A benchmark of Ruby refinements activation cost: `using` slows down with the number of classes loaded in the process on Ruby < 4.0 (bug #21201, fixed by the refinement-cache table)

refinements_activation_bench.rb

# frozen_string_literal: true

# Activating a refinement (`using`) on Ruby < 4.0 triggers a full object-space
# scan to invalidate method caches, so its cost grows with the number of classes
# already loaded in the process.
# Ref: https://bugs.ruby-lang.org/issues/21201
#
# This benchmark generates N classes, then loads a refinement-heavy gem
# (`anyway_config`). Compare a small and a large N, with refinements on and off:
#
#   N=4000 ruby refinements_activation_bench.rb
#   N=30000 ruby refinements_activation_bench.rb
#
#   NOUSING=1 N=4000 ruby refinements_activation_bench.rb
#   NOUSING=1 N=30000 ruby refinements_activation_bench.rb
#
# 4,000 and 30,000 mirror a real Rails monolith's class count
# right after `rails/all` and after the full gem list, respectively.

require "bundler/inline"

gemfile do
  source "https://rubygems.org"
  gem "benchmark"
  gem "anyway_config", "~> 2.8", require: false
end

require "benchmark"

N = ENV.fetch("N", "5").to_i

N.times do |i|
  class_name = "Class#{i + 1}"
  eval <<-RUBY
    class #{class_name}
      def initialize
        @name = '#{class_name}'
      end

      def greet
        puts "Hello from \#{@name}!"
      end
    end
  RUBY
end

if ENV["NOUSING"] == "1"
  class Module
    def using(*) = nil
    def refine(*, &) = nil
  end
end

res = Benchmark.measure do
  require "anyway_config"
end

puts res.real

results.md

Ruby 4.0.6 vs. 3.3.8

• 4.0.6 is 40-67% faster overall across all test sizes.
• The gap is most significant at N=100,000 (67.3% faster) and smallest at N=5,000 (39.7% faster).

using overhead (refinements on vs. NOUSING=1)

• At N=5,000: 4.0.6 barely notices (+4.3%), while 3.3.8 already pays +57.8%.
• At N=30,000: 4.0.6 stays efficient (+11.0%), while 3.3.8 more than doubles the load time (+115.6%).
• At N=100,000: both versions pay noticeably, but 3.3.8 is dramatically worse (+16.9% vs. +228.0%).

Environment

Measured on x86_64-linux, best of five runs per configuration:

• ruby 3.3.8 (2025-04-09 revision b200bad6cd)
• ruby 4.0.6 (2026-07-14 revision 03b6d3f889) +PRISM