bare · adv-contradiction-semantic-control-scope · seed 2

Write it to BigQuery.

Reasoning:

- **Purpose match.** A nightly rollup produces aggregated data for analysis and reporting. That's exactly what the analytics warehouse exists for. BigQuery's columnar storage and MPP engine are built for scanning and aggregating large tables, whereas PostgreSQL is tuned for transactional, row-level access.

- **Protect production.** Running heavy aggregation reads and bulk writes against your OLTP PostgreSQL instance competes with live application traffic for CPU, I/O, cache, and locks. Keeping analytical workloads off the production database avoids latency spikes and bloat (dead tuples, vacuum pressure) from large periodic writes.

- **Source of truth stays clean.** PostgreSQL should hold authoritative operational state. Derived, recomputable rollups don't belong there; if they're wrong you just re-run the job against the warehouse.

- **Consumers.** Analysts, BI tools, and dashboards typically already connect to BigQuery. Putting rollups there makes them queryable alongside the raw event/warehouse data.

When you might also write (a copy) back to PostgreSQL:

- The aggregates need to be served to end users inside the application at low latency (e.g., an in-product "usage this month" widget).
- Another operational service needs them transactionally.

In that case, the canonical rollup still lives in BigQuery, and you push a small, query-optimized subset into a dedicated table (or a separate reporting database/replica) in PostgreSQL — not the core transactional tables.