firestore: arrays vs sub collection of documents performance

I would generally go with the approach of "Every chat room has a subcollection of messages. And every new message is a separate document in this subcollection." This has several advantages: It's easy to add or edit individual messages, and you can perform a number of different queries (like "Grab the 20 most recent messages")

The biggest drawback, I suppose, is that if you find that new users are frequently going to be entering your chat and will want to see the entire chat history of the room up until they joined, that would result in a large number of database reads. Realistically, though, I don't know how often that would happen in real life, and you could mitigate this by using pagination to grab your historical chat in batches.