Reads directly after writes to a table are stale when using read replicas if the update takes too long

We are experiencing an issue when an import of the big courses are failing with error/restore_controller_dbops_nonexisting error when we are using read replica set up.

To replicate it you'd need a decent course that would take quite a while to restore. And import it to an empty course. 

It works ok if we exclude backup_controllers table from hitting the read replica database.