Help debugging IPv6 latency on Ruby Windows

I am experiencing this issue on a Windows machine. Original issue filed with Windows Ruby installer: https://github.com/oneclick/rubyinstaller2/issues/264

It seems that the change from TcpSocket to Socket (https://github.com/ruby/net-http/pull/10) released in Ruby 3.1 is now causing timeout latency for me when accessing IPv6 sites using Net::HTTP. Socket.tcp waits 60 sec attempting connect to IPv6, times out, then immediately connects to IPv4. TCPSocket.open does not have this issue.

TCPSocket.open("google.com", 80, nil, nil) # fast

Socket.tcp("google.com", 80, nil, nil, connect_timeout: 60) # times out after 60 sec on IPv6, then falls back to IPv4

TCPSocket.open is fast on both Ruby 3.0 and 3.1, while Socket.tcp is slow on both Ruby 3.0 and 3.1. However, since Net::HTTP switched to Socket in 3.1 it has the end-result of making Net::HTTP slow overall.

The issue may be Windows-specific, since I do not see when using Linux via WSL Ubuntu on the same Windows machine. There's a lot more details in the rubyinstaller2 issue, however, I'm wondering if someone can provide some insight on how to debug this further.

Script to reproduce issue:

require 'uri'
require 'net/http'
require 'openssl'

module HTTPS_IPv4_IPv6
  class << self

    def run(uri_str)
      uri = URI(uri_str)

      puts uri_str

      t_st = Process.clock_gettime(Process::CLOCK_MONOTONIC)

      Net::HTTP.start(uri.host, uri.port, use_ssl: true, verify_mode: OpenSSL::SSL::VERIFY_PEER) do |http|
        req = Net::HTTP::Get.new uri.request_uri
        resp = http.request req
        case resp
        when Net::HTTPSuccess
          body = resp.body
          t_end = Process.clock_gettime(Process::CLOCK_MONOTONIC)
          puts "body bytesize #{body.bytesize}"
          puts format("response time %6.3f sec", t_end - t_st)
          puts http.instance_variable_get(:@socket).io.to_io.remote_address.inspect, ''
        when Net::HTTPRedirection
          puts "#{resp.class}\n   Redirect: #{resp['location']}"
          run resp['location']
        else
          puts resp.class
        end
      end
    end

  end
end

puts '', RUBY_DESCRIPTION, ''

HTTPS_IPv4_IPv6.run 'https://www.google.com'

HTTPS_IPv4_IPv6.run 'https://global.jd.com'

ruby 3.0.0p0 (2020-12-25 revision 95aff21468) [x64-mingw32]

https://www.google.com
body bytesize 15110
response time  2.390 sec
#<Addrinfo: 142.251.42.196:443 TCP>

https://global.jd.com  -- IPv4 only
body bytesize 20592
response time  0.633 sec
#<Addrinfo: 14.0.43.163:443 TCP>

Beginne mit dem bereitgestellten Reproduktionsskript unter Windows mit Ruby 3.0 und 3.1 und vergleiche Net::HTTP, Socket.tcp und TCPSocket.open für IPv6- und IPv4-Ziele. Verfolge den Verbindungsweg rund um die Änderung an Socket.tcp und vergleiche ihn mit dem verlinkten Bericht von rubyinstaller2; abgeschlossen ist die Untersuchung, wenn die Windows-spezifische Ursache oder die nächste Debugging-Grenze dokumentiert ist.