Selecting appropriate attributes when configuring automated watchdogs (monitors) is a recurring challenge in operating large cloud services. These decisions are often made manually based on experience, leading to missed incidents or excessive alert noise. Thus, there is a need to automate the monitor configuration setting in a structured manner. In this paper, we present a deployable, data-driven system that recommends monitoring attributes (dimensions) by learning from the configurations and relationships between historical monitors in production. Our approach models monitor entities as a heterogeneous interaction graph and leverages structural and textual information available in real monitoring systems. The system is designed for sparse, large-scale network and integrates into an existing monitor creation workflow. Experiments on production data show significant improvements over prior approaches, and user studies with service owners highlight the usefulness of the recommendations. Beyond the modeling approach, we share lessons learned from deploying the system in practice, including challenges related to scalability, explainability, and user adoption. These insights may inform similar efforts to apply learning-based recommendation in operational software engineering contexts.