As long as the client and the server don't close the TCP session, the inner channel should stay up.
The TCP connection times out.
The inner channel doesn't support fragmentation. If there is a gateway in the path which doesn't support for the MTU size of the in-path interface, we expect to see an ICMP Fragmentation Required packet with the right MTU size. If that packet doesn't get delivered, the TCP session will time out.
The TCP session gets reset.
A firewall in between the CSH and SSH has noted that the final ACK of the setup of the original TCP session didn't arrive. After a timeout the firewall will send a TCP RST to the client and server, intercepted by the CSH and SSH which will terminate the TCP session. This will affect long-lived TCP sessions most.
With the Full Transparency or Port Transparency WAN visibility, an IPS/IDS will find out that the protocol spoken on the inner channel isn't the protocol expected and it will terminate the TCP session.
The traffic must be suitable for optimization.
Bad data reduction.
Encrypted traffic. If it is SSL encrypted it could be optimized with SSL pre-optimization. If anything else, no luck. Check the documentation of the application to see if this can be disabled.
Compressed traffic. Check the documentation of the application if this can be disabled.
Bad traffic scenario.
If the traffic between the client and server is not a stream of lots of traffic in burst but more a ping-pong, the neural framing algorithm will cause extra delay here.
If the network isn't happy, the optimized traffic won't perform well neither.
Speed / duplex mismatch.
Congestion.
QoS not enabled.
QoS bandwidth misconfigured.
Large Fat Network.
WAN link doesn't get filled. Use HS-TCP or MX-TCP.
WAN bandwidth admission control.