The problem isn't when the host is down, it is when nginx can't lookup the address for the host. If your DNS records are OK and your resolvers are working this isn't a problem, but if nginx (re)starts while you are having connectivity issues or if a DNS record that your config relies upon has been dropped it won't start any service until that one is edited.
Workarounds include using addresses not names (not practical at scale), running a local resolver that will hand out the last known address if it can't see the resolvers it normally forwards requests to or gets an NXDOMAIN response (extra hassle and technically breaks DNS so make sure other things don't use that resolver), or just ignoring the problem because it very rarely happens (though for some setups it is more likely than in others).
In my (current) uses of nginx it isn't a problem at all, though a concern springs to mind that might affect me later: if it only fails on startup, does that imply it isn't ever refreshing name->address mappings at all in normal operation so that if a server it is proxying for moves address that proxy config will start to fail until nginx is restarted? Any nginx experts want to comment so I can be lazy and not test or otherwise research this myself?
Huh. I don't think I've ever encountered this because I've only used nginx to server content or applications. I've always had it behind haproxy, which does have an option to allow haproxy to startup even if a backend host name doesn't resolve.
You describe an environment that is quite scaled up. I would assume that if there is scale, there is money. And a license to the commercial version is what, 2 salaried hours a year? It is easy to save time and money. I am sorry, but I cannot see this as a big problem.
We'd strayed off topic and were discussing an issue in nginx. Basic per-instance subscriptions for that are $2,500/yr each for small numbers of instances which is somewhat more than a couple of salaried hours.
Still not a large amount for a company with a profit making product, of course, but companies with profit making products are fewer and farther between than companies would like, and for a one-man-band or other small business trying to boot-strap something bigger, there isn't the same drop/ocean ratio.
Also, the implied problem (this still needs confirming/contradicting, so I might be barking up the wrong end of the stick) that it may not be updating name->address resolution regularly during normal operation (otherwise why does this only cause failure on start-up?) might not by fixed by the extra modules enabled by the paid licence.
Workarounds include using addresses not names (not practical at scale), running a local resolver that will hand out the last known address if it can't see the resolvers it normally forwards requests to or gets an NXDOMAIN response (extra hassle and technically breaks DNS so make sure other things don't use that resolver), or just ignoring the problem because it very rarely happens (though for some setups it is more likely than in others).
In my (current) uses of nginx it isn't a problem at all, though a concern springs to mind that might affect me later: if it only fails on startup, does that imply it isn't ever refreshing name->address mappings at all in normal operation so that if a server it is proxying for moves address that proxy config will start to fail until nginx is restarted? Any nginx experts want to comment so I can be lazy and not test or otherwise research this myself?