Collections (entities), fields, types, constraints (required, unique, default, enum), indexes, and relationships (ref vs embed). All of these are hard to change after code is written.

Use ref for one-to-many with large N (e.g., user and messages). Use embed for small, tightly coupled data that always loads together (e.g., an address inside a user). Embedding is fast but bloats documents; ref keeps them small.

Based on the queries you will run. If you often look up by email, add a unique index on email. If you often check if a request from A to B exists, add a compound unique index on (fromUserId, toUserId). Indexes serve queries.