In order for a software system to self-adapt, it often needs to be aware of its behavior. A typical way of achieving this is by means of the runtime collection of execution traces, which requires the interception of the execution of, e.g., methods and record information about them. Although this is simple in theory, in practice, it can be very costly, because it might have a significant impact on the application performance or require huge amounts of memory or storage. This becomes a significant issue mainly in real-time applications, which are time-sensitive and must often meet deadlines in resource-constrained environments. We thus in this paper propose a two-phase tracing framework to cope with the monitoring overhead. In its first phase, the application is monitored in a lightweight fashion providing enough information for the second phase, which identifies an adaptive configuration and samples traces according to the current configuration. The adaptive configuration is determined by a set of criteria, specified through a proposed domain-specific language. We empirically evaluate our framework by instantiating it as a reduced-overhead monitoring solution, integrated into an existing automated application-level caching approach. We demonstrate that our solution reduces the overhead caused by monitoring, without compromising the performance improvements provided by the caching approach.