The head node has to do all the processing of what changes to make, for each server it runs against. Pull based systems have the destination server do that, ma

The head node has to do all the processing of what changes to make, for each server it runs against. Pull based systems have the destination server do that, making them scale horizontally natively. Presuming there's no need for orchestration between nodes (which pull-based configs make very hard/impossible), pull-based systems parallelize perfectly where push-based systems cannot parallelize. SSH is also very slow as far as these things go. At some point the head node just gets saturated on doing SSH encryption or the NIC.