Make pg more Ruby-3.0 fiber scheduler friendly

[larskanis]

This was raised by @ioquatix in ffi/ffi#797 (comment) :

In order to support the upcoming Ruby 3.0 fiber scheduler, you need to do only one thing:

When you would call a blocking function that does I/O, instead call IO#wait_readable or IO#wait_writable respectively.

If you want to support Ruby 2.x, then allow the user to inject a custom IO class, which responds to .new and implements #wait_readable and #wait_writable.

You can check how to do it here:

https://github.com/socketry/db-postgres/blob/6fbdd03dd6b04c57c5b7995b60e3f626c674e609/lib/db/postgres/native/connection.rb#L106-L139

We will be augmenting the C interface with some improvements for invoking IO#wait_readable/wait_writable but it hasn't landed yet.

Feature description is here: https://bugs.ruby-lang.org/issues/16786

[ioquatix]

Awesome, let me know if I can provide any further help :)

cc @tenderlove

[jasl]

I just wrote a PoC

class PG::Connection
  module FiberScheduler
    [
      :exec, :query, :exec_params, :prepare,
      :exec_prepared, :describe_portal, :describe_prepared
    ].each do |method|
      module_eval <<-CODE, __FILE__, __LINE__ + 1
        def #{method}
          Fiber.scheduler.io_wait(socket_io, IO::WRITABLE, 5)
          super
          Fiber.scheduler.io_wait(socket_obj, IO::READABLE, 5)
        end
      CODE
    end
  end
  include FiberScheduler
end

then do set evt as Fiber scheduler in spec_helpers

It's dirty, but pass all tests (except connection_spec and connection_sync_spec tests)

Any directive or guidance?

A question, I'm stucking on adding evt to Gemfile but failed to require it for 10 minutes, then I found I need to declare it in Rakefile, any historical reason for using Hoe not use standard Gemfile directly?

[jasl]

@larskanis @ioquatix could you take a look PR #364 ?

I'm done a PR:

• Change behavior of PG::Connection.async_api = false
  • It will raise error if Fiber Scheduler set, because the scheduler leading blocking IO will stucking,
• Update rake spec, test both with and without fiber on Ruby 3.0 (ruby-head)

Todo:

• OS thread support test cases disabled for now, they will hanging, I don't know how to fix

Help needed:

• I don't know my work is right and covered all cases, although test cases pass

[ioquatix]

If the scheduler is set you should use non-blocking event driven I/O. There shouldn't be a need to add PG::Connection.use_fiber= or any other "configuration".

[jasl]

If the scheduler is set you should use non-blocking event driven I/O. There shouldn't be a need to add PG::Connection.use_fiber= or any other "configuration".

You're right, I redo my work, only 2 test cases I don't know how to fix https://github.com/ged/ruby-pg/blob/master/spec/pg/connection_spec.rb#L1670-L1695

[jasl]

One pain point I've met is somewhere hanging when fiber scheduler enabled, but I don't know where.
Any sugguest about debugging this situation?

[ioquatix]

Fiber#backtrace may help. If you use async you can print out all the backtraces using Reactor#print_hierarchy.