Writes go to the primary. Reads are served by replicas that trail behind by the configured lag.
Write a new balance and read it back immediately. If the read lands on a replica before replication catches up, you will see the old value — a stale read, with both machines behaving perfectly correctly.
Raise the lag and the window widens. Route reads to the primary and it closes, at the cost of the capacity you added replicas to get.
Account balancelag 120ms
Raise the replication lag in the inspector, write, then read immediately. The wider the lag, the easier the stale read is to catch — and under real load, lag grows exactly when you least want it to.