Nothing here is lost or corrupted. Watch only which replica answers each read. A user watched their own successful write disappear. No data was lost at any point. Any read-replica setup with a naive load balancer — the most common production consistency complaint there is