To ensure the high availability of modern online systems, effective maintenance is of critical importance. Today’s software maintenance techniques for online systems heavily rely on metrics, which are time series data that can describe the real-time state of a system from various perspectives. Typically, software engineers generate dashboards with metrics to aid software maintenance. Though several attempts have been devoted to metric analysis for automatic software maintenance, the primary step, i.e., dashboard generation, remains manual to a large extent. In this paper, we develop a metric recommendation service, which can automate the dashboard generation practice and greatly ease the burden in maintaining an online system. Specifically, we analyze the needs of two essential steps of online system maintenance, i.e., anomaly detection and fault diagnosis, and design metric recommendation mechanisms for them respectively. Graph learning techniques are employed in the automation of metric recommendation. Our experiments demonstrate that the proposed approach can achieve an F1-score of 0.912 in selecting metrics for anomaly detection, and an accuracy of 0.859 in retrieving metrics for faults diagnosis, which significantly outperforms the compared baselines.