Performance overhead of Ruby TracePoint API

benchmarks.rb

######
# Simplest case - one method that does nothing, with and without TP
######

require 'benchmark/ips'

def noop; end

Benchmark.ips { |x| x.report('basic', 'noop') }; 1

trace = TracePoint.new(:call) { |tp| }
trace.enable do 
  Benchmark.ips { |x| x.report('trace', 'noop') }; 1
end

# RESULTS
# => basic     31.860M (± 0.5%) i/s -    161.220M in   5.060445s
# => trace     12.411M (± 0.6%) i/s -     62.084M in   5.002639s

######
# Comparison of various ways to intercept calls, 
# with a more realistic method chain
######

# Playing with the numbers affects results.
# More method calls = worse relative tracepoint performance,
# but only up to a value of 10 or so.
METHOD_CALLS = 10
# More allocations or work per method = better relative tracepoint performance
ALLOCATIONS_PER_METHOD = 5

module M
  (1..METHOD_CALLS).each do |n|
    eval <<~RUBY
      def method#{n}
        Array.new(ALLOCATIONS_PER_METHOD) { Object.new }
        #{"method#{n + 1}" if n < METHOD_CALLS}
      end
    RUBY
  end
end

object = Object.new.extend(M)

object_with_prepend = Object.new.extend(M)
object_with_prepend.singleton_class.prepend Module.new {
  def method1
    $callers << caller_locations(1, 1)
    super
  end
}

object_with_alias = Object.new.extend(M)
object_with_alias.instance_eval do
  alias original_method1 method1

  def method1
    $callers << caller_locations(1, 1)
    original_method1
  end
end

object_with_rebind = Object.new.extend(M)
original_method1 = M.instance_method(:method1)
object_with_rebind.define_singleton_method(:method1) do
  $callers << caller_locations(1, 1)
  original_method1.bind_call(self)
end

trace = TracePoint.new(:call) { |tp| $callers << tp }

Benchmark.ips do |x|
  x.report('default') do
    object.method1
  end

  x.report('prepended') do
    $callers = []
    object_with_prepend.method1
  end

  x.report('aliased') do
    $callers = []
    object_with_alias.method1
  end

  x.report('rebound') do
    $callers = []
    object_with_rebind.method1
  end

  x.report('traced') do
    $callers = []
    trace.enable
    object.method1
    trace.disable
  end

  x.compare!
end; 1

# RESULTS
#           default:   214501.5 i/s
#           aliased:   200091.9 i/s - 1.07x  slower
#         prepended:   200087.1 i/s - 1.07x  slower
#           rebound:   196798.0 i/s - 1.09x  slower
#            traced:   174653.6 i/s - 1.23x  slower