[PROF-4081] Make threads with empty backtrace visible

lib/ddtrace/profiling/collectors/stack.rb

@@ -19,6 +19,7 @@ class Stack < Worker # rubocop:disable Metrics/ClassLength
         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
@@ -119,6 +120,26 @@ 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)]

spec/ddtrace/profiling/collectors/stack_spec.rb

@@ -308,13 +308,13 @@
     let(:last_wall_time) { 42 }
     let(:current_wall_time) { 123 }
 
-    context 'when the backtrace is empty' do
+    context 'when the backtrace is nil' do
       let(:backtrace) { nil }
 
       it { is_expected.to be nil }
     end
 
-    context 'when the backtrace is not empty' do
+    context 'when the backtrace is not nil' do
       let(:backtrace) do
         Array.new(backtrace_size) do
           instance_double(
@@ -470,6 +470,20 @@
           end
         end
       end
+
+      context 'when the backtrace is empty' do
+        let(:backtrace) { [] }
+
+        it 'builds an event that includes a includes a synthetic placeholder frame to mark execution in native code' do
+          is_expected.to have_attributes(
+            total_frame_count: 1,
+            frames: [Datadog::Profiling::BacktraceLocation.new('', 0, 'In native code')],
+            timestamp: kind_of(Float),
+            thread_id: thread.object_id,
+            wall_time_interval_ns: current_wall_time - last_wall_time,
+          )
+        end
+      end
     end
 
     context 'Process::Waiter crash regression tests' do