Merge branch 'open-telemetry:main' into barebones

Author: nagkumar91

.github/component_owners.yml

@@ -9,8 +9,12 @@ components:
   instrumentation/opentelemetry-instrumentation-asyncio:
     - bourbonkk
 
+  instrumentation/opentelemetry-instrumentation-botocore:
+    - lukeina2z
+    - yiyuan-he
+
   instrumentation/opentelemetry-instrumentation-pymssql:
-     - guillaumep
+    - guillaumep
 
   instrumentation/opentelemetry-instrumentation-urllib:
     - shalevr
@@ -43,4 +47,4 @@ components:
 
   instrumentation-genai/opentelemetry-instrumentation-langchain:
     - zhirafovod
-    - wrisa
+    - wrisa

instrumentation-genai/opentelemetry-instrumentation-vertexai/CHANGELOG.md

@@ -7,9 +7,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
-- Start making changes to implement the big semantic convention changes made in https://github.com/open-telemetry/semantic-conventions/pull/2179.
-Now only a single event (`gen_ai.client.inference.operation.details`) is used to capture Chat History. These changes will be opt-in,
-users will need to set the environment variable OTEL_SEMCONV_STABILITY_OPT_IN to `gen_ai_latest_experimental` to see them ([#3386](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3386)).
+- Update instrumentation to use the latest semantic convention changes made in https://github.com/open-telemetry/semantic-conventions/pull/2179.
+Now only a single event and span (`gen_ai.client.inference.operation.details`) are used to capture prompt and response content. These changes are opt-in,
+users will need to set the environment variable OTEL_SEMCONV_STABILITY_OPT_IN to `gen_ai_latest_experimental` to see them ([#3799](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3799)) and ([#3709](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3709)).
 - Implement uninstrument for `opentelemetry-instrumentation-vertexai`
   ([#3328](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3328))
 - VertexAI support for async calling

instrumentation-genai/opentelemetry-instrumentation-vertexai/examples/manual/.env

@@ -6,3 +6,17 @@ OTEL_SERVICE_NAME=opentelemetry-python-vertexai
 
 # Change to 'false' to hide prompt and completion content
 OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
+
+# Alternatively set this env var to enable the latest semantic conventions:
+OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental
+
+# When using the latest experimental flag this env var controls which telemetry signals will have prompt and response content included in them.
+# Choices are NO_CONTENT, SPAN_ONLY, EVENT_ONLY, SPAN_AND_EVENT.
+OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_AND_EVENT
+
+# Optional hook that will upload prompt and response content to some external destination.
+# For example fsspec.
+OTEL_INSTRUMENTATION_GENAI_COMPLETION_HOOK = "upload"
+
+# Required if using a completion hook. The path to upload content to for example gs://my_bucket.
+OTEL_INSTRUMENTATION_GENAI_UPLOAD_BASE_PATH = "gs://my_bucket"

instrumentation-genai/opentelemetry-instrumentation-vertexai/examples/zero-code/.env

@@ -12,3 +12,17 @@ OTEL_PYTHON_LOGGING_AUTO_INSTRUMENTATION_ENABLED=true
 
 # Change to 'false' to hide prompt and completion content
 OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
+
+# Alternatively set this env var to enable the latest semantic conventions:
+OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental
+
+# When using the latest experimental flag this env var controls which telemetry signals will have prompt and response content included in them.
+# Choices are NO_CONTENT, SPAN_ONLY, EVENT_ONLY, SPAN_AND_EVENT.
+OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_AND_EVENT
+
+# Optional hook that will upload prompt and response content to some external destination.
+# For example fsspec.
+OTEL_INSTRUMENTATION_GENAI_COMPLETION_HOOK = "upload"
+
+# Required if using a completion hook. The path to upload content to for example gs://my_bucket.
+OTEL_INSTRUMENTATION_GENAI_UPLOAD_BASE_PATH = "gs://my_bucket"

instrumentation-genai/opentelemetry-instrumentation-vertexai/src/opentelemetry/instrumentation/vertexai/__init__.py

@@ -60,6 +60,7 @@
 from opentelemetry.instrumentation.vertexai.utils import is_content_enabled
 from opentelemetry.semconv.schemas import Schemas
 from opentelemetry.trace import get_tracer
+from opentelemetry.util.genai.completion_hook import load_completion_hook
 
 
 def _methods_to_wrap(
@@ -109,6 +110,9 @@ def instrumentation_dependencies(self) -> Collection[str]:
 
     def _instrument(self, **kwargs: Any):
         """Enable VertexAI instrumentation."""
+        completion_hook = (
+            kwargs.get("completion_hook") or load_completion_hook()
+        )
         sem_conv_opt_in_mode = _OpenTelemetrySemanticConventionStability._get_opentelemetry_stability_opt_in_mode(
             _OpenTelemetryStabilitySignalType.GEN_AI,
         )
@@ -141,6 +145,7 @@ def _instrument(self, **kwargs: Any):
                 event_logger,
                 is_content_enabled(sem_conv_opt_in_mode),
                 sem_conv_opt_in_mode,
+                completion_hook,
             )
         elif sem_conv_opt_in_mode == _StabilityMode.GEN_AI_LATEST_EXPERIMENTAL:
             # Type checker now knows it's the other literal
@@ -149,6 +154,7 @@ def _instrument(self, **kwargs: Any):
                 event_logger,
                 is_content_enabled(sem_conv_opt_in_mode),
                 sem_conv_opt_in_mode,
+                completion_hook,
             )
         else:
             raise RuntimeError(f"{sem_conv_opt_in_mode} mode not supported")

instrumentation-genai/opentelemetry-instrumentation-vertexai/src/opentelemetry/instrumentation/vertexai/patch.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 from contextlib import contextmanager
+from dataclasses import asdict
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -27,22 +28,32 @@
     overload,
 )
 
-from opentelemetry._events import EventLogger
+from opentelemetry._events import Event, EventLogger
 from opentelemetry.instrumentation._semconv import (
     _StabilityMode,
 )
 from opentelemetry.instrumentation.vertexai.utils import (
     GenerateContentParams,
-    create_operation_details_event,
+    _map_finish_reason,
+    convert_content_to_message_parts,
     get_genai_request_attributes,
     get_genai_response_attributes,
     get_server_attributes,
     get_span_name,
     request_to_events,
     response_to_events,
 )
+from opentelemetry.semconv._incubating.attributes import (
+    gen_ai_attributes as GenAI,
+)
 from opentelemetry.trace import SpanKind, Tracer
-from opentelemetry.util.genai.types import ContentCapturingMode
+from opentelemetry.util.genai.completion_hook import CompletionHook
+from opentelemetry.util.genai.types import (
+    ContentCapturingMode,
+    InputMessage,
+    OutputMessage,
+)
+from opentelemetry.util.genai.utils import gen_ai_json_dumps
 
 if TYPE_CHECKING:
     from google.cloud.aiplatform_v1.services.prediction_service import client
@@ -110,6 +121,7 @@ def __init__(
         sem_conv_opt_in_mode: Literal[
             _StabilityMode.GEN_AI_LATEST_EXPERIMENTAL
         ],
+        completion_hook: CompletionHook,
     ) -> None: ...
 
     @overload
@@ -119,6 +131,7 @@ def __init__(
         event_logger: EventLogger,
         capture_content: bool,
         sem_conv_opt_in_mode: Literal[_StabilityMode.DEFAULT],
+        completion_hook: CompletionHook,
     ) -> None: ...
 
     def __init__(
@@ -130,11 +143,13 @@ def __init__(
             Literal[_StabilityMode.DEFAULT],
             Literal[_StabilityMode.GEN_AI_LATEST_EXPERIMENTAL],
         ],
+        completion_hook: CompletionHook,
     ) -> None:
         self.tracer = tracer
         self.event_logger = event_logger
         self.capture_content = capture_content
         self.sem_conv_opt_in_mode = sem_conv_opt_in_mode
+        self.completion_hook = completion_hook
 
     @contextmanager
     def _with_new_instrumentation(
@@ -146,40 +161,88 @@ def _with_new_instrumentation(
         kwargs: Any,
     ):
         params = _extract_params(*args, **kwargs)
-        api_endpoint: str = instance.api_endpoint  # type: ignore[reportUnknownMemberType]
-        span_attributes = {
-            **get_genai_request_attributes(False, params),
-            **get_server_attributes(api_endpoint),
-        }
-
-        span_name = get_span_name(span_attributes)
-
+        request_attributes = get_genai_request_attributes(True, params)
         with self.tracer.start_as_current_span(
-            name=span_name,
+            name=f"{GenAI.GenAiOperationNameValues.CHAT.value} {request_attributes.get(GenAI.GEN_AI_REQUEST_MODEL, '')}".strip(),
             kind=SpanKind.CLIENT,
-            attributes=span_attributes,
         ) as span:
 
             def handle_response(
                 response: prediction_service.GenerateContentResponse
                 | prediction_service_v1beta1.GenerateContentResponse
                 | None,
             ) -> None:
-                if span.is_recording() and response:
-                    # When streaming, this is called multiple times so attributes would be
-                    # overwritten. In practice, it looks the API only returns the interesting
-                    # attributes on the last streamed response. However, I couldn't find
-                    # documentation for this and setting attributes shouldn't be too expensive.
-                    span.set_attributes(
-                        get_genai_response_attributes(response)
-                    )
-                self.event_logger.emit(
-                    create_operation_details_event(
-                        api_endpoint=api_endpoint,
-                        params=params,
-                        capture_content=capture_content,
-                        response=response,
+                attributes = (
+                    get_server_attributes(instance.api_endpoint)  # type: ignore[reportUnknownMemberType]
+                    | request_attributes
+                    | get_genai_response_attributes(response)
+                )
+                system_instructions, inputs, outputs = [], [], []
+                if params.system_instruction:
+                    system_instructions = convert_content_to_message_parts(
+                        params.system_instruction
                     )
+                if params.contents:
+                    inputs = [
+                        InputMessage(
+                            role=content.role,
+                            parts=convert_content_to_message_parts(content),
+                        )
+                        for content in params.contents
+                    ]
+                if response:
+                    outputs = [
+                        OutputMessage(
+                            finish_reason=_map_finish_reason(
+                                candidate.finish_reason
+                            ),
+                            role=candidate.content.role,
+                            parts=convert_content_to_message_parts(
+                                candidate.content
+                            ),
+                        )
+                        for candidate in response.candidates
+                    ]
+                content_attributes = {
+                    k: [asdict(x) for x in v]
+                    for k, v in [
+                        (
+                            GenAI.GEN_AI_SYSTEM_INSTRUCTIONS,
+                            system_instructions,
+                        ),
+                        (GenAI.GEN_AI_INPUT_MESSAGES, inputs),
+                        (GenAI.GEN_AI_OUTPUT_MESSAGES, outputs),
+                    ]
+                    if v
+                }
+                if span.is_recording():
+                    span.set_attributes(attributes)
+                    if capture_content in (
+                        ContentCapturingMode.SPAN_AND_EVENT,
+                        ContentCapturingMode.SPAN_ONLY,
+                    ):
+                        span.set_attributes(
+                            {
+                                k: gen_ai_json_dumps(v)
+                                for k, v in content_attributes.items()
+                            }
+                        )
+                event = Event(
+                    name="gen_ai.client.inference.operation.details",
+                )
+                event.attributes = attributes
+                if capture_content in (
+                    ContentCapturingMode.SPAN_AND_EVENT,
+                    ContentCapturingMode.EVENT_ONLY,
+                ):
+                    event.attributes |= content_attributes
+                self.event_logger.emit(event)
+                self.completion_hook.on_completion(
+                    inputs=inputs,
+                    outputs=outputs,
+                    system_instruction=system_instructions,
+                    span=span,
+                    log_record=event,
                 )
 
             yield handle_response