Scaling from 0 to 1M users: what architecture changes were necessary?

It's always fascinating to look back at the architectural decisions made when a system goes from zero to millions of users. I've seen it happen a few times, and the journey is never linear. Initially, a single PostgreSQL database with a simple cache might be more than enough. But then, the first bottleneck hits, usually the database CPU or I/O, and you add read replicas. Next, the caching layer becomes critical, moving from simple in-memory caches to distributed systems like Redis. At some point, the single relational database becomes too much to handle, forcing a move to sharding or specialized NoSQL stores for different data types. Often, what started as a monolithic application needs to be broken down into microservices to allow independent scaling and development velocity. The key is recognizing when to make these shifts, and what decisions from the early days need to be revisited or completely re-architected. What were some of the most critical architectural inflection points you experienced going from a handful of users to over a million?