We use latches to sequence concurrent, conflicting requests on the leaseholder. A write request will acquire write latches, which will block any read requests with higher timestamps than the write. That's because these reads need to see the MVCC value written by the write. Latches won't be released until the write has been committed to the leaseholder's log and then applied to its state machine. So, in the example in the thread, all future read requests will wait on these latches, which means we won't serve a stale read.