Feature freshness and TTL semantics

For a latest-observation feature, event age measures how long ago the selected observation occurred. Availability age measures how long ago that version became usable on the relevant path. Pipeline delay compares those two timestamps, while cache age starts when the cache entry was filled. They answer different questions. A profile attribute that changes rarely can remain valid for months; a market quote may be unusable after seconds. Define feature-specific acceptable age and behavior when timestamps are absent or clocks are inconsistent.

Aggregates need additional care. The timestamp on a window result may identify its window endpoint, not the age of the newest contributing event or the completeness of upstream ingestion. A zero count for a recent interval can be correct when there were no events, or wrong when the source stalled. Monitor source progress, ingestion lag, watermark assumptions, and expected update cadence alongside the value’s timestamp. Do not declare every unchanged value stale or every recently recomputed value informed by recent complete source data.

A cache lifetime often expires an entry relative to insertion or refresh, while a data-validity rule can reject a feature relative to its event timestamp. Redis EXPIRE supplies key-expiration behavior; it does not inspect a value’s embedded source timestamp. Rewriting an old observation with a new expiration can therefore keep old information available. Physical deletion timing, read-time filtering, registry-cache refresh, and source retention are distinct policies. Check the exact backend contract instead of assuming every parameter named ttl operates at the same layer.

Feast’s historical feature-view TTL bounds how far back retrieval searches relative to each entity-row timestamp. That is different from measuring age relative to the current time when generating a dataset. Online validity and physical retention also depend on implementation and configuration; verify them for the selected provider and version. A historical TTL does not make an upstream stream healthy or certify an online serving-age objective. A request should apply the declared missing or stale fallback policy using the metadata and guarantees it actually has.

Track age distributions, missing timestamps, update failures, source progress, and stale/default responses at the point where models consume values. A global average can hide hot entities that update promptly and rare entities that are never materialized. Preserve the model and feature definition versions so an alert can identify which consumer policy was violated. Decide explicitly whether serving an old but interpretable value is preferable to a default or failure; that tradeoff belongs to the model’s tested contract rather than an accidental cache behavior.

The LRU instrument illustrates that residency and hits do not contain age information. Reset, set both C=4 and s=4, predict four misses and thirteen hits after seventeen requests, then Step once. Separately Reset, set both C=2 and s=4, predict seventeen misses after seventeen requests, then Step once. Reset starts with sixteen requests and Step adds one. This toy tracks addresses and recency only: it has no TTL or source timestamps. A high hit count therefore cannot be interpreted as fresh feature observations, just as an actual cache hit alone cannot establish upstream recency.