Techniques for scalable performance monitoring using dynamic flow sampling are described. According to one approach, a method comprises intercepting, at an intermediary network device, one or more packets traveling between a source device and a destination device; identifying, at the intermediary network device, a traffic flow based on the one or more packets; determining, at the intermediary network device, whether to collect one or more metrics from the traffic flow based on one or more performance factors of the intermediary network device; in response to a determination to collect the one or more metrics from the traffic flow, the intermediary network collecting the one or more metrics from subsequently intercepted packets belonging to the traffic flow; wherein the method is performed by one or more computing devices.