The fast and intelligent reconfigurability of recon-figurable add-drop multiplexers (ROADMs) in metropolitan area networks (MANs) has gained much attention recently due to technical advancements in artificial intelligence and fast optical switching. However, it is challenging to realize submillisecond-level automatic reconfiguration for MANs under fast time-varying traffic pattern, because of the latency of the wavelength scheduling and traffic cognition lag. On the one hand, the latency for wavelength scheduling takes tens of millisecond for the most-used ROADMs; On the other hand, the lag involved in the traffic cognition weakens the advantage of fast wavelength scheduling. To view of these problems, this article proposes a fast-reconfigurable MAN architecture with closed control plane targeted to the submillisecond-level reconfiguration. The proposed architecture reduces the reconfigurable latency for both the data plane and the control plane. Furthermore, we design a latency estimator based on graph neural network (GNN) for congestion awareness, and develop a fast-reconfigurable ROADM based on semiconductor optical amplifier. We evaluate the estimator and proposed architecture under various scenarios. The results show that the GNN-based estimator can achieve high precision in the latency estimation.