See ticket:1948#comment:1 : the problem is solved for shadow servers by skipping "polling" refresh frames until the packet send queue is (almost) empty. For desktop servers, we're dealing with damage events that aren't tagged as "polling" refreshes, and can't be dropped. We'll need to accumulate them and process them later when congestion occurs.
this ticket has been moved to: https://github.com/Xpra-org/xpra/issues/1949