I spoke about this exact thing at a conference (HPTS’19) a while back. This can work, but introduces modal behaviors into systems that make reasoning about availability very difficult and tends to cause meta stable behaviors and long outages.
The feedback loop is replicas slow -> traffic increases to primary -> primary slows -> replicas slow, etc. The only way out of this loop is to shed traffic.
Comments
I spoke about this exact thing at a conference (HPTS’19) a while back. This can work, but introduces modal behaviors into systems that make reasoning about availability very difficult and tends to cause meta stable behaviors and long outages.
The feedback loop is replicas slow -> traffic increases to primary -> primary slows -> replicas slow, etc. The only way out of this loop is to shed traffic.