Distributed databases are fundamental to modern infrastructure, but face complex performance anomalies due to their multi-node architecture. Existing root cause analysis methods, particularly those based on causal inference, struggle with the scale and complexity of distributed environments. They suffer from high computational costs and exhibit low accuracy in diagnosing high-dimensional runtime monitoring metrics. To address these challenges, we propose DBRooter, a diagnostic framework for distributed databases. DBRooter introduces dependency-guided causal modeling for distributed metrics that efficiently constructs local causal graphs in parallel and integrates them with a node dependency graph derived from query plans. In addition, it employs a workload-aware anomaly metric ranking method using structural causal models to quantitatively assess causal effects and identify root causes. The experimental results show that DBRooter effectively diagnoses anomalies in individual queries, transactions, and entire workloads and outperforms the state-of-the-art baselines. Compared to the second-best method, AC@1, AC@3, and AC@5 are improved up to 10%, 15%, and 12%.