Or did you mean failing to resolve some internal service's hostname?
Yeah, something more like that where the internal service is somehow 'sharded' due to some overly complicated distributed database nonsense, and there's a DNS lookup that is failing. Of course that'd mean the DNS lookup wasn't cached, so you're taking that normal latency on every single hit, which would be terrible architecture. The curl-vs-wget performance isn't explained by that though (although that's a bit weird in and of itself, and might suggest that they had to allow that for some internal tool that they didn't want to punish).
> glibc defaults to 5 sec,
The timeout being close to 5 seconds is what made me wonder about it. Its just off though.