A Queen node is one process. Every write is decided once, on the leader, and then applied on every node. These pages follow a write through the process, from the request to the disk.
The rule behind the whole design is: one leader per tenant. A tenant lives in one raft group, and one leader orders all its writes. So a transaction, even across ten queues, is one log entry: no coordinator, no two-phase commit. The price is that one tenant’s writes do not spread over more nodes. To use more nodes for writes, run more raft groups (see Many tenants on one cluster): with as many groups as nodes, every node leads one.
