Most of what I read uses the term real-time for anything from a few seconds of delay to whatever is presented before the morning standup, and the database setup you end up with depends heavily on which end of that range you need.

The usual options are running reports against a read replica, using CDC to populate a separate analytical store with something like Debezium, an HTAP database that handles both workloads, or one of the newer designs like Databricks, where transactional and analytical compute share a single storage layer.

For those of you who have had to make this call, I'd like to know how fresh the analytics had to be in the end, which of these you chose, and what caused trouble once it was in production, whether that was replica lag, schema changes, backfills, or cost. If you decided a nightly load was good enough, I'd like to hear that reasoning, too.