PostHog: analytics on billions of events

Product analytics asks questions row databases are bad at: count events matching a filter over months, grouped by a property, for one customer among thousands. PostHog started on Postgres and moved to ClickHouse, a column store that reads only the columns a query touches and compresses them well.

Their posts are useful because they cover the costs, not only the speed-up. Events arrive through Kafka rather than direct inserts, so spikes and database outages are absorbed. Changing existing rows in a column store is expensive, and doing it often caused real trouble. Properties stored as one JSON blob had to be parsed on every query until the most used ones were given columns of their own. And joining events to people at query time was slow enough that they now copy person data onto each event as it arrives, accepting that a person's later changes no longer rewrite history.