I used feeltty to measure the interactivity of a Coder SSH session vs. SSH directly into the box.

Coder is substantially slower in our coder-ts (Docker-based) template. I get baseline latencies in my AWS workspace.

$ feeltty ssh coder.poll
connect              2990.36792ms
input sample size    32
input mean           118.698ms
input stddev         274.132ms

$ feeltty ssh <redacted>
connect              1199.80383ms
input sample size    32
input mean           55.526ms
input stddev         43.138ms

Oddly, I see equal throughputs as tested by ssh <host> -- cat /dev/urandom | pv > /dev/null.

Perhaps #3505 will fix this. (cc @kylecarbs, @bpmct)