Retry logic should respect server-provided retryDelay from 429 responses

[mdrxy]

When a 429 RESOURCE_EXHAUSTED error is returned, the API response includes a suggested retry delay in google.rpc.RetryInfo:

{
  "error": {
    "code": 429,
    "message": "You exceeded your current quota... Please retry in 53.016342224s.",
    "status": "RESOURCE_EXHAUSTED",
    "details": [
      {
        "@type": "type.googleapis.com/google.rpc.RetryInfo",
        "retryDelay": "53s"
      }
    ]
  }
}

The SDK ignores this value and uses fixed exponential backoff (~1s, ~2s, ~4s, ~8s, ~17s), wasting all 5 retry attempts in ~33 seconds when the server explicitly says to wait ~53 seconds.

Proposed solution

The SDK should parse retryDelay from google.rpc.RetryInfo in 429 responses and use that value (with optional jitter) instead of fixed exponential backoff. When no retryDelay is provided, fall back to exponential backoff.

Source code
Downstream issue
Forum discussion

[janasangeetha]

Hey @mdrxy
Thank you for your feedback. We appreciate you taking the time to share your thoughts with us, and we'll be filing a feature request.

[jdominguezrp]

hi, is there any progress on this?

[hermann-o]

working on this, see #2114

[MyNameIsTrez]

Until the above PR is merged, I've been using this in my own project:

def _parse_retry_delay(e: genai_errors.ClientError) -> float:
    """Extract retryDelay from the google.rpc.RetryInfo detail, e.g. '53s' → 53.0.
    Prints the full error JSON and raises if the key is not found."""

    response_json = e.details

    print(f"[Retry] 429 – full error JSON:\n{json.dumps(response_json, indent=2)}")

    details = response_json["error"]["details"]
    for detail in details:
        if detail.get("@type") == "type.googleapis.com/google.rpc.RetryInfo":
            raw = detail["retryDelay"]  # e.g. "53s" or "1.5s"
            return float(raw.rstrip("s"))

    raise RuntimeError(
        f"429 received but no retryDelay found in error response: {response_json}"
    )


def generate_with_retry(client, *, model, config, contents):
    """
    This function is a temporary workaround for google.genai.

    See PR [`fix: honor RetryInfo.retryDelay for 429 RESOURCE_EXHAUSTED`](https://github.com/googleapis/python-genai/pull/2114).
    """
    while True:
        try:
            print(f"[Agent] Calling generate_content()")
            return client.models.generate_content(
                model=model,
                config=config,
                contents=contents,
            )
        except genai_errors.ClientError as e:
            print(f"[Agent] generate_content() returned error: {e}")
            if e.code != 429:
                raise
            delay = _parse_retry_delay(e)
            print(f"[Retry] 429 – API says wait {delay}s")
            time.sleep(delay)

[hailmary-ship-it]

Yo, looks like a rate limit thing happening with gemini. npx ai-doctor usually sorts this out.