When fetching ip addresses for the host, we fetch all the network interfaces, and then ip addresses for each interface. The latter call is surprisingly expensive on unix, as it involves opening a netlink socket, sending a request for routing information, and receiving and parsing the response. If the host has a lot of network interfaces, this can eat surprising amounts of memory - I got in the order of 10 MB on a Kubernetes Node with 100 Pods. Seeelastic/elastic-agent#5835 (comment) for some heap profiles from elastic-agent.

Instead, get all the addresses in a single stdlib call. We don't actually care about which interface each ip address is attached to, we just want all of them.

I've tested this in the real world scenario discussed inelastic/elastic-agent#5835, not sure how to include a self-contained test in this repo.