Add streaming timing metrics to generic stream wrappers

instrumentation/opentelemetry-instrumentation-genai-openai/.changelog/13.added

@@ -0,0 +1 @@
+Record `gen_ai.client.operation.time_to_first_chunk` and `gen_ai.client.operation.time_per_output_chunk` metrics for chat completion streams.

instrumentation/opentelemetry-instrumentation-genai-openai/src/opentelemetry/instrumentation/genai/openai/chat_wrappers.py

@@ -182,7 +182,11 @@ def __init__(
         invocation: InferenceInvocation,
         capture_content: bool,
     ) -> None:
-        super().__init__(stream)
+        super().__init__(
+            stream,
+            start_time_s=invocation.monotonic_start_s,
+            timing_target=invocation,
+        )
         self._self_invocation = invocation
         self._self_choice_buffers = []
         self._self_capture_content = capture_content
@@ -203,7 +207,11 @@ def __init__(
         invocation: InferenceInvocation,
         capture_content: bool,
     ) -> None:
-        super().__init__(stream)
+        super().__init__(
+            stream,
+            start_time_s=invocation.monotonic_start_s,
+            timing_target=invocation,
+        )
         self._self_invocation = invocation
         self._self_choice_buffers = []
         self._self_capture_content = capture_content

instrumentation/opentelemetry-instrumentation-genai-openai/tests/test_chat_metrics.py

@@ -254,3 +254,172 @@ async def test_async_chat_completion_metrics(
     assert_all_metric_attributes(
         output_token_usage, latest_experimental_enabled
     )
+
+
+# TTFC and per-output-chunk histograms have a different attribute shape than
+# the duration/token histograms: streaming responses in the recorded cassettes
+# do not include system_fingerprint or service_tier, so we assert only the
+# core gen_ai.* attributes that should always be populated.
+def assert_streaming_metric_attributes(
+    data_point, latest_experimental_enabled, expected_request_model
+):
+    assert GenAIAttributes.GEN_AI_OPERATION_NAME in data_point.attributes
+    assert (
+        data_point.attributes[GenAIAttributes.GEN_AI_OPERATION_NAME]
+        == GenAIAttributes.GenAiOperationNameValues.CHAT.value
+    )
+
+    provider_name_attr_name = (
+        "gen_ai.provider.name"
+        if latest_experimental_enabled
+        else GenAIAttributes.GEN_AI_SYSTEM
+    )
+    assert provider_name_attr_name in data_point.attributes
+    assert (
+        data_point.attributes[provider_name_attr_name]
+        == GenAIAttributes.GenAiSystemValues.OPENAI.value
+    )
+
+    assert GenAIAttributes.GEN_AI_REQUEST_MODEL in data_point.attributes
+    assert (
+        data_point.attributes[GenAIAttributes.GEN_AI_REQUEST_MODEL]
+        == expected_request_model
+    )
+    assert GenAIAttributes.GEN_AI_RESPONSE_MODEL in data_point.attributes
+    assert data_point.attributes[GenAIAttributes.GEN_AI_RESPONSE_MODEL]
+
+    assert (
+        data_point.attributes[ServerAttributes.SERVER_ADDRESS]
+        == "api.openai.com"
+    )
+
+
+def test_chat_completion_streaming_metrics(
+    metric_reader, openai_client, instrument_with_content, vcr
+):
+    """Regression test for the openai_v2 sync chat stream wrapper wiring.
+
+    Exercises the actual ChatStreamWrapper path so that removing
+    timing_target=invocation in chat_wrappers.py would cause this test to
+    fail, not just the util-layer tests.
+    """
+    if not is_experimental_mode():
+        pytest.skip("new stream wrapper only")
+
+    latest_experimental_enabled = is_experimental_mode()
+    request_model = "gpt-4"
+
+    with vcr.use_cassette("test_chat_completion_streaming.yaml"):
+        response = openai_client.chat.completions.create(
+            messages=USER_ONLY_PROMPT,
+            model=request_model,
+            stream=True,
+            stream_options={"include_usage": True},
+        )
+        for _ in response:
+            pass
+
+    metrics = metric_reader.get_metrics_data().resource_metrics
+    assert len(metrics) == 1
+    metric_data = metrics[0].scope_metrics[0].metrics
+
+    ttfc_metric = next(
+        (
+            m
+            for m in metric_data
+            if m.name == "gen_ai.client.operation.time_to_first_chunk"
+        ),
+        None,
+    )
+    assert ttfc_metric is not None
+    assert len(ttfc_metric.data.data_points) == 1
+    ttfc_point = ttfc_metric.data.data_points[0]
+    assert ttfc_point.count == 1
+    assert ttfc_point.sum >= 0
+    assert_streaming_metric_attributes(
+        ttfc_point, latest_experimental_enabled, request_model
+    )
+
+    chunk_metric = next(
+        (
+            m
+            for m in metric_data
+            if m.name == "gen_ai.client.operation.time_per_output_chunk"
+        ),
+        None,
+    )
+    assert chunk_metric is not None
+    assert len(chunk_metric.data.data_points) == 1
+    chunk_point = chunk_metric.data.data_points[0]
+    assert chunk_point.count >= 1
+    assert chunk_point.sum >= 0
+    assert_streaming_metric_attributes(
+        chunk_point, latest_experimental_enabled, request_model
+    )
+
+
+@pytest.mark.asyncio()
+async def test_async_chat_completion_streaming_metrics(
+    metric_reader, async_openai_client, instrument_with_content, vcr
+):
+    """Regression test for the openai_v2 async chat stream wrapper wiring.
+
+    The async path has separate __init__ wiring from the sync path in
+    chat_wrappers.py, so it needs its own coverage. Removing
+    timing_target=invocation in AsyncChatStreamWrapper would still pass
+    every util-layer test, but would silently break TTFC and per-output-chunk
+    metrics for async OpenAI streaming.
+    """
+    if not is_experimental_mode():
+        pytest.skip("new stream wrapper only")
+
+    latest_experimental_enabled = is_experimental_mode()
+    request_model = "gpt-4"
+
+    with vcr.use_cassette("test_async_chat_completion_streaming.yaml"):
+        response = await async_openai_client.chat.completions.create(
+            messages=USER_ONLY_PROMPT,
+            model=request_model,
+            stream=True,
+            stream_options={"include_usage": True},
+        )
+        async for _ in response:
+            pass
+
+    metrics = metric_reader.get_metrics_data().resource_metrics
+    assert len(metrics) == 1
+    metric_data = metrics[0].scope_metrics[0].metrics
+
+    ttfc_metric = next(
+        (
+            m
+            for m in metric_data
+            if m.name == "gen_ai.client.operation.time_to_first_chunk"
+        ),
+        None,
+    )
+    assert ttfc_metric is not None
+    assert len(ttfc_metric.data.data_points) == 1
+    ttfc_point = ttfc_metric.data.data_points[0]
+    assert ttfc_point.count == 1
+    assert ttfc_point.sum >= 0
+    assert_streaming_metric_attributes(
+        ttfc_point, latest_experimental_enabled, request_model
+    )
+
+    chunk_metric = next(
+        (
+            m
+            for m in metric_data
+            if m.name == "gen_ai.client.operation.time_per_output_chunk"
+        ),
+        None,
+    )
+    assert chunk_metric is not None
+    assert len(chunk_metric.data.data_points) == 1
+    chunk_point = chunk_metric.data.data_points[0]
+    assert chunk_point.count >= 1
+    assert chunk_point.sum >= 0
+    assert_streaming_metric_attributes(
+        chunk_point, latest_experimental_enabled, request_model
+    )

instrumentation/opentelemetry-instrumentation-google-genai/tests/generate_content/cassettes/test_upload_hook_non_streaming[NO_CONTENT-vertexaiapi-async-gemini-2.5-flash-enable_completion_hook].yaml

@@ -0,0 +1,102 @@
+interactions:
+- request:
+    body: |-
+      {
+        "contents": [
+          {
+            "parts": [
+              {
+                "text": "Create a haiku about Open Telemetry."
+              }
+            ],
+            "role": "user"
+          }
+        ]
+      }
+    headers:
+      accept:
+      - '*/*'
+      accept-encoding:
+      - identity
+      connection:
+      - keep-alive
+      content-length:
+      - '93'
+      content-type:
+      - application/json
+      host:
+      - us-central1-aiplatform.googleapis.com
+      user-agent:
+      - google-genai-sdk/1.32.0 gl-python/3.10.18
+      x-goog-api-client:
+      - <REDACTED>
+      x-goog-user-project:
+      - <REDACTED>
+    method: POST
+    uri: https://test-location-aiplatform.googleapis.com/v1beta1/projects/test-project/locations/test-location/publishers/google/models/gemini-2.5-flash:generateContent
+  response:
+    body:
+      string: |-
+        {
+          "candidates": [
+            {
+              "content": {
+                "role": "model",
+                "parts": [
+                  {
+                    "text": "Open data streams,\nMetrics, logs, and traces flow,\nClearly see inside."
+                  }
+                ]
+              },
+              "finishReason": "STOP",
+              "avgLogprobs": -5.934557172987196
+            }
+          ],
+          "usageMetadata": {
+            "promptTokenCount": 8,
+            "candidatesTokenCount": 18,
+            "totalTokenCount": 459,
+            "trafficType": "ON_DEMAND",
+            "promptTokensDetails": [
+              {
+                "modality": "TEXT",
+                "tokenCount": 8
+              }
+            ],
+            "candidatesTokensDetails": [
+              {
+                "modality": "TEXT",
+                "tokenCount": 18
+              }
+            ],
+            "thoughtsTokenCount": 433
+          },
+          "modelVersion": "gemini-2.5-flash",
+          "createTime": "2025-10-10T16:32:45.350496Z",
+          "responseId": "LTXpaKCyFdPlnvgPuajSiQQ"
+        }
+    headers:
+      Accept-Ranges:
+      - none
+      Content-Type:
+      - application/json; charset=UTF-8
+      Date:
+      - Fri, 10 Oct 2025 16:32:59 GMT
+      Server:
+      - scaffolding on HTTPServer2
+      Transfer-Encoding:
+      - chunked
+      Vary:
+      - X-Origin
+      - Referer
+      - Origin,Accept-Encoding
+      X-Content-Type-Options:
+      - nosniff
+      X-Frame-Options:
+      - SAMEORIGIN
+      X-XSS-Protection:
+      - '0'
+    status:
+      code: 200
+      message: OK
+version: 1