🐛 Bug Report: openai, anthropic, google-generativeai: a call cancelled while waiting for the response (asyncio timeout/cancel, Ctrl-C) never ends its span

未关闭
#4,576 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@chengwudi1 已经在做这个了。

开始于 2026年10月9日。

  • #4587 来自 @chengwudi1 —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
58/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python
领域
observability

调研方向

Start with the cited wrappers in chat_wrappers.py, completion_wrappers.py, and embeddings_wrappers.py, then compare the Anthropic and Google Generative AI wrappers identified in the issue. Check how MCP session span cancellation was handled in #4517 and run the supplied async timeout reproducer for each SDK. Done means cancelled or interrupted calls end their spans as errors, propagate the exception unchanged, and have regression coverage.

由索引模型根据 Issue 内容生成。

描述

Which component is this bug for?

All Packages

📜 Description

When a call is cancelled while it waits for the server (for example asyncio.wait_for(...) times out, the task is cancelled, or the user presses Ctrl-C), the LLM span is never ended. An unended span is never exported, so the call disappears from the trace, even though the request was sent and may be billed.

The wrappers start the span with tracer.start_span(...) and end it only in except Exception or after a successful response. asyncio.CancelledError and KeyboardInterrupt are BaseException, so they pass through without ending the span:

  • openai chat_wrapper / achat_wrapper: L104, L203 (same pattern in completion_wrappers.py and embeddings_wrappers.py)
  • anthropic _wrap / _awrap: L569, L701
  • google-generativeai _awrap / _wrap: L261, L340

Both streaming and non-streaming calls are affected, because the cancellation happens before the stream object exists. The same fix was made for the MCP session span in #4517.

👟 Reproduction steps
pip install openai anthropic google-genai opentelemetry-sdk opentelemetry-instrumentation-openai opentelemetry-instrumentation-anthropic opentelemetry-instrumentation-google-generativeai
python repro.py openai
python repro.py anthropic
python repro.py google
import asyncio, gc, logging, sys
from importlib.metadata import version

import httpx
from opentelemetry.sdk.trace import SpanProcessor, TracerProvider

logging.disable(logging.WARNING)
SDK = sys.argv[1] if len(sys.argv) > 1 else "openai"


class Spans(SpanProcessor):
    def __init__(self): self.open, self.ended = set(), []
    def on_start(self, span, parent_context=None): self.open.add(span.context.span_id)
    def on_end(self, span):
        self.open.discard(span.context.span_id)
        self.ended.append(f"{span.name} [{span.status.status_code.name}]")


spans = Spans()
provider = TracerProvider()
provider.add_span_processor(spans)


def slow_http(hx=httpx):                           # a server that answers after 5 s
    async def slow_server(request):
        await asyncio.sleep(5)
        return hx.Response(200, json={})
    return hx.AsyncClient(transport=hx.MockTransport(slow_server))


M = [{"role": "user", "content": "hi"}]
if SDK == "openai":
    import openai
    from opentelemetry.instrumentation.openai import OpenAIInstrumentor
    OpenAIInstrumentor().instrument(tracer_provider=provider)
    client = openai.AsyncOpenAI(api_key="x", base_url="http://mock/v1", max_retries=0, http_client=slow_http())
    calls = {"create": lambda: client.chat.completions.create(model="m", messages=M),
             "create(stream=True)": lambda: client.chat.completions.create(model="m", messages=M, stream=True)}
    pkg = ("openai", "opentelemetry-instrumentation-openai")
elif SDK == "anthropic":
    import anthropic, anthropic._base_client as base
    from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor
    AnthropicInstrumentor().instrument(tracer_provider=provider)
    hx = getattr(base, "httpx2", None) or httpx            # newer anthropic releases use httpx2
    client = anthropic.AsyncAnthropic(api_key="x", base_url="http://mock", max_retries=0, http_client=slow_http(hx))
    calls = {"create": lambda: client.messages.create(model="m", max_tokens=50, messages=M),
             "create(stream=True)": lambda: client.messages.create(model="m", max_tokens=50, messages=M, stream=True)}
    pkg = ("anthropic", "opentelemetry-instrumentation-anthropic")
else:
    from google import genai
    from google.genai import types
    from opentelemetry.instrumentation.google_generativeai import GoogleGenerativeAiInstrumentor
    GoogleGenerativeAiInstrumentor().instrument(tracer_provider=provider)
    client = genai.Client(api_key="x", http_options=types.HttpOptions(base_url="http://mock", httpx_async_client=slow_http()))
    calls = {"generate_content": lambda: client.aio.models.generate_content(model="m", contents="hi"),
             "generate_content_stream": lambda: client.aio.models.generate_content_stream(model="m", contents="hi")}
    pkg = ("google-genai", "opentelemetry-instrumentation-google-generativeai")


async def main():
    print(" | ".join(f"{p} {version(p)}" for p in pkg))
    for name, call in calls.items():
        spans.ended.clear()
        try:
            await asyncio.wait_for(call(), timeout=1.0)    # the app gives up after 1 s
        except TimeoutError:
            pass
        gc.collect()
        await asyncio.sleep(0.1)
        print(f"{name:25} timed out -> ended spans: {spans.ended or 'none'}; spans never ended: {len(spans.open)}")
        spans.open.clear()

asyncio.run(main())

The sync clients behave the same way when the call is interrupted with Ctrl-C (KeyboardInterrupt) while waiting for the response.

👍 Expected behavior

The span ends with status ERROR and the exception (CancelledError / KeyboardInterrupt) recorded. The exception still propagates to the caller unchanged.

Suggested fix: catch BaseException instead of Exception around the wrapped call (or end the span in a finally block).

👎 Actual Behavior with Screenshots

The span is never ended, so it is never exported:

openai 3.26.0 | opentelemetry-instrumentation-openai 0.62.4
create                    timed out -> ended spans: none; spans never ended: 1
create(stream=True)       timed out -> ended spans: none; spans never ended: 1
anthropic 1.11.0 | opentelemetry-instrumentation-anthropic 0.62.4
create                    timed out -> ended spans: none; spans never ended: 1
create(stream=True)       timed out -> ended spans: none; spans never ended: 1
google-genai 2.28.0 | opentelemetry-instrumentation-google-generativeai 0.62.4
generate_content          timed out -> ended spans: none; spans never ended: 1
generate_content_stream   timed out -> ended spans: none; spans never ended: 1
🤖 Python Version

3.12

📃 Provide any additional context for the Bug.

The span-ending gap is in the request wrappers. A stream that is cancelled in the middle of reading is also not recorded as an error: the stream wrappers catch only Exception as well.

👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
Are you willing to submit PR?

Yes I am willing to submit a PR!

主要语言
Python
星标
7.5k
派生
1.2k
平均合并
1 天 23 小时
30 天内合并 PR
3

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

traceloop/openllmetry 的其他 Issue

查看 traceloop/openllmetry 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。