As network scale and business concurrency continue to grow and data centers expand across multiple regions, cross-domain data processing has become the norm. In current distributed databases, leader-based consensus protocols face the challenge of high latency when processing cross-domain requests. To address this, we present P-Raft, a Raft-based protocol tailored for cross-domain sites that shortens the cross-domain commit’s critical path and employs machine-learning–based, pro-active leader migration to adapt to network topology and load shifts, thereby improving performance and reducing global latency. Under representative read/write-balanced workloads, P-Raft reduces average latency by 75.19% and 72.61% relative to Raft and EPaxos, respectively; compared with the state-of-the-art leader-management approach GeoLM, P-Raft reduces average latency by 63.89%.