Abstraction Platform/Runtime And Artifact Procedures

This family builds, materializes, refreshes, and inspects abstraction artifacts and scenario-aware runtime outputs.

Feature snapshots are abstraction artifacts, not database archive snapshots. A materialized feature snapshot now writes each computed feature into a stream-backed feature history keyed by feature name, subject identity, and snapshot timestamp. Snapshot artifacts still provide contract metadata for listing, freshness, and grouped reads, but the canonical value surface is the feature stream: timestamp, subject_id, and value.

Inline abstraction feature reads use the same stream. RETURN actor(Member:'member-102').churn_risk returns the latest materialized value for that subject, and ASOF <timestamp> RETURN actor(Member:'member-102').churn_risk returns the latest stream value at or before the query timestamp. Use nexyron.feature_history when the caller needs every (timestamp, subject_id, value) row rather than one latest value.

The backend abstraction and feature generation flow also registers one automatic daily feature stream contract containing the finalized active feature set across all generated abstractions. The automatic 30-second materialization loop is currently disabled, so the backend-owned contract remains available without starting background materialization. Manual snapshot backfills are capped at 30 materialization points as well.

Automatic feature stream artifacts are retained to the same 30-day active window. The service prunes stale automatic materializations on startup, during the recurring stream-retention maintenance pass, and around manual automatic backfill runs. Pruning uses nexyron.snapshot_materialized_drop, so clustered cleanup follows the normal write path and removes the corresponding stream-backed feature history points instead of only hiding artifact metadata.

That endpoint reports the active abstraction and feature counts, ready and missing stream-days, stale materialization counts, the manual backfill window, and the last ten automatic stream materializations. POST /system/ml-studio/feature-snapshots/cleanup?database=<name> runs the 30-day automatic retention pass immediately, while DELETE /system/ml-studio/feature-snapshots?database=<name> clears all automatic feature stream materializations for the database. These are feature-stream maintenance APIs, not database archive snapshot APIs.

Pages: