There are several heuristic mechanisms in Lightstreamer governing the automatic choice between streaming and polling modes. Did you observe the two different behaviors under the same exact conditions (that is, same machine, same browser, same proxy, same firewall, etc.)? If this is the case, then an explanation could be found in the adaptive mechanism of the Lightstreamer client library. If the client CPU is saturated and cannot process the updates fast enough, it will switch to long polling mode to find the right pace. Did you observe any CPU peak during the state transition?