The comment at [1] hints at a solution: in the response of the write request return the id of the transaction or its commit position in the TX log (LSN). When routing a subsequent read request to a replica, the system can either wait until the transaction is present on the replica, or it redirect to to the primary node. Discussed a similar solution a while ago in a talk [2], in the context of serving denormalized data views from a cache.

[1] https://news.ycombinator.com/item?id=46073630 [2] https://speakerdeck.com/gunnarmorling/keep-your-cache-always...