Repository Public
1# 0003 — PostgreSQL native partitioning instead of TimescaleDB; no Redis23**Context (observed).** PostgreSQL 17 (Homebrew) is present on 8 nodes; TimescaleDB is installed nowhere. Expected volume: ~29 nodes × ~25 series every 10 s ≈ 6 M rows/day.45**Decision.** One PostgreSQL database holds inventory, identity, actions, audit and metrics. Metrics use declarative range partitioning: `samples_raw` (daily partitions, 7 days), `samples_1m` (30 days) and `samples_1h` (400 days), filled by idempotent roll-up jobs driven by watermarks. Retention = dropping partitions. Coordination between services uses durable tables plus `LISTEN/NOTIFY`; rate limits are in-process (single API instance). Redis is not used.67**Why.** Building and maintaining a PostgreSQL extension on the host for this volume adds an operational dependency without need; the "or equivalent" of the mandate is met by native partitions + roll-ups. Fewer moving parts is the robust choice on a single host.89**Revisit when** raw ingestion exceeds ~50 M rows/day or queries need continuous aggregates beyond the two fixed tiers.10