ddtrace 0.42.0 → 0.54.2

data/lib/ddtrace/pin.rb

@@ -1,4 +1,6 @@
+# typed: true
 require 'ddtrace/patcher'
+require 'ddtrace/utils/only_once'
 
 # \Datadog global namespace that includes all tracing functionality for Tracer and Span classes.
 module Datadog
@@ -7,19 +9,15 @@ module Datadog
   # This is useful if you wanted to, say, trace two different
   # database clusters.
   class Pin
+    DEPRECATION_WARN_ONLY_ONCE = Datadog::Utils::OnlyOnce.new
+
     def self.get_from(obj)
       return nil unless obj.respond_to? :datadog_pin
+
       obj.datadog_pin
     end
 
-    attr_accessor :app
-    attr_accessor :app_type
-    attr_accessor :config
-    attr_accessor :name
-    attr_accessor :service_name
-    attr_accessor :tags
-    attr_reader :tracer
-    attr_accessor :writer
+    attr_accessor :app, :app_type, :config, :name, :service_name, :tags, :writer
 
     alias service= service_name=
     alias service service_name
@@ -42,6 +40,7 @@ module Datadog
 
     def enabled?
       return tracer.enabled if tracer
+
       false
     end
 
@@ -79,59 +78,8 @@ module Datadog
       ).freeze
 
     def deprecation_warning
-      log_deprecation_warning('Datadog::Pin.new')
-    end
-
-    include Datadog::Patcher
-
-    def log_deprecation_warning(method_name)
-      # Only log each deprecation warning once (safeguard against log spam)
-      do_once(method_name) do
-        Datadog.logger.warn("#{method_name}:#{DEPRECATION_WARNING}")
-      end
-    end
-  end
-
-  # Modification to Pin which logs deprecation warnings if accessed.
-  # Will be used by integrations which are phasing out the direct use of #datadog_pin.
-  module DeprecatedPin
-    include Datadog::Patcher
-
-    DEPRECATION_WARNING = %(
-      Use of Datadog::Pin is DEPRECATED.
-      Upgrade to the configuration API using the migration guide here:
-      https://github.com/DataDog/dd-trace-rb/releases/tag/v0.11.0).freeze
-
-    # Raise a deprecation warning when #datadog_pin or #datadog_pin= is accessed.
-    def onto(obj)
-      obj.instance_exec(self) do |pin|
-        @datadog_deprecated_pin = pin
-
-        unless respond_to? :datadog_pin=
-          def datadog_pin=(pin)
-            @datadog_deprecated_pin.log_deprecation_warning('#datadog_pin=')
-            @datadog_pin = pin
-          end
-        end
-
-        unless respond_to? :datadog_pin
-          def datadog_pin
-            @datadog_deprecated_pin.log_deprecation_warning('#datadog_pin')
-            @datadog_pin
-          end
-        end
-
-        # Set instance variable to avoid deprecation warnings
-        @datadog_pin = @datadog_deprecated_pin
-      end
-
-      self
-    end
-
-    def log_deprecation_warning(method_name)
-      # Only log each deprecation warning once (safeguard against log spam)
-      do_once(method_name) do
-        Datadog.logger.warn("#{method_name}:#{DEPRECATION_WARNING}")
+      DEPRECATION_WARN_ONLY_ONCE.run do
+        Datadog.logger.warn("Datadog::Pin.new:#{DEPRECATION_WARNING}")
       end
     end
   end

data/lib/ddtrace/pipeline/span_filter.rb

@@ -1,3 +1,4 @@
+# typed: true
 module Datadog
   module Pipeline
     # SpanFilter implements a processor that filters entire span subtrees
@@ -10,7 +11,7 @@ module Datadog
         @criteria = filter || block
       end
 
-      # Note: this SpanFilter implementation only handles traces in which child spans appear
+      # 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)

data/lib/ddtrace/pipeline/span_processor.rb

@@ -1,3 +1,4 @@
+# typed: true
 module Datadog
   module Pipeline
     # SpanProcessor

data/lib/ddtrace/pipeline.rb

@@ -1,3 +1,4 @@
+# typed: true
 module Datadog
   # Pipeline
   module Pipeline

data/lib/ddtrace/profiling/backtrace_location.rb

@@ -0,0 +1,33 @@
+# typed: true
+module Datadog
+  module Profiling
+    # Entity class used to represent an entry in a stack trace.
+    # Its fields are a simplified struct version of `Thread::Backtrace::Location`.
+    class BacktraceLocation
+      attr_reader \
+        :base_label,
+        :lineno,
+        :path,
+        :hash
+
+      def initialize(
+        base_label,
+        lineno,
+        path
+      )
+        @base_label = base_label
+        @lineno = lineno
+        @path = path
+        @hash = [base_label, lineno, path].hash
+      end
+
+      def ==(other)
+        hash == other.hash
+      end
+
+      def eql?(other)
+        hash == other.hash
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/buffer.rb

@@ -0,0 +1,42 @@
+# 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

@@ -0,0 +1,297 @@
+# typed: true
+
+require 'ddtrace/profiling/native_extension'
+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
+        SYNTHETIC_STACK_IN_NATIVE_CODE = [BacktraceLocation.new('', 0, 'In native code').freeze].freeze
+
+        # 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,
+          :cpu_time_provider
+
+        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,
+          cpu_time_provider: Datadog::Profiling::NativeExtension,
+          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
+          # Only set the provider if it's able to work in the current Ruby/OS combo
+          @cpu_time_provider = cpu_time_provider unless cpu_time_provider.cpu_time_ns_for(thread_api.current).nil?
+
+          # Workers::Async::Thread settings
+          self.fork_policy = fork_policy
+
+          # Workers::IntervalLoop settings
+          self.loop_base_interval = interval
+
+          # Workers::Polling settings
+          self.enabled = enabled
+
+          # 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 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?
+
+          # Having empty locations means that the thread is alive, but we don't know what it's doing:
+          #
+          # 1. It can be starting up
+          #    ```
+          #    > Thread.new { sleep }.backtrace
+          #    => [] # <-- note the thread hasn't actually started running sleep yet, we got there first
+          #    ```
+          # 2. It can be running native code
+          #    ```
+          #    > t = Process.detach(fork { sleep })
+          #    => #<Process::Waiter:0x00007ffe7285f7a0 run>
+          #    > t.backtrace
+          #    => [] # <-- this can happen even minutes later, e.g. it's not a race as in 1.
+          #    ```
+          #    This effect has been observed in threads created by the Iodine web server and the ffi gem
+          #
+          # To give customers visibility into these threads, we replace the empty stack with one containing a
+          # synthetic placeholder frame, so that these threads are properly represented in the UX.
+          locations = SYNTHETIC_STACK_IN_NATIVE_CODE if locations.empty?
+
+          # 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.object_id
+          root_span_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,
+            root_span_id,
+            span_id,
+            trace_resource,
+            cpu_time,
+            wall_time_interval_ns
+          )
+        end
+
+        def get_cpu_time_interval!(thread)
+          return unless cpu_time_provider
+
+          current_cpu_time_ns = cpu_time_provider.cpu_time_ns_for(thread)
+
+          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
+
+        # 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)
+          # Process::Waiter crash workaround:
+          #
+          # This is a workaround for a Ruby VM segfault (usually something like
+          # "[BUG] Segmentation fault at 0x0000000000000008") in the affected Ruby versions.
+          # See https://bugs.ruby-lang.org/issues/17807 for details.
+          #
+          # In those Ruby versions, there's a very special subclass of `Thread` called `Process::Waiter` that causes VM
+          # crashes whenever something tries to read its instance or thread variables. This subclass of thread only
+          # shows up when the `Process.detach` API gets used.
+          # In the specs you'll find crash regression tests that include a way of reproducing it.
+          #
+          # As workaround 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

@@ -0,0 +1,46 @@
+# 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(start: flush.start, finish: flush.finish)
+          end
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/event.rb

@@ -0,0 +1,14 @@
+# typed: true
+module Datadog
+  module Profiling
+    # Describes a sample of some data obtained from the runtime.
+    class Event
+      attr_reader \
+        :timestamp
+
+      def initialize(timestamp = nil)
+        @timestamp = timestamp || Time.now.utc.to_f
+      end
+    end
+  end
+end

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

@@ -0,0 +1,81 @@
+# 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,
+          :root_span_id,
+          :span_id,
+          :trace_resource
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          root_span_id,
+          span_id,
+          trace_resource
+        )
+          super(timestamp)
+
+          @frames = frames
+          @total_frame_count = total_frame_count
+          @thread_id = thread_id
+          @root_span_id = root_span_id
+          @span_id = span_id
+          @trace_resource = trace_resource
+
+          @hash = [
+            thread_id,
+            root_span_id,
+            span_id,
+            # trace_resource is deliberately not included -- events that share the same (root_span_id, span_id) 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,
+          root_span_id,
+          span_id,
+          trace_resource,
+          cpu_time_interval_ns,
+          wall_time_interval_ns
+        )
+          super(
+            timestamp,
+            frames,
+            total_frame_count,
+            thread_id,
+            root_span_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

@@ -0,0 +1,24 @@
+# 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