datadog 2.0.0.beta1

data/lib/datadog/profiling/component.rb

@@ -0,0 +1,418 @@
+# frozen_string_literal: true
+
+module Datadog
+  module Profiling
+    # Responsible for wiring up the Profiler for execution
+    module Component
+      # Passing in a `nil` tracer is supported and will disable the following profiling features:
+      # * Code Hotspots panel in the trace viewer, as well as scoping a profile down to a span
+      # * Endpoint aggregation in the profiler UX, including normalization (resource per endpoint call)
+      def self.build_profiler_component(settings:, agent_settings:, optional_tracer:) # rubocop:disable Metrics/MethodLength
+        return [nil, { profiling_enabled: false }] unless settings.profiling.enabled
+
+        # Workaround for weird dependency direction: the Core::Configuration::Components class currently has a
+        # dependency on individual products, in this case the Profiler.
+        # (Note "currently": in the future we want to change this so core classes don't depend on specific products)
+        #
+        # If the current file included a `require 'datadog/profiler'` at its beginning, we would generate circular
+        # requires when used from profiling:
+        #
+        # datadog/profiling
+        #     └─requires─> datadog/core
+        #                      └─requires─> datadog/core/configuration/components
+        #                                       └─requires─> datadog/profiling       # Loop!
+        #
+        # ...thus in #1998 we removed such a require.
+        #
+        # On the other hand, if datadog/core is loaded by a different product and no general `require 'datadog'` is
+        # done, then profiling may not be loaded, and thus to avoid this issue we do a require here (which is a
+        # no-op if profiling is already loaded).
+        require_relative '../profiling'
+
+        return [nil, { profiling_enabled: false }] unless Profiling.supported?
+
+        # Activate forking extensions
+        Profiling::Tasks::Setup.new.run
+
+        # NOTE: Please update the Initialization section of ProfilingDevelopment.md with any changes to this method
+
+        no_signals_workaround_enabled = no_signals_workaround_enabled?(settings)
+        timeline_enabled = settings.profiling.advanced.timeline_enabled
+        allocation_profiling_enabled = enable_allocation_profiling?(settings)
+        heap_sample_every = get_heap_sample_every(settings)
+        heap_profiling_enabled = enable_heap_profiling?(settings, allocation_profiling_enabled, heap_sample_every)
+        heap_size_profiling_enabled = enable_heap_size_profiling?(settings, heap_profiling_enabled)
+
+        overhead_target_percentage = valid_overhead_target(settings.profiling.advanced.overhead_target_percentage)
+        upload_period_seconds = [60, settings.profiling.advanced.upload_period_seconds].max
+
+        recorder = Datadog::Profiling::StackRecorder.new(
+          cpu_time_enabled: RUBY_PLATFORM.include?('linux'), # Only supported on Linux currently
+          alloc_samples_enabled: allocation_profiling_enabled,
+          heap_samples_enabled: heap_profiling_enabled,
+          heap_size_enabled: heap_size_profiling_enabled,
+          heap_sample_every: heap_sample_every,
+          timeline_enabled: timeline_enabled,
+        )
+        thread_context_collector = build_thread_context_collector(settings, recorder, optional_tracer, timeline_enabled)
+        worker = Datadog::Profiling::Collectors::CpuAndWallTimeWorker.new(
+          gc_profiling_enabled: enable_gc_profiling?(settings),
+          no_signals_workaround_enabled: no_signals_workaround_enabled,
+          thread_context_collector: thread_context_collector,
+          dynamic_sampling_rate_overhead_target_percentage: overhead_target_percentage,
+          allocation_profiling_enabled: allocation_profiling_enabled,
+        )
+
+        internal_metadata = {
+          no_signals_workaround_enabled: no_signals_workaround_enabled,
+          timeline_enabled: timeline_enabled,
+          heap_sample_every: heap_sample_every,
+        }.freeze
+
+        exporter = build_profiler_exporter(settings, recorder, worker, internal_metadata: internal_metadata)
+        transport = build_profiler_transport(settings, agent_settings)
+        scheduler = Profiling::Scheduler.new(exporter: exporter, transport: transport, interval: upload_period_seconds)
+
+        [Profiling::Profiler.new(worker: worker, scheduler: scheduler), { profiling_enabled: true }]
+      end
+
+      private_class_method def self.build_thread_context_collector(settings, recorder, optional_tracer, timeline_enabled)
+        Datadog::Profiling::Collectors::ThreadContext.new(
+          recorder: recorder,
+          max_frames: settings.profiling.advanced.max_frames,
+          tracer: optional_tracer,
+          endpoint_collection_enabled: settings.profiling.advanced.endpoint.collection.enabled,
+          timeline_enabled: timeline_enabled,
+        )
+      end
+
+      private_class_method def self.build_profiler_exporter(settings, recorder, worker, internal_metadata:)
+        info_collector = Profiling::Collectors::Info.new(settings)
+        code_provenance_collector =
+          (Profiling::Collectors::CodeProvenance.new if settings.profiling.advanced.code_provenance_enabled)
+
+        Profiling::Exporter.new(
+          pprof_recorder: recorder,
+          worker: worker,
+          info_collector: info_collector,
+          code_provenance_collector: code_provenance_collector,
+          internal_metadata: internal_metadata,
+        )
+      end
+
+      private_class_method def self.build_profiler_transport(settings, agent_settings)
+        settings.profiling.exporter.transport ||
+          Profiling::HttpTransport.new(
+            agent_settings: agent_settings,
+            site: settings.site,
+            api_key: settings.api_key,
+            upload_timeout_seconds: settings.profiling.upload.timeout_seconds,
+          )
+      end
+
+      private_class_method def self.enable_gc_profiling?(settings)
+        # See comments on the setting definition for more context on why it exists.
+        if settings.profiling.advanced.force_enable_gc_profiling
+          if Gem::Version.new(RUBY_VERSION) >= Gem::Version.new('3')
+            Datadog.logger.debug(
+              'Profiling time/resources spent in Garbage Collection force enabled. Do not use Ractors in combination ' \
+              'with this option as profiles will be incomplete.'
+            )
+          end
+
+          true
+        else
+          false
+        end
+      end
+
+      private_class_method def self.get_heap_sample_every(settings)
+        heap_sample_rate = settings.profiling.advanced.experimental_heap_sample_rate
+
+        raise ArgumentError, "Heap sample rate must be a positive integer. Was #{heap_sample_rate}" if heap_sample_rate <= 0
+
+        heap_sample_rate
+      end
+
+      private_class_method def self.enable_allocation_profiling?(settings)
+        unless settings.profiling.allocation_enabled
+          # Allocation profiling disabled, short-circuit out
+          return false
+        end
+
+        # Allocation sampling is safe and supported on Ruby 2.x, but has a few caveats on Ruby 3.x.
+
+        # SEVERE - All configurations
+        # Ruby 3.2.0 to 3.2.2 have a bug in the newobj tracepoint (https://bugs.ruby-lang.org/issues/19482,
+        # https://github.com/ruby/ruby/pull/7464) that makes this crash in any configuration. This bug is
+        # fixed on Ruby versions 3.2.3 and 3.3.0.
+        if RUBY_VERSION.start_with?('3.2.') && RUBY_VERSION < '3.2.3'
+          Datadog.logger.warn(
+            'Allocation profiling is not supported in Ruby versions 3.2.0, 3.2.1 and 3.2.2 and will be forcibly '\
+            'disabled. This is due to a VM bug that can lead to crashes (https://bugs.ruby-lang.org/issues/19482). '\
+            'Other Ruby versions do not suffer from this issue.'
+          )
+          return false
+        end
+
+        # SEVERE - Only with Ractors
+        # On Ruby versions 3.0 (all), 3.1.0 to 3.1.3, and 3.2.0 to 3.2.2 allocation profiling can trigger a VM bug
+        # that causes a segmentation fault during garbage collection of Ractors
+        # (https://bugs.ruby-lang.org/issues/18464). We don't recommend using this feature on such Rubies.
+        # This bug is fixed on Ruby versions 3.1.4, 3.2.3 and 3.3.0.
+        if RUBY_VERSION.start_with?('3.0.') ||
+            (RUBY_VERSION.start_with?('3.1.') && RUBY_VERSION < '3.1.4') ||
+            (RUBY_VERSION.start_with?('3.2.') && RUBY_VERSION < '3.2.3')
+          Datadog.logger.warn(
+            "Current Ruby version (#{RUBY_VERSION}) has a VM bug where enabling allocation profiling while using "\
+            'Ractors may cause unexpected issues, including crashes (https://bugs.ruby-lang.org/issues/18464). '\
+            'This does not happen if Ractors are not used.'
+          )
+        # ANNOYANCE - Only with Ractors
+        # On all known versions of Ruby 3.x, due to https://bugs.ruby-lang.org/issues/19112, when a ractor gets
+        # garbage collected, Ruby will disable all active tracepoints, which this feature internally relies on.
+        elsif RUBY_VERSION.start_with?('3.')
+          Datadog.logger.warn(
+            'In all known versions of Ruby 3.x, using Ractors may result in allocation profiling unexpectedly ' \
+            'stopping (https://bugs.ruby-lang.org/issues/19112). Note that this stop has no impact in your ' \
+            'application stability or performance. This does not happen if Ractors are not used.'
+          )
+        end
+
+        Datadog.logger.debug('Enabled allocation profiling')
+
+        true
+      end
+
+      private_class_method def self.enable_heap_profiling?(settings, allocation_profiling_enabled, heap_sample_rate)
+        heap_profiling_enabled = settings.profiling.advanced.experimental_heap_enabled
+
+        return false unless heap_profiling_enabled
+
+        if RUBY_VERSION.start_with?('2.') && RUBY_VERSION < '2.7'
+          Datadog.logger.warn(
+            'Heap profiling currently relies on features introduced in Ruby 2.7 and will be forcibly disabled. '\
+            'Please upgrade to Ruby >= 2.7 in order to use this feature.'
+          )
+          return false
+        end
+
+        if RUBY_VERSION < '3.1'
+          Datadog.logger.debug(
+            "Current Ruby version (#{RUBY_VERSION}) supports forced object recycling which has a bug that the " \
+            'heap profiler is forced to work around to remain accurate. This workaround requires force-setting '\
+            "the SEEN_OBJ_ID flag on objects that should have it but don't. Full details can be found in " \
+            'https://github.com/DataDog/dd-trace-rb/pull/3360. This workaround should be safe but can be ' \
+            'bypassed by disabling the heap profiler or upgrading to Ruby >= 3.1 where forced object recycling ' \
+            'was completely removed (https://bugs.ruby-lang.org/issues/18290).'
+          )
+        end
+
+        unless allocation_profiling_enabled
+          raise ArgumentError,
+            'Heap profiling requires allocation profiling to be enabled'
+        end
+
+        Datadog.logger.warn(
+          "Enabled experimental heap profiling: heap_sample_rate=#{heap_sample_rate}. This is experimental, not " \
+          'recommended, and will increase overhead!'
+        )
+
+        true
+      end
+
+      private_class_method def self.enable_heap_size_profiling?(settings, heap_profiling_enabled)
+        heap_size_profiling_enabled = settings.profiling.advanced.experimental_heap_size_enabled
+
+        return false unless heap_profiling_enabled && heap_size_profiling_enabled
+
+        Datadog.logger.warn(
+          'Enabled experimental heap size profiling. This is experimental, not recommended, and will increase overhead!'
+        )
+
+        true
+      end
+
+      private_class_method def self.no_signals_workaround_enabled?(settings) # rubocop:disable Metrics/MethodLength
+        setting_value = settings.profiling.advanced.no_signals_workaround_enabled
+        legacy_ruby_that_should_use_workaround = RUBY_VERSION.start_with?('2.3.', '2.4.', '2.5.')
+
+        unless [true, false, :auto].include?(setting_value)
+          Datadog.logger.error(
+            "Ignoring invalid value for profiling no_signals_workaround_enabled setting: #{setting_value.inspect}. " \
+            'Valid options are `true`, `false` or (default) `:auto`.'
+          )
+
+          setting_value = :auto
+        end
+
+        if setting_value == false
+          if legacy_ruby_that_should_use_workaround
+            Datadog.logger.warn(
+              'The profiling "no signals" workaround has been disabled via configuration on a legacy Ruby version ' \
+              '(< 2.6). This is not recommended ' \
+              'in production environments, as due to limitations in Ruby APIs, we suspect it may lead to crashes ' \
+              'in very rare situations. Please report any issues you run into to Datadog support or ' \
+              'via <https://github.com/datadog/dd-trace-rb/issues/new>!'
+            )
+          else
+            Datadog.logger.warn('Profiling "no signals" workaround disabled via configuration')
+          end
+
+          return false
+        end
+
+        if setting_value == true
+          Datadog.logger.warn(
+            'Profiling "no signals" workaround enabled via configuration. Profiling data will have lower quality.'
+          )
+
+          return true
+        end
+
+        # Setting is in auto mode. Let's probe to see if we should enable it:
+
+        # We don't warn users in this situation because "upgrade your Ruby" is not a great warning
+        return true if legacy_ruby_that_should_use_workaround
+
+        if Gem.loaded_specs['mysql2'] && incompatible_libmysqlclient_version?(settings)
+          Datadog.logger.warn(
+            'Enabling the profiling "no signals" workaround because an incompatible version of the mysql2 gem is ' \
+            'installed. Profiling data will have lower quality. ' \
+            'To fix this, upgrade the libmysqlclient in your OS image to version 8.0.0 or above.'
+          )
+          return true
+        end
+
+        if Gem.loaded_specs['rugged']
+          Datadog.logger.warn(
+            'Enabling the profiling "no signals" workaround because the rugged gem is installed. ' \
+            'This is needed because some operations on this gem are currently incompatible with the normal working mode ' \
+            'of the profiler, as detailed in <https://github.com/datadog/dd-trace-rb/issues/2721>. ' \
+            'Profiling data will have lower quality.'
+          )
+          return true
+        end
+
+        if (defined?(::PhusionPassenger) || Gem.loaded_specs['passenger']) && incompatible_passenger_version?
+          Datadog.logger.warn(
+            'Enabling the profiling "no signals" workaround because an incompatible version of the passenger gem is ' \
+            'installed. Profiling data will have lower quality.' \
+            'To fix this, upgrade the passenger gem to version 6.0.19 or above.'
+          )
+          return true
+        end
+
+        false
+      end
+
+      # Versions of libmysqlclient prior to 8.0.0 are known to have buggy handling of system call interruptions.
+      # The profiler can sometimes cause system call interruptions, and so this combination can cause queries to fail.
+      #
+      # See https://bugs.mysql.com/bug.php?id=83109 and
+      # https://docs.datadoghq.com/profiler/profiler_troubleshooting/ruby/#unexpected-run-time-failures-and-errors-from-ruby-gems-that-use-native-extensions-in-dd-trace-rb-1110
+      # for details.
+      #
+      # The `mysql2` gem's `info` method can be used to determine which `libmysqlclient` version is in use, and thus to
+      # detect if it's safe for the profiler to use signals or if we need to employ a fallback.
+      private_class_method def self.incompatible_libmysqlclient_version?(settings)
+        return true if settings.profiling.advanced.skip_mysql2_check
+
+        Datadog.logger.debug(
+          'Requiring `mysql2` to check if the `libmysqlclient` version it uses is compatible with profiling'
+        )
+
+        begin
+          require 'mysql2'
+
+          # The mysql2-aurora gem likes to monkey patch itself in replacement of Mysql2::Client, and uses
+          # `method_missing` to delegate to the original BUT unfortunately does not implement `respond_to_missing?` and
+          # thus our `respond_to?(:info)` below was failing.
+          #
+          # But on the bright side, the gem does stash a reference to the original Mysql2::Client class in a constant,
+          # so if that constant exists, we use that for our probing.
+          mysql2_client_class =
+            if defined?(Mysql2::Aurora::ORIGINAL_CLIENT_CLASS)
+              Mysql2::Aurora::ORIGINAL_CLIENT_CLASS
+            elsif defined?(Mysql2::Client)
+              Mysql2::Client
+            end
+
+          return true unless mysql2_client_class && mysql2_client_class.respond_to?(:info)
+
+          info = mysql2_client_class.info
+          libmysqlclient_version = Gem::Version.new(info[:version])
+
+          compatible =
+            libmysqlclient_version >= Gem::Version.new('8.0.0') ||
+            looks_like_mariadb?(info, libmysqlclient_version)
+
+          Datadog.logger.debug(
+            "The `mysql2` gem is using #{compatible ? 'a compatible' : 'an incompatible'} version of " \
+            "the `libmysqlclient` library (#{libmysqlclient_version})"
+          )
+
+          !compatible
+        rescue StandardError, LoadError => e
+          Datadog.logger.warn(
+            'Failed to probe `mysql2` gem information. ' \
+            "Cause: #{e.class.name} #{e.message} Location: #{Array(e.backtrace).first}"
+          )
+
+          true
+        end
+      end
+
+      # See https://github.com/datadog/dd-trace-rb/issues/2976 for details.
+      private_class_method def self.incompatible_passenger_version?
+        if Gem.loaded_specs['passenger']
+          Gem.loaded_specs['passenger'].version < Gem::Version.new('6.0.19')
+        else
+          true
+        end
+      end
+
+      private_class_method def self.valid_overhead_target(overhead_target_percentage)
+        if overhead_target_percentage > 0 && overhead_target_percentage <= 20
+          overhead_target_percentage
+        else
+          Datadog.logger.error(
+            'Ignoring invalid value for profiling overhead_target_percentage setting: ' \
+            "#{overhead_target_percentage.inspect}. Falling back to default value."
+          )
+
+          2.0
+        end
+      end
+
+      # To add just a bit more complexity to our detection code, in https://github.com/DataDog/dd-trace-rb/issues/3334
+      # a user reported that our code was incorrectly flagging the mariadb variant of libmysqlclient as being
+      # incompatible. In fact we have no reports of the mariadb variant needing the "no signals" workaround,
+      # so we flag it as compatible when it's in use.
+      #
+      # A problem is that there doesn't seem to be an obvious way to query the mysql2 gem on which kind of
+      # libmysqlclient it's using, so we detect it by looking at the version.
+      #
+      # The info method for mysql2 with mariadb looks something like this:
+      # `{:id=>30308, :version=>"3.3.8", :header_version=>"11.2.2"}`
+      #
+      # * The version seems to come from https://github.com/mariadb-corporation/mariadb-connector-c and the latest
+      # one is 3.x.
+      # * The header_version is what people usually see as the "mariadb version"
+      #
+      # As a comparison, for libmysql the info looks like:
+      # * `{:id=>80035, :version=>"8.0.35", :header_version=>"8.0.35"}`
+      #
+      # Thus our detection is version 4 or older, because libmysqlclient 4 is almost 20 years old so it's most probably
+      # not that one + header_version being 10 or newer, since according to https://endoflife.date/mariadb that's a
+      # sane range for modern mariadb releases.
+      private_class_method def self.looks_like_mariadb?(info, libmysqlclient_version)
+        header_version = Gem::Version.new(info[:header_version]) if info[:header_version]
+
+        !!(header_version &&
+          libmysqlclient_version < Gem::Version.new('5.0.0') &&
+          header_version >= Gem::Version.new('10.0.0'))
+      end
+    end
+  end
+end

data/lib/datadog/profiling/exporter.rb

@@ -0,0 +1,103 @@
+# frozen_string_literal: true
+
+require_relative 'ext'
+require_relative 'tag_builder'
+
+module Datadog
+  module Profiling
+    # Exports profiling data gathered by the multiple recorders in a `Flush`.
+    #
+    # @ivoanjo: Note that the recorder that gathers pprof data is special, since we use its start/finish/empty? to
+    # decide if there's data to flush, as well as the timestamp for that data.
+    # I could've made the whole design more generic, but I'm unsure if we'll ever have more than a handful of
+    # recorders, so I've decided to make it specific until we actually need to support more recorders.
+    #
+    class Exporter
+      # Profiles with duration less than this will not be reported
+      PROFILE_DURATION_THRESHOLD_SECONDS = 1
+
+      private
+
+      attr_reader \
+        :pprof_recorder,
+        :code_provenance_collector, # The code provenance collector acts both as collector and as a recorder
+        :minimum_duration_seconds,
+        :time_provider,
+        :last_flush_finish_at,
+        :created_at,
+        :internal_metadata,
+        :info_json
+
+      public
+
+      def initialize(
+        pprof_recorder:,
+        worker:,
+        info_collector:,
+        code_provenance_collector:,
+        internal_metadata:,
+        minimum_duration_seconds: PROFILE_DURATION_THRESHOLD_SECONDS,
+        time_provider: Time
+      )
+        @pprof_recorder = pprof_recorder
+        @worker = worker
+        @code_provenance_collector = code_provenance_collector
+        @minimum_duration_seconds = minimum_duration_seconds
+        @time_provider = time_provider
+        @last_flush_finish_at = nil
+        @created_at = time_provider.now.utc
+        @internal_metadata = internal_metadata
+        # NOTE: At the time of this comment collected info does not change over time so we'll hardcode
+        #       it on startup to prevent serializing the same info on every flush.
+        @info_json = JSON.fast_generate(info_collector.info).freeze
+      end
+
+      def flush
+        worker_stats = @worker.stats_and_reset_not_thread_safe
+        start, finish, compressed_pprof = pprof_recorder.serialize
+        @last_flush_finish_at = finish
+
+        return if compressed_pprof.nil? # We don't want to report empty profiles
+
+        if duration_below_threshold?(start, finish)
+          Datadog.logger.debug('Skipped exporting profiling events as profile duration is below minimum')
+          return
+        end
+
+        uncompressed_code_provenance = code_provenance_collector.refresh.generate_json if code_provenance_collector
+
+        Flush.new(
+          start: start,
+          finish: finish,
+          pprof_file_name: Datadog::Profiling::Ext::Transport::HTTP::PPROF_DEFAULT_FILENAME,
+          pprof_data: compressed_pprof.to_s,
+          code_provenance_file_name: Datadog::Profiling::Ext::Transport::HTTP::CODE_PROVENANCE_FILENAME,
+          code_provenance_data: uncompressed_code_provenance,
+          tags_as_array: Datadog::Profiling::TagBuilder.call(settings: Datadog.configuration).to_a,
+          internal_metadata: internal_metadata.merge(
+            {
+              worker_stats: worker_stats,
+              gc: GC.stat,
+            }
+          ),
+          info_json: info_json,
+        )
+      end
+
+      def can_flush?
+        !duration_below_threshold?(last_flush_finish_at || created_at, time_provider.now.utc)
+      end
+
+      def reset_after_fork
+        @last_flush_finish_at = time_provider.now.utc
+        nil
+      end
+
+      private
+
+      def duration_below_threshold?(start, finish)
+        (finish - start) < minimum_duration_seconds
+      end
+    end
+  end
+end

data/lib/datadog/profiling/ext/forking.rb

@@ -0,0 +1,98 @@
+# frozen_string_literal: true
+
+module Datadog
+  module Profiling
+    module Ext
+      # Monkey patches `Kernel#fork`, adding a `Kernel#at_fork` callback mechanism which is used to restore
+      # profiling abilities after the VM forks.
+      #
+      # Known limitations: Does not handle `BasicObject`s that include `Kernel` directly; e.g.
+      # `Class.new(BasicObject) { include(::Kernel); def call; fork { }; end }.new.call`.
+      #
+      # This will be fixed once we moved to hooking into `Process._fork`
+      module Forking
+        def self.supported?
+          Process.respond_to?(:fork)
+        end
+
+        def self.apply!
+          return false unless supported?
+
+          [
+            ::Process.singleton_class, # Process.fork
+            ::Kernel.singleton_class,  # Kernel.fork
+            ::Object,                  # fork without explicit receiver (it's defined as a method in ::Kernel)
+            # Note: Modifying Object as we do here is irreversible. During tests, this
+            # change will stick around even if we otherwise stub `Process` and `Kernel`
+          ].each { |target| target.prepend(Kernel) }
+
+          ::Process.singleton_class.prepend(ProcessDaemonPatch)
+        end
+
+        # Extensions for kernel
+        #
+        # TODO: Consider hooking into `Process._fork` on Ruby 3.1+ instead, see
+        #       https://github.com/ruby/ruby/pull/5017 and https://bugs.ruby-lang.org/issues/17795
+        module Kernel
+          def fork
+            # If a block is provided, it must be wrapped to trigger callbacks.
+            child_block = if block_given?
+                            proc do
+                              # Trigger :child callback
+                              datadog_at_fork_blocks[:child].each(&:call) if datadog_at_fork_blocks.key?(:child)
+
+                              # Invoke original block
+                              yield
+                            end
+                          end
+
+            # Start fork
+            # If a block is provided, use the wrapped version.
+            result = child_block.nil? ? super : super(&child_block)
+
+            # Trigger correct callbacks depending on whether we're in the parent or child.
+            # If we're in the fork, result = nil: trigger child callbacks.
+            # If we're in the parent, result = fork PID: trigger parent callbacks.
+            datadog_at_fork_blocks[:child].each(&:call) if result.nil? && datadog_at_fork_blocks.key?(:child)
+
+            # Return PID from #fork
+            result
+          end
+
+          def at_fork(stage, &block)
+            raise ArgumentError, 'Bad \'stage\' for ::at_fork' unless stage == :child
+
+            datadog_at_fork_blocks[stage] = [] unless datadog_at_fork_blocks.key?(stage)
+            datadog_at_fork_blocks[stage] << block
+          end
+
+          module_function
+
+          def datadog_at_fork_blocks
+            # Blocks should be shared across all users of this module,
+            # e.g. Process#fork, Kernel#fork, etc. should all invoke the same callbacks.
+            # rubocop:disable Style/ClassVars
+            @@datadog_at_fork_blocks ||= {}
+            # rubocop:enable Style/ClassVars
+          end
+        end
+
+        # A call to Process.daemon ( https://rubyapi.org/3.1/o/process#method-c-daemon ) forks the current process and
+        # keeps executing code in the child process, killing off the parent, thus effectively replacing it.
+        #
+        # This monkey patch makes the `Kernel#at_fork` mechanism defined above also work in this situation.
+        module ProcessDaemonPatch
+          def daemon(*args)
+            datadog_at_fork_blocks = Datadog::Profiling::Ext::Forking::Kernel.datadog_at_fork_blocks
+
+            result = super
+
+            datadog_at_fork_blocks[:child].each(&:call) if datadog_at_fork_blocks.key?(:child)
+
+            result
+          end
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/ext.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+
+module Datadog
+  module Profiling
+    module Ext
+      ENV_ENABLED = 'DD_PROFILING_ENABLED'
+      ENV_UPLOAD_TIMEOUT = 'DD_PROFILING_UPLOAD_TIMEOUT'
+      ENV_MAX_FRAMES = 'DD_PROFILING_MAX_FRAMES'
+      ENV_AGENTLESS = 'DD_PROFILING_AGENTLESS'
+      ENV_ENDPOINT_COLLECTION_ENABLED = 'DD_PROFILING_ENDPOINT_COLLECTION_ENABLED'
+
+      module Transport
+        module HTTP
+          FORM_FIELD_TAG_ENV = 'env'
+          FORM_FIELD_TAG_HOST = 'host'
+          FORM_FIELD_TAG_LANGUAGE = 'language'
+          FORM_FIELD_TAG_PID = 'process_id'
+          FORM_FIELD_TAG_PROFILER_VERSION = 'profiler_version'
+          FORM_FIELD_TAG_RUNTIME = 'runtime'
+          FORM_FIELD_TAG_RUNTIME_ENGINE = 'runtime_engine'
+          FORM_FIELD_TAG_RUNTIME_ID = 'runtime-id'
+          FORM_FIELD_TAG_RUNTIME_PLATFORM = 'runtime_platform'
+          FORM_FIELD_TAG_RUNTIME_VERSION = 'runtime_version'
+          FORM_FIELD_TAG_SERVICE = 'service'
+          FORM_FIELD_TAG_VERSION = 'version'
+          TAG_GIT_REPOSITORY_URL = 'git.repository_url'
+          TAG_GIT_COMMIT_SHA = 'git.commit.sha'
+
+          PPROF_DEFAULT_FILENAME = 'rubyprofile.pprof'
+          CODE_PROVENANCE_FILENAME = 'code-provenance.json'
+        end
+      end
+    end
+  end
+end