ddtrace 0.53.0 → 1.0.0.beta1

data/lib/ddtrace/pipeline/span_filter.rb

@@ -1,39 +0,0 @@
-# typed: true
-module Datadog
-  module Pipeline
-    # SpanFilter implements a processor that filters entire span subtrees
-    class SpanFilter
-      def initialize(filter = nil, &block)
-        callable = filter || block
-
-        raise(ArgumentError) unless callable.respond_to?(:call)
-
-        @criteria = filter || block
-      end
-
-      # NOTE: this SpanFilter implementation only handles traces in which child spans appear
-      # after parent spans in the trace array. If in the future child spans can be before
-      # parent spans, then the code below will need to be updated.
-      def call(trace)
-        deleted = Set.new
-
-        trace.delete_if do |span|
-          if deleted.include?(span.parent)
-            deleted << span
-            true
-          else
-            drop = drop_it?(span)
-            deleted << span if drop
-            drop
-          end
-        end
-      end
-
-      private
-
-      def drop_it?(span)
-        @criteria.call(span) rescue false
-      end
-    end
-  end
-end

data/lib/ddtrace/pipeline/span_processor.rb

@@ -1,21 +0,0 @@
-# typed: true
-module Datadog
-  module Pipeline
-    # SpanProcessor
-    class SpanProcessor
-      def initialize(operation = nil, &block)
-        callable = operation || block
-
-        raise(ArgumentError) unless callable.respond_to?(:call)
-
-        @operation = operation || block
-      end
-
-      def call(trace)
-        trace.each do |span|
-          @operation.call(span) rescue next
-        end
-      end
-    end
-  end
-end

data/lib/ddtrace/pipeline.rb

@@ -1,47 +0,0 @@
-# typed: true
-module Datadog
-  # Pipeline
-  module Pipeline
-    require_relative 'pipeline/span_filter'
-    require_relative 'pipeline/span_processor'
-
-    @mutex = Mutex.new
-    @processors = []
-
-    def self.before_flush(*processors, &processor_block)
-      processors = [processor_block] if processors.empty?
-
-      @mutex.synchronize do
-        @processors.concat(processors)
-      end
-    end
-
-    def self.process!(traces)
-      @mutex.synchronize do
-        traces
-          .map(&method(:apply_processors!))
-          .select(&:any?)
-      end
-    end
-
-    def self.processors=(value)
-      @processors = value
-    end
-
-    def self.apply_processors!(trace)
-      result = @processors.inject(trace) do |current_trace, processor|
-        processor.call(current_trace)
-      end
-
-      result || []
-    rescue => e
-      Datadog.logger.debug(
-        "trace dropped entirely due to `Pipeline.before_flush` error: #{e}"
-      )
-
-      []
-    end
-
-    private_class_method :apply_processors!
-  end
-end

data/lib/ddtrace/profiling/buffer.rb

@@ -1,42 +0,0 @@
-# typed: true
-require 'ddtrace/buffer'
-require 'ddtrace/utils/string_table'
-require 'ddtrace/utils/object_set'
-
-module Datadog
-  module Profiling
-    # Profiling buffer that stores profiling events. The buffer has a maximum size and when
-    # the buffer is full, a random event is discarded. This class is thread-safe.
-    class Buffer < Datadog::ThreadSafeBuffer
-      def initialize(*args)
-        super
-        @caches = {}
-        @string_table = Utils::StringTable.new
-      end
-
-      def cache(cache_name)
-        synchronize do
-          @caches[cache_name] ||= Utils::ObjectSet.new
-        end
-      end
-
-      def string_table
-        synchronize do
-          @string_table
-        end
-      end
-
-      protected
-
-      def drain!
-        items = super
-
-        # Clear caches
-        @caches = {}
-        @string_table = Utils::StringTable.new
-
-        items
-      end
-    end
-  end
-end

data/lib/ddtrace/profiling/collectors/stack.rb

@@ -1,301 +0,0 @@
-# typed: true
-require 'ddtrace/profiling/backtrace_location'
-require 'ddtrace/profiling/events/stack'
-require 'ddtrace/utils/only_once'
-require 'ddtrace/utils/time'
-require 'ddtrace/worker'
-require 'ddtrace/workers/polling'
-
-module Datadog
-  module Profiling
-    module Collectors
-      # Collects stack trace samples from Ruby threads for both CPU-time (if available) and wall-clock.
-      # Runs on its own background thread.
-      #
-      class Stack < Worker # rubocop:disable Metrics/ClassLength
-        include Workers::Polling
-
-        DEFAULT_MAX_TIME_USAGE_PCT = 2.0
-        MIN_INTERVAL = 0.01
-        THREAD_LAST_CPU_TIME_KEY = :datadog_profiler_last_cpu_time
-        THREAD_LAST_WALL_CLOCK_KEY = :datadog_profiler_last_wall_clock
-
-        # This default was picked based on the current sampling performance and on expected concurrency on an average
-        # Ruby MRI application. Lowering this optimizes for latency (less impact each time we sample), and raising
-        # optimizes for coverage (less chance to miss what a given thread is doing).
-        DEFAULT_MAX_THREADS_SAMPLED = 16
-
-        attr_reader \
-          :recorder,
-          :max_frames,
-          :trace_identifiers_helper,
-          :ignore_thread,
-          :max_time_usage_pct,
-          :thread_api
-
-        def initialize(
-          recorder,
-          max_frames:,
-          trace_identifiers_helper:, # Usually an instance of Datadog::Profiling::TraceIdentifiers::Helper
-          ignore_thread: nil,
-          max_time_usage_pct: DEFAULT_MAX_TIME_USAGE_PCT,
-          max_threads_sampled: DEFAULT_MAX_THREADS_SAMPLED,
-          thread_api: Thread,
-          fork_policy: Workers::Async::Thread::FORK_POLICY_RESTART, # Restart in forks by default
-          interval: MIN_INTERVAL,
-          enabled: true
-        )
-          @recorder = recorder
-          @max_frames = max_frames
-          @trace_identifiers_helper = trace_identifiers_helper
-          @ignore_thread = ignore_thread
-          @max_time_usage_pct = max_time_usage_pct
-          @max_threads_sampled = max_threads_sampled
-          @thread_api = thread_api
-
-          # Workers::Async::Thread settings
-          self.fork_policy = fork_policy
-
-          # Workers::IntervalLoop settings
-          self.loop_base_interval = interval
-
-          # Workers::Polling settings
-          self.enabled = enabled
-
-          @warn_about_missing_cpu_time_instrumentation_only_once = Datadog::Utils::OnlyOnce.new
-
-          # Cache this proc, since it's pretty expensive to keep recreating it
-          @build_backtrace_location = method(:build_backtrace_location).to_proc
-          # Cache this buffer, since it's pretty expensive to keep accessing it
-          @stack_sample_event_recorder = recorder[Events::StackSample]
-          # See below for details on why this is needed
-          @needs_process_waiter_workaround =
-            Gem::Version.new(RUBY_VERSION) >= Gem::Version.new('2.3') &&
-            Gem::Version.new(RUBY_VERSION) < Gem::Version.new('2.7')
-        end
-
-        def start
-          reset_cpu_time_tracking
-          perform
-        end
-
-        def perform
-          collect_and_wait
-        end
-
-        def loop_back_off?
-          false
-        end
-
-        def collect_and_wait
-          run_time = Datadog::Utils::Time.measure do
-            collect_events
-          end
-
-          # Update wait time to throttle profiling
-          self.loop_wait_time = compute_wait_time(run_time)
-        end
-
-        def collect_events
-          events = []
-          current_wall_time_ns = get_current_wall_time_timestamp_ns
-
-          # Collect backtraces from each thread
-          threads_to_sample.each do |thread|
-            next unless thread.alive?
-            next if ignore_thread.is_a?(Proc) && ignore_thread.call(thread)
-
-            event = collect_thread_event(thread, current_wall_time_ns)
-            events << event unless event.nil?
-          end
-
-          # Send events to recorder
-          recorder.push(events) unless events.empty?
-
-          events
-        end
-
-        def collect_thread_event(thread, current_wall_time_ns)
-          locations = thread.backtrace_locations
-          return if locations.nil?
-
-          # Get actual stack size then trim the stack
-          stack_size = locations.length
-          locations = locations[0..(max_frames - 1)]
-
-          # Convert backtrace locations into structs
-          locations = convert_backtrace_locations(locations)
-
-          thread_id = thread.respond_to?(:pthread_thread_id) ? thread.pthread_thread_id : thread.object_id
-          trace_id, span_id, trace_resource = trace_identifiers_helper.trace_identifiers_for(thread)
-          cpu_time = get_cpu_time_interval!(thread)
-          wall_time_interval_ns =
-            get_elapsed_since_last_sample_and_set_value(thread, THREAD_LAST_WALL_CLOCK_KEY, current_wall_time_ns)
-
-          Events::StackSample.new(
-            nil,
-            locations,
-            stack_size,
-            thread_id,
-            trace_id,
-            span_id,
-            trace_resource,
-            cpu_time,
-            wall_time_interval_ns
-          )
-        end
-
-        def get_cpu_time_interval!(thread)
-          # Return if we can't get the current CPU time
-          unless thread.respond_to?(:cpu_time_instrumentation_installed?) && thread.cpu_time_instrumentation_installed?
-            warn_about_missing_cpu_time_instrumentation(thread)
-            return
-          end
-
-          current_cpu_time_ns = thread.cpu_time(:nanosecond)
-
-          # NOTE: This can still be nil even when all of the checks above passed because of a race: there's a bit of
-          # initialization that needs to be done by the thread itself, and it's possible for us to try to sample
-          # *before* the thread had time to finish the initialization
-          return unless current_cpu_time_ns
-
-          get_elapsed_since_last_sample_and_set_value(thread, THREAD_LAST_CPU_TIME_KEY, current_cpu_time_ns)
-        end
-
-        def compute_wait_time(used_time)
-          # We took used_time to get the last sample.
-          #
-          # What we're computing here is -- if used_time corresponds to max_time_usage_pct of the time we should
-          # spend working, how much is (100% - max_time_usage_pct) of the time?
-          #
-          # For instance, if we took 10ms to sample, and max_time_usage_pct is 1%, then the other 99% is 990ms, which
-          # means we need to sleep for 990ms to guarantee that we don't spend more than 1% of the time working.
-          used_time_ns = used_time * 1e9
-          interval = (used_time_ns / (max_time_usage_pct / 100.0)) - used_time_ns
-          [interval / 1e9, MIN_INTERVAL].max
-        end
-
-        # Convert backtrace locations into structs
-        # Re-use old backtrace location objects if they already exist in the buffer
-        def convert_backtrace_locations(locations)
-          locations.collect do |location|
-            # Re-use existing BacktraceLocation if identical copy, otherwise build a new one.
-            @stack_sample_event_recorder.cache(:backtrace_locations).fetch(
-              # Function name
-              location.base_label,
-              # Line number
-              location.lineno,
-              # Filename
-              location.path,
-              # Build function
-              &@build_backtrace_location
-            )
-          end
-        end
-
-        def build_backtrace_location(_id, base_label, lineno, path)
-          string_table = @stack_sample_event_recorder.string_table
-
-          Profiling::BacktraceLocation.new(
-            string_table.fetch_string(base_label),
-            lineno,
-            string_table.fetch_string(path)
-          )
-        end
-
-        private
-
-        def warn_about_missing_cpu_time_instrumentation(thread)
-          @warn_about_missing_cpu_time_instrumentation_only_once.run do
-            # Is the profiler thread instrumented? If it is, then we know instrumentation is available, but seems to be
-            # missing on this thread we just found.
-            #
-            # As far as we know, it can be missing due to one the following:
-            #
-            # a) The thread was started before we installed our instrumentation.
-            #    In this case, the fix is to make sure ddtrace gets loaded before any other parts of the application.
-            #
-            # b) The thread was started using the Ruby native APIs (e.g. from a C extension such as ffi).
-            #    Known cases right now that trigger this are the ethon/typhoeus gems.
-            #    We currently have no solution for this case; these threads will always be missing our CPU instrumentation.
-            #
-            # c) The thread was started with `Thread.start`/`Thread.fork` and hasn't yet enabled the instrumentation.
-            #    When threads are started using these APIs, there's a small time window during which the thread has started
-            #    but our code to apply the instrumentation hasn't run yet; in these cases it's just a matter of allowing
-            #    it to run and our instrumentation to be applied.
-            #
-            if thread_api.current.respond_to?(:cpu_time) && thread_api.current.cpu_time
-              Datadog.logger.debug(
-                "Thread ('#{thread}') is missing profiling instrumentation; other threads should be unaffected"
-              )
-            end
-          end
-        end
-
-        # If the profiler is started for a while, stopped and then restarted OR whenever the process forks, we need to
-        # clean up any leftover per-thread counters, so that the first sample after starting doesn't end up with:
-        #
-        # a) negative time: At least on my test docker container, and on the reliability environment, after the process
-        #    forks, the cpu time reference changes and (old cpu time - new cpu time) can be < 0
-        #
-        # b) large amount of time: if the profiler was started, then stopped for some amount of time, and then
-        #    restarted, we don't want the first sample to be "blamed" for multiple minutes of CPU time
-        #
-        # By resetting the last cpu time seen, we start with a clean slate every time we start the stack collector.
-        def reset_cpu_time_tracking
-          thread_api.list.each do |thread|
-            # See below for details on why this is needed
-            next if @needs_process_waiter_workaround && thread.is_a?(::Process::Waiter)
-
-            thread.thread_variable_set(THREAD_LAST_CPU_TIME_KEY, nil)
-            thread.thread_variable_set(THREAD_LAST_WALL_CLOCK_KEY, nil)
-          end
-        end
-
-        def get_elapsed_since_last_sample_and_set_value(thread, key, current_value)
-          # See cthread.rb for more details, but this is a workaround for https://bugs.ruby-lang.org/issues/17807 ;
-          # using all thread_variable related methods on these instances also triggers a crash and for now we just
-          # skip it for the affected Rubies
-          return 0 if @needs_process_waiter_workaround && thread.is_a?(::Process::Waiter)
-
-          last_value = thread.thread_variable_get(key) || current_value
-          thread.thread_variable_set(key, current_value)
-
-          current_value - last_value
-        end
-
-        # Whenever there are more than max_threads_sampled active, we only sample a subset of them.
-        # We do this to avoid impacting the latency of the service being profiled. We want to avoid doing
-        # a big burst of work all at once (sample everything), and instead do a little work each time
-        # (sample a bit by bit).
-        #
-        # Because we pick the threads to sample randomly, we'll eventually sample all threads -- just not at once.
-        # Notice also that this will interact with our dynamic sampling mechanism -- if samples are faster, we take
-        # them more often, if they are slower, we take them less often -- which again means that over a longer period
-        # we should take sample roughly the same samples.
-        #
-        # One downside of this approach is that if there really are many threads, the resulting wall clock times
-        # in a one minute profile may "drift" around the 60 second mark, e.g. maybe we only sampled a thread once per
-        # second and only 59 times, so we'll report 59s, but on the next report we'll include the missing one, so
-        # then the result will be 61s. I've observed 60 +- 1.68 secs for an app with ~65 threads, given the
-        # default maximum of 16 threads. This seems a reasonable enough margin of error given the improvement to
-        # latency (especially on such a large application! -> even bigger latency impact if we tried to sample all
-        # threads).
-        #
-        def threads_to_sample
-          all_threads = thread_api.list
-
-          if all_threads.size > @max_threads_sampled
-            all_threads.sample(@max_threads_sampled)
-          else
-            all_threads
-          end
-        end
-
-        def get_current_wall_time_timestamp_ns
-          Datadog::Utils::Time.get_time(:nanosecond)
-        end
-      end
-    end
-  end
-end

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

@@ -1,46 +0,0 @@
-# typed: true
-require 'set'
-require 'time'
-
-require 'ddtrace/profiling/flush'
-require 'ddtrace/profiling/pprof/template'
-
-module Datadog
-  module Profiling
-    module Encoding
-      module Profile
-        # Encodes gathered data into the pprof format
-        module Protobuf
-          module_function
-
-          def encode(flush)
-            return unless flush
-
-            # Create a pprof template from the list of event types
-            event_classes = flush.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) }
-
-            Datadog.logger.debug do
-              max_events = Datadog.configuration.profiling.advanced.max_events
-              events_sampled =
-                if flush.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})"
-            end
-
-            # Build the profile and encode it
-            template.to_pprof
-          end
-        end
-      end
-    end
-  end
-end

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

@@ -1,81 +0,0 @@
-# typed: true
-require 'ddtrace/profiling/event'
-
-module Datadog
-  module Profiling
-    module Events
-      # Describes a stack profiling event
-      class Stack < Event
-        attr_reader \
-          :hash,
-          :frames,
-          :total_frame_count,
-          :thread_id,
-          :trace_id,
-          :span_id,
-          :trace_resource
-
-        def initialize(
-          timestamp,
-          frames,
-          total_frame_count,
-          thread_id,
-          trace_id,
-          span_id,
-          trace_resource
-        )
-          super(timestamp)
-
-          @frames = frames
-          @total_frame_count = total_frame_count
-          @thread_id = thread_id
-          @trace_id = trace_id
-          @span_id = span_id
-          @trace_resource = trace_resource
-
-          @hash = [
-            thread_id,
-            trace_id,
-            span_id,
-            # trace_resource is deliberately not included -- events that share the same (trace_id, span_id)
-            # trace_resource might not match between pairs, but they refer to the same trace.
-            frames.collect(&:hash),
-            total_frame_count
-          ].hash
-        end
-      end
-
-      # Describes a stack sample
-      class StackSample < Stack
-        attr_reader \
-          :cpu_time_interval_ns,
-          :wall_time_interval_ns
-
-        def initialize(
-          timestamp,
-          frames,
-          total_frame_count,
-          thread_id,
-          trace_id,
-          span_id,
-          trace_resource,
-          cpu_time_interval_ns,
-          wall_time_interval_ns
-        )
-          super(
-            timestamp,
-            frames,
-            total_frame_count,
-            thread_id,
-            trace_id,
-            span_id,
-            trace_resource
-          )
-
-          @cpu_time_interval_ns = cpu_time_interval_ns
-          @wall_time_interval_ns = wall_time_interval_ns
-        end
-      end
-    end
-  end
-end

data/lib/ddtrace/profiling/exporter.rb

@@ -1,24 +0,0 @@
-# typed: true
-require 'ddtrace/profiling/transport/io/client'
-
-module Datadog
-  module Profiling
-    # Writes profiling data to a given transport
-    class Exporter
-      attr_reader \
-        :transport
-
-      def initialize(transport)
-        unless transport.is_a?(Profiling::Transport::Client)
-          raise ArgumentError, 'Unsupported transport for profiling exporter.'
-        end
-
-        @transport = transport
-      end
-
-      def export(flush)
-        transport.send_profiling_flush(flush)
-      end
-    end
-  end
-end

data/lib/ddtrace/profiling/ext/cpu.rb

@@ -1,67 +0,0 @@
-# typed: true
-module Datadog
-  module Profiling
-    module Ext
-      # Monkey patches Ruby's `Thread` with our `Ext::CThread` to enable CPU-time profiling
-      module CPU
-        # We cannot apply our CPU extension if a broken rollbar is around because that can cause customer apps to fail
-        # with a SystemStackError: stack level too deep.
-        #
-        # This occurs whenever our extensions to Thread are applied BEFORE rollbar applies its own. This happens
-        # because a loop forms: our extension tries to call Thread#initialize, but it's intercepted by rollbar, which
-        # then tries to call the original Thread#initialize as well, but instead alls our extension, leading to stack
-        # exhaustion.
-        #
-        # See https://github.com/rollbar/rollbar-gem/pull/1018 for more details on the issue
-        ROLLBAR_INCOMPATIBLE_VERSIONS = Gem::Requirement.new('<= 3.1.1')
-
-        def self.supported?
-          unsupported_reason.nil?
-        end
-
-        def self.apply!
-          return false unless supported?
-
-          # Applying CThread to Thread will ensure any new threads
-          # will provide a thread/clock ID for CPU timing.
-          require 'ddtrace/profiling/ext/cthread'
-          ::Thread.prepend(Profiling::Ext::CThread)
-          ::Thread.singleton_class.prepend(Datadog::Profiling::Ext::WrapThreadStartFork)
-        end
-
-        def self.unsupported_reason
-          # NOTE: Only the first matching reason is returned, so try to keep a nice order on reasons -- e.g. tell users
-          # first that they can't use this on macOS before telling them that they have the wrong ffi version
-
-          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"
-          elsif Gem::Specification.find_all_by_name('rollbar', ROLLBAR_INCOMPATIBLE_VERSIONS).any?
-            'You have an incompatible rollbar gem version installed; ensure that you have rollbar >= 3.1.2 by ' \
-            "adding `gem 'rollbar', '>= 3.1.2'` to your Gemfile or gems.rb file. " \
-            'See https://github.com/rollbar/rollbar-gem/pull/1018 for details'
-          elsif Gem::Specification.find_all_by_name('logging').any? && logging_inherit_context_enabled?
-            'The `logging` gem is installed and its thread inherit context feature is enabled. ' \
-            "Please add LOGGING_INHERIT_CONTEXT=false to your application's environment variables to disable the " \
-            'conflicting `logging` gem feature. ' \
-            'See https://github.com/TwP/logging/pull/230 for details'
-          end
-        end
-
-        private_class_method def self.logging_inherit_context_enabled?
-          # The logging gem provides a mechanism to disable the conflicting behavior, see
-          # https://github.com/TwP/logging/blob/ae9872d093833b2a5a34cbe1faa4e895a81f6845/lib/logging/diagnostic_context.rb#L418
-          # Here we check if the behavior is enabled
-          inherit_context_configuration = ENV['LOGGING_INHERIT_CONTEXT']
-
-          inherit_context_configuration.nil? || !%w[false no 0].include?(inherit_context_configuration.downcase)
-        end
-      end
-    end
-  end
-end