feat(bedrock): Set gen_ai.response.time_to_first_chunk

sentry_sdk/consts.py

@@ -646,6 +646,12 @@ class SPANDATA:
     Example: [{"role": "assistant", "parts": [{"type": "text", "content": "The weather in Paris is currently rainy with a temperature of 57°F."}], "finish_reason": "stop"}]
     """
 
+    GEN_AI_RESPONSE_TIME_TO_FIRST_CHUNK = "gen_ai.response.time_to_first_chunk"
+    """
+    Time in seconds when the first response content chunk arrived in streaming responses.
+    Example: 0.6853435
+    """
+
     GEN_AI_RESPONSE_TIME_TO_FIRST_TOKEN = "gen_ai.response.time_to_first_token"
     """
     The time it took to receive the first token from the model.

sentry_sdk/integrations/boto3/_client.py

@@ -1,3 +1,4 @@
+import time
 from contextlib import contextmanager
 from typing import TYPE_CHECKING
 
@@ -116,6 +117,8 @@ def sentry_patched_make_api_call(
 
         attributes: "Attributes" = {}
         try:
+            if service_ext is not None:
+                service_ext.request_start_time = time.perf_counter()
             with span_ctx:
                 try:
                     parsed = orig_make_api_call(self, operation_name, api_params)

sentry_sdk/integrations/boto3/_services/base.py

@@ -17,9 +17,12 @@ class _ServiceExtension:
     to use the generic instrumentation.
     """
 
-    __slots__ = ()
+    __slots__ = "request_start_time"
     skip_if_no_active_span = True
 
+    def __init__(self) -> None:
+        self.request_start_time: "Optional[float]" = None
+
     def get_span_op(self, ctx: "AwsCallContext") -> "Optional[str]":
         """Return an optional `sentry.op` override, or `None` to keep the default."""
         return None

sentry_sdk/integrations/boto3/_services/bedrock.py

@@ -1,5 +1,6 @@
 import json
 import sys
+import time
 from typing import TYPE_CHECKING, cast
 
 import sentry_sdk
@@ -44,6 +45,10 @@
     from sentry_sdk.integrations.boto3._context import AwsCallContext
     from sentry_sdk.tracing import Span
 
+    class _PatchedEventStream(EventStream[Any]):
+        _span: "Union[Span, StreamedSpan]"
+        _request_start_time: "Optional[float]"
+
 
 def _finish_span(
     stream: "EventStream[Any]",
@@ -89,6 +94,8 @@ def sentry_patched_iter(
             yield from original_iter(self)
             return
 
+        self = cast("_PatchedEventStream", self)
+
         set_on_span = (
             span.set_attribute if isinstance(span, StreamedSpan) else span.set_data
         )
@@ -97,6 +104,7 @@ def sentry_patched_iter(
         parts: "dict[int, Union[TextPart, ReasoningPart, ToolCallPart, ToolCallResponsePart, UriPart, BlobPart]]" = {}
 
         try:
+            start_time: "Optional[float]" = self._request_start_time
             for event in original_iter(self):
                 if "metadata" in event:
                     usage = event["metadata"]["usage"]
@@ -132,6 +140,13 @@ def sentry_patched_iter(
                     yield event
                     continue
 
+                if start_time is not None and "contentBlockDelta" in event:
+                    set_on_span(
+                        SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_CHUNK,
+                        time.perf_counter() - start_time,
+                    )
+                    start_time = None
+
                 if "messageStop" in event:
                     finish_reasons: "list[str]" = [event["messageStop"]["stopReason"]]
                     set_on_span(SPANDATA.GEN_AI_RESPONSE_FINISH_REASONS, finish_reasons)
@@ -602,9 +617,12 @@ def _transform_tool_definitions(
 
 
 class _BedrockExtension(_ServiceExtension):
-    __slots__ = ()
+    __slots__ = "request_start_time"
     skip_if_no_active_span = False
 
+    def __init__(self) -> None:
+        self.request_start_time: "Optional[float]" = None
+
     def get_span_op(self, ctx: "AwsCallContext") -> "Optional[str]":
         if ctx.operation_name != "Converse" and ctx.operation_name != "ConverseStream":
             return None
@@ -743,6 +761,7 @@ def on_success_response(
         if ctx.operation_name == "ConverseStream":
             stream = cast("ConverseStreamResponseTypeDef", response)["stream"]
             stream._span = span  # type: ignore[attr-defined]
+            stream._request_start_time = self.request_start_time  # type: ignore[attr-defined]
             return
 
         if isinstance(span, StreamedSpan):

tests/integrations/boto3/test_bedrock.py

@@ -995,6 +995,7 @@ def test_converse_stream(
         assert span["attributes"][SPANDATA.GEN_AI_RESPONSE_FINISH_REASONS] == [
             "end_turn"
         ]
+        assert SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_CHUNK in span["attributes"]
 
         assert span["attributes"][SPANDATA.GEN_AI_USAGE_OUTPUT_TOKENS] == 20
         assert span["attributes"][SPANDATA.GEN_AI_USAGE_TOTAL_TOKENS] == 30
@@ -1035,6 +1036,7 @@ def test_converse_stream(
         assert span["data"][SPANDATA.GEN_AI_REQUEST_TEMPERATURE] == 0.7
         assert span["data"][SPANDATA.GEN_AI_REQUEST_TOP_P] == 0.9
         assert span["data"][SPANDATA.GEN_AI_RESPONSE_FINISH_REASONS] == ["end_turn"]
+        assert SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_CHUNK in span["data"]
 
         assert span["data"][SPANDATA.GEN_AI_USAGE_OUTPUT_TOKENS] == 20
         assert span["data"][SPANDATA.GEN_AI_USAGE_TOTAL_TOKENS] == 30