Two databases is not a failure of nerve. It is an admission that the workloads have nothing in common.
Playback events arrive at a rate that makes a relational database sad and answer questions that make a relational database sadder: distinct viewers per title per hour, buffering ratio by region, completion curves.
PostgreSQL keeps everything with a lifecycle — subscriptions, entitlements, catalogue. ClickHouse keeps everything append-only. Nothing is written to both, and nothing joins across them at query time; the reporting layer asks each what it is good at.
Query planning matters more than schema. A MergeTree ordered by (title_id, event_time) answers the product questions in milliseconds and the finance questions in seconds — reversing the key reverses which team complains.
Materialised views did the rest. Pre-aggregating per hour cut dashboard latency by an order of magnitude, at the cost of a schema that now has opinions about what people will ask.