Packet loss can seriously degrade the performance of TCP flows, especially that of small latency-sensitive flows. Managing packet loss efficiently is critical for maintaining good network performance. However, the loss recovery mechanisms of TCP by nature incur high latency due to (i) long loss detection time, (ii) inaccurate loss notification, and (iii) one-size-fits-all loss reaction. It is desirable and essential to design a fast, accurate, and fine-grained packet loss management mechanism to reduce the latency introduced by packet loss. We present DeepSight, an in-network packet loss management mechanism for TCP applications in cloud networks. DeepSight addresses the problems of TCP by exploring the packet processing capabilities of existing programmable switches. It adopts a switch-end collaboration design which offloads the loss detection and recovery from the end to the switch data plane. We implement DeepSight on both the software simulator and hardware testbed. Extensive experiments show that DeepSight achieves faster loss detection time and smaller latency when managing packet loss compared with TCP’s loss recovery mechanisms.