Data-driven functions for network operation and management are based upon AI/ML methods whose models are usually trained offline with measurement data collected through monitoring. Online learning provides an alternative with the prospect of shorter learning times and lower overhead, suitable for edge or other resource-constraint environments. We propose an approach to online learning that involves a cache of fixed size to store measurement samples and periodic re-computation of ML models. Key to this approach are sample selection algorithms that decide which samples are stored in the cache and which are evicted. We present and evaluate four sample selection algorithms, all of which are derived from well-studied algorithms, and we specifically argue that feature selection algorithms can be used for our purpose. We perform an extensive evaluation of these algorithms for the task of performance prediction using data from an in-house testbed. We find that one of them (RR-SS) leads to models that achieve a prediction accuracy close to that obtained through offline learning, but at a much lower cost.