This was closed by #10, correct?

yes it was. I've actually got equivalent pull requests at ruby/net-smtp#21, ruby/net-pop#3, and ruby/net-ftp#5, @deivid-rodriguez any chance I could get you to give those a quick code review? That seemed to be the impetus behind getting #10 accepted...

#10 Was reverted in #74, it re-opening this issue as it would still be nice to get rid of use of the problematic Timeout.timeout for the reasons stated above

Based on @hsbt's comment here, it sounds like we were waiting for Addrinfo to support a getaddrinfo_a method. But Socket.tcp does already have a resolv_timeout option. Is it not sufficient for this use? It's the timeout being passed to Addrinfo.getaddrinfo.

In the context of this use case, is there a reason continuing to use Addrinfo.getaddrinfo with a timeout isn't sufficient?

As #74 says, resolv_timeout has no effect and so is not usable.

Just to reiterate though, Timeout.timeout has no effect on whatsoever on DNS lookups, so I'm not sure what we gained by going back to the previous implementation. Either way we can't place a timeout on DNS lookups and have to default to the operating system's timeout. See https://bugs.ruby-lang.org/issues/17528. This is the case whether we use Socket.tcp, and its still the case now that we've reverted the change:

#set the nameserver to an address that doesnt exist, then run the following command:
bin $ time ./ruby -e 'require "net/http"; c = Net::HTTP.new("www.ffefecfceaefefxx.com"); c.open_timeout = 2; response = c.request_get("/"); puts response.code'
/Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1271:in `initialize': Failed to open TCP connection to www.ffefecfceaefefxx.com:80 (execution expired) (Net::OpenTimeout)
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1271:in `open'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1271:in `block in connect'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/timeout.rb:189:in `block in timeout'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/timeout.rb:196:in `timeout'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1269:in `connect'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1248:in `do_start'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1237:in `start'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1817:in `request'
	from /Users/mohamed/.rvm/rubies/ruby-3.2.2/lib/ruby/3.2.0/net/http.rb:1727:in `request_get'
	from -e:1:in `<main>'

real	0m30.106s
user	0m10.066s
sys	0m41.866s

If you use resolv-replace (the only way to interrupt DNS lookup currently AFAIK) Timeout.timeout will work. So that's a clear advantage.

Also see https://bugs.ruby-lang.org/issues/19430, that would solve it but it's not implemented yet.

I have been investigating around this issue for improving net/http's Ractor support (context: https://bugs.ruby-lang.org/issues/21309). I'd like to see Timeout.timeout removed since it raises in non-main Ractors, rendering net/http unusable unless open_timeout: nil is specified.

Since the time this ticket was opened, some things have changed and some things have not.

• Since Ruby 3.3, Addrinfo.getaddrinfo is interruptible (https://bugs.ruby-lang.org/issues/19965).
  • This means that the open_timeout timeout actually covers DNS resolution too.
  • This does not mean that getaddrinfo has a proper way for timeouting (without Timeout.timeout) https://bugs.ruby-lang.org/issues/19430 .
• Since Ruby 3.4, Net::HTTP benefits from the Happy Eyeballs v2 (RFC8305) implementation in CRuby, since it uses TCPSocket under the hoods.
• As a pleasant side-effect, TCPSocket.open's resolv_timeout now has actual effect, even in environments without resolv-replace. (@shioimm please correct me here if I am wrong)
  • This can be observed by ruby -v -rsocket -e "p TCPSocket.open('example.com', 80, resolv_timeout: 0.00001)".
    • On Ruby 3.4, this should raise an Errno::ETIMEDOUT on most network connections, but not when RUBY_TCP_NO_FAST_FALLBACK=1.
• Given that, it is still not possible to mimic the current Timeout.timeout behavior in Net::HTTP.new using TCPSocket's timeouts. Net::HTTP's timeout operates on the entire time TCPSocket.open consumed, which is roughly the sum of DNS resolution time and connect(2) time = resolv_timeout + connect_timeout.

I have been thinking of ideas to replace Timeout.timeout with a more resource-friendly and Ractor-compatible timeout implementation.

• (1) Add an open_timeout (resolv_timeout + connect_timeout) to TCPSocket.open.
  • Very clean from net/http's perspective, allowing direct mapping from Net::HTTP.new's open_timeout.
  • Could be a bit complex to implement, and using the new option would make new versions of net/http incompatible with Ruby <= 3.4.
• (2a) Replace TCPSocket.open(dns_name) with Addrinfo.getaddrinfo(dns_name) + TCPSocket.open(ip, connect_timeout:)
• (2b) Or use IO.select with Socket#connect_nonblock.
  • Even if this was implementable, net/http will lose benefits from HEv2, which is very unideal.
• (3) Use a Ractor-compatible Timeout and keep the code mostly as-is.
  • Although this won't solve the original issue, it is OK from the Ractor-compatibility perspective.
  • Also, I hear there are plans for making the timeout library compatible with Ractors.

I'm under the impression that option (1) is the cleanest but not straightforward to implement. I'll try to learn if (3) is doable.
Please let me know if there are other ways to accomplish this.