Since a node can process any number of jobs concurrently, there is a risk that it can be overwhelmed, and its performance may degrade suddenly or it could even crash, for instance because of an out of memory condition.

We propose to implement a pluggable mechanism in the node to alert the driver that it cannot accept any more jobs when a given condition is true. The same mechanism would send another alert when the condition is no lnger true, so it can resume taking additional jobs.