Repository navigation
live.connect() (Vertex AI) resolves ADC and refreshes the OAuth token synchronously on the event loop on every connect #3056
Description
Environment
- google-genai 2.24.0, google-auth 2.58.0, Python 3.12
- Vertex AI backend (
GOOGLE_GENAI_USE_VERTEXAI=true), service-account ADC - Live API with many concurrent sessions per process (telephony voice agent)
What happens
On Vertex, AsyncLive.connect() (google/genai/live.py, ~L1055-1069) runs this before it opens the websocket:
if not self._api_client._credentials:
creds, _ = google.auth.default(scopes=[...]) # sync, on the event loop
else:
creds = self._api_client._credentials
if not (creds.token and creds.valid):
auth_req = _auth_requests().Request()
creds.refresh(auth_req) # sync HTTP, on the event loopTwo problems:
- Blocking I/O inside an async API.
creds.refresh()makes a synchronousrequestscall to the token endpoint, on the event loop. Every other coroutine in the process stalls until it returns. For us that is live audio for every other call on the pod. - No caching when the client was built without
credentials=. The resolved credentials are a local variable and are never stored on_api_client. Everyconnect()therefore runsgoogle.auth.default()again and refreshes a fresh, token-less credentials object. The sync refresh runs on every connect, not once an hour.
The async REST path already handles both. _api_client._async_access_token() loads auth via asyncio.to_thread(load_auth, ...) under an async lock and caches it in self._credentials. async_get_token_from_credentials() refreshes via asyncio.to_thread(refresh_auth, ...), only when expired.
Repro
import asyncio, time
from google import genai
async def ticker(stop, worst):
while not stop.is_set():
t = time.perf_counter(); await asyncio.sleep(0.01)
worst[0] = max(worst[0], (time.perf_counter() - t - 0.01) * 1000)
async def main():
client = genai.Client() # Vertex via env, no credentials= passed
for i in range(3):
stop, worst = asyncio.Event(), [0.0]
tk = asyncio.create_task(ticker(stop, worst))
await asyncio.sleep(0.05) # let the ticker start before connect() blocks
async with client.aio.live.connect(
model="gemini-live-2.5-flash-native-audio",
config={"response_modalities": ["AUDIO"]},
):
pass
stop.set(); await tk
print(f"connect #{i+1}: max event-loop stall {worst[0]:.0f} ms")
await client.aio.aclose()
asyncio.run(main())Output (Vertex, us-central1). I also counted service_account.Credentials.refresh calls:
connect #1: max event-loop stall 517 ms, token refreshes during connect: 1
connect #2: max event-loop stall 410 ms, token refreshes during connect: 1
connect #3: max event-loop stall 415 ms, token refreshes during connect: 1
Expected: auth resolved once per client, refresh only near expiry, and never on the event loop. Stalls should be a few ms.
Suggested fix
In AsyncLive.connect(), reuse the async REST path: await self._api_client._async_access_token(). That loads ADC off-loop, caches it on the client, and refreshes off-loop under the async auth lock. At minimum, store the resolved credentials on _api_client and wrap google.auth.default() and creds.refresh() in asyncio.to_thread.
Workaround
We build one genai.Client(credentials=creds) per process, in a thread. We keep creds FRESH by calling creds.refresh() via asyncio.to_thread whenever creds.token_state is not TokenState.FRESH. valid turns False at the same REFRESH_THRESHOLD, so connect() never takes the sync branch. With the same ticker, max stall per connect drops to 9–28 ms.
Activity
hossiendehghan989 commentedon Oct 6, 2026
I independently traced the current
mainimplementation and the existing async-auth tests. The report is consistent with the source:AsyncLive.connect()still callsgoogle.auth.default()andCredentials.refresh()directly, whileBaseApiClient._async_access_token()already provides the desired off-event-loop loading, caching, and per-loop refresh lock.One focused regression seam that would make this especially actionable is an
AsyncLive.connect()integration test with the websocket handshake mocked:credentials=,BaseApiClient._async_access_tokento return a deterministic token,google.auth.defaultandCredentials.refreshto fail the test if called directly,That test would verify the behavior at the call site rather than only testing
_async_access_token()in isolation (the latter is already covered bytest_async_access_token_locking). It would also catch regressions where the implementation accidentally reintroduces synchronous ADC resolution in the Live path while the shared auth helper remains correct.The proposed direction—routing Vertex Live authentication through
_async_access_token()—therefore looks narrowly scoped and consistent with the SDK’s existing auth design. The main compatibility points to preserve are explicit credentials, API-key authentication, custom base URLs, and the currentAuthorizationheader precedence.Venkaiahbabuneelam commentedon Oct 6, 2026
Hi @sandeep2rawat, Thank you for bringing this to our attention. We truly appreciate you flagging this issue regarding Vertex AI live.connect() freezing the event loop during token refresh, we will file a bug internally.
Thanks