In distributed key-value stores, scheduling the service sequence of key-value operations can effectively reduce the completion time of user requests and improve user experience. In response to the difficulty of uniformly scheduling requests from different clients, this paper proposes the adaptive centralized key-value scheduler (AC-KVS). Specifically, AC-KVS aggregates requests from different clients in the programmable switches and schedules the waiting requests according to the load of different servers. Meanwhile, packets directly sent to the servers are scheduled by the shortest bottleneck keyset first method. In this way, the key-value operations of the same request will complete simultaneously as far as possible, and thus the average completion time of requests will be reduced. Simulation results show that AC-KVS reduces the average completion time by 9.3% compared to FIFO and 10% compared to SRPT with the light-tailed traffic, respectively. Under the heavy-tailed traffic, AC-KVS reduces the tail completion time by 13% compared to the state-of-the-art algorithm SRPT.