Feature Store  ·  View 18 of 21  ·  Operations

The Feature Lifecycle Loop

Define, materialise, serve, observe, attribute, revise — and the one field that lets the loop close.

Editable source SVG draw.io All views
Define one artefact Materialise two plans Serve value + age + status Observe skew + drift Attribute data or pipeline? Revise or retire new version, pins notified Feature lifecycle compiled plans version-stamped values 1% sampled vectors verdict cause new version The Feature Lifecycle Loop Application we own Security / platform Decision point The loop only closes because every served value carries the definition version that produced it. v 1.0 · owner Data Platform Architecture · date 2026-09

Decisions

  • The loop only closes because every served value carries the definition version that produced it. Without that field, Observe produces a number and Attribute produces an argument.
  • Attribute is a decision node, not a step: separating a changed world from a changed pipeline is the judgement the whole loop exists to support.
  • Revision creates a new version and notifies pinned consumers. Retirement is refused while a production consumer remains.

Assumptions

  • Nightly skew replay per onboarded model; skew tolerance ≤ 0.1% of sampled requests per feature.
  • Lifecycle states: experimental → production → deprecated → retired.

Risks

  • Nightly replay dates a regression to a day, not an hour. For a fast-moving regression at peak, the loop is too slow and the per-consumer health view is the only live signal.