<p>Task scheduling in Fog computing is a challenging problem due to the heterogeneous, resource-constrained, and distributed nature of Fog Computing Nodes (FCNs), compounded by the growing resource demands of diverse latency-sensitive Internet of Things (IoT) applications. This article presents DUAL, an efficient task scheduling approach designed to address this problem in Fog computing environments, providing an end-to-end solution for latency-sensitive applications. It operates at two levels: application scheduling and tuple scheduling. At the application scheduling level, DUAL employs Deep Reinforcement Learning (DRL) to solve the application scheduling problem in a dynamic and heterogeneous Fog computing environment. While at the tuple scheduling level, it leverages Lyapunov Drift Plus-Penalty (LDPP) function. The proposed approach enhances efficiency by improving resource utilization while fulfilling the Quality of Service (QoS) requirements of end users. DUAL is implemented in the iFogSim simulator for validation, and its performance is evaluated using key QoS metrics, including resource utilization, latency, throughput, network usage, and task drop rate. The results are then compared with benchmark approaches such as DRL-based, heuristic-based, hybrid optimization-based, and energy-aware based scheduling. The findings demonstrate that DUAL offers better resource utilization and a significant improvement in QoS. It achieves a 15.2% average relative increase in resource utilization, a 39.3% average relative reduction in latency, an 11.9% average relative increase in throughput, a 16.2% average relative reduction in network usage, and a 54% average relative reduction in task drop rate.</p>