There is a configurable global connection limit which is useful on platforms (such as macOS) which have abnormally low maximum file descriptor limits.
The UDP NAT binds a socket per internal source address, and expires it after it has been idle for 60s.
Unfortunately a UDP protocol which sends UDP from random source ports and which doesn't expect replies (such as that in envoy, as used by kubeflow) can quickly exhaust the global limit which then breaks other things, such as port forwards.
This PR adds a new limit on the total number of active UDP NAT table entries and sockets, and expires the oldest when the new limit is hit. This should keep the network usable even when using one of these chatty UDP protocols.
There is a configurable global connection limit which is useful on platforms (such as macOS) which have abnormally low maximum file descriptor limits.
The UDP NAT binds a socket per internal source address, and expires it after it has been idle for 60s.
Unfortunately a UDP protocol which sends UDP from random source ports and which doesn't expect replies (such as that in
envoy, as used bykubeflow) can quickly exhaust the global limit which then breaks other things, such as port forwards.This PR adds a new limit on the total number of active UDP NAT table entries and sockets, and expires the oldest when the new limit is hit. This should keep the network usable even when using one of these chatty UDP protocols.