ddtrace 0.54.2 → 1.8.0

data/lib/datadog/profiling/pprof/stack_sample.rb

@@ -0,0 +1,141 @@
+# typed: true
+
+require_relative '../ext'
+require_relative '../events/stack'
+require_relative 'builder'
+require_relative 'converter'
+
+module Datadog
+  module Profiling
+    module Pprof
+      # Builds a profile from a StackSample
+      #
+      # NOTE: This class may appear stateless but is in fact stateful; a new instance should be created for every
+      # encoded profile.
+      class StackSample < Converter
+        SAMPLE_TYPES = {
+          cpu_time_ns: [
+            Profiling::Ext::Pprof::VALUE_TYPE_CPU,
+            Profiling::Ext::Pprof::VALUE_UNIT_NANOSECONDS
+          ],
+          wall_time_ns: [
+            Profiling::Ext::Pprof::VALUE_TYPE_WALL,
+            Profiling::Ext::Pprof::VALUE_UNIT_NANOSECONDS
+          ]
+        }.freeze
+
+        def self.sample_value_types
+          SAMPLE_TYPES
+        end
+
+        def initialize(*_)
+          super
+
+          @most_recent_trace_samples = {}
+          @processed_unique_stacks = 0
+          @processed_with_trace = 0
+        end
+
+        def add_events!(stack_samples)
+          new_samples = build_samples(stack_samples)
+          builder.samples.concat(new_samples)
+        end
+
+        def stack_sample_group_key(stack_sample)
+          # We want to make sure we have the most recent sample for any trace.
+          # (This is done here to save an iteration over all samples.)
+          update_most_recent_trace_sample(stack_sample)
+
+          stack_sample.hash
+        end
+
+        # Track the most recent sample for each trace (identified by root span id)
+        def update_most_recent_trace_sample(stack_sample)
+          return unless stack_sample.root_span_id && stack_sample.trace_resource
+
+          # Update trace resource with most recent value
+          if (most_recent_trace_sample = @most_recent_trace_samples[stack_sample.root_span_id])
+            if most_recent_trace_sample.timestamp < stack_sample.timestamp
+              @most_recent_trace_samples[stack_sample.root_span_id] = stack_sample
+            end
+          else
+            # Add trace resource
+            @most_recent_trace_samples[stack_sample.root_span_id] = stack_sample
+          end
+        end
+
+        def build_samples(stack_samples)
+          groups = group_events(stack_samples, &method(:stack_sample_group_key))
+          groups.collect do |_group_key, group|
+            @processed_unique_stacks += 1
+            build_sample(group.sample, group.values)
+          end
+        end
+
+        def build_sample(stack_sample, values)
+          locations = builder.build_locations(
+            stack_sample.frames,
+            stack_sample.total_frame_count
+          )
+
+          Perftools::Profiles::Sample.new(
+            location_id: locations.collect { |location| location['id'.freeze] },
+            value: values,
+            label: build_sample_labels(stack_sample)
+          )
+        end
+
+        def build_event_values(stack_sample)
+          no_value = Profiling::Ext::Pprof::SAMPLE_VALUE_NO_VALUE
+          values = super(stack_sample)
+          values[sample_value_index(:cpu_time_ns)] = stack_sample.cpu_time_interval_ns || no_value
+          values[sample_value_index(:wall_time_ns)] = stack_sample.wall_time_interval_ns || no_value
+          values
+        end
+
+        def build_sample_labels(stack_sample)
+          labels = [
+            Perftools::Profiles::Label.new(
+              key: builder.string_table.fetch(Profiling::Ext::Pprof::LABEL_KEY_THREAD_ID),
+              str: builder.string_table.fetch(stack_sample.thread_id.to_s)
+            )
+          ]
+
+          root_span_id = stack_sample.root_span_id || 0
+          span_id = stack_sample.span_id || 0
+
+          if root_span_id != 0 && span_id != 0
+            @processed_with_trace += 1
+
+            labels << Perftools::Profiles::Label.new(
+              key: builder.string_table.fetch(Profiling::Ext::Pprof::LABEL_KEY_LOCAL_ROOT_SPAN_ID),
+              str: builder.string_table.fetch(root_span_id.to_s)
+            )
+
+            labels << Perftools::Profiles::Label.new(
+              key: builder.string_table.fetch(Profiling::Ext::Pprof::LABEL_KEY_SPAN_ID),
+              str: builder.string_table.fetch(span_id.to_s)
+            )
+
+            # Use most up-to-date trace resource, if available.
+            # Otherwise, use the trace resource provided.
+            trace_resource = @most_recent_trace_samples.fetch(stack_sample.root_span_id, stack_sample).trace_resource
+
+            if trace_resource && !trace_resource.empty?
+              labels << Perftools::Profiles::Label.new(
+                key: builder.string_table.fetch(Profiling::Ext::Pprof::LABEL_KEY_TRACE_ENDPOINT),
+                str: builder.string_table.fetch(trace_resource)
+              )
+            end
+          end
+
+          labels
+        end
+
+        def debug_statistics
+          "unique stacks: #{@processed_unique_stacks}, of which had active traces: #{@processed_with_trace}"
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/pprof/string_table.rb

@@ -0,0 +1,12 @@
+# typed: strict
+
+require_relative '../../core/utils/string_table'
+
+module Datadog
+  module Profiling
+    module Pprof
+      # Tracks strings and returns IDs
+      class StringTable < Core::Utils::StringTable; end
+    end
+  end
+end

data/lib/datadog/profiling/pprof/template.rb

@@ -0,0 +1,120 @@
+# typed: false
+
+require_relative 'payload'
+require_relative 'message_set'
+require_relative 'builder'
+
+require_relative '../events/stack'
+require_relative 'stack_sample'
+
+module Datadog
+  module Profiling
+    module Pprof
+      # Converts a collection of profiling events into a Perftools::Profiles::Profile
+      class Template
+        DEFAULT_MAPPINGS = {
+          Events::StackSample => Pprof::StackSample
+        }.freeze
+
+        attr_reader \
+          :builder,
+          :converters,
+          :sample_type_mappings
+
+        def self.for_event_classes(event_classes)
+          # Build a map of event class --> converter class
+          mappings = event_classes.each_with_object({}) do |event_class, m|
+            converter_class = DEFAULT_MAPPINGS[event_class]
+            raise NoProfilingEventConversionError, event_class unless converter_class
+
+            m[event_class] = converter_class
+          end
+
+          new(mappings)
+        end
+
+        def initialize(mappings)
+          @builder = Builder.new
+          @converters = Hash.new { |_h, event_class| raise NoProfilingEventConversionError, event_class }
+          @sample_type_mappings = Hash.new { |_h, type| raise UnknownSampleTypeMappingError, type }
+
+          # Add default mapping
+          builder.mappings.fetch($PROGRAM_NAME, &builder.method(:build_mapping))
+
+          # Combine all sample types from each converter class
+          types = mappings.values.each_with_object({}) do |converter_class, t|
+            t.merge!(converter_class.sample_value_types)
+          end
+
+          # Build the sample types into sample type objects
+          types.each do |type_name, type_args|
+            index = nil
+
+            sample_type = builder.sample_types.fetch(*type_args) do |id, type, unit|
+              index = id
+              builder.build_value_type(type, unit)
+            end
+
+            # Create mapping between the type and index to which its assigned.
+            # Do this for faster lookup while building profile sample values.
+            sample_type_mappings[type_name] = index || builder.sample_types.messages.index(sample_type)
+          end
+
+          # Freeze them so they can't be modified.
+          # We don't want the number of sample types to vary between samples within the same profile.
+          builder.sample_types.freeze
+          sample_type_mappings.freeze
+
+          # Add converters
+          mappings.each do |event_class, converter_class|
+            converters[event_class] = converter_class.new(builder, sample_type_mappings)
+          end
+
+          converters.freeze
+        end
+
+        def add_events!(event_class, events)
+          converters[event_class].add_events!(events)
+        end
+
+        def debug_statistics
+          converters.values.map(&:debug_statistics).join(', ')
+        end
+
+        def to_pprof(start:, finish:)
+          profile = builder.build_profile(start: start, finish: finish)
+          data = builder.encode_profile(profile)
+          types = sample_type_mappings.keys
+
+          Payload.new(data, types)
+        end
+
+        # Error when an unknown event type is given to be converted
+        class NoProfilingEventConversionError < ArgumentError
+          attr_reader :type
+
+          def initialize(type)
+            @type = type
+          end
+
+          def message
+            "Profiling event type '#{type}' cannot be converted to pprof."
+          end
+        end
+
+        # Error when the mapping of a sample type to value index is unknown
+        class UnknownSampleTypeMappingError < ArgumentError
+          attr_reader :type
+
+          def initialize(type)
+            @type = type
+          end
+
+          def message
+            "Mapping for sample value type '#{type}' is unknown."
+          end
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/preload.rb

@@ -0,0 +1,5 @@
+# typed: strict
+
+require_relative '../../ddtrace'
+
+Datadog::Profiling.start_if_enabled

data/lib/datadog/profiling/profiler.rb

@@ -0,0 +1,41 @@
+# typed: true
+
+module Datadog
+  module Profiling
+    # Profiling entry point, which coordinates collectors and a scheduler
+    class Profiler
+      include Datadog::Core::Utils::Forking
+
+      attr_reader \
+        :collectors,
+        :scheduler
+
+      def initialize(collectors, scheduler)
+        @collectors = collectors
+        @scheduler = scheduler
+      end
+
+      def start
+        after_fork! do
+          collectors.each(&:reset_after_fork)
+          scheduler.reset_after_fork
+        end
+
+        collectors.each(&:start)
+        scheduler.start
+      end
+
+      def shutdown!
+        Datadog.logger.debug('Shutting down profiler')
+
+        collectors.each do |collector|
+          collector.enabled = false
+          collector.stop(true)
+        end
+
+        scheduler.enabled = false
+        scheduler.stop(true)
+      end
+    end
+  end
+end

data/lib/datadog/profiling/scheduler.rb

@@ -0,0 +1,130 @@
+# typed: true
+
+require_relative '../core/utils/time'
+
+require_relative '../core/worker'
+require_relative '../core/workers/polling'
+
+module Datadog
+  module Profiling
+    # Periodically (every DEFAULT_INTERVAL_SECONDS) takes a profile from the `Exporter` and reports it using the
+    # configured transport. Runs on its own background thread.
+    class Scheduler < Core::Worker
+      include Core::Workers::Polling
+
+      DEFAULT_INTERVAL_SECONDS = 60
+      MINIMUM_INTERVAL_SECONDS = 0
+
+      # We sleep for at most this duration seconds before reporting data to avoid multi-process applications all
+      # reporting profiles at the exact same time
+      DEFAULT_FLUSH_JITTER_MAXIMUM_SECONDS = 3
+
+      private
+
+      attr_reader \
+        :exporter,
+        :transport
+
+      public
+
+      def initialize(
+        exporter:,
+        transport:,
+        fork_policy: Core::Workers::Async::Thread::FORK_POLICY_RESTART, # Restart in forks by default
+        interval: DEFAULT_INTERVAL_SECONDS,
+        enabled: true
+      )
+        @exporter = exporter
+        @transport = transport
+
+        # Workers::Async::Thread settings
+        self.fork_policy = fork_policy
+
+        # Workers::IntervalLoop settings
+        self.loop_base_interval = interval
+
+        # Workers::Polling settings
+        self.enabled = enabled
+      end
+
+      def start
+        perform
+      end
+
+      def perform
+        # A profiling flush may be called while the VM is shutting down, to report the last profile. When we do so,
+        # we impose a strict timeout. This means this last profile may or may not be sent, depending on if the flush can
+        # successfully finish in the strict timeout.
+        # This can be somewhat confusing (why did it not get reported?), so let's at least log what happened.
+        interrupted = true
+
+        begin
+          flush_and_wait
+          interrupted = false
+        ensure
+          Datadog.logger.debug('#flush was interrupted or failed before it could complete') if interrupted
+        end
+      end
+
+      # Configure Workers::IntervalLoop to not report immediately when scheduler starts
+      #
+      # When a scheduler gets created (or reset), we don't want it to immediately try to flush; we want it to wait for
+      # the loop wait time first. This avoids an issue where the scheduler reported a mostly-empty profile if the
+      # application just started but this thread took a bit longer so there's already profiling data in the exporter.
+      def loop_wait_before_first_iteration?
+        true
+      end
+
+      def work_pending?
+        exporter.can_flush?
+      end
+
+      def reset_after_fork
+        exporter.reset_after_fork
+      end
+
+      private
+
+      def flush_and_wait
+        run_time = Core::Utils::Time.measure do
+          flush_events
+        end
+
+        # Update wait time to try to wake consistently on time.
+        # Don't drop below the minimum interval.
+        self.loop_wait_time = [loop_base_interval - run_time, MINIMUM_INTERVAL_SECONDS].max
+      end
+
+      def flush_events
+        # Collect data to be exported
+        flush = exporter.flush
+
+        return false unless flush
+
+        # Sleep for a bit to cause misalignment between profilers in multi-process applications
+        #
+        # When not being run in a loop, it means the scheduler has not been started or was stopped, and thus
+        # a) it's being shut down (and is trying to report the last profile)
+        # b) it's being run as a one-shot, usually in a test
+        # ...so in those cases we don't sleep
+        #
+        # During PR review (https://github.com/DataDog/dd-trace-rb/pull/1807) we discussed the possible alternative of
+        # just sleeping before starting the scheduler loop. We ended up not going with that option to avoid the first
+        # profile containing up to DEFAULT_INTERVAL_SECONDS + DEFAULT_FLUSH_JITTER_MAXIMUM_SECONDS instead of the
+        # usual DEFAULT_INTERVAL_SECONDS size.
+        if run_loop?
+          jitter_seconds = rand * DEFAULT_FLUSH_JITTER_MAXIMUM_SECONDS # floating point number between (0.0...maximum)
+          sleep(jitter_seconds)
+        end
+
+        begin
+          transport.export(flush)
+        rescue StandardError => e
+          Datadog.logger.error("Unable to report profile. Cause: #{e} Location: #{Array(e.backtrace).first}")
+        end
+
+        true
+      end
+    end
+  end
+end

data/lib/datadog/profiling/stack_recorder.rb

@@ -0,0 +1,79 @@
+# typed: false
+
+module Datadog
+  module Profiling
+    # Stores stack samples in a native libdatadog data structure and expose Ruby-level serialization APIs
+    # Note that `record_sample` is only accessible from native code.
+    # Methods prefixed with _native_ are implemented in `stack_recorder.c`
+    class StackRecorder
+      def initialize
+        # This mutex works in addition to the fancy C-level mutexes we have in the native side (see the docs there).
+        # It prevents multiple Ruby threads calling serialize at the same time -- something like
+        # `10.times { Thread.new { stack_recorder.serialize } }`.
+        # This isn't something we expect to happen normally, but because it would break the assumptions of the
+        # C-level mutexes (that there is a single serializer thread), we add it here as an extra safeguard against it
+        # accidentally happening.
+        @no_concurrent_synchronize_mutex = Mutex.new
+      end
+
+      def serialize
+        status, result = @no_concurrent_synchronize_mutex.synchronize { self.class._native_serialize(self) }
+
+        if status == :ok
+          start, finish, encoded_pprof = result
+
+          Datadog.logger.debug { "Encoded profile covering #{start.iso8601} to #{finish.iso8601}" }
+
+          [start, finish, encoded_pprof]
+        else
+          error_message = result
+
+          Datadog.logger.error("Failed to serialize profiling data: #{error_message}")
+
+          nil
+        end
+      end
+
+      def serialize!
+        status, result = @no_concurrent_synchronize_mutex.synchronize { self.class._native_serialize(self) }
+
+        if status == :ok
+          _start, _finish, encoded_pprof = result
+
+          encoded_pprof
+        else
+          error_message = result
+
+          raise("Failed to serialize profiling data: #{error_message}")
+        end
+      end
+
+      def clear
+        status, result = @no_concurrent_synchronize_mutex.synchronize { self.class._native_clear(self) }
+
+        if status == :ok
+          finish_timestamp = result
+
+          Datadog.logger.debug { "Cleared profile at #{finish_timestamp}" }
+
+          finish_timestamp
+        else
+          error_message = result
+
+          Datadog.logger.error("Failed to clear profiling data: #{error_message}")
+
+          nil
+        end
+      end
+
+      # Used only for Ruby 2.2 which doesn't have the native `rb_time_timespec_new` API; called from native code
+      def self.ruby_time_from(timespec_seconds, timespec_nanoseconds)
+        Time.at(0).utc + timespec_seconds + (timespec_nanoseconds.to_r / 1_000_000_000)
+      end
+
+      def reset_after_fork
+        self.class._native_reset_after_fork(self)
+      end
+    end
+  end
+end

data/lib/datadog/profiling/tag_builder.rb

@@ -0,0 +1,53 @@
+# typed: true
+
+require_relative '../core/utils'
+
+module Datadog
+  module Profiling
+    # Builds a hash of default plus user tags to be included in a profile
+    module TagBuilder
+      include Datadog::Profiling::Ext::Transport::HTTP # Tag name constants
+
+      def self.call(
+        settings:,
+        # Unified service tagging
+        env: settings.env,
+        service: settings.service,
+        version: settings.version,
+        # Other metadata
+        host: Core::Environment::Socket.hostname,
+        language: Core::Environment::Identity.lang,
+        pid: Process.pid.to_s,
+        profiler_version: Core::Environment::Identity.tracer_version,
+        runtime_engine: Core::Environment::Identity.lang_engine,
+        runtime_id: Core::Environment::Identity.id,
+        runtime_platform: Core::Environment::Identity.lang_platform,
+        runtime_version: Core::Environment::Identity.lang_version,
+        # User-provided tags
+        user_tags: settings.tags
+      )
+        tags = {
+          # When changing or adding these, make sure they are kept in sync with
+          # https://docs.google.com/spreadsheets/d/1LOGMf4c4Avbtn36uZ2SWvhIGKRPLM1BoWkUP4JYj7hA/ (Datadog internal link)
+          FORM_FIELD_TAG_HOST => host,
+          FORM_FIELD_TAG_LANGUAGE => language,
+          FORM_FIELD_TAG_PID => pid,
+          FORM_FIELD_TAG_PROFILER_VERSION => profiler_version,
+          FORM_FIELD_TAG_RUNTIME => language, # This is known to be repeated from language, above
+          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,
+        }
+        tags[FORM_FIELD_TAG_ENV] = env if env
+        tags[FORM_FIELD_TAG_SERVICE] = service if service
+        tags[FORM_FIELD_TAG_VERSION] = version if version
+
+        # Make sure everything is an utf-8 string, to avoid encoding issues in native code/libddprof/further downstream
+        user_tags.merge(tags).map do |key, value|
+          [Datadog::Core::Utils.utf8_encode(key), Datadog::Core::Utils.utf8_encode(value)]
+        end.to_h
+      end
+    end
+  end
+end

data/lib/datadog/profiling/tasks/exec.rb

@@ -0,0 +1,50 @@
+# typed: true
+
+module Datadog
+  module Profiling
+    module Tasks
+      # Wraps command with Datadog tracing
+      class Exec
+        attr_reader :args
+
+        def initialize(args)
+          @args = args
+        end
+
+        def run
+          set_rubyopt!
+          exec_with_error_handling(args)
+        end
+
+        def rubyopts
+          [
+            '-rdatadog/profiling/preload'
+          ]
+        end
+
+        private
+
+        def set_rubyopt!
+          existing_rubyopt = ENV['RUBYOPT']
+
+          ENV['RUBYOPT'] = existing_rubyopt ? "#{existing_rubyopt} #{rubyopts.join(' ')}" : rubyopts.join(' ')
+        end
+
+        # If there's an error here, rather than throwing a cryptic stack trace, let's instead have clearer messages, and
+        # follow the same status codes as the shell uses
+        # See also:
+        # * https://www.gnu.org/software/bash/manual/html_node/Exit-Status.html
+        # * https://github.com/rubygems/rubygems/blob/dd93966cac224532035deda533cba2685dfa30cc/bundler/lib/bundler/cli/exec.rb#L45
+        def exec_with_error_handling(args)
+          Kernel.exec(*args)
+        rescue Errno::ENOENT => e
+          Kernel.warn "ddtracerb exec failed: #{e.class.name} #{e.message} (command was '#{args.join(' ')}')"
+          Kernel.exit 127
+        rescue Errno::EACCES, Errno::ENOEXEC => e
+          Kernel.warn "ddtracerb exec failed: #{e.class.name} #{e.message} (command was '#{args.join(' ')}')"
+          Kernel.exit 126
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/tasks/help.rb

@@ -0,0 +1,18 @@
+# typed: true
+
+module Datadog
+  module Profiling
+    module Tasks
+      # Prints help message for usage of `ddtrace`
+      class Help
+        def run
+          puts %(
+  Usage: ddtracerb [command] [arguments]
+    exec [command]: Executes command with tracing & profiling preloaded.
+    help:           Prints this help message.
+          )
+        end
+      end
+    end
+  end
+end