AI-driven routing pipeline in software-defined networks using DQL: a mini review

State-of-the-art data center networks are experiencing an increase in dynamic traffic. Even minor inefficiencies cause latency, congestion, and high costs. Software-defined networking (SDN) provides centralized programmability, but classical algorithms such as Dijkstra and Equal-Cost Multi-Path (ECMP) fall short because they cannot adapt in real time. To overcome this limitation, Reinforcement Learning (RL), particularly Q-learning, adds adaptability; however, scalability remains a challenge. DQL addresses this by using neural networks to approximate the Q-function, enabling SDN controllers to learn routing strategies directly from live network states. This Mini Review brings together recent DQL approaches for SDN. We examine architectures, algorithmic variants, and emulation environments (such as Mininet with Ryu). In addition, we introduce a structured taxonomy, with a practice-oriented synthesis of empirical trade-offs and deployment issues. The focus is on trade-offs, throughput, latency, and convergence. Reported studies show that DQL typically improves throughput by about 15-22 percent and reduces delays by roughly 10-12 percent compared with ECMP. These gains, however, come at the cost of longer training, inference delays, and scalability hurdles. Unlike prior surveys, this review makes three distinct contributions: a structured taxonomy, with a practice-oriented synthesis of empirical trade-offs and deployment issues. We also highlight emerging directions: federated learning, graph-based neural models, and explainable AI, which may help transition DQL from promising simulations to production-ready SDN solutions.