As a nascent technique, Mobile Edge Computing (MEC) is mushrooming with a broad application prospect. By transferring abundant computing and storage resources from cloud to edge servers close to users, it allows services to be hosted on edge servers, which greatly reduces service latency. However, due to environmental dynamics, the edge services' reliability fluctuates in real-time. When the reliability of an edge service degrades severely, it may be suffering an anomaly, which can seriously impact its real-time performance and users' quality of experience. To ensure the real-time performance of edge services, this paper presents CDCrest, a concept drift-based checkpoint restart approach for edge service rejuvenation. CDCrest employs L-Detection, a concept drift-based approach to detect edge services' runtime reliability anomalies. L-Detection leverages a sliding window and Locality Sensitive Hashing (LSH)-based sampling to extract features from an edge service's real-time and historic reliability data streams. Then, it calculates the real-time Distribution Change Degree (DCD) based on Jensen-Shannon (JS) divergence to infer whether the edge service is suffering a reliability anomaly. Once an anomaly for an edge service is identified, CDCrest employs a checkpoint restart mechanism to ensure the rapid rejuvenation of the edge service. Extensive experiments conducted based on a popular real-world dataset demonstrate the effectiveness and efficiency of CDCrest against the state-of-the-art approaches.