Hacktoberfest 2026: as issues que os mantenedores marcaram para outubro, abertas e boas para iniciantes. Ver issues do Hacktoberfest

bug(openai): a stream that fails mid-iteration ends the generation without level ERROR

Aberta
  • Algo errado?
#2,039 0 comentários 0 reações 1 responsável Ver no GitHub

Mantenedores costumam responder em até 1 dia

@hassiebp já está trabalhando nisso.

Desde 22/10/2026.

  • #2056 de @Boulea7 — aberto

Avaliação

Esta issue ainda não foi avaliada.

Descrição

bug integration-openai sdk-python
Describe the bug

When an OpenAI chat stream fails after the first chunk, the generation ends with no level and no status_message. The caller gets openai.APIError, while Langfuse shows a normal generation holding the partial output. The same failure before the stream starts is recorded as ERROR.

The cause is in langfuse/openai.py. _wrap catches only what create() raises. Exceptions raised while iterating go through traced_iterator's finally (line 1112) into _finalize_stream_response, which calls generation.end() with no error. The async path (_instrument_openai_async_stream) and LangfuseResponseGeneratorSync/Async do the same.

To reproduce
import httpx
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from langfuse import Langfuse
from langfuse.openai import OpenAI

exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
Langfuse(public_key="pk", secret_key="sk", base_url="http://127.0.0.1:9", tracer_provider=provider)

SSE = (
    b'data: {"id":"c1","object":"chat.completion.chunk","created":1,"model":"gpt-4o-mini",'
    b'"choices":[{"index":0,"delta":{"role":"assistant","content":"Hel"},"finish_reason":null}]}\n\n'
    b'data: {"error":{"message":"The server had an error while processing your request.","type":"server_error"}}\n\n'
)
client = OpenAI(api_key="x", http_client=httpx.Client(transport=httpx.MockTransport(
    lambda r: httpx.Response(200, headers={"content-type": "text/event-stream"}, content=SSE))))

stream = client.chat.completions.create(name="failed-stream", model="gpt-4o-mini",
                                        messages=[{"role": "user", "content": "hi"}], stream=True)
try:
    for chunk in stream:
        pass
except Exception as e:
    print("caller saw  :", type(e).__name__, e)

a = next(s for s in exporter.get_finished_spans() if s.name == "failed-stream").attributes
print("level       :", a.get("langfuse.observation.level"))
print("status_msg  :", a.get("langfuse.observation.status_message"))
print("output      :", a.get("langfuse.observation.output"))

$ python lf_stream_error.py 2>/dev/null
caller saw  : APIError The server had an error while processing your request.
level       : None
status_msg  : None
output      : Hel

With the transport answering 500 to create(stream=True) instead, the generation records ERROR | Error code: 500 - {'error': {'message': 'server_error'}}. The AsyncOpenAI version of the failing stream also prints level: None.

Expected

level="ERROR" and status_message=str(error) on the generation, with the partial output and any usage kept. The OpenAI integration docs say Langfuse "automatically tracks and monitors OpenAI API errors" in level and statusMessage, and test_chat_completion_error_marks_generation_error pins that for the non-streaming case.

SDK versions

langfuse 5.0.2 at 81c457c, openai 3.35.1, opentelemetry-sdk 1.45.1, Python 3.12. Not checked: the Responses API, where a response.failed event raises nothing.

@hassiebp I'd catch the exception in each iterator wrapper, set the error on the generation, and re-raise. A test using the SSE body above can live next to the existing break test in tests/unit/test_openai.py. Shall I open that PR?

Linguagem predominante
Python
Estrelas
502
Forks
381
Merge médio
1d 1h
PRs com merge (30d)
64

Preparar o ambiente

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de langfuse/langfuse-python

Todas as issues de langfuse/langfuse-python

Issues semelhantes

Mais issues de Python

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.