When designing the maintenance of multi-component aircraft systems, we consider parameters such as safety margins (used when component replacements are scheduled), and reliability thresholds (used to define data-driven Remaining-Useful-Life prognostics of components). We propose Gaussian process learning and novel adaptive sampling techniques to efficiently optimize these design parameters. We illustrate our approach for aircraft landing gear bakes. Data-driven, Remaining-Useful-Life prognostics for brakes are obtained using a Bayesian linear regression. Pareto optimal safety margins for scheduling brake replacements are identified, together with Pareto optimal reliability thresholds for prognostics.