Use Netlink to Fetch TCP states

[kustodian]

Currently tcpstat collector uses /proc/net/tcp and /proc/net/tcp6 to fetch TCP states, which can be (as the documentation states) slow on high traffic hosts. To resolve these performance problems it should use Netlink instead, because Netlink is a few times faster. After this is implemented, tcpstat collector should probably be turned on by default.

[SuperQ]

Seems like a good idea, we need to be careful as we've seen issues with netlink causing crashes due to CGO interactions in the wifi collector.

[iyashu]

Is anyone already working on this? @discordianfish @SuperQ

We need to enable tcp connection stats (including their states) from node-exporter for k8s cluster to track connection leaks & related fd usage. If any one haven't started working on it, I can start exploring netlink and ways to integrate it with tcpstat.

[SuperQ]

No, I don't think anyone is working on this. Thanks for volunteering.

One note. If possible, this should be implemented so that it replaces the collector completely. Keeping two ways of getting the same metrics is not a great idea. I guess it depends on if the transition can be done cleanly.

[ojagodzinski]

same thing for collecting conntrack stats on newer systems

• https://cateee.net/lkddb/web-lkddb/NF_CONNTRACK_PROCFS.html
• https://serverfault.com/a/1080121
• https://github.com/szuecs/netlink-conntrack-status