Revert "fix: tracer leak and cost tracking guards (#15084)" (#15317)

This reverts commit 70c1225.

Author: axiomofjoy

src/phoenix/server/api/routers/agents.py

@@ -84,53 +84,53 @@ async def _stream_single_chat_completion(
     span_cost_calculator: SpanCostCalculator,
     otel_context: OtelContext,
 ) -> ChatStream:
-    async with Tracer(span_cost_calculator=span_cost_calculator) as tracer:
-        try:
-            messages = prompt_chat_template_to_playground_messages(
-                input.prompt_version.template.to_orm()
-            )
-            if template_options := input.template:
-                messages = formatted_messages(
-                    messages=messages,
-                    template_format=template_options.format,
-                    template_variables=cast(Mapping[str, Any], template_options.variables),
-                )
-            invocation_parameters = input.prompt_version.invocation_parameters.to_orm()
-            tools = input.prompt_version.tools.to_orm() if input.prompt_version.tools else None
-            response_format = (
-                input.prompt_version.response_format.to_orm()
-                if input.prompt_version.response_format
-                else None
-            )
-
-            async for chunk in llm_client.chat_completion_create(
+    tracer = Tracer(span_cost_calculator=span_cost_calculator)
+    try:
+        messages = prompt_chat_template_to_playground_messages(
+            input.prompt_version.template.to_orm()
+        )
+        if template_options := input.template:
+            messages = formatted_messages(
                 messages=messages,
-                tools=tools,
-                response_format=response_format,
-                invocation_parameters=invocation_parameters,
-                tracer=tracer,
-                otel_context=otel_context,
-                stream_model_output=input.stream_model_output,
-            ):
-                chunk.repetition_number = repetition_number
-                yield chunk
-        except Exception as error:
-            yield ChatCompletionSubscriptionError(
-                message=str(error),
-                repetition_number=repetition_number,
+                template_format=template_options.format,
+                template_variables=cast(Mapping[str, Any], template_options.variables),
             )
+        invocation_parameters = input.prompt_version.invocation_parameters.to_orm()
+        tools = input.prompt_version.tools.to_orm() if input.prompt_version.tools else None
+        response_format = (
+            input.prompt_version.response_format.to_orm()
+            if input.prompt_version.response_format
+            else None
+        )
 
-        db_traces = tracer.get_db_traces(project_id=project_id)
-        async with db() as session:
-            session.add_all(db_traces)
-            await session.flush()
-        if db_traces and db_traces[0].spans:
-            db_span = db_traces[0].spans[0]
-            yield ChatCompletionSubscriptionResult(
-                span=Span(id=db_span.id, db_record=db_span),
-                repetition_number=repetition_number,
-            )
-            on_span_insertion()
+        async for chunk in llm_client.chat_completion_create(
+            messages=messages,
+            tools=tools,
+            response_format=response_format,
+            invocation_parameters=invocation_parameters,
+            tracer=tracer,
+            otel_context=otel_context,
+            stream_model_output=input.stream_model_output,
+        ):
+            chunk.repetition_number = repetition_number
+            yield chunk
+    except Exception as error:
+        yield ChatCompletionSubscriptionError(
+            message=str(error),
+            repetition_number=repetition_number,
+        )
+
+    db_traces = tracer.get_db_traces(project_id=project_id)
+    async with db() as session:
+        session.add_all(db_traces)
+        await session.flush()
+    if db_traces and db_traces[0].spans:
+        db_span = db_traces[0].spans[0]
+        yield ChatCompletionSubscriptionResult(
+            span=Span(id=db_span.id, db_record=db_span),
+            repetition_number=repetition_number,
+        )
+        on_span_insertion()
 
 
 async def _cleanup_chat_completion_resources(

src/phoenix/server/api/subscriptions.py

@@ -34,14 +34,7 @@ def calculate_cost(
         attributes: Mapping[str, Any],
         tokens: float,
     ) -> float:
-        v = get_attribute_value(attributes, self.key)
-        # OTLP preserves the type the client sent, so this attribute can be a
-        # string or a list. Comparing one against the threshold raises, and span
-        # ingestion catches per span, so the span forfeits its whole cost record.
-        # Non-numeric values bill at the base rate, as `get_aggregated_tokens`
-        # does for the same attributes. `bool` is an `int` subclass, never a
-        # token count.
-        if not isinstance(v, (int, float)) or isinstance(v, bool) or not v:
+        if not (v := get_attribute_value(attributes, self.key)):
             return tokens * self.base_rate
         if v > self.threshold:
             return tokens * self.new_rate