Due to their thermal inertia, buildings equipped with electric heating and cooling systems can help to stabilize the electricity grid by shifting their load in time, and can thus facilitate energy flexible urban energy systems with the right control system in place. Because of minimum capacity requirements, they can often only participate in demand response schemes, such as secondary frequency reserves through aggregation. Such an aggregation could also take the form of entire district heating and cooling systems with connected buildings that are supplied by large-scale heat pumps and chillers. However, there is a lack of studies investigating the control of such configurations, both in simulation and in application. We present a two-level control scheme based on robust Model Predictive Control with affine policies to offer frequency reserves with a district system, where we exploit the thermal inertia of buffer storage tanks and a subset of the connected buildings. We leverage data-driven model generation methods to overcome the bottleneck of physics-based building modeling. In a numerical case study based on one-year historical data of a real system, we compare the approach to a situation where only the buffer storage is used for flexibility and demonstrate that the reserves offered increase substantially if the inertia of a subset of the connected buildings is also exploited. Furthermore, we validate the control approach in a first-of-its-kind experiment on the actual system, where we show that reserves can be offered by the district system without compromising the comfort in the connected buildings.