Adding the support tracing of child models invoked from a BLS model

docs/user_guide/trace.md

@@ -448,11 +448,6 @@ class TritonPythonModel:
       ...
       for request in requests:
         ...
-        # Create an InferenceRequest object. `model_name`,
-        # `requested_output_names`, and `inputs` are the required arguments and
-        # must be provided when constructing an InferenceRequest object. Make
-        # sure to replace `inputs` argument with a list of `pb_utils.Tensor`
-        # objects.
         inference_request = pb_utils.InferenceRequest(
             model_name='model_name',
             requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],

qa/L0_trace/opentelemetry_unittest.py

@@ -37,6 +37,7 @@
 import tritonclient.http as httpclient
 
 EXPECTED_NUM_SPANS = 16
+NO_PARENT_SPAN = "0000000000000000"
 
 
 class OpenTelemetryTest(tu.TestResultCollector):
@@ -155,7 +156,7 @@ def _check_parent(self, child_span, parent_span):
         self.assertEqual(child_span["trace_id"], parent_span["trace_id"])
         self.assertNotEqual(
             child_span["parent_span_id"],
-            "0000000000000000",
+            NO_PARENT_SPAN,
             "child span does not have parent span id specified",
         )
         self.assertEqual(
@@ -262,7 +263,7 @@ def send_bls_request(model_name="simple"):
     with httpclient.InferenceServerClient("localhost:8000") as client:
         inputs = prepare_data(httpclient)
         inputs.append(httpclient.InferInput("MODEL_NAME", [1], "BYTES"))
-        inputs[2].set_data_from_numpy(np.array([model_name], dtype=np.object_))
+        inputs[-1].set_data_from_numpy(np.array([model_name], dtype=np.object_))
         client.infer("bls_simple", inputs)
 
 

qa/L0_trace/test.sh

@@ -669,19 +669,25 @@ $TRACE_SUMMARY -t bls_trace.log > summary_bls.log
 
 if [ `grep -c "COMPUTE_INPUT_END" summary_bls.log` != "2" ]; then
     cat summary_bls.log
-    echo -e "\n***\n*** Test Failed\n***"
+    echo -e "\n***\n*** Test Failed: Unexpected number of traced "COMPUTE_INPUT_END" events.\n***"
     RET=1
 fi
 
-if [ `grep -c ^simple summary_bls.log` != "1" ]; then
+if [ `grep -c ^ensemble_add_sub_int32_int32_int32 summary_bls.log` != "1" ]; then
     cat summary_bls.log
-    echo -e "\n***\n*** Test Failed\n***"
+    echo -e "\n***\n*** Test Failed: BLS child ensemble model wasn't traced. \n***"
     RET=1
 fi
 
-if [ `grep -c 'parent id' bls_trace.log` == "1" ]; then
+if [ `grep -c ^simple summary_bls.log` != "1" ]; then
     cat summary_bls.log
-    echo -e "\n***\n*** Test Failed\n***"
+    echo -e "\n***\n*** Test Failed: ensemble's model 'simple' wasn't traced. \n***"
+    RET=1
+fi
+
+if [ `grep -c 'parent id' bls_trace.log` != "2" ]; then
+    cat bls_trace.log
+    echo -e "\n***\n*** Test Failed: Unexpected number of 'parent id' fields. \n***"
     RET=1
 fi
 
@@ -760,9 +766,8 @@ fi
 
 set +e
 # Preparing traces for unittest.
-# Note: need to run this separately, to speed up trace collection.
-# Otherwise internal (opentelemetry_unittest.OpenTelemetryTest.setUp) check
-# will slow down collection.
+# Note: running this separately, so that I could extract spans with `grep`
+# from server log later.
 python -c 'import opentelemetry_unittest; \
         opentelemetry_unittest.prepare_traces()' >>$CLIENT_LOG 2>&1
 

src/tracer.cc

@@ -61,8 +61,6 @@ TraceManager::Create(
   *manager = new TraceManager(
       level, rate, count, log_frequency, filepath, mode, config_map);
 
-  (*manager)->InitTracer(config_map);
-
   return nullptr;  // success
 }
 
@@ -86,6 +84,8 @@ TraceManager::TraceManager(
       false /*filepath_specified*/, false /*mode_specified*/,
       false /*config_map_specified*/));
   trace_files_.emplace(filepath, file);
+
+  InitTracer(config_map);
 }
 
 TRITONSERVER_Error*
@@ -357,58 +357,76 @@ TraceManager::Trace::CaptureTimestamp(
 void
 TraceManager::InitTracer(const triton::server::TraceConfigMap& config_map)
 {
-  if (global_setting_->mode_ == TRACE_MODE_TRITON) {
-    return;
-  }
+  switch (global_setting_->mode_) {
+    case TRACE_MODE_OPENTELEMETRY: {
 #if !defined(_WIN32) && defined(TRITON_ENABLE_TRACING)
-  otlp::OtlpHttpExporterOptions opts;
-  otel_resource::ResourceAttributes attributes = {};
-  attributes[otel_resource::SemanticConventions::kServiceName] =
-      "triton-inference-server";
-  auto mode_key = std::to_string(TRACE_MODE_OPENTELEMETRY);
-  auto otel_options_it = config_map.find(mode_key);
-  if (otel_options_it != config_map.end()) {
-    for (const auto& setting : otel_options_it->second) {
-      // FIXME add more configuration options of OTLP HTTP Exporter
-      if (setting.first == "url") {
-        opts.url = setting.second;
+      otlp::OtlpHttpExporterOptions opts;
+      otel_resource::ResourceAttributes attributes = {};
+      attributes[otel_resource::SemanticConventions::kServiceName] =
+          "triton-inference-server";
+      auto mode_key = std::to_string(TRACE_MODE_OPENTELEMETRY);
+      auto otel_options_it = config_map.find(mode_key);
+      if (otel_options_it != config_map.end()) {
+        for (const auto& setting : otel_options_it->second) {
+          // FIXME add more configuration options of OTLP HTTP Exporter
+          if (setting.first == "url") {
+            opts.url = setting.second;
+          }
+          if (setting.first == "resource") {
+            auto pos = setting.second.find('=');
+            auto key = setting.second.substr(0, pos);
+            auto value = setting.second.substr(pos + 1);
+            attributes[key] = value;
+          }
+        }
       }
-      if (setting.first == "resource") {
-        auto pos = setting.second.find('=');
-        auto key = setting.second.substr(0, pos);
-        auto value = setting.second.substr(pos + 1);
-        attributes[key] = value;
+      auto exporter = otlp::OtlpHttpExporterFactory::Create(opts);
+      auto test_exporter = triton::server::GetEnvironmentVariableOrDefault(
+          "TRITON_OPENTELEMETRY_TEST", "false");
+      if (test_exporter != "false") {
+        exporter = opentelemetry::exporter::trace::OStreamSpanExporterFactory::
+            Create();
       }
+      auto processor = otel_trace_sdk::SimpleSpanProcessorFactory::Create(
+          std::move(exporter));
+      auto resource = otel_resource::Resource::Create(attributes);
+      std::shared_ptr<otel_trace_api::TracerProvider> provider =
+          otel_trace_sdk::TracerProviderFactory::Create(
+              std::move(processor), resource);
+
+      otel_trace_api::Provider::SetTracerProvider(provider);
+      break;
+#else
+      LOG_ERROR << "Unsupported trace mode: "
+                << TraceManager::InferenceTraceModeString(
+                       global_setting_->mode_);
+      break;
+#endif
     }
+    default:
+      return;
   }
-  auto exporter = otlp::OtlpHttpExporterFactory::Create(opts);
-  auto test_exporter = triton::server::GetEnvironmentVariableOrDefault(
-      "TRITON_OPENTELEMETRY_TEST", "false");
-  if (test_exporter != "false") {
-    exporter =
-        opentelemetry::exporter::trace::OStreamSpanExporterFactory::Create();
-  }
-  auto processor =
-      otel_trace_sdk::SimpleSpanProcessorFactory::Create(std::move(exporter));
-  auto resource = otel_resource::Resource::Create(attributes);
-  std::shared_ptr<otel_trace_api::TracerProvider> provider =
-      otel_trace_sdk::TracerProviderFactory::Create(
-          std::move(processor), resource);
-
-  otel_trace_api::Provider::SetTracerProvider(provider);
-#endif
 }
 
 void
 TraceManager::CleanupTracer()
 {
-  if (global_setting_->mode_ == TRACE_MODE_TRITON) {
-    return;
-  }
+  switch (global_setting_->mode_) {
+    case TRACE_MODE_OPENTELEMETRY: {
 #if !defined(_WIN32) && defined(TRITON_ENABLE_TRACING)
-  std::shared_ptr<otel_trace_api::TracerProvider> none;
-  otel_trace_api::Provider::SetTracerProvider(none);
+      std::shared_ptr<otel_trace_api::TracerProvider> none;
+      otel_trace_api::Provider::SetTracerProvider(none);
+      break;
+#else
+      LOG_ERROR << "Unsupported trace mode: "
+                << TraceManager::InferenceTraceModeString(
+                       global_setting_->mode_);
+      break;
 #endif
+    }
+    default:
+      return;
+  }
 }
 
 #ifndef _WIN32