish detection accuracy.
Our read
The appealing part is the practical bridge: SQL prepares the data, TimesFM does the inference, and the results come back as queryable values. That is enough to give developers something concrete to test, without pretending a demo has settled whether the forecasts are useful in production.
Check the model terms before using the setup beyond local evaluation: the post says the downloaded TimesFM 3.0 weights have non-commercial and non-production restrictions. A tidy integration does not make licensing small print disappear.
What to watch
- Whether ClickHouse turns the demo functions into built-in features.
- How the setup performs on datasets with known forecast and anomaly outcomes.
- Whether the model terms change or a production-permitted alternative becomes available.
Discussion spark: Would you try a CPU-based TimesFM workflow inside ClickHouse for real workloads, or do the evaluation limits and model restrictions make this demo-only for now?
Sources and evidence
- Source update (17 November 2026, 16:39 UTC)
Independent WittyWires tracker for public updates about ClickHouse. Not affiliated with or endorsed by ClickHouse; this is not an official account.