ddtrace 0.40.0 → 0.54.1

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

@@ -0,0 +1,140 @@
+# typed: true
+require 'ddtrace/ext/profiling'
+require 'ddtrace/profiling/events/stack'
+require 'ddtrace/profiling/pprof/builder'
+require 'ddtrace/profiling/pprof/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: [
+            Datadog::Ext::Profiling::Pprof::VALUE_TYPE_CPU,
+            Datadog::Ext::Profiling::Pprof::VALUE_UNIT_NANOSECONDS
+          ],
+          wall_time_ns: [
+            Datadog::Ext::Profiling::Pprof::VALUE_TYPE_WALL,
+            Datadog::Ext::Profiling::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 = Datadog::Ext::Profiling::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(Datadog::Ext::Profiling::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(Datadog::Ext::Profiling::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(Datadog::Ext::Profiling::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(Datadog::Ext::Profiling::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/ddtrace/profiling/pprof/string_table.rb

@@ -0,0 +1,11 @@
+# typed: strict
+require 'ddtrace/utils/string_table'
+
+module Datadog
+  module Profiling
+    module Pprof
+      # Tracks strings and returns IDs
+      class StringTable < Utils::StringTable; end
+    end
+  end
+end

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

@@ -0,0 +1,119 @@
+# typed: false
+require 'ddtrace/profiling/pprof/payload'
+require 'ddtrace/profiling/pprof/message_set'
+require 'ddtrace/profiling/pprof/builder'
+
+require 'ddtrace/profiling/events/stack'
+require 'ddtrace/profiling/pprof/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/ddtrace/profiling/preload.rb

@@ -0,0 +1,4 @@
+# typed: strict
+require 'ddtrace'
+
+Datadog.profiler.start if Datadog.profiler

data/lib/ddtrace/profiling/profiler.rb

@@ -0,0 +1,31 @@
+# typed: true
+module Datadog
+  # Profiling entry point, which coordinates collectors and a scheduler
+  class Profiler
+    attr_reader \
+      :collectors,
+      :scheduler
+
+    def initialize(collectors, scheduler)
+      @collectors = collectors
+      @scheduler = scheduler
+    end
+
+    def start
+      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

data/lib/ddtrace/profiling/recorder.rb

@@ -0,0 +1,96 @@
+# typed: true
+require 'ddtrace/profiling/buffer'
+require 'ddtrace/profiling/flush'
+
+module Datadog
+  module Profiling
+    # Stores profiling events gathered by `Collector`s
+    class Recorder
+      attr_reader :max_size
+
+      def initialize(event_classes, max_size, last_flush_time: Time.now.utc)
+        @buffers = {}
+        @last_flush_time = last_flush_time
+        @max_size = max_size
+
+        # Add a buffer for each class
+        event_classes.each do |event_class|
+          @buffers[event_class] = Profiling::Buffer.new(max_size)
+        end
+
+        # Event classes can only be added ahead of time
+        @buffers.freeze
+      end
+
+      def [](event_class)
+        @buffers[event_class]
+      end
+
+      def push(events)
+        if events.is_a?(Array)
+          # Push multiple events
+          event_class = events.first.class
+          raise UnknownEventError, event_class unless @buffers.key?(event_class)
+
+          @buffers[event_class].concat(events)
+        else
+          # Push single event
+          event_class = events.class
+          raise UnknownEventError, event_class unless @buffers.key?(event_class)
+
+          @buffers[event_class].push(events)
+        end
+      end
+
+      def flush
+        event_count = 0
+
+        event_groups, start, finish = update_time do
+          @buffers.collect do |event_class, buffer|
+            events = buffer.pop
+            next if events.empty?
+
+            event_count += events.length
+            EventGroup.new(event_class, events)
+          end.compact
+        end
+
+        Flush.new(
+          start,
+          finish,
+          event_groups,
+          event_count
+        )
+      end
+
+      # NOTE: Remember that if the recorder is being accessed by multiple threads, this is an inherently racy operation.
+      def empty?
+        @buffers.values.all?(&:empty?)
+      end
+
+      # Error when event of an unknown type is used with the Recorder
+      class UnknownEventError < StandardError
+        attr_reader :event_class
+
+        def initialize(event_class)
+          @event_class = event_class
+        end
+
+        def message
+          @message ||= "Unknown event class '#{event_class}' for profiling recorder."
+        end
+      end
+
+      private
+
+      def update_time
+        start = @last_flush_time
+        result = yield
+        @last_flush_time = Time.now.utc
+
+        # Return event groups, start time, finish time
+        [result, start, @last_flush_time]
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/scheduler.rb

@@ -0,0 +1,134 @@
+# typed: true
+require 'ddtrace/utils/time'
+
+require 'ddtrace/worker'
+require 'ddtrace/workers/polling'
+
+module Datadog
+  module Profiling
+    # Periodically (every DEFAULT_INTERVAL_SECONDS) takes data from the `Recorder` and pushes them to all configured
+    # `Exporter`s. Runs on its own background thread.
+    class Scheduler < Worker
+      include Workers::Polling
+
+      DEFAULT_INTERVAL_SECONDS = 60
+      MINIMUM_INTERVAL_SECONDS = 0
+
+      # Profiles with duration less than this will not be reported
+      PROFILE_DURATION_THRESHOLD_SECONDS = 1
+
+      private_constant :DEFAULT_INTERVAL_SECONDS, :MINIMUM_INTERVAL_SECONDS, :PROFILE_DURATION_THRESHOLD_SECONDS
+
+      attr_reader \
+        :exporters,
+        :recorder
+
+      def initialize(
+        recorder,
+        exporters,
+        fork_policy: Workers::Async::Thread::FORK_POLICY_RESTART, # Restart in forks by default
+        interval: DEFAULT_INTERVAL_SECONDS,
+        enabled: true
+      )
+        @recorder = recorder
+        @exporters = [exporters].flatten
+
+        # 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
+
+      def loop_back_off?
+        false
+      end
+
+      def after_fork
+        # Clear recorder's buffers by flushing events.
+        # Objects from parent process will copy-on-write,
+        # and we don't want to send events for the wrong process.
+        recorder.flush
+      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 samples in the recorder.
+      def loop_wait_before_first_iteration?
+        true
+      end
+
+      def work_pending?
+        !recorder.empty?
+      end
+
+      private
+
+      def flush_and_wait
+        run_time = Datadog::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
+        # Get events from recorder
+        flush = recorder.flush
+
+        if duration_below_threshold?(flush)
+          Datadog.logger.debug do
+            "Skipped exporting profiling events as profile duration is below minimum (#{flush.event_count} events skipped)"
+          end
+
+          return flush
+        end
+
+        # Send events to each exporter
+        if flush.event_count > 0
+          exporters.each do |exporter|
+            begin
+              exporter.export(flush)
+            rescue StandardError => e
+              Datadog.logger.error(
+                "Unable to export #{flush.event_count} profiling events. Cause: #{e} Location: #{Array(e.backtrace).first}"
+              )
+            end
+          end
+        end
+
+        flush
+      end
+
+      def duration_below_threshold?(flush)
+        (flush.finish - flush.start) < PROFILE_DURATION_THRESHOLD_SECONDS
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/tasks/setup.rb

@@ -0,0 +1,90 @@
+# typed: false
+require 'ddtrace/utils/only_once'
+require 'ddtrace/profiling'
+require 'ddtrace/profiling/ext/forking'
+
+module Datadog
+  module Profiling
+    module Tasks
+      # Takes care of loading our extensions/monkey patches to handle fork() and validating if CPU-time profiling is usable
+      class Setup
+        ACTIVATE_EXTENSIONS_ONLY_ONCE = Datadog::Utils::OnlyOnce.new
+
+        def run
+          ACTIVATE_EXTENSIONS_ONLY_ONCE.run do
+            begin
+              check_if_cpu_time_profiling_is_supported
+              activate_forking_extensions
+              setup_at_fork_hooks
+            rescue StandardError, ScriptError => e
+              Datadog.logger.warn do
+                "Profiler extensions unavailable. Cause: #{e.message} Location: #{Array(e.backtrace).first}"
+              end
+            end
+          end
+        end
+
+        private
+
+        def activate_forking_extensions
+          if Ext::Forking.supported?
+            Ext::Forking.apply!
+          elsif Datadog.configuration.profiling.enabled
+            Datadog.logger.debug('Profiler forking extensions skipped; forking not supported.')
+          end
+        rescue StandardError, ScriptError => e
+          Datadog.logger.warn do
+            "Profiler forking extensions unavailable. Cause: #{e.message} Location: #{Array(e.backtrace).first}"
+          end
+        end
+
+        def check_if_cpu_time_profiling_is_supported
+          unsupported = cpu_time_profiling_unsupported_reason
+
+          if unsupported
+            Datadog.logger.info do
+              'CPU time profiling skipped because native CPU time is not supported: ' \
+              "#{unsupported}. Profiles containing 'Wall time' data will still be reported."
+            end
+          end
+        end
+
+        def setup_at_fork_hooks
+          if Process.respond_to?(:at_fork)
+            Process.at_fork(:child) do
+              begin
+                # When Ruby forks, clock IDs for each of the threads
+                # will change. We can only update these IDs from the
+                # execution context of the thread that owns it.
+                # This hook will update the IDs for the main thread
+                # after a fork occurs.
+                Thread.current.send(:update_native_ids) if Thread.current.respond_to?(:update_native_ids, true)
+
+                # Restart profiler, if enabled
+                Datadog.profiler.start if Datadog.profiler
+              rescue StandardError => e
+                Datadog.logger.warn do
+                  "Error during post-fork hooks. Cause: #{e.message} Location: #{Array(e.backtrace).first}"
+                end
+              end
+            end
+          end
+        end
+
+        def cpu_time_profiling_unsupported_reason
+          # NOTE: Only the first matching reason is returned, so try to keep a nice order on reasons
+
+          if RUBY_ENGINE == 'jruby'
+            'JRuby is not supported'
+          elsif RUBY_PLATFORM.include?('darwin')
+            'Feature requires Linux; macOS is not supported'
+          elsif RUBY_PLATFORM =~ /(mswin|mingw)/
+            'Feature requires Linux; Windows is not supported'
+          elsif !RUBY_PLATFORM.include?('linux')
+            "Feature requires Linux; #{RUBY_PLATFORM} is not supported"
+          end
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/trace_identifiers/ddtrace.rb

@@ -0,0 +1,42 @@
+# typed: true
+# frozen_string_literal: true
+
+require 'ddtrace/ext/http'
+
+module Datadog
+  module Profiling
+    module TraceIdentifiers
+      # Used by Datadog::Profiling::TraceIdentifiers::Helper to get the trace identifiers (root span id and span id)
+      # for a given thread, if there is an active trace for that thread in the supplied tracer object.
+      class Ddtrace
+        def initialize(tracer:)
+          @tracer = (tracer if tracer.respond_to?(:call_context))
+        end
+
+        def trace_identifiers_for(thread)
+          return unless @tracer
+
+          context = @tracer.call_context(thread)
+          return unless context
+
+          span, root_span = context.current_span_and_root_span
+          return unless span && root_span
+
+          root_span_id = root_span.span_id || 0
+          span_id = span.span_id || 0
+
+          [root_span_id, span_id, maybe_extract_resource(root_span)] if root_span_id != 0 && span_id != 0
+        end
+
+        private
+
+        # NOTE: Currently we're only interested in HTTP service endpoints. Over time, this list may be expanded.
+        # Resources MUST NOT include personal identifiable information (PII); this should not be the case with
+        # ddtrace integrations, but worth mentioning just in case :)
+        def maybe_extract_resource(root_span)
+          root_span.resource if root_span.span_type == Datadog::Ext::HTTP::TYPE_INBOUND
+        end
+      end
+    end
+  end
+end