Cyber-Physical Systems (CPS) enable various new applications, including self-driving cars, drones, implantable medical devices, smart cars, distributed transportation systems, smart grids, and planetary robots. CPS sense the physical environment, process data in real-time, control the actuators, and guarantee the timing of the whole execution chain for ensuring safety. Since CPS are tightly coupled with the physical world, anomalies such as hardware failures and timing errors may cause significant damage to life and/or property. Common practices addressing those failures tend to use redundant hardware to overprovision resources. However, many CPS are targeted towards large-scale cost-sensitive markets that have stringent space and bill-of-material constraints that cannot afford overprovisioning. In this talk, I will discuss my research on a software-level redundancy approach providing many of the same benefits of using redundant hardware while maintaining lower costs and a higher level of flexibility. I will first present a framework called SAFER (System-level Architecture for Failure Evasion in Real-time applications) that incorporates configurable software mechanisms and policies to tolerate failures of critical CPS resources while meeting their timing constraints. I will then present how to guarantee timeliness by devising new computational models reflecting the timing nature of CPS. I will describe the schedulability analysis and runtime support for such models with and without resource failures. Finally, I will show how the proposed approaches make an autonomous vehicle dependable and conclude my talk with future directions towards large-scale CPS.