ddtrace 1.12.1 → 1.23.2

data/ext/{ddtrace_profiling_native_extension → datadog_profiling_native_extension}/collectors_thread_context.c

@@ -3,6 +3,7 @@
 #include "collectors_thread_context.h"
 #include "clock_id.h"
 #include "collectors_stack.h"
+#include "collectors_gc_profiling_helper.h"
 #include "helpers.h"
 #include "libdatadog_helpers.h"
 #include "private_vm_api_access.h"
@@ -37,24 +38,29 @@
 // When `thread_context_collector_on_gc_start` gets called, the current cpu and wall-time get recorded to the thread
 // context: `cpu_time_at_gc_start_ns` and `wall_time_at_gc_start_ns`.
 //
-// While these fields are set, regular samples (if any) do not account for any time that passes after these two
-// timestamps.
+// While `cpu_time_at_gc_start_ns` is set, regular samples (if any) do not account for cpu-time any time that passes
+// after this timestamp. The idea is that this cpu-time will be blamed separately on GC, and not on the user thread.
+// Wall-time accounting is not affected by this (e.g. we still record 60 seconds every 60 seconds).
 //
-// (Regular samples can still account for the time between the previous sample and the start of GC.)
+// (Regular samples can still account for the cpu-time between the previous sample and the start of GC.)
 //
-// When `thread_context_collector_on_gc_finish` gets called, the current cpu and wall-time again get recorded to the
-// thread context: `cpu_time_at_gc_finish_ns` and `wall_time_at_gc_finish_ns`.
+// When `thread_context_collector_on_gc_finish` gets called, the cpu-time and wall-time spent during GC gets recorded
+// into the global gc_tracking structure, and further samples are not affected. (The `cpu_time_at_previous_sample_ns`
+// of the thread that did GC also gets adjusted to avoid double-accounting.)
 //
-// Finally, when `thread_context_collector_sample_after_gc` gets called, the following happens:
+// Finally, when `thread_context_collector_sample_after_gc` gets called, a sample gets recorded with a stack having
+// a single placeholder `Garbage Collection` frame. This sample gets
+// assigned the cpu-time and wall-time that was recorded between calls to `on_gc_start` and `on_gc_finish`, as well
+// as metadata for the last GC.
 //
-// 1. A sample gets taken, using the special `SAMPLE_IN_GC` sample type, which produces a stack with a placeholder
-// `Garbage Collection` frame as the latest frame. This sample gets assigned the cpu-time and wall-time period that was
-// recorded between calls to `on_gc_start` and `on_gc_finish`.
-//
-// 2. The thread is no longer marked as being in gc (all gc tracking fields get reset back to `INVALID_TIME`).
-//
-// 3. The `cpu_time_at_previous_sample_ns` and `wall_time_at_previous_sample_ns` get updated with the elapsed time in
-// GC, so that all time is accounted for -- e.g. the next sample will not get "blamed" by time spent in GC.
+// Note that the Ruby GC does not usually do all of the GC work in one go. Instead, it breaks it up into smaller steps
+// so that the application can keep doing user work in between GC steps.
+// The `on_gc_start` / `on_gc_finish` will trigger each time the VM executes these smaller steps, and on a benchmark
+// that executes `Object.new` in a loop, I measured more than 50k of this steps per second (!!).
+// Creating these many events for every GC step is a lot of overhead, so instead `on_gc_finish` coalesces time
+// spent in GC and only flushes it at most every 10 ms/every complete GC collection. This reduces the amount of
+// individual GC events we need to record. We use the latest GC metadata for this event, reflecting the last GC that
+// happened in the coalesced period.
 //
 // In an earlier attempt at implementing this functionality (https://github.com/DataDog/dd-trace-rb/pull/2308), we
 // discovered that we needed to factor the sampling work away from `thread_context_collector_on_gc_finish` and into a
@@ -63,11 +69,12 @@
 // allowed to happen during Ruby's garbage collection start/finish hooks.
 // ---
 
-#define INVALID_TIME -1
 #define THREAD_ID_LIMIT_CHARS 44 // Why 44? "#{2**64} (#{2**64})".size + 1 for \0
+#define THREAD_INVOKE_LOCATION_LIMIT_CHARS 512
 #define IS_WALL_TIME true
 #define IS_NOT_WALL_TIME false
 #define MISSING_TRACER_CONTEXT_KEY 0
+#define TIME_BETWEEN_GC_EVENTS_NS MILLIS_AS_NS(10)
 
 static ID at_active_span_id;  // id of :@active_span in Ruby
 static ID at_active_trace_id; // id of :@active_trace in Ruby
@@ -75,6 +82,9 @@ static ID at_id_id;           // id of :@id in Ruby
 static ID at_resource_id;     // id of :@resource in Ruby
 static ID at_root_span_id;    // id of :@root_span in Ruby
 static ID at_type_id;         // id of :@type in Ruby
+static ID at_otel_values_id;  // id of :@otel_values in Ruby
+static ID at_parent_span_id_id; // id of :@parent_span_id in Ruby
+static ID at_datadog_trace_id;  // id of :@datadog_trace in Ruby
 
 // Contains state for a single ThreadContext instance
 struct thread_context_collector_state {
@@ -99,6 +109,16 @@ struct thread_context_collector_state {
   VALUE thread_list_buffer;
   // Used to omit endpoint names (retrieved from tracer) from collected data
   bool endpoint_collection_enabled;
+  // Used to omit timestamps / timeline events from collected data
+  bool timeline_enabled;
+  // Used to omit class information from collected allocation data
+  bool allocation_type_enabled;
+  // Used when calling monotonic_to_system_epoch_ns
+  monotonic_to_system_epoch_state time_converter_state;
+  // Used to identify the main thread, to give it a fallback name
+  VALUE main_thread;
+  // Used when extracting trace identifiers from otel spans. Lazily initialized.
+  VALUE otel_current_span_key;
 
   struct stats {
     // Track how many garbage collection samples we've taken.
@@ -106,26 +126,31 @@ struct thread_context_collector_state {
     // See thread_context_collector_on_gc_start for details
     unsigned int gc_samples_missed_due_to_missing_context;
   } stats;
+
+  struct {
+    unsigned long accumulated_cpu_time_ns;
+    unsigned long accumulated_wall_time_ns;
+
+    long wall_time_at_previous_gc_ns; // Will be INVALID_TIME unless there's accumulated time above
+    long wall_time_at_last_flushed_gc_event_ns; // Starts at 0 and then will always be valid
+  } gc_tracking;
 };
 
 // Tracks per-thread state
 struct per_thread_context {
   char thread_id[THREAD_ID_LIMIT_CHARS];
   ddog_CharSlice thread_id_char_slice;
+  char thread_invoke_location[THREAD_INVOKE_LOCATION_LIMIT_CHARS];
+  ddog_CharSlice thread_invoke_location_char_slice;
   thread_cpu_time_id thread_cpu_time_id;
   long cpu_time_at_previous_sample_ns;  // Can be INVALID_TIME until initialized or if getting it fails for another reason
   long wall_time_at_previous_sample_ns; // Can be INVALID_TIME until initialized
 
   struct {
-    // Both of these fields are set by on_gc_start and kept until sample_after_gc is called.
+    // Both of these fields are set by on_gc_start and kept until on_gc_finish is called.
     // Outside of this window, they will be INVALID_TIME.
     long cpu_time_at_start_ns;
     long wall_time_at_start_ns;
-
-    // Both of these fields are set by on_gc_finish and kept until sample_after_gc is called.
-    // Outside of this window, they will be INVALID_TIME.
-    long cpu_time_at_finish_ns;
-    long wall_time_at_finish_ns;
   } gc_tracking;
 };
 
@@ -148,7 +173,9 @@ static VALUE _native_initialize(
   VALUE recorder_instance,
   VALUE max_frames,
   VALUE tracer_context_key,
-  VALUE endpoint_collection_enabled
+  VALUE endpoint_collection_enabled,
+  VALUE timeline_enabled,
+  VALUE allocation_type_enabled
 );
 static VALUE _native_sample(VALUE self, VALUE collector_instance, VALUE profiler_overhead_stack_thread);
 static VALUE _native_on_gc_start(VALUE self, VALUE collector_instance);
@@ -168,16 +195,19 @@ static void trigger_sample_for_thread(
   VALUE stack_from_thread,
   struct per_thread_context *thread_context,
   sample_values values,
-  sample_type type
+  long current_monotonic_wall_time_ns,
+  ddog_CharSlice *ruby_vm_type,
+  ddog_CharSlice *class_name
 );
 static VALUE _native_thread_list(VALUE self);
 static struct per_thread_context *get_or_create_context_for(VALUE thread, struct thread_context_collector_state *state);
 static struct per_thread_context *get_context_for(VALUE thread, struct thread_context_collector_state *state);
-static void initialize_context(VALUE thread, struct per_thread_context *thread_context);
+static void initialize_context(VALUE thread, struct per_thread_context *thread_context, struct thread_context_collector_state *state);
 static VALUE _native_inspect(VALUE self, VALUE collector_instance);
 static VALUE per_thread_context_st_table_as_ruby_hash(struct thread_context_collector_state *state);
 static int per_thread_context_as_ruby_hash(st_data_t key_thread, st_data_t value_context, st_data_t result_hash);
 static VALUE stats_as_ruby_hash(struct thread_context_collector_state *state);
+static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *state);
 static void remove_context_for_dead_threads(struct thread_context_collector_state *state);
 static int remove_if_dead_thread(st_data_t key_thread, st_data_t value_context, st_data_t _argument);
 static VALUE _native_per_thread_context(VALUE self, VALUE collector_instance);
@@ -185,11 +215,22 @@ static long update_time_since_previous_sample(long *time_at_previous_sample_ns,
 static long cpu_time_now_ns(struct per_thread_context *thread_context);
 static long thread_id_for(VALUE thread);
 static VALUE _native_stats(VALUE self, VALUE collector_instance);
+static VALUE _native_gc_tracking(VALUE self, VALUE collector_instance);
 static void trace_identifiers_for(struct thread_context_collector_state *state, VALUE thread, struct trace_identifiers *trace_identifiers_result);
-static bool is_type_web(VALUE root_span_type);
+static bool should_collect_resource(VALUE root_span);
 static VALUE _native_reset_after_fork(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
 static VALUE thread_list(struct thread_context_collector_state *state);
-static VALUE _native_sample_allocation(VALUE self, VALUE collector_instance, VALUE sample_weight);
+static VALUE _native_sample_allocation(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE sample_weight, VALUE new_object);
+static VALUE _native_new_empty_thread(VALUE self);
+static ddog_CharSlice ruby_value_type_to_class_name(enum ruby_value_type type);
+static void ddtrace_otel_trace_identifiers_for(
+  struct thread_context_collector_state *state,
+  VALUE *active_trace,
+  VALUE *root_span,
+  VALUE *numeric_span_id,
+  VALUE active_span,
+  VALUE otel_values
+);
 
 void collectors_thread_context_init(VALUE profiling_module) {
   VALUE collectors_module = rb_define_module_under(profiling_module, "Collectors");
@@ -207,17 +248,19 @@ void collectors_thread_context_init(VALUE profiling_module) {
   // https://bugs.ruby-lang.org/issues/18007 for a discussion around this.
   rb_define_alloc_func(collectors_thread_context_class, _native_new);
 
-  rb_define_singleton_method(collectors_thread_context_class, "_native_initialize", _native_initialize, 5);
+  rb_define_singleton_method(collectors_thread_context_class, "_native_initialize", _native_initialize, 7);
   rb_define_singleton_method(collectors_thread_context_class, "_native_inspect", _native_inspect, 1);
   rb_define_singleton_method(collectors_thread_context_class, "_native_reset_after_fork", _native_reset_after_fork, 1);
   rb_define_singleton_method(testing_module, "_native_sample", _native_sample, 2);
-  rb_define_singleton_method(testing_module, "_native_sample_allocation", _native_sample_allocation, 2);
+  rb_define_singleton_method(testing_module, "_native_sample_allocation", _native_sample_allocation, 3);
   rb_define_singleton_method(testing_module, "_native_on_gc_start", _native_on_gc_start, 1);
   rb_define_singleton_method(testing_module, "_native_on_gc_finish", _native_on_gc_finish, 1);
   rb_define_singleton_method(testing_module, "_native_sample_after_gc", _native_sample_after_gc, 1);
   rb_define_singleton_method(testing_module, "_native_thread_list", _native_thread_list, 0);
   rb_define_singleton_method(testing_module, "_native_per_thread_context", _native_per_thread_context, 1);
   rb_define_singleton_method(testing_module, "_native_stats", _native_stats, 1);
+  rb_define_singleton_method(testing_module, "_native_gc_tracking", _native_gc_tracking, 1);
+  rb_define_singleton_method(testing_module, "_native_new_empty_thread", _native_new_empty_thread, 0);
 
   at_active_span_id = rb_intern_const("@active_span");
   at_active_trace_id = rb_intern_const("@active_trace");
@@ -225,6 +268,11 @@ void collectors_thread_context_init(VALUE profiling_module) {
   at_resource_id = rb_intern_const("@resource");
   at_root_span_id = rb_intern_const("@root_span");
   at_type_id = rb_intern_const("@type");
+  at_otel_values_id = rb_intern_const("@otel_values");
+  at_parent_span_id_id = rb_intern_const("@parent_span_id");
+  at_datadog_trace_id = rb_intern_const("@datadog_trace");
+
+  gc_profiling_init();
 }
 
 // This structure is used to define a Ruby object that stores a pointer to a struct thread_context_collector_state
@@ -249,6 +297,8 @@ static void thread_context_collector_typed_data_mark(void *state_ptr) {
   rb_gc_mark(state->recorder_instance);
   st_foreach(state->hash_map_per_thread_context, hash_map_per_thread_context_mark, 0 /* unused */);
   rb_gc_mark(state->thread_list_buffer);
+  rb_gc_mark(state->main_thread);
+  rb_gc_mark(state->otel_current_span_key);
 }
 
 static void thread_context_collector_typed_data_free(void *state_ptr) {
@@ -285,6 +335,9 @@ static int hash_map_per_thread_context_free_values(DDTRACE_UNUSED st_data_t _thr
 static VALUE _native_new(VALUE klass) {
   struct thread_context_collector_state *state = ruby_xcalloc(1, sizeof(struct thread_context_collector_state));
 
+  // Note: Any exceptions raised from this note until the TypedData_Wrap_Struct call will lead to the state memory
+  // being leaked.
+
   // Update this when modifying state struct
   state->sampling_buffer = NULL;
   state->hash_map_per_thread_context =
@@ -294,6 +347,13 @@ static VALUE _native_new(VALUE klass) {
   state->tracer_context_key = MISSING_TRACER_CONTEXT_KEY;
   state->thread_list_buffer = rb_ary_new();
   state->endpoint_collection_enabled = true;
+  state->timeline_enabled = true;
+  state->allocation_type_enabled = true;
+  state->time_converter_state = (monotonic_to_system_epoch_state) MONOTONIC_TO_SYSTEM_EPOCH_INITIALIZER;
+  state->main_thread = rb_thread_main();
+  state->otel_current_span_key = Qnil;
+  state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
+  state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = 0;
 
   return TypedData_Wrap_Struct(klass, &thread_context_collector_typed_data, state);
 }
@@ -304,9 +364,13 @@ static VALUE _native_initialize(
   VALUE recorder_instance,
   VALUE max_frames,
   VALUE tracer_context_key,
-  VALUE endpoint_collection_enabled
+  VALUE endpoint_collection_enabled,
+  VALUE timeline_enabled,
+  VALUE allocation_type_enabled
 ) {
   ENFORCE_BOOLEAN(endpoint_collection_enabled);
+  ENFORCE_BOOLEAN(timeline_enabled);
+  ENFORCE_BOOLEAN(allocation_type_enabled);
 
   struct thread_context_collector_state *state;
   TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
@@ -319,6 +383,8 @@ static VALUE _native_initialize(
   // hash_map_per_thread_context is already initialized, nothing to do here
   state->recorder_instance = enforce_recorder_instance(recorder_instance);
   state->endpoint_collection_enabled = (endpoint_collection_enabled == Qtrue);
+  state->timeline_enabled = (timeline_enabled == Qtrue);
+  state->allocation_type_enabled = (allocation_type_enabled == Qtrue);
 
   if (RTEST(tracer_context_key)) {
     ENFORCE_TYPE(tracer_context_key, T_SYMBOL);
@@ -433,7 +499,11 @@ void update_metrics_and_sample(
   long wall_time_elapsed_ns = update_time_since_previous_sample(
     &thread_context->wall_time_at_previous_sample_ns,
     current_monotonic_wall_time_ns,
-    thread_context->gc_tracking.wall_time_at_start_ns,
+    // We explicitly pass in `INVALID_TIME` as an argument for `gc_start_time_ns` here because we don't want wall-time
+    // accounting to change during GC.
+    // E.g. if 60 seconds pass in the real world, 60 seconds of wall-time are recorded, regardless of the thread doing
+    // GC or not.
+    INVALID_TIME,
     IS_WALL_TIME
   );
 
@@ -442,14 +512,16 @@ void update_metrics_and_sample(
     thread_being_sampled,
     stack_from_thread,
     thread_context,
-    (sample_values) {.cpu_time_ns = cpu_time_elapsed_ns, .cpu_samples = 1, .wall_time_ns = wall_time_elapsed_ns},
-    SAMPLE_REGULAR
+    (sample_values) {.cpu_time_ns = cpu_time_elapsed_ns, .cpu_or_wall_samples = 1, .wall_time_ns = wall_time_elapsed_ns},
+    current_monotonic_wall_time_ns,
+    NULL,
+    NULL
   );
 }
 
 // This function gets called when Ruby is about to start running the Garbage Collector on the current thread.
 // It updates the per_thread_context of the current thread to include the current cpu/wall times, to be used to later
-// create a stack sample that blames the cpu/wall time spent from now until the end of the garbage collector work.
+// create an event including the cpu/wall time spent in garbage collector work.
 //
 // Safety: This function gets called while Ruby is doing garbage collection. While Ruby is doing garbage collection,
 // *NO ALLOCATION* is allowed. This function, and any it calls must never trigger memory or object allocation.
@@ -474,27 +546,14 @@ void thread_context_collector_on_gc_start(VALUE self_instance) {
     return;
   }
 
-  // If these fields are set, there's an existing GC sample that still needs to be written out by `sample_after_gc`.
-  //
-  // When can this happen? Because we don't have precise control over when `sample_after_gc` gets called (it will be
-  // called sometime after GC finishes), there is no way to guarantee that Ruby will not trigger more than one GC cycle
-  // before we can actually run that method.
-  //
-  // We handle this by collapsing multiple GC cycles into one. That is, if the following happens:
-  // `on_gc_start` (time=0) -> `on_gc_finish` (time=1) -> `on_gc_start` (time=2) -> `on_gc_finish` (time=3) -> `sample_after_gc`
-  // then we just use time=0 from the first on_gc_start and time=3 from the last on_gc_finish, e.g. we behave as if
-  // there was a single, longer GC period.
-  if (thread_context->gc_tracking.cpu_time_at_finish_ns != INVALID_TIME &&
-    thread_context->gc_tracking.wall_time_at_finish_ns != INVALID_TIME) return;
-
-  // Here we record the wall-time first and in on_gc_finish we record it second to avoid having wall-time be slightly < cpu-time
+  // Here we record the wall-time first and in on_gc_finish we record it second to try to avoid having wall-time be slightly < cpu-time
   thread_context->gc_tracking.wall_time_at_start_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
   thread_context->gc_tracking.cpu_time_at_start_ns = cpu_time_now_ns(thread_context);
 }
 
 // This function gets called when Ruby has finished running the Garbage Collector on the current thread.
-// It updates the per_thread_context of the current thread to include the current cpu/wall times, to be used to later
-// create a stack sample that blames the cpu/wall time spent from the start of garbage collector work until now.
+// It records the cpu/wall-time observed during GC, which will be used to later
+// create an event including the cpu/wall time spent from the start of garbage collector work until now.
 //
 // Safety: This function gets called while Ruby is doing garbage collection. While Ruby is doing garbage collection,
 // *NO ALLOCATION* is allowed. This function, and any it calls must never trigger memory or object allocation.
@@ -502,9 +561,9 @@ void thread_context_collector_on_gc_start(VALUE self_instance) {
 //
 // Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
 // Assumption 2: This function is called from the main Ractor (if Ruby has support for Ractors).
-void thread_context_collector_on_gc_finish(VALUE self_instance) {
+bool thread_context_collector_on_gc_finish(VALUE self_instance) {
   struct thread_context_collector_state *state;
-  if (!rb_typeddata_is_kind_of(self_instance, &thread_context_collector_typed_data)) return;
+  if (!rb_typeddata_is_kind_of(self_instance, &thread_context_collector_typed_data)) return false;
   // This should never fail the the above check passes
   TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
 
@@ -512,29 +571,70 @@ void thread_context_collector_on_gc_finish(VALUE self_instance) {
 
   // If there was no previously-existing context for this thread, we won't allocate one (see safety). We keep a metric for
   // how often this happens -- see on_gc_start.
-  if (thread_context == NULL) return;
+  if (thread_context == NULL) return false;
 
-  if (thread_context->gc_tracking.cpu_time_at_start_ns == INVALID_TIME &&
-    thread_context->gc_tracking.wall_time_at_start_ns == INVALID_TIME) {
+  long cpu_time_at_start_ns = thread_context->gc_tracking.cpu_time_at_start_ns;
+  long wall_time_at_start_ns = thread_context->gc_tracking.wall_time_at_start_ns;
+
+  if (cpu_time_at_start_ns == INVALID_TIME && wall_time_at_start_ns == INVALID_TIME) {
     // If this happened, it means that on_gc_start was either never called for the thread OR it was called but no thread
     // context existed at the time. The former can be the result of a bug, but since we can't distinguish them, we just
     // do nothing.
-    return;
+    return false;
   }
 
-  // Here we record the wall-time second and in on_gc_start we record it first to avoid having wall-time be slightly < cpu-time
-  thread_context->gc_tracking.cpu_time_at_finish_ns = cpu_time_now_ns(thread_context);
-  thread_context->gc_tracking.wall_time_at_finish_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+  // Mark thread as no longer in GC
+  thread_context->gc_tracking.cpu_time_at_start_ns = INVALID_TIME;
+  thread_context->gc_tracking.wall_time_at_start_ns = INVALID_TIME;
+
+  // Here we record the wall-time second and in on_gc_start we record it first to try to avoid having wall-time be slightly < cpu-time
+  long cpu_time_at_finish_ns = cpu_time_now_ns(thread_context);
+  long wall_time_at_finish_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+
+  // If our end timestamp is not OK, we bail out
+  if (wall_time_at_finish_ns == 0) return false;
+
+  long gc_cpu_time_elapsed_ns = cpu_time_at_finish_ns - cpu_time_at_start_ns;
+  long gc_wall_time_elapsed_ns = wall_time_at_finish_ns - wall_time_at_start_ns;
+
+  // Wall-time can go backwards if the system clock gets changed (and we observed spurious jumps back on macOS as well)
+  // so let's ensure we don't get negative values for time deltas.
+  gc_cpu_time_elapsed_ns = long_max_of(gc_cpu_time_elapsed_ns, 0);
+  gc_wall_time_elapsed_ns = long_max_of(gc_wall_time_elapsed_ns, 0);
+
+  if (state->gc_tracking.wall_time_at_previous_gc_ns == INVALID_TIME) {
+    state->gc_tracking.accumulated_cpu_time_ns = 0;
+    state->gc_tracking.accumulated_wall_time_ns = 0;
+  }
+
+  state->gc_tracking.accumulated_cpu_time_ns += gc_cpu_time_elapsed_ns;
+  state->gc_tracking.accumulated_wall_time_ns += gc_wall_time_elapsed_ns;
+  state->gc_tracking.wall_time_at_previous_gc_ns = wall_time_at_finish_ns;
+
+  // Update cpu-time accounting so it doesn't include the cpu-time spent in GC during the next sample
+  // We don't update the wall-time because we don't subtract the wall-time spent in GC (see call to
+  // `update_time_since_previous_sample` for wall-time in `update_metrics_and_sample`).
+  if (thread_context->cpu_time_at_previous_sample_ns != INVALID_TIME) {
+    thread_context->cpu_time_at_previous_sample_ns += gc_cpu_time_elapsed_ns;
+  }
+
+  // Let the caller know if it should schedule a flush or not. Returning true every time would cause a lot of overhead
+  // on the application (see GC tracking introduction at the top of the file), so instead we try to accumulate a few
+  // samples first.
+  bool over_flush_time_treshold =
+    (wall_time_at_finish_ns - state->gc_tracking.wall_time_at_last_flushed_gc_event_ns) >= TIME_BETWEEN_GC_EVENTS_NS;
+
+  if (over_flush_time_treshold) {
+    return true;
+  } else {
+    return gc_profiling_has_major_gc_finished();
+  }
 }
 
-// This function gets called shortly after Ruby has finished running the Garbage Collector.
+// This function gets called after one or more GC work steps (calls to on_gc_start/on_gc_finish).
 // It creates a new sample including the cpu and wall-time spent by the garbage collector work, and resets any
 // GC-related tracking.
 //
-// Specifically, it will search for thread(s) which have gone through a cycle of on_gc_start/on_gc_finish
-// and thus have cpu_time_at_start_ns, cpu_time_at_finish_ns, wall_time_at_start_ns, wall_time_at_finish_ns
-// set on their context.
-//
 // Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
 // Assumption 2: This function is allowed to raise exceptions. Caller is responsible for handling them, if needed.
 // Assumption 3: Unlike `on_gc_start` and `on_gc_finish`, this method is allowed to allocate memory as needed.
@@ -543,67 +643,45 @@ VALUE thread_context_collector_sample_after_gc(VALUE self_instance) {
   struct thread_context_collector_state *state;
   TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-  VALUE threads = thread_list(state);
-  bool sampled_any_thread = false;
-
-  const long thread_count = RARRAY_LEN(threads);
-  for (long i = 0; i < thread_count; i++) {
-    VALUE thread = RARRAY_AREF(threads, i);
-    struct per_thread_context *thread_context = get_or_create_context_for(thread, state);
+  if (state->gc_tracking.wall_time_at_previous_gc_ns == INVALID_TIME) {
+    rb_raise(rb_eRuntimeError, "BUG: Unexpected call to sample_after_gc without valid GC information available");
+  }
 
-    if (
-      thread_context->gc_tracking.cpu_time_at_start_ns == INVALID_TIME ||
-      thread_context->gc_tracking.cpu_time_at_finish_ns == INVALID_TIME ||
-      thread_context->gc_tracking.wall_time_at_start_ns == INVALID_TIME ||
-      thread_context->gc_tracking.wall_time_at_finish_ns == INVALID_TIME
-    ) continue; // Ignore threads with no/incomplete garbage collection data
-
-    sampled_any_thread = true;
-
-    long gc_cpu_time_elapsed_ns =
-      thread_context->gc_tracking.cpu_time_at_finish_ns - thread_context->gc_tracking.cpu_time_at_start_ns;
-    long gc_wall_time_elapsed_ns =
-      thread_context->gc_tracking.wall_time_at_finish_ns - thread_context->gc_tracking.wall_time_at_start_ns;
-
-    // We don't expect non-wall time to go backwards, so let's flag this as a bug
-    if (gc_cpu_time_elapsed_ns < 0) rb_raise(rb_eRuntimeError, "BUG: Unexpected negative gc_cpu_time_elapsed_ns between samples");
-    // Wall-time can actually go backwards (e.g. when the system clock gets set) so we can't assume time going backwards
-    // was a bug.
-    // @ivoanjo: I've also observed time going backwards spuriously on macOS, see discussion on
-    // https://github.com/DataDog/dd-trace-rb/pull/2336.
-    if (gc_wall_time_elapsed_ns < 0) gc_wall_time_elapsed_ns = 0;
-
-    if (thread_context->gc_tracking.wall_time_at_start_ns == 0 && thread_context->gc_tracking.wall_time_at_finish_ns != 0) {
-      // Avoid using wall-clock if we got 0 for a start (meaning there was an error) but not 0 for finish so we don't
-      // come up with a crazy value for the frame
-      rb_raise(rb_eRuntimeError, "BUG: Unexpected zero value for gc_tracking.wall_time_at_start_ns");
-    }
+  int max_labels_needed_for_gc = 7; // Magic number gets validated inside gc_profiling_set_metadata
+  ddog_prof_Label labels[max_labels_needed_for_gc];
+  uint8_t label_pos = gc_profiling_set_metadata(labels, max_labels_needed_for_gc);
 
-    trigger_sample_for_thread(
-      state,
-      /* thread: */  thread,
-      /* stack_from_thread: */ thread,
-      thread_context,
-      (sample_values) {.cpu_time_ns = gc_cpu_time_elapsed_ns, .cpu_samples = 1, .wall_time_ns = gc_wall_time_elapsed_ns},
-      SAMPLE_IN_GC
-    );
+  ddog_prof_Slice_Label slice_labels = {.ptr = labels, .len = label_pos};
 
-    // Mark thread as no longer in GC
-    thread_context->gc_tracking.cpu_time_at_start_ns = INVALID_TIME;
-    thread_context->gc_tracking.cpu_time_at_finish_ns = INVALID_TIME;
-    thread_context->gc_tracking.wall_time_at_start_ns = INVALID_TIME;
-    thread_context->gc_tracking.wall_time_at_finish_ns = INVALID_TIME;
+  // The end_timestamp_ns is treated specially by libdatadog and that's why it's not added as a ddog_prof_Label
+  int64_t end_timestamp_ns = 0;
 
-    // Update counters so that they won't include the time in GC during the next sample
-    if (thread_context->cpu_time_at_previous_sample_ns != INVALID_TIME) {
-      thread_context->cpu_time_at_previous_sample_ns += gc_cpu_time_elapsed_ns;
-    }
-    if (thread_context->wall_time_at_previous_sample_ns != INVALID_TIME) {
-      thread_context->wall_time_at_previous_sample_ns += gc_wall_time_elapsed_ns;
-    }
+  if (state->timeline_enabled) {
+    end_timestamp_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, state->gc_tracking.wall_time_at_previous_gc_ns);
   }
 
-  if (sampled_any_thread) state->stats.gc_samples++;
+  record_placeholder_stack(
+    state->sampling_buffer,
+    state->recorder_instance,
+    (sample_values) {
+      // This event gets both a regular cpu/wall-time duration, as a normal cpu/wall-time sample would, as well as a
+      // timeline duration.
+      // This is done to enable two use-cases:
+      // * regular cpu/wall-time makes this event show up as a regular stack in the flamegraph
+      // * the timeline duration is used when the event shows up in the timeline
+      .cpu_time_ns = state->gc_tracking.accumulated_cpu_time_ns,
+      .cpu_or_wall_samples = 1,
+      .wall_time_ns = state->gc_tracking.accumulated_wall_time_ns,
+      .timeline_wall_time_ns = state->gc_tracking.accumulated_wall_time_ns,
+    },
+    (sample_labels) {.labels = slice_labels, .state_label = NULL, .end_timestamp_ns = end_timestamp_ns},
+    DDOG_CHARSLICE_C("Garbage Collection")
+  );
+
+  state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = state->gc_tracking.wall_time_at_previous_gc_ns;
+  state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
+
+  state->stats.gc_samples++;
 
   // Return a VALUE to make it easier to call this function from Ruby APIs that expect a return value (such as rb_rescue2)
   return Qnil;
@@ -615,12 +693,17 @@ static void trigger_sample_for_thread(
   VALUE stack_from_thread, // This can be different when attributing profiler overhead using a different stack
   struct per_thread_context *thread_context,
   sample_values values,
-  sample_type type
+  long current_monotonic_wall_time_ns,
+  // These two labels are only used for allocation profiling; @ivoanjo: may want to refactor this at some point?
+  ddog_CharSlice *ruby_vm_type,
+  ddog_CharSlice *class_name
 ) {
   int max_label_count =
     1 + // thread id
     1 + // thread name
     1 + // profiler overhead
+    2 + // ruby vm type and allocation class
+    1 + // state (only set for cpu/wall-time samples)
     2;  // local root span id and span id
   ddog_prof_Label labels[max_label_count];
   int label_pos = 0;
@@ -636,6 +719,19 @@ static void trigger_sample_for_thread(
       .key = DDOG_CHARSLICE_C("thread name"),
       .str = char_slice_from_ruby_string(thread_name)
     };
+  } else if (thread == state->main_thread) { // Threads are often not named, but we can have a nice fallback for this special thread
+    ddog_CharSlice main_thread_name = DDOG_CHARSLICE_C("main");
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("thread name"),
+      .str = main_thread_name
+    };
+  } else {
+    // For other threads without name, we use the "invoke location" (first file:line of the block used to start the thread), if any.
+    // This is what Ruby shows in `Thread#to_s`.
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("thread name"),
+      .str = thread_context->thread_invoke_location_char_slice // This is an empty string if no invoke location was available
+    };
   }
 
   struct trace_identifiers trace_identifiers_result = {.valid = false, .trace_endpoint = Qnil};
@@ -670,21 +766,56 @@ static void trigger_sample_for_thread(
     };
   }
 
+  if (ruby_vm_type != NULL) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("ruby vm type"),
+      .str = *ruby_vm_type
+    };
+  }
+
+  if (class_name != NULL) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("allocation class"),
+      .str = *class_name
+    };
+  }
+
+  // This label is handled specially:
+  // 1. It's only set for cpu/wall-time samples
+  // 2. We set it here to its default state of "unknown", but the `Collectors::Stack` may choose to override it with
+  //    something more interesting.
+  ddog_prof_Label *state_label = NULL;
+  if (values.cpu_or_wall_samples > 0) {
+    state_label = &labels[label_pos++];
+    *state_label = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("state"),
+      .str = DDOG_CHARSLICE_C("unknown"),
+      .num = 0, // This shouldn't be needed but the tracer-2.7 docker image ships a buggy gcc that complains about this
+    };
+  }
+
   // The number of times `label_pos++` shows up in this function needs to match `max_label_count`. To avoid "oops I
   // forgot to update max_label_count" in the future, we've also added this validation.
-  // @ivoanjo: I wonder if C compilers are smart enough to statically prove when this check never triggers happens and
-  // remove it entirely.
+  // @ivoanjo: I wonder if C compilers are smart enough to statically prove this check never triggers unless someone
+  // changes the code erroneously and remove it entirely?
   if (label_pos > max_label_count) {
     rb_raise(rb_eRuntimeError, "BUG: Unexpected label_pos (%d) > max_label_count (%d)", label_pos, max_label_count);
   }
 
+  ddog_prof_Slice_Label slice_labels = {.ptr = labels, .len = label_pos};
+
+  // The end_timestamp_ns is treated specially by libdatadog and that's why it's not added as a ddog_prof_Label
+  int64_t end_timestamp_ns = 0;
+  if (state->timeline_enabled && current_monotonic_wall_time_ns != INVALID_TIME) {
+    end_timestamp_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, current_monotonic_wall_time_ns);
+  }
+
   sample_thread(
     stack_from_thread,
     state->sampling_buffer,
     state->recorder_instance,
     values,
-    (ddog_prof_Slice_Label) {.ptr = labels, .len = label_pos},
-    type
+    (sample_labels) {.labels = slice_labels, .state_label = state_label, .end_timestamp_ns = end_timestamp_ns}
   );
 }
 
@@ -704,7 +835,7 @@ static struct per_thread_context *get_or_create_context_for(VALUE thread, struct
     thread_context = (struct per_thread_context*) value_context;
   } else {
     thread_context = ruby_xcalloc(1, sizeof(struct per_thread_context));
-    initialize_context(thread, thread_context);
+    initialize_context(thread, thread_context, state);
     st_insert(state->hash_map_per_thread_context, (st_data_t) thread, (st_data_t) thread_context);
   }
 
@@ -722,10 +853,57 @@ static struct per_thread_context *get_context_for(VALUE thread, struct thread_co
   return thread_context;
 }
 
-static void initialize_context(VALUE thread, struct per_thread_context *thread_context) {
+#define LOGGING_GEM_PATH "/lib/logging/diagnostic_context.rb"
+
+// The `logging` gem monkey patches thread creation, which makes the `invoke_location_for` useless, since every thread
+// will point to the `logging` gem. When that happens, we avoid using the invoke location.
+//
+// TODO: This approach is a bit brittle, since it matches on the specific gem path, and only works for the `logging`
+// gem.
+// In the future we should probably explore a more generic fix (e.g. using Thread.method(:new).source_location or
+// something like that to detect redefinition of the `Thread` methods). One difficulty of doing it is that we need
+// to either run Ruby code during sampling (not great), or otherwise use some of the VM private APIs to detect this.
+//
+static bool is_logging_gem_monkey_patch(VALUE invoke_file_location) {
+  int logging_gem_path_len = strlen(LOGGING_GEM_PATH);
+  char *invoke_file = StringValueCStr(invoke_file_location);
+  int invoke_file_len = strlen(invoke_file);
+
+  if (invoke_file_len < logging_gem_path_len) return false;
+
+  return strncmp(invoke_file + invoke_file_len - logging_gem_path_len, LOGGING_GEM_PATH, logging_gem_path_len) == 0;
+}
+
+static void initialize_context(VALUE thread, struct per_thread_context *thread_context, struct thread_context_collector_state *state) {
   snprintf(thread_context->thread_id, THREAD_ID_LIMIT_CHARS, "%"PRIu64" (%lu)", native_thread_id_for(thread), (unsigned long) thread_id_for(thread));
   thread_context->thread_id_char_slice = (ddog_CharSlice) {.ptr = thread_context->thread_id, .len = strlen(thread_context->thread_id)};
 
+  int invoke_line_location;
+  VALUE invoke_file_location = invoke_location_for(thread, &invoke_line_location);
+  if (invoke_file_location != Qnil) {
+    if (!is_logging_gem_monkey_patch(invoke_file_location)) {
+      snprintf(
+        thread_context->thread_invoke_location,
+        THREAD_INVOKE_LOCATION_LIMIT_CHARS,
+        "%s:%d",
+        StringValueCStr(invoke_file_location),
+        invoke_line_location
+      );
+    } else {
+      snprintf(thread_context->thread_invoke_location, THREAD_INVOKE_LOCATION_LIMIT_CHARS, "%s", "(Unnamed thread)");
+    }
+  } else if (thread != state->main_thread) {
+    // If the first function of a thread is native code, there won't be an invoke location, so we use this fallback.
+    // NOTE: In the future, I wonder if we could take the pointer to the native function, and try to see if there's a native
+    // symbol attached to it.
+    snprintf(thread_context->thread_invoke_location, THREAD_INVOKE_LOCATION_LIMIT_CHARS, "%s", "(Unnamed thread from native code)");
+  }
+
+  thread_context->thread_invoke_location_char_slice = (ddog_CharSlice) {
+    .ptr = thread_context->thread_invoke_location,
+    .len = strlen(thread_context->thread_invoke_location)
+  };
+
   thread_context->thread_cpu_time_id = thread_cpu_time_id_for(thread);
 
   // These will get initialized during actual sampling
@@ -734,9 +912,7 @@ static void initialize_context(VALUE thread, struct per_thread_context *thread_c
 
   // These will only be used during a GC operation
   thread_context->gc_tracking.cpu_time_at_start_ns = INVALID_TIME;
-  thread_context->gc_tracking.cpu_time_at_finish_ns = INVALID_TIME;
   thread_context->gc_tracking.wall_time_at_start_ns = INVALID_TIME;
-  thread_context->gc_tracking.wall_time_at_finish_ns = INVALID_TIME;
 }
 
 static VALUE _native_inspect(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
@@ -753,6 +929,16 @@ static VALUE _native_inspect(DDTRACE_UNUSED VALUE _self, VALUE collector_instanc
   rb_str_concat(result, rb_sprintf(" sample_count=%u", state->sample_count));
   rb_str_concat(result, rb_sprintf(" stats=%"PRIsVALUE, stats_as_ruby_hash(state)));
   rb_str_concat(result, rb_sprintf(" endpoint_collection_enabled=%"PRIsVALUE, state->endpoint_collection_enabled ? Qtrue : Qfalse));
+  rb_str_concat(result, rb_sprintf(" timeline_enabled=%"PRIsVALUE, state->timeline_enabled ? Qtrue : Qfalse));
+  rb_str_concat(result, rb_sprintf(" allocation_type_enabled=%"PRIsVALUE, state->allocation_type_enabled ? Qtrue : Qfalse));
+  rb_str_concat(result, rb_sprintf(
+    " time_converter_state={.system_epoch_ns_reference=%ld, .delta_to_epoch_ns=%ld}",
+    state->time_converter_state.system_epoch_ns_reference,
+    state->time_converter_state.delta_to_epoch_ns
+  ));
+  rb_str_concat(result, rb_sprintf(" main_thread=%"PRIsVALUE, state->main_thread));
+  rb_str_concat(result, rb_sprintf(" gc_tracking=%"PRIsVALUE, gc_tracking_as_ruby_hash(state)));
+  rb_str_concat(result, rb_sprintf(" otel_current_span_key=%"PRIsVALUE, state->otel_current_span_key));
 
   return result;
 }
@@ -772,15 +958,14 @@ static int per_thread_context_as_ruby_hash(st_data_t key_thread, st_data_t value
 
   VALUE arguments[] = {
     ID2SYM(rb_intern("thread_id")),                       /* => */ rb_str_new2(thread_context->thread_id),
+    ID2SYM(rb_intern("thread_invoke_location")),          /* => */ rb_str_new2(thread_context->thread_invoke_location),
     ID2SYM(rb_intern("thread_cpu_time_id_valid?")),       /* => */ thread_context->thread_cpu_time_id.valid ? Qtrue : Qfalse,
     ID2SYM(rb_intern("thread_cpu_time_id")),              /* => */ CLOCKID2NUM(thread_context->thread_cpu_time_id.clock_id),
     ID2SYM(rb_intern("cpu_time_at_previous_sample_ns")),  /* => */ LONG2NUM(thread_context->cpu_time_at_previous_sample_ns),
     ID2SYM(rb_intern("wall_time_at_previous_sample_ns")), /* => */ LONG2NUM(thread_context->wall_time_at_previous_sample_ns),
 
     ID2SYM(rb_intern("gc_tracking.cpu_time_at_start_ns")),   /* => */ LONG2NUM(thread_context->gc_tracking.cpu_time_at_start_ns),
-    ID2SYM(rb_intern("gc_tracking.cpu_time_at_finish_ns")),  /* => */ LONG2NUM(thread_context->gc_tracking.cpu_time_at_finish_ns),
     ID2SYM(rb_intern("gc_tracking.wall_time_at_start_ns")),  /* => */ LONG2NUM(thread_context->gc_tracking.wall_time_at_start_ns),
-    ID2SYM(rb_intern("gc_tracking.wall_time_at_finish_ns")), /* => */ LONG2NUM(thread_context->gc_tracking.wall_time_at_finish_ns)
   };
   for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(context_as_hash, arguments[i], arguments[i+1]);
 
@@ -798,6 +983,19 @@ static VALUE stats_as_ruby_hash(struct thread_context_collector_state *state) {
   return stats_as_hash;
 }
 
+static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *state) {
+  // Update this when modifying state struct (gc_tracking inner struct)
+  VALUE result = rb_hash_new();
+  VALUE arguments[] = {
+    ID2SYM(rb_intern("accumulated_cpu_time_ns")),               /* => */ ULONG2NUM(state->gc_tracking.accumulated_cpu_time_ns),
+    ID2SYM(rb_intern("accumulated_wall_time_ns")),              /* => */ ULONG2NUM(state->gc_tracking.accumulated_wall_time_ns),
+    ID2SYM(rb_intern("wall_time_at_previous_gc_ns")),           /* => */ LONG2NUM(state->gc_tracking.wall_time_at_previous_gc_ns),
+    ID2SYM(rb_intern("wall_time_at_last_flushed_gc_event_ns")), /* => */ LONG2NUM(state->gc_tracking.wall_time_at_last_flushed_gc_event_ns),
+  };
+  for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(result, arguments[i], arguments[i+1]);
+  return result;
+}
+
 static void remove_context_for_dead_threads(struct thread_context_collector_state *state) {
   st_foreach(state->hash_map_per_thread_context, remove_if_dead_thread, 0 /* unused */);
 }
@@ -900,8 +1098,6 @@ VALUE enforce_thread_context_collector_instance(VALUE object) {
 
 // This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
-//
-// Returns the whole contents of the per_thread_context structs being tracked.
 static VALUE _native_stats(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
   struct thread_context_collector_state *state;
   TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
@@ -909,6 +1105,15 @@ static VALUE _native_stats(DDTRACE_UNUSED VALUE _self, VALUE collector_instance)
   return stats_as_ruby_hash(state);
 }
 
+// This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
+// It SHOULD NOT be used for other purposes.
+static VALUE _native_gc_tracking(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  return gc_tracking_as_ruby_hash(state);
+}
+
 // Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
 static void trace_identifiers_for(struct thread_context_collector_state *state, VALUE thread, struct trace_identifiers *trace_identifiers_result) {
   if (state->tracer_context_key == MISSING_TRACER_CONTEXT_KEY) return;
@@ -921,10 +1126,19 @@ static void trace_identifiers_for(struct thread_context_collector_state *state,
 
   VALUE root_span = rb_ivar_get(active_trace, at_root_span_id /* @root_span */);
   VALUE active_span = rb_ivar_get(active_trace, at_active_span_id /* @active_span */);
-  if (root_span == Qnil || active_span == Qnil) return;
+  // Note: On Ruby 3.x `rb_attr_get` is exactly the same as `rb_ivar_get`. For Ruby 2.x, the difference is that
+  // `rb_ivar_get` can trigger "warning: instance variable @otel_values not initialized" if warnings are enabled and
+  // opentelemetry is not in use, whereas `rb_attr_get` does the lookup without generating the warning.
+  VALUE otel_values = rb_attr_get(active_trace, at_otel_values_id /* @otel_values */);
+
+  VALUE numeric_span_id = Qnil;
+
+  if (otel_values != Qnil) ddtrace_otel_trace_identifiers_for(state, &active_trace, &root_span, &numeric_span_id, active_span, otel_values);
+
+  if (root_span == Qnil || (active_span == Qnil && numeric_span_id == Qnil)) return;
 
   VALUE numeric_local_root_span_id = rb_ivar_get(root_span, at_id_id /* @id */);
-  VALUE numeric_span_id = rb_ivar_get(active_span, at_id_id /* @id */);
+  if (active_span != Qnil && numeric_span_id == Qnil) numeric_span_id = rb_ivar_get(active_span, at_id_id /* @id */);
   if (numeric_local_root_span_id == Qnil || numeric_span_id == Qnil) return;
 
   trace_identifiers_result->local_root_span_id = NUM2ULL(numeric_local_root_span_id);
@@ -932,10 +1146,7 @@ static void trace_identifiers_for(struct thread_context_collector_state *state,
 
   trace_identifiers_result->valid = true;
 
-  if (!state->endpoint_collection_enabled) return;
-
-  VALUE root_span_type = rb_ivar_get(root_span, at_type_id /* @type */);
-  if (root_span_type == Qnil || !is_type_web(root_span_type)) return;
+  if (!state->endpoint_collection_enabled || !should_collect_resource(root_span)) return;
 
   VALUE trace_resource = rb_ivar_get(active_trace, at_resource_id /* @resource */);
   if (RB_TYPE_P(trace_resource, T_STRING)) {
@@ -946,11 +1157,32 @@ static void trace_identifiers_for(struct thread_context_collector_state *state,
   }
 }
 
-static bool is_type_web(VALUE root_span_type) {
+// We opt-in to collecting the resource for spans of types:
+// * 'web', for web requests
+// * 'proxy', used by the rack integration with request_queuing: true (e.g. also represents a web request)
+// * 'worker', used for sidekiq and similar background job processors
+//
+// 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 :)
+static bool should_collect_resource(VALUE root_span) {
+  VALUE root_span_type = rb_ivar_get(root_span, at_type_id /* @type */);
+  if (root_span_type == Qnil) return false;
   ENFORCE_TYPE(root_span_type, T_STRING);
 
-  return RSTRING_LEN(root_span_type) == strlen("web") &&
-    (memcmp("web", StringValuePtr(root_span_type), strlen("web")) == 0);
+  int root_span_type_length = RSTRING_LEN(root_span_type);
+  const char *root_span_type_value = StringValuePtr(root_span_type);
+
+  bool is_web_request =
+    (root_span_type_length == strlen("web") && (memcmp("web", root_span_type_value, strlen("web")) == 0)) ||
+    (root_span_type_length == strlen("proxy") && (memcmp("proxy", root_span_type_value, strlen("proxy")) == 0));
+
+  if (is_web_request) return true;
+
+  bool is_worker_request =
+    (root_span_type_length == strlen("worker") && (memcmp("worker", root_span_type_value, strlen("worker")) == 0));
+
+  return is_worker_request;
 }
 
 // After the Ruby VM forks, this method gets called in the child process to clean up any leftover state from the parent.
@@ -977,25 +1209,194 @@ static VALUE thread_list(struct thread_context_collector_state *state) {
   return result;
 }
 
-void thread_context_collector_sample_allocation(VALUE self_instance, unsigned int sample_weight) {
+void thread_context_collector_sample_allocation(VALUE self_instance, unsigned int sample_weight, VALUE new_object) {
   struct thread_context_collector_state *state;
   TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
 
   VALUE current_thread = rb_thread_current();
 
+  enum ruby_value_type type = rb_type(new_object);
+
+  // Tag samples with the VM internal types
+  ddog_CharSlice ruby_vm_type = ruby_value_type_to_char_slice(type);
+
+  // Since this is stack allocated, be careful about moving it
+  ddog_CharSlice class_name;
+  ddog_CharSlice *optional_class_name = NULL;
+  char imemo_type[100];
+
+  if (state->allocation_type_enabled) {
+    optional_class_name = &class_name;
+
+    if (
+      type == RUBY_T_OBJECT   ||
+      type == RUBY_T_CLASS    ||
+      type == RUBY_T_MODULE   ||
+      type == RUBY_T_FLOAT    ||
+      type == RUBY_T_STRING   ||
+      type == RUBY_T_REGEXP   ||
+      type == RUBY_T_ARRAY    ||
+      type == RUBY_T_HASH     ||
+      type == RUBY_T_STRUCT   ||
+      type == RUBY_T_BIGNUM   ||
+      type == RUBY_T_FILE     ||
+      type == RUBY_T_DATA     ||
+      type == RUBY_T_MATCH    ||
+      type == RUBY_T_COMPLEX  ||
+      type == RUBY_T_RATIONAL ||
+      type == RUBY_T_NIL      ||
+      type == RUBY_T_TRUE     ||
+      type == RUBY_T_FALSE    ||
+      type == RUBY_T_SYMBOL   ||
+      type == RUBY_T_FIXNUM
+    ) {
+      VALUE klass = rb_class_of(new_object);
+
+      // Ruby sometimes plays a bit fast and loose with some of its internal objects, e.g.
+      // `rb_str_tmp_frozen_acquire` allocates a string with no class (klass=0).
+      // Thus, we need to make sure there's actually a class before getting its name.
+
+      if (klass != 0) {
+        const char *name = rb_obj_classname(new_object);
+        size_t name_length = name != NULL ? strlen(name) : 0;
+
+        if (name_length > 0) {
+          class_name = (ddog_CharSlice) {.ptr = name, .len = name_length};
+        } else {
+          // @ivoanjo: I'm not sure this can ever happen, but just-in-case
+          class_name = ruby_value_type_to_class_name(type);
+        }
+      } else {
+        // Fallback for objects with no class
+        class_name = ruby_value_type_to_class_name(type);
+      }
+    } else if (type == RUBY_T_IMEMO) {
+      const char *imemo_string = imemo_kind(new_object);
+      if (imemo_string != NULL) {
+        snprintf(imemo_type, 100, "(VM Internal, T_IMEMO, %s)", imemo_string);
+        class_name = (ddog_CharSlice) {.ptr = imemo_type, .len = strlen(imemo_type)};
+      } else { // Ruby < 3
+        class_name = DDOG_CHARSLICE_C("(VM Internal, T_IMEMO)");
+      }
+    } else {
+      class_name = ruby_vm_type; // For other weird internal things we just use the VM type
+    }
+  }
+
+  track_object(state->recorder_instance, new_object, sample_weight, optional_class_name);
+
   trigger_sample_for_thread(
     state,
     /* thread: */  current_thread,
     /* stack_from_thread: */ current_thread,
     get_or_create_context_for(current_thread, state),
     (sample_values) {.alloc_samples = sample_weight},
-    SAMPLE_REGULAR
+    INVALID_TIME, // For now we're not collecting timestamps for allocation events, as per profiling team internal discussions
+    &ruby_vm_type,
+    optional_class_name
   );
 }
 
 // This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
-static VALUE _native_sample_allocation(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE sample_weight) {
-  thread_context_collector_sample_allocation(collector_instance, NUM2UINT(sample_weight));
+static VALUE _native_sample_allocation(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE sample_weight, VALUE new_object) {
+  thread_context_collector_sample_allocation(collector_instance, NUM2UINT(sample_weight), new_object);
   return Qtrue;
 }
+
+static VALUE new_empty_thread_inner(DDTRACE_UNUSED void *arg) { return Qnil; }
+
+// This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
+// It SHOULD NOT be used for other purposes.
+// (It creates an empty native thread, so we can test our native thread naming fallback)
+static VALUE _native_new_empty_thread(DDTRACE_UNUSED VALUE self) {
+  return rb_thread_create(new_empty_thread_inner, NULL);
+}
+
+static ddog_CharSlice ruby_value_type_to_class_name(enum ruby_value_type type) {
+  switch (type) {
+    case(RUBY_T_OBJECT  ): return DDOG_CHARSLICE_C("Object");
+    case(RUBY_T_CLASS   ): return DDOG_CHARSLICE_C("Class");
+    case(RUBY_T_MODULE  ): return DDOG_CHARSLICE_C("Module");
+    case(RUBY_T_FLOAT   ): return DDOG_CHARSLICE_C("Float");
+    case(RUBY_T_STRING  ): return DDOG_CHARSLICE_C("String");
+    case(RUBY_T_REGEXP  ): return DDOG_CHARSLICE_C("Regexp");
+    case(RUBY_T_ARRAY   ): return DDOG_CHARSLICE_C("Array");
+    case(RUBY_T_HASH    ): return DDOG_CHARSLICE_C("Hash");
+    case(RUBY_T_STRUCT  ): return DDOG_CHARSLICE_C("Struct");
+    case(RUBY_T_BIGNUM  ): return DDOG_CHARSLICE_C("Integer");
+    case(RUBY_T_FILE    ): return DDOG_CHARSLICE_C("File");
+    case(RUBY_T_DATA    ): return DDOG_CHARSLICE_C("(VM Internal, T_DATA)");
+    case(RUBY_T_MATCH   ): return DDOG_CHARSLICE_C("MatchData");
+    case(RUBY_T_COMPLEX ): return DDOG_CHARSLICE_C("Complex");
+    case(RUBY_T_RATIONAL): return DDOG_CHARSLICE_C("Rational");
+    case(RUBY_T_NIL     ): return DDOG_CHARSLICE_C("NilClass");
+    case(RUBY_T_TRUE    ): return DDOG_CHARSLICE_C("TrueClass");
+    case(RUBY_T_FALSE   ): return DDOG_CHARSLICE_C("FalseClass");
+    case(RUBY_T_SYMBOL  ): return DDOG_CHARSLICE_C("Symbol");
+    case(RUBY_T_FIXNUM  ): return DDOG_CHARSLICE_C("Integer");
+                  default: return DDOG_CHARSLICE_C("(VM Internal, Missing class)");
+  }
+}
+
+static VALUE get_otel_current_span_key(struct thread_context_collector_state *state) {
+  if (state->otel_current_span_key == Qnil) {
+    VALUE datadog_module = rb_const_get(rb_cObject, rb_intern("Datadog"));
+    VALUE opentelemetry_module = rb_const_get(datadog_module, rb_intern("OpenTelemetry"));
+    VALUE api_module = rb_const_get(opentelemetry_module, rb_intern("API"));
+    VALUE context_module = rb_const_get(api_module, rb_intern_const("Context"));
+    VALUE current_span_key = rb_const_get(context_module, rb_intern_const("CURRENT_SPAN_KEY"));
+
+    if (current_span_key == Qnil) {
+      rb_raise(rb_eRuntimeError, "Unexpected: Missing Datadog::OpenTelemetry::API::Context::CURRENT_SPAN_KEY");
+    }
+
+    state->otel_current_span_key = current_span_key;
+  }
+
+  return state->otel_current_span_key;
+}
+
+// This method gets used when ddtrace is being used indirectly via the otel APIs. Information gets stored slightly
+// differently, and this codepath handles it.
+static void ddtrace_otel_trace_identifiers_for(
+  struct thread_context_collector_state *state,
+  VALUE *active_trace,
+  VALUE *root_span,
+  VALUE *numeric_span_id,
+  VALUE active_span,
+  VALUE otel_values
+) {
+  VALUE resolved_numeric_span_id =
+    active_span == Qnil ?
+      // For traces started from otel spans, the span id will be empty, and the @parent_span_id has the right value
+      rb_ivar_get(*active_trace, at_parent_span_id_id /* @parent_span_id */) :
+      // Regular span created by ddtrace
+      rb_ivar_get(active_span, at_id_id /* @id */);
+
+  if (resolved_numeric_span_id == Qnil) return;
+
+  VALUE otel_current_span_key = get_otel_current_span_key(state);
+  VALUE current_trace = *active_trace;
+
+  // ddtrace uses a different structure when spans are created from otel, where each otel span will have a unique ddtrace
+  // trace and span representing it. Each ddtrace trace is then connected to the previous otel span, forming a linked
+  // list. The local root span is going to be the trace/span we find at the end of this linked list.
+  while (otel_values != Qnil) {
+    VALUE otel_span = rb_hash_lookup(otel_values, otel_current_span_key);
+    if (otel_span == Qnil) break;
+    VALUE next_trace = rb_ivar_get(otel_span, at_datadog_trace_id);
+    if (next_trace == Qnil) break;
+
+    current_trace = next_trace;
+    otel_values = rb_ivar_get(current_trace, at_otel_values_id /* @otel_values */);
+  }
+
+  // We found the last trace in the linked list. This contains the local root span
+  VALUE resolved_root_span = rb_ivar_get(current_trace, at_root_span_id /* @root_span */);
+  if (resolved_root_span == Qnil) return;
+
+  *root_span = resolved_root_span;
+  *active_trace = current_trace;
+  *numeric_span_id = resolved_numeric_span_id;
+}