The indexes in the Realtime Database are created in-memory and on-the-fly - and they are 'garbage collected' when memory is needed for something else.

This makes it expensive (meaning high CPU usage, not costly) to query large collections of data in the Realtime Database.

Collections can also become so large that they are not queryable through the SDKs, but only using the firebase cli tool. One hope would be that with an actual persisted index, it would always be possible to iterate through a collection of nodes, no matter the number of nodes in the collection.