ddtrace 0.51.1 → 0.54.2

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

@@ -1,3 +1,6 @@
+# typed: true
+
+require 'ddtrace/profiling/native_extension'
 require 'ddtrace/profiling/backtrace_location'
 require 'ddtrace/profiling/events/stack'
 require 'ddtrace/utils/only_once'
@@ -11,36 +14,51 @@ module Datadog
       # Collects stack trace samples from Ruby threads for both CPU-time (if available) and wall-clock.
       # Runs on its own background thread.
       #
-      # rubocop:disable Metrics/ClassLength
-      class Stack < Worker
+      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
+          :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
@@ -51,16 +69,17 @@ module Datadog
           # 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
-          @last_wall_time = Datadog::Utils::Time.get_time
           reset_cpu_time_tracking
           perform
         end
@@ -69,10 +88,6 @@ module Datadog
           collect_and_wait
         end
 
-        def loop_back_off?
-          false
-        end
-
         def collect_and_wait
           run_time = Datadog::Utils::Time.measure do
             collect_events
@@ -84,24 +99,14 @@ module Datadog
 
         def collect_events
           events = []
-
-          # Compute wall time interval
-          current_wall_time = Datadog::Utils::Time.get_time
-          last_wall_time = if instance_variable_defined?(:@last_wall_time)
-                             @last_wall_time
-                           else
-                             current_wall_time
-                           end
-
-          wall_time_interval_ns = ((current_wall_time - last_wall_time).round(9) * 1e9).to_i
-          @last_wall_time = current_wall_time
+          current_wall_time_ns = get_current_wall_time_timestamp_ns
 
           # Collect backtraces from each thread
-          thread_api.list.each do |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, wall_time_interval_ns)
+            event = collect_thread_event(thread, current_wall_time_ns)
             events << event unless event.nil?
           end
 
@@ -111,10 +116,30 @@ module Datadog
           events
         end
 
-        def collect_thread_event(thread, wall_time_interval_ns)
+        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)]
@@ -122,52 +147,33 @@ module Datadog
           # 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 = get_trace_identifiers(thread)
+          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,
-            trace_id,
+            root_span_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
+          return unless cpu_time_provider
 
-          current_cpu_time_ns = thread.cpu_time(:nanosecond)
+          current_cpu_time_ns = cpu_time_provider.cpu_time_ns_for(thread)
 
-          # 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
 
-          last_cpu_time_ns = (thread[THREAD_LAST_CPU_TIME_KEY] || current_cpu_time_ns)
-          interval = current_cpu_time_ns - last_cpu_time_ns
-
-          # Update CPU time for thread
-          thread[THREAD_LAST_CPU_TIME_KEY] = current_cpu_time_ns
-
-          # Return interval
-          interval
-        end
-
-        def get_trace_identifiers(thread)
-          return unless thread.is_a?(::Thread)
-          return unless Datadog.respond_to?(:tracer) && Datadog.tracer.respond_to?(:active_correlation)
-
-          identifier = Datadog.tracer.active_correlation(thread)
-          [identifier.trace_id, identifier.span_id]
+          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)
@@ -213,38 +219,11 @@ module Datadog
 
         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 the per-thread cpu time counters we keep, so that the first sample after starting doesn't end up with:
+        # 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 clock reference changes and (old cpu time - new cpu time) can be < 0
+        #    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
@@ -252,11 +231,67 @@ module Datadog
         # 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|
-            thread[THREAD_LAST_CPU_TIME_KEY] = nil if thread[THREAD_LAST_CPU_TIME_KEY]
+            # 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
-      # rubocop:enable Metrics/ClassLength
     end
   end
 end

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

@@ -1,3 +1,4 @@
+# typed: true
 require 'set'
 require 'time'
 
@@ -23,12 +24,20 @@ module Datadog
             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} (#{template.debug_statistics})"
+              "events: #{flush.event_count} (#{events_sampled}#{template.debug_statistics})"
             end
 
             # Build the profile and encode it
-            template.to_pprof
+            template.to_pprof(start: flush.start, finish: flush.finish)
           end
         end
       end

data/lib/ddtrace/profiling/event.rb

@@ -1,3 +1,4 @@
+# typed: true
 module Datadog
   module Profiling
     # Describes a sample of some data obtained from the runtime.

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

@@ -1,3 +1,4 @@
+# typed: true
 require 'ddtrace/profiling/event'
 
 module Datadog
@@ -6,37 +7,40 @@ module Datadog
       # Describes a stack profiling event
       class Stack < Event
         attr_reader \
-          :frames,
           :hash,
-          :span_id,
-          :thread_id,
+          :frames,
           :total_frame_count,
-          :trace_id
+          :thread_id,
+          :root_span_id,
+          :span_id,
+          :trace_resource
 
         def initialize(
           timestamp,
           frames,
           total_frame_count,
           thread_id,
-          trace_id,
-          span_id
+          root_span_id,
+          span_id,
+          trace_resource
         )
           super(timestamp)
 
           @frames = frames
           @total_frame_count = total_frame_count
           @thread_id = thread_id
-          @trace_id = trace_id
+          @root_span_id = root_span_id
           @span_id = span_id
+          @trace_resource = trace_resource
 
           @hash = [
             thread_id,
-            trace_id,
+            root_span_id,
             span_id,
-            [
-              frames.collect(&:hash),
-              total_frame_count
-            ]
+            # 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
@@ -52,8 +56,9 @@ module Datadog
           frames,
           total_frame_count,
           thread_id,
-          trace_id,
+          root_span_id,
           span_id,
+          trace_resource,
           cpu_time_interval_ns,
           wall_time_interval_ns
         )
@@ -62,41 +67,15 @@ module Datadog
             frames,
             total_frame_count,
             thread_id,
-            trace_id,
-            span_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
-
-      # Describes a stack sample with exception
-      class StackExceptionSample < Stack
-        attr_reader \
-          :exception
-
-        def initialize(
-          timestamp,
-          frames,
-          total_frame_count,
-          thread_id,
-          trace_id,
-          span_id,
-          exception
-        )
-          super(
-            timestamp,
-            frames,
-            total_frame_count,
-            thread_id,
-            trace_id,
-            span_id
-          )
-
-          @exception = exception
-        end
-      end
     end
   end
 end

data/lib/ddtrace/profiling/exporter.rb

@@ -1,3 +1,4 @@
+# typed: true
 require 'ddtrace/profiling/transport/io/client'
 
 module Datadog

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

@@ -1,3 +1,4 @@
+# typed: false
 module Datadog
   module Profiling
     module Ext

data/lib/ddtrace/profiling/flush.rb

@@ -1,3 +1,4 @@
+# typed: false
 require 'datadog/core/environment/identity'
 require 'datadog/core/environment/socket'
 

data/lib/ddtrace/profiling/native_extension.rb

@@ -0,0 +1,40 @@
+# typed: true
+module Datadog
+  module Profiling
+    # This module contains classes and methods which are implemented using native code in the
+    # ext/ddtrace_profiling_native_extension folder, as well as some Ruby-level utilities that don't make sense to
+    # write using C
+    module NativeExtension
+      private_class_method def self.working?
+        native_working?
+      end
+
+      unless singleton_class.private_method_defined?(:native_working?)
+        private_class_method def self.native_working?
+          false
+        end
+      end
+
+      unless singleton_class.method_defined?(:clock_id_for)
+        def self.clock_id_for(_)
+          nil
+        end
+      end
+
+      def self.cpu_time_ns_for(thread)
+        clock_id =
+          begin
+            clock_id_for(thread)
+          rescue Errno::ESRCH
+            nil
+          end
+
+        begin
+          ::Process.clock_gettime(clock_id, :nanosecond) if clock_id
+        rescue Errno::EINVAL
+          nil
+        end
+      end
+    end
+  end
+end

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

@@ -1,8 +1,10 @@
+# typed: true
 # frozen_string_literal: true
 
 require 'ddtrace/profiling/flush'
 require 'ddtrace/profiling/pprof/message_set'
 require 'ddtrace/profiling/pprof/string_table'
+require 'ddtrace/utils/time'
 
 module Datadog
   module Profiling
@@ -46,14 +48,19 @@ module Datadog
           Perftools::Profiles::Profile.encode(profile).force_encoding(DEFAULT_ENCODING)
         end
 
-        def build_profile
+        def build_profile(start:, finish:)
+          start_ns = Datadog::Utils::Time.as_utc_epoch_ns(start)
+          finish_ns = Datadog::Utils::Time.as_utc_epoch_ns(finish)
+
           Perftools::Profiles::Profile.new(
             sample_type: @sample_types.messages,
             sample: @samples,
             mapping: @mappings.messages,
             location: @locations.values,
             function: @functions.messages,
-            string_table: @string_table.strings
+            string_table: @string_table.strings,
+            time_nanos: start_ns,
+            duration_nanos: finish_ns - start_ns,
           )
         end
 

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

@@ -1,3 +1,4 @@
+# typed: true
 require 'ddtrace/ext/profiling'
 
 module Datadog
@@ -24,20 +25,19 @@ module Datadog
           # [key, EventGroup]
           event_groups = {}
 
+          # Aggregate each event into a group
+          # with identical properties, but different values.
           events.each do |event|
             key = yield(event)
-            values = build_sample_values(event)
+            values = build_event_values(event)
 
             unless key.nil?
               if event_groups.key?(key)
-                # Update values for group
-                group_values = event_groups[key].values
-                group_values.each_with_index do |group_value, i|
-                  group_values[i] = group_value + values[i]
-                end
+                # Update existing group from event
+                update_group(event_groups[key], event, values)
               else
                 # Add new group
-                event_groups[key] = EventGroup.new(event, values)
+                event_groups[key] = new_group(event, values)
               end
             end
           end
@@ -56,7 +56,7 @@ module Datadog
           index
         end
 
-        def build_sample_values(stack_sample)
+        def build_event_values(event)
           # Build a value array that matches the length of the sample types
           # Populate all values with "no value" by default
           Array.new(@sample_type_mappings.length, Datadog::Ext::Profiling::Pprof::SAMPLE_VALUE_NO_VALUE)
@@ -68,7 +68,7 @@ module Datadog
 
         # Represents a grouped event
         # 'sample' is an example event object from the group.
-        # 'values' is the the summation of the group's sample values
+        # 'values' is the summation of the group's sample values
         EventGroup = Struct.new(:sample, :values)
 
         # Error when the mapping of a sample type to value index is unknown
@@ -83,6 +83,20 @@ module Datadog
             "Mapping for sample value type '#{type}' to index is unknown."
           end
         end
+
+        protected
+
+        def new_group(event, values)
+          EventGroup.new(event, values)
+        end
+
+        def update_group(event_group, event, values)
+          # Update values for group
+          group_values = event_group.values
+          group_values.each_with_index do |group_value, i|
+            group_values[i] = group_value + values[i]
+          end
+        end
       end
     end
   end

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

@@ -1,3 +1,4 @@
+# typed: true
 require 'ddtrace/utils/object_set'
 
 module Datadog
@@ -5,7 +6,9 @@ module Datadog
     module Pprof
       # Acts as a unique dictionary of protobuf messages
       class MessageSet < Utils::ObjectSet
-        alias_method :messages, :objects
+        def messages
+          objects
+        end
       end
     end
   end

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

@@ -1,10 +1,11 @@
+# typed: true
 module Datadog
   module Profiling
     module Pprof
       # Pprof output data.
       # Includes encoded data and list of types.
       Payload = Struct.new(:data, :types) do
-        def initialize(*args)
+        def initialize(data, types)
           super
           self.types = types || []
         end

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

@@ -1,3 +1,4 @@
+# typed: false
 # Generated by the protocol buffer compiler.  DO NOT EDIT!
 # source: lib/ddtrace/profiling/pprof/pprof.proto