ddtrace 1.0.0 → 1.10.1

data/lib/datadog/profiling/component.rb

@@ -0,0 +1,161 @@
+# frozen_string_literal: true
+
+module Datadog
+  module Profiling
+    # Profiling component
+    module Component
+      def build_profiler(settings, agent_settings, tracer)
+        return 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 'ddtrace'` 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 unless Profiling.supported?
+
+        unless defined?(Profiling::Tasks::Setup)
+          # In #1545 a user reported a NameError due to this constant being uninitialized
+          # I've documented my suspicion on why that happened in
+          # https://github.com/DataDog/dd-trace-rb/issues/1545#issuecomment-856049025
+          #
+          # > Thanks for the info! It seems to feed into my theory: there's two moments in the code where we check if
+          # > profiler is "supported": 1) when loading ddtrace (inside preload) and 2) when starting the profile
+          # > after Datadog.configure gets run.
+          # > The problem is that the code assumes that both checks 1) and 2) will always reach the same conclusion:
+          # > either profiler is supported, or profiler is not supported.
+          # > In the problematic case, it looks like in your case check 1 decides that profiler is not
+          # > supported => doesn't load it, and then check 2 decides that it is => assumes it is loaded and tries to
+          # > start it.
+          #
+          # I was never able to validate if this was the issue or why exactly .supported? would change its mind BUT
+          # just in case it happens again, I've left this check which avoids breaking the user's application AND
+          # would instead direct them to report it to us instead, so that we can investigate what's wrong.
+          #
+          # TODO: As of June 2021, most checks in .supported? are related to the google-protobuf gem; so it's
+          # very likely that it was the origin of the issue we saw. Thus, if, as planned we end up moving away from
+          # protobuf OR enough time has passed and no users saw the issue again, we can remove this check altogether.
+          Datadog.logger.error(
+            'Profiling was marked as supported and enabled, but setup task was not loaded properly. ' \
+            'Please report this at https://github.com/DataDog/dd-trace-rb/blob/master/CONTRIBUTING.md#found-a-bug'
+          )
+
+          return
+        end
+
+        # Load extensions needed to support some of the Profiling features
+        Profiling::Tasks::Setup.new.run
+
+        # NOTE: Please update the Initialization section of ProfilingDevelopment.md with any changes to this method
+
+        if settings.profiling.advanced.force_enable_new_profiler
+          print_new_profiler_warnings
+
+          recorder = Datadog::Profiling::StackRecorder.new(
+            cpu_time_enabled: RUBY_PLATFORM.include?('linux'), # Only supported on Linux currently
+            alloc_samples_enabled: false, # Always disabled for now -- work in progress
+          )
+          collector = Datadog::Profiling::Collectors::CpuAndWallTimeWorker.new(
+            recorder: recorder,
+            max_frames: settings.profiling.advanced.max_frames,
+            tracer: tracer,
+            gc_profiling_enabled: should_enable_gc_profiling?(settings),
+            allocation_counting_enabled: settings.profiling.advanced.allocation_counting_enabled,
+          )
+        else
+          trace_identifiers_helper = Profiling::TraceIdentifiers::Helper.new(
+            tracer: tracer,
+            endpoint_collection_enabled: settings.profiling.advanced.endpoint.collection.enabled
+          )
+
+          recorder = build_profiler_old_recorder(settings)
+          collector = build_profiler_oldstack_collector(settings, recorder, trace_identifiers_helper)
+        end
+
+        exporter = build_profiler_exporter(settings, recorder)
+        transport = build_profiler_transport(settings, agent_settings)
+        scheduler = Profiling::Scheduler.new(exporter: exporter, transport: transport)
+
+        Profiling::Profiler.new([collector], scheduler)
+      end
+
+      private
+
+      def build_profiler_old_recorder(settings)
+        Profiling::OldRecorder.new([Profiling::Events::StackSample], settings.profiling.advanced.max_events)
+      end
+
+      def build_profiler_exporter(settings, recorder)
+        code_provenance_collector =
+          (Profiling::Collectors::CodeProvenance.new if settings.profiling.advanced.code_provenance_enabled)
+
+        Profiling::Exporter.new(pprof_recorder: recorder, code_provenance_collector: code_provenance_collector)
+      end
+
+      def build_profiler_oldstack_collector(settings, old_recorder, trace_identifiers_helper)
+        Profiling::Collectors::OldStack.new(
+          old_recorder,
+          trace_identifiers_helper: trace_identifiers_helper,
+          max_frames: settings.profiling.advanced.max_frames
+        )
+      end
+
+      def 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
+
+      def should_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
+
+      def print_new_profiler_warnings
+        if Gem::Version.new(RUBY_VERSION) >= Gem::Version.new('2.6')
+          Datadog.logger.warn(
+            'New Ruby profiler has been force-enabled. This is a beta feature. Please report any issues ' \
+            'you run into to Datadog support or via <https://github.com/datadog/dd-trace-rb/issues/new>!'
+          )
+        else
+          # For more details on the issue, see the "BIG Issue" comment on `gvl_owner` function in
+          # `private_vm_api_access.c`.
+          Datadog.logger.warn(
+            'New Ruby profiler has been force-enabled 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>!'
+          )
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/encoding/profile.rb

@@ -1,11 +1,6 @@
-# typed: true
-
-require 'set'
 require 'time'
 
-require 'datadog/core'
-require 'datadog/profiling/flush'
-require 'datadog/profiling/pprof/template'
+require_relative '../pprof/template'
 
 module Datadog
   module Profiling
@@ -15,31 +10,29 @@ module Datadog
         module Protobuf
           module_function
 
-          def encode(flush)
-            return unless flush
-
+          def encode(event_count:, event_groups:, start:, finish:)
             # Create a pprof template from the list of event types
-            event_classes = flush.event_groups.collect(&:event_class).uniq
+            event_classes = event_groups.collect(&:event_class).uniq
             template = Pprof::Template.for_event_classes(event_classes)
 
             # Add all events to the pprof
-            flush.event_groups.each { |event_group| template.add_events!(event_group.event_class, event_group.events) }
+            event_groups.each { |event_group| template.add_events!(event_group.event_class, event_group.events) }
 
             Datadog.logger.debug do
               max_events = Datadog.configuration.profiling.advanced.max_events
               events_sampled =
-                if flush.event_count == max_events
+                if event_count == max_events
                   'max events limit hit, events were sampled [profile will be biased], '
                 else
                   ''
                 end
 
-              "Encoding profile covering #{flush.start.iso8601} to #{flush.finish.iso8601}, " \
-              "events: #{flush.event_count} (#{events_sampled}#{template.debug_statistics})"
+              "Encoding profile covering #{start.iso8601} to #{finish.iso8601}, " \
+              "events: #{event_count} (#{events_sampled}#{template.debug_statistics})"
             end
 
             # Build the profile and encode it
-            template.to_pprof(start: flush.start, finish: flush.finish)
+            template.to_pprof(start: start, finish: finish)
           end
         end
       end

data/lib/datadog/profiling/event.rb

@@ -1,5 +1,3 @@
-# typed: true
-
 module Datadog
   module Profiling
     # Describes a sample of some data obtained from the runtime.

data/lib/datadog/profiling/events/stack.rb

@@ -1,6 +1,4 @@
-# typed: true
-
-require 'datadog/profiling/event'
+require_relative '../event'
 
 module Datadog
   module Profiling

data/lib/datadog/profiling/exporter.rb

@@ -1,24 +1,82 @@
-# typed: true
-
-require 'datadog/profiling/transport/io/client'
+require_relative 'ext'
+require_relative 'tag_builder'
 
 module Datadog
   module Profiling
-    # Writes profiling data to a given transport
+    # 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 \
-        :transport
+        :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
+
+      public
+
+      def initialize(
+        pprof_recorder:,
+        code_provenance_collector:,
+        minimum_duration_seconds: PROFILE_DURATION_THRESHOLD_SECONDS,
+        time_provider: Time
+      )
+        @pprof_recorder = pprof_recorder
+        @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
+      end
+
+      def flush
+        start, finish, uncompressed_pprof = pprof_recorder.serialize
+        @last_flush_finish_at = finish
 
-      def initialize(transport)
-        unless transport.is_a?(Profiling::Transport::Client)
-          raise ArgumentError, 'Unsupported transport for profiling exporter.'
+        return if uncompressed_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
 
-        @transport = transport
+        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: uncompressed_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,
+        )
+      end
+
+      def can_flush?
+        !duration_below_threshold?(last_flush_finish_at || created_at, time_provider.now.utc)
       end
 
-      def export(flush)
-        transport.send_profiling_flush(flush)
+      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

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

@@ -1,10 +1,13 @@
-# typed: false
-
 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)
@@ -13,48 +16,34 @@ module Datadog
         def self.apply!
           return false unless supported?
 
-          modules = [::Process, ::Kernel]
-          # TODO: Ruby < 2.3 doesn't support Binding#receiver.
-          #       Remove "else #eval" clause when Ruby < 2.3 support is dropped.
-          # NOTE: Modifying the "main" object as we do here is (as far as I know) irreversible. During tests, this change
-          #       will stick around even if we otherwise stub `Process` and `Kernel`.
-          modules << (TOPLEVEL_BINDING.respond_to?(:receiver) ? TOPLEVEL_BINDING.receiver : TOPLEVEL_BINDING.eval('self'))
-
-          # Patch top-level binding, Kernel, Process.
-          # NOTE: We could instead do Kernel.module_eval { def fork; ... end }
-          #       however, this method rewrite is more invasive and irreversible.
-          #       It could also have collisions with other libraries that patch.
-          #       Opt to modify the inheritance of each relevant target instead.
-          modules.each do |mod|
-            clazz = if mod.class <= Module
-                      mod.singleton_class
-                    else
-                      mod.class
-                    end
-
-            clazz.prepend(Kernel)
-          end
+          [
+            ::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
-          FORK_STAGES = [:prepare, :parent, :child].freeze
-
           def fork
             # If a block is provided, it must be wrapped to trigger callbacks.
             child_block = if block_given?
                             proc do
                               # Trigger :child callback
-                              at_fork_blocks[:child].each(&:call) if at_fork_blocks.key?(:child)
+                              ddtrace_at_fork_blocks[:child].each(&:call) if ddtrace_at_fork_blocks.key?(:child)
 
                               # Invoke original block
                               yield
                             end
                           end
 
-            # Trigger :prepare callback
-            at_fork_blocks[:prepare].each(&:call) if at_fork_blocks.key?(:prepare)
-
             # Start fork
             # If a block is provided, use the wrapped version.
             result = child_block.nil? ? super : super(&child_block)
@@ -62,37 +51,45 @@ module Datadog
             # 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.
-            # rubocop:disable Style/IfInsideElse
-            if result.nil?
-              # Trigger :child callback
-              at_fork_blocks[:child].each(&:call) if at_fork_blocks.key?(:child)
-            else
-              # Trigger :parent callback
-              at_fork_blocks[:parent].each(&:call) if at_fork_blocks.key?(:parent)
-            end
-            # rubocop:enable Style/IfInsideElse
+            ddtrace_at_fork_blocks[:child].each(&:call) if result.nil? && ddtrace_at_fork_blocks.key?(:child)
 
             # Return PID from #fork
             result
           end
 
-          def at_fork(stage = :prepare, &block)
-            raise ArgumentError, 'Bad \'stage\' for ::at_fork' unless FORK_STAGES.include?(stage)
+          def at_fork(stage, &block)
+            raise ArgumentError, 'Bad \'stage\' for ::at_fork' unless stage == :child
 
-            at_fork_blocks[stage] = [] unless at_fork_blocks.key?(stage)
-            at_fork_blocks[stage] << block
+            ddtrace_at_fork_blocks[stage] = [] unless ddtrace_at_fork_blocks.key?(stage)
+            ddtrace_at_fork_blocks[stage] << block
           end
 
           module_function
 
-          def at_fork_blocks
+          def ddtrace_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
-            @@at_fork_blocks ||= {}
+            @@ddtrace_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)
+            ddtrace_at_fork_blocks = Datadog::Profiling::Ext::Forking::Kernel.ddtrace_at_fork_blocks
+
+            result = super
+
+            ddtrace_at_fork_blocks[:child].each(&:call) if ddtrace_at_fork_blocks.key?(:child)
+
+            result
+          end
+        end
       end
     end
   end

data/lib/datadog/profiling/ext.rb

@@ -1,5 +1,3 @@
-# typed: true
-
 module Datadog
   module Profiling
     module Ext
@@ -9,6 +7,7 @@ module Datadog
       ENV_AGENTLESS = 'DD_PROFILING_AGENTLESS'.freeze
       ENV_ENDPOINT_COLLECTION_ENABLED = 'DD_PROFILING_ENDPOINT_COLLECTION_ENABLED'.freeze
 
+      # TODO: Consider removing this once the Ruby-based pprof encoding is removed and replaced by libdatadog
       module Pprof
         LABEL_KEY_LOCAL_ROOT_SPAN_ID = 'local root span id'.freeze
         LABEL_KEY_SPAN_ID = 'span id'.freeze
@@ -22,15 +21,10 @@ module Datadog
 
       module Transport
         module HTTP
-          URI_TEMPLATE_DD_API = 'https://intake.profile.%s/'.freeze
-
-          FORM_FIELD_RECORDING_START = 'start'.freeze
-          FORM_FIELD_RECORDING_END = 'end'.freeze
-          FORM_FIELD_FAMILY = 'family'.freeze
           FORM_FIELD_TAG_ENV = 'env'.freeze
           FORM_FIELD_TAG_HOST = 'host'.freeze
           FORM_FIELD_TAG_LANGUAGE = 'language'.freeze
-          FORM_FIELD_TAG_PID = 'pid'.freeze
+          FORM_FIELD_TAG_PID = 'process_id'.freeze
           FORM_FIELD_TAG_PROFILER_VERSION = 'profiler_version'.freeze
           FORM_FIELD_TAG_RUNTIME = 'runtime'.freeze
           FORM_FIELD_TAG_RUNTIME_ENGINE = 'runtime_engine'.freeze
@@ -39,17 +33,9 @@ module Datadog
           FORM_FIELD_TAG_RUNTIME_VERSION = 'runtime_version'.freeze
           FORM_FIELD_TAG_SERVICE = 'service'.freeze
           FORM_FIELD_TAG_VERSION = 'version'.freeze
-          FORM_FIELD_TAGS = 'tags'.freeze
-          FORM_FIELD_INTAKE_VERSION = 'version'.freeze
-
-          HEADER_CONTENT_TYPE = 'Content-Type'.freeze
-          HEADER_CONTENT_TYPE_OCTET_STREAM = 'application/octet-stream'.freeze
-
-          FORM_FIELD_PPROF_DATA = 'data[rubyprofile.pprof]'.freeze
-          PPROF_DEFAULT_FILENAME = 'rubyprofile.pprof.gz'.freeze
 
-          FORM_FIELD_CODE_PROVENANCE_DATA = 'data[code-provenance.json]'.freeze
-          CODE_PROVENANCE_FILENAME = 'code-provenance.json.gz'.freeze
+          PPROF_DEFAULT_FILENAME = 'rubyprofile.pprof'.freeze
+          CODE_PROVENANCE_FILENAME = 'code-provenance.json'.freeze
         end
       end
     end

data/lib/datadog/profiling/flush.rb

@@ -1,69 +1,36 @@
-# typed: false
-
-require 'datadog/core/environment/identity'
-require 'datadog/core/environment/socket'
-
 module Datadog
   module Profiling
+    # Represents a collection of events of a specific type being flushed.
+    EventGroup = Struct.new(:event_class, :events)
+
     # Entity class used to represent metadata for a given profile
-    Flush = Struct.new(
-      :start,
-      :finish,
-      :event_groups,
-      :event_count,
-      :code_provenance,
-      :runtime_id,
-      :service,
-      :env,
-      :version,
-      :host,
-      :language,
-      :runtime_engine,
-      :runtime_platform,
-      :runtime_version,
-      :profiler_version,
-      :tags
-    ) do
+    class Flush
+      attr_reader \
+        :start,
+        :finish,
+        :pprof_file_name,
+        :pprof_data, # gzipped pprof bytes
+        :code_provenance_file_name,
+        :code_provenance_data, # gzipped json bytes
+        :tags_as_array
+
       def initialize(
         start:,
         finish:,
-        event_groups:,
-        event_count:,
-        code_provenance:,
-        runtime_id: Core::Environment::Identity.id,
-        service: Datadog.configuration.service,
-        env: Datadog.configuration.env,
-        version: Datadog.configuration.version,
-        host: Core::Environment::Socket.hostname,
-        language: Core::Environment::Identity.lang,
-        runtime_engine: Core::Environment::Identity.lang_engine,
-        runtime_platform: Core::Environment::Identity.lang_platform,
-        runtime_version: Core::Environment::Identity.lang_version,
-        profiler_version: Core::Environment::Identity.tracer_version,
-        tags: Datadog.configuration.tags
+        pprof_file_name:,
+        pprof_data:,
+        code_provenance_file_name:,
+        code_provenance_data:,
+        tags_as_array:
       )
-        super(
-          start,
-          finish,
-          event_groups,
-          event_count,
-          code_provenance,
-          runtime_id,
-          service,
-          env,
-          version,
-          host,
-          language,
-          runtime_engine,
-          runtime_platform,
-          runtime_version,
-          profiler_version,
-          tags,
-        )
+        @start = start
+        @finish = finish
+        @pprof_file_name = pprof_file_name
+        @pprof_data = pprof_data
+        @code_provenance_file_name = code_provenance_file_name
+        @code_provenance_data = code_provenance_data
+        @tags_as_array = tags_as_array
       end
     end
-
-    # Represents a collection of events of a specific type being flushed.
-    EventGroup = Struct.new(:event_class, :events).freeze
   end
 end