Traffic surges, like the Slashdot effect, occur when a web application is overloaded by a huge number of requests, potentially leading to unavailability. Unfortunately, such traffic variations are generally totally unplanned, of great amplitude, within a very short period, and a variable delay to return to a normal regime. In this report, we introduce PeakForecast as an elastic middleware solution to detect and absorb a traffic surge. In particular, PeakForecast can, from a trace of queries received in the last seconds, minutes or hours, to detect if the underlying system is facing a traffic surge or not, and then estimate the future traffic using a forecast model with an acceptable precision, thereby calculating the number of resources required to absorb the remaining traffic to come. We validate our solution by experimental results demonstrating that it can provide instantaneous elasticity of resources for traffic surges observed on the Japanese version of Wikipedia during the Fukushima Daiichi nuclear disaster in March 2011.
