Define your feature views (batch, stream, or on-demand) in Python using the Tecton SDK, specifying the data sources, transformations, and materialization schedules
Group the feature views needed for a given model into a FeatureService object, which provides a single retrieval endpoint returning a feature vector composed from all specified views
Apply the feature definitions to your Tecton workspace with tecton apply to materialize features into the Online Feature Store
In your inference application, call the Tecton Feature Server HTTP API or SDK to fetch real-time feature values by passing entity keys (e.g., user_id, item_id) to the FeatureService endpoint
Validate returned feature vectors for expected schema and freshness using Tecton's feature monitoring to detect drift or stale features before they reach the model
Use the Tecton SDK's get_online_features() method or the HTTP API for sub-10ms p99 latency retrieval from the Online Feature Store in production
Known gotchas
On-demand feature views compute features at request time using request-time data; if the request payload is missing required fields, the Feature Server returns an error rather than defaulting, causing inference failures
Materialization schedules for batch feature views introduce a freshness lag; if your model depends on features that are only updated hourly or daily, real-time predictions will use data that may be hours old
Tecton's online store keys are case-sensitive entity key strings; mismatched casing between feature definition entity names and the keys passed at inference time results in null feature values being returned silently
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp