Feature #14430: net/http: use Socket.tcp with connect_timeout, instead of TCPSocket.open wrapped in Timeout.timeout

• Status: Open
• Assignee: -
• Target version: -

[carl.hoerberg (Carl Hörberg)]

Instead of using TCPSocket.open, wrapped in Timeout.timeout, that will create a temporary thread. By using Socket.tcp with the connect_timeout argument for open_timeout the connection opening is much more efficient as the kernels timeout is used instead.

PR at: https://github.com/ruby/ruby/pull/1806

[normalperson (Eric Wong)]

[e-mail] wrote:

Instead of using TCPSocket.open, wrapped in Timeout.timeout,
that will create a temporary thread. By using Socket.tcp with
the connect_timeout argument for open_timeout the connection
opening is much more efficient as the kernels timeout is used
instead.

Unfortunately, we can't do this, yet. The Addrinfo calls use
getaddrinfo(3) which doesn't support timeout natively.

My goals for later this year is:

1. implement Timeout in the VM itself so it doesn't need to create
   a temporary thread.

2. update resolv-replace.rb to cover Addrinfo cases, including
   nscd cache lookup for glibc compatibility.

[shevegen (Robert A. Heiler)]

implement Timeout in the VM itself so it doesn't need
to create a temporary thread.

There be dragons hiding in the VM.

\o/

[carl.hoerberg (Carl Hörberg)]

normalperson (Eric Wong) wrote:

Unfortunately, we can't do this, yet. The Addrinfo calls use
getaddrinfo(3) which doesn't support timeout natively.

Right, good catch.

2. update resolv-replace.rb to cover Addrinfo cases, including
   nscd cache lookup for glibc compatibility.

Looking into this i realize that /etc/nsswitch.conf has be to taken into account first. And then maybe nscd, but very few distros (if any?) install it by default anymore, and it generally seems to have a very bad rep "around the internets".

But which one do you suggest? Go down the route of Socket.tcp, but make the DNS resolving interruptible/async (without doing it in a separate threads) or to implement thread-less timeout in the VM and continue to rely on getinfoaddr from glibc?

[Glass_saga (Masaki Matsushita)]

• Related to Feature #15553: Addrinfo.getaddrinfo supports timeout added

[byroot (Jean Boussier)]

We've been recently bitten by this.

Does the merge of [#15553] makes https://github.com/ruby/ruby/pull/1806 functional?

[carl.hoerberg (Carl Hörberg)]

The PR is updated, based on master and adds Net::HTTP#resolve_timeout, which delegates to Socket.tcp resolve_timeout, based on https://bugs.ruby-lang.org/issues/16381 but reraiseing Errno::ETIMEOUT as Net::OpenTimeout