nilalytics v0.3.0

Highlights

• Subject partition + person bucketing for user_events. The curated table is now laid out subject › date › bucket(person_id) and sorted by person_id, so category, time-window, and single-person reads all prune well. person_id is hashed into N buckets (bucket(N, person_id)) — DuckLake's recommended way to partition a high-cardinality key without the small-files problem. Tune with NILA_USER_EVENTS_BUCKETS (default 16).
• Subjects. Every event is classified into a low-cardinality subject — errors / activities / ai_usage / traceability / other. An explicit nila.subject attribute wins; otherwise it's derived (severity/exception → errors, gen_ai/llm → ai_usage, identify/audit → traceability). Unknown values clamp to other.
• Identity lookup by any key. Users are defined by any stable, unique key (email / account id / phone), hashed into a person_id. New nilalytics identify <key> prints that id, and nilalytics query user --key <value> [days] hashes it for you.
• New reads. query subject <name> [days], a subject breakdown in query user_events, and the emitter now sends sample ai_usage events.

Upgrade note

The user_events table gains a subject column and a new partition layout. On first start after upgrade, the server rebuilds it from otlp_logs (the source of truth) automatically.

Full changelog: v0.2.0...v0.3.0