httpx.TimeoutException is not retried despite HttpRetryOptions being configured #2337

Is this a client library issue or a product issue?

This is a client library issue (google-genai). The SDK's retry predicate only matches errors.APIError subclasses, meaning lower-level HTTP transport exceptions (httpx.TimeoutException and its subclasses) are silently dropped and never retried - even when the user has explicitly configured HttpRetryOptions.

Environment details

• Programming language: Python
• OS: macOS (darwin 24.6.0), also reproducible on any OS
• Language runtime version: Python 3.12.12
• Package version: google-genai==1.73.1

Steps to reproduce

The following self-contained script has no external dependencies, no credentials, and makes no network calls

import contextlib
import httpx
import tenacity
from google.genai._api_client import retry_args
from google.genai.errors import ServerError
from google.genai.types import HttpRetryOptions

_ATTEMPTS = 3
_OPTIONS = HttpRetryOptions(
    attempts=_ATTEMPTS,
    http_status_codes=[408, 429, 500, 502, 503, 504],
    initial_delay=0.0,
    max_delay=0.0,
)


async def _run(fn) -> int:
    call_count = 0

    async def wrapped():
        nonlocal call_count
        call_count += 1
        return await fn()

    kwargs = retry_args(_OPTIONS)
    async_retry = tenacity.AsyncRetrying(**kwargs)
    with contextlib.suppress(httpx.TimeoutException, Exception):
        await async_retry(wrapped)
    return call_count


# ✅ Works as expected: 500 ServerError is retried 3 times.
async def raises_500():
    raise ServerError(500, {"error": {"code": 500, "message": "oops", "status": "INTERNAL"}})

# ❌Bug: httpx.ReadTimeout is never retried regardless of `attempts`.
async def raises_timeout():
    raise httpx.ReadTimeout("simulated network stall")

Expected behavior:
raises_timeout() should be retried _ATTEMPTS (3) times. Google's own Vertex AI retry strategy docs explicitly list "Socket timeouts and TCP disconnects" as retryable errors: https://cloud.google.com/vertex-ai/generative-ai/docs/retry-strategy

Actual behavior:
raises_timeout() is called exactly once and the exception propagates immediately, completely bypassing the configured HttpRetryOptions.

Root cause:
In _api_client.py, the tenacity retry predicate is:

retry = tenacity.retry_if_exception(
    lambda e: isinstance(e, errors.APIError) and e.code in retriable_codes,
)

httpx.TimeoutException (and its subclasses ReadTimeout, ConnectTimeout, etc.) do not inherit from errors.APIError, so the predicate always returns False and tenacity stops after the first attempt.

Why this matters in production:
LLM inference calls can take 30–60+ seconds. In that window, cloud load balancers, NAT gateways, and firewalls may silently drop idle TCP connections, causing the httpx client to raise ReadTimeout mid-request. Without retry coverage for these exceptions, callers experience hard failures from transient infrastructure noise - the exact scenario HttpRetryOptions is supposed to protect against.

Suggested fix:
Extend the retry predicate in retry_args() to also cover httpx.TimeoutException and httpx.ConnectError:

retry = tenacity.retry_if_exception(
    lambda e: (isinstance(e, errors.APIError) and e.code in retriable_codes)
              or isinstance(e, (httpx.TimeoutException, httpx.ConnectError))
)