🐛 Bug Report: llama streams: close() before reading never ends the span; aclose() leaves the HTTP response (connection) open

Aperta
#4,608 4 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

@Suddha-Abhitej ci sta già lavorando.

Dal 13/10/2026.

  • #4617 di @Suddha-Abhitej — aperta
  • #4664 di @Mike-Animal-Counseling — chiusa senza merge

Valutazione

Difficoltà
3/5
Tempo stimato
1-2 giorni
Idoneità per principianti
68/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
ollama, python

Direzione di ricerca

Start in opentelemetry/instrumentation/ollama/__init__.py at _accumulate_streaming_response and _aaccumulate_streaming_response, then check the Ollama streaming tests. Reproduce the sync close() before reading and async aclose() after one chunk using the supplied script. Done means both paths end their spans and the async path closes the underlying HTTP response promptly.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Which component is this bug for?

Ollama Instrumentation

📜 Description

For stream=True, the Ollama instrumentation returns its own generator, _accumulate_streaming_response / _aaccumulate_streaming_response (opentelemetry/instrumentation/ollama/__init__.py, lines 90 and 161 at 0.62.4), which iterates over the Ollama stream and ends the span in its finally. Two problems follow from that:

  1. Sync, closed before reading. The span is created before the generator, but a generator that is closed before its first next() never runs its body, so the try / finally that ends the span is never entered. stream.close() before reading (the app decides it does not need the answer) leaves the span open forever: it is never exported and stays referenced.

  2. Async, aclose(). await stream.aclose() closes the instrumentation's async generator (its finally ends the span), but not the Ollama async stream it iterates over: async generators are not closed when an enclosing async for is interrupted. The Ollama stream holds the httpx response, so the response, and its connection, stays open until the stream is garbage-collected. Without the instrumentation, aclose() closes it immediately. In a server that cancels streamed answers this holds connections from the pool.

A fix would be to end the span (and close the wrapped stream) in the wrapper's close path: e.g. a small wrapper class whose close() / aclose() end the span if the iteration never started, and to call response.close() / await response.aclose() in the finally blocks.

👟 Reproduction steps

pip install ollama opentelemetry-instrumentation-ollama opentelemetry-sdk httpx

import asyncio, gc, json, httpx, ollama
from importlib.metadata import version
from opentelemetry.sdk.trace import SpanProcessor, TracerProvider
from opentelemetry.instrumentation.ollama import OllamaInstrumentor

class Spans(SpanProcessor):
    def __init__(self): self.started, self.ended = {}, {}
    def on_start(self, span, parent_context=None): self.started[span.context.span_id] = span.name
    def on_end(self, span): self.ended[span.context.span_id] = span.name
    def open(self): return [n for k, n in self.started.items() if k not in self.ended]

spans = Spans()
provider = TracerProvider()
provider.add_span_processor(spans)
OllamaInstrumentor().instrument(tracer_provider=provider)
CLOSED = []

class Body(httpx.SyncByteStream, httpx.AsyncByteStream):     # NDJSON body; records whether the client closed it
    def __init__(self, lines): self.lines = lines
    def __iter__(self): yield from self.lines
    async def __aiter__(self):
        for x in self.lines: yield x
    def close(self): CLOSED.append(True)
    async def aclose(self): CLOSED.append(True)

def respond(request):            # local stand-in for the Ollama server: a 3-part streamed answer
    chat = request.url.path.endswith("/api/chat")
    parts = [{"model": "m", "created_at": "2026-01-01T00:00:00Z", "done": False,
              **({"message": {"role": "assistant", "content": w}} if chat else {"response": w})} for w in ("Hel", "lo", "!")]
    parts.append({"model": "m", "created_at": "2026-01-01T00:00:00Z", "done": True, "done_reason": "stop",
                  **({"message": {"role": "assistant", "content": ""}} if chat else {"response": ""})})
    return httpx.Response(200, headers={"content-type": "application/x-ndjson"},
                          stream=Body([(json.dumps(p) + "\n").encode() for p in parts]))

print(f"ollama {version('ollama')} | opentelemetry-instrumentation-ollama {version('opentelemetry-instrumentation-ollama')}")
client = ollama.Client(host="http://mock", transport=httpx.MockTransport(respond))
stream = client.generate(model="m", prompt="hi", stream=True)
stream.close()                                   # the app decides it does not need the answer
gc.collect()
print("sync:  close() before reading     -> spans never ended:", spans.open())

async def main():
    aclient = ollama.AsyncClient(host="http://mock", transport=httpx.MockTransport(respond))
    stream = await aclient.chat(model="m", messages=[{"role": "user", "content": "hi"}], stream=True)
    await stream.__anext__()
    await stream.aclose()                        # the app stops reading after the first part
    print("async: aclose() after 1 chunk     -> HTTP response closed:", bool(CLOSED))
asyncio.run(main())
👍 Expected behavior

close() before reading ends the span (the call was made, it was just not read), and aclose() closes the underlying Ollama stream, so its HTTP response is closed right away, as it is without the instrumentation.

👎 Actual Behavior with Screenshots
ollama 0.6.3 | opentelemetry-instrumentation-ollama 0.62.4
sync:  close() before reading     -> spans never ended: ['ollama.completion']
async: aclose() after 1 chunk     -> HTTP response closed: False

The same script without OllamaInstrumentor().instrument(...) prints HTTP response closed: True.

🤖 Python Version

3.12

📃 Provide any additional context for the Bug.

opentelemetry-instrumentation-ollama 0.62.4 (the latest release), ollama 0.6.3. Not covered by the open PRs #4127 (exceptions) or #4223 (semantic conventions).

👀 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?

None

Lingua principale
Python
Stelle
7.5k
Fork
1.2k
Merge medio
2g 6h
PR unite (30g)
2

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di traceloop/openllmetry

Tutte le issue di traceloop/openllmetry

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.