A network is congested when offered traffic persistently exceeds the capacity of some shared resource, usually an output link or queue.

As queues build, latency rises. When buffers overflow, packets are dropped. If senders respond to loss by blindly injecting still more traffic, retransmissions can consume capacity without delivering useful new data. In extreme cases, useful throughput falls even while the network carries enormous traffic: congestion collapse.

Congestion is different from receiver overload. A receiver may have plenty of buffer space while an intermediate bottleneck link is saturated.

A transport therefore needs feedback about path conditions and a rule for adapting its sending rate. Loss, delay and explicit congestion signals can all provide information.

The central principle is cooperative: endpoints sharing a network must react to congestion rather than treating bandwidth as an unlimited private resource.