Feature definitions and entity keys

An entity can be a user, merchant, device, session, or a composite such as user within merchant category. Feast distinguishes an entity name from the physical join keys used to retrieve values. Include tenancy and domain where identifiers are not globally unique. Preserve types and canonical encodings: a string identifier with leading zeros may not equal an integer with the same displayed digits. Normalization rules must be shared by writers, historical joins, and online requests or values can silently attach to the wrong subject.

Naively concatenating composite fields can create collisions. The tuples (1,23) and (12,3) both become the string 123 if separators and lengths are omitted. Use an unambiguous typed encoding rather than relying on a delimiter that might itself occur in data. Hashing an already ambiguous string cannot recover lost distinctions. A key migration therefore changes a data contract, not merely a storage optimization. Plan how old and new requests resolve identity, and avoid accidentally mixing namespaces during a staged migration.

A definition should name the source and its version, selection predicates, deduplication rule, window boundaries, aggregation, units, output type, missing-value policy, and transformation revision. For a purchase rate, distinguish events from successful payments and define whether canceled purchases are retracted. State the timestamp that anchors the window and whether the upper boundary includes an event exactly at prediction time. A feature reference used by a model must resolve to the intended definition, even when another team publishes a similarly named field.

Schema compatibility is broader than numerical type compatibility. Changing cents to dollars or a count to a rate can leave the field represented as Float32 while completely changing its meaning. A model’s feature bundle should fix ordering and definitions instead of depending on whichever view is currently active. Feast feature services group features used by model versions, but application governance must still bind the intended transformation and source contracts. Publishing a definition does not itself compute historical values, complete a backfill, or prove that every online key is populated.

A missing feature may mean no qualifying activity, no source record, an expired value, a failed lookup, or an unmaterialized key. These are not interchangeable states. Define which cases legitimately map to zero and which trigger a default, mask, fallback model, or explicit failure. Attach event and availability metadata when required to assess the observation. Validate ranges and units using domain knowledge; a schema parser cannot tell whether a plausible-looking value came from the wrong customer or a future correction.

The fixed sigmoid instrument illustrates why interpretation changes matter downstream. Reset, set both x=1 and w=1, predict σ(0.5)≈0.62246, then Play or Step. Separately Reset, set both x=2 and w=1, predict σ(1.5)≈0.81757, then Play or Step. The model computes only σ(wx−0.5); doubling a feature representation without adapting the consumer changes its result. This is not a feature-store lookup, trained predictor, or key validator. It provides a concrete consequence of changing an input’s scale while keeping its consumer rule unchanged.