`AsyncAnthropicBedrock` blocks the event loop during AWS credential resolution or signing

[jjurm]

Description

AsyncAnthropicBedrock performs AWS credential resolution and SigV4 signing synchronously on the event-loop thread.

Expected: AsyncAnthropicBedrock should not execute potentially blocking credential resolution or refresh on the event-loop thread.

Reproduction

import asyncio
import time
from unittest.mock import patch

from anthropic import AsyncAnthropicBedrock


async def heartbeat():
    while True:
        await asyncio.sleep(0.5)
        print("heartbeat")


def slow_get_credentials(*args, **kwargs):
    time.sleep(5)
    raise RuntimeError()


async def main():
    client = AsyncAnthropicBedrock(aws_region="us-east-1")

    with patch(
        "botocore.session.Session.get_credentials",
        side_effect=slow_get_credentials,
    ):
        await asyncio.gather(
            heartbeat(),
            client.messages.create(
                model="anthropic.claude-3-haiku-20240307-v1:0",
                max_tokens=1,
                messages=[{"role": "user", "content": "hi"}],
            ),
        )


asyncio.run(main())

No heartbeat output appears during the five-second credential call.

Cause: AsyncAnthropicBedrock._prepare_request() calls the synchronous credential/signing path directly.

[jjurm]

For reference, openai-python's AsyncBedrockOpenAI (also SigV4-signed against boto3 credentials, so subject to the same class of blocking-refresh risk) handles this correctly, see openai/providers/bedrock.py's prepare_async_request.

[nightcityblade]

Hi, I'd like to work on this. I'll submit a focused PR with regression coverage shortly.

[nightcityblade]

I’m stepping back from this issue and unclaiming it so someone else can take it. I did not open a PR or push any changes.

[IgorGanapolsky]

Operational reliability note for anyone hitting this on a busy agent loop:

When credential resolution / SigV4 signing runs synchronously on the asyncio loop thread, every other coroutine shares that stall — heartbeats, tool timeouts, and "are we still alive?" checks all go quiet at once. That shows up as a flaky agent even when the Bedrock call itself eventually succeeds.

Practical isolation pattern while the SDK path is fixed:

1. Treat credential refresh as a blocking boundary — run it in asyncio.to_thread(...) (or a small executor) and only await the future from the loop.
2. Put a wall-clock budget around the refresh (e.g. 2–5s) so a hung IMDS / SSO path fails closed instead of parking the whole process.
3. Emit a structured terminal event on timeout: {stage: "aws_cred_resolve", status: "timeout|error", duration_ms} so dashboards can distinguish "model slow" from "identity plane blocked the loop".
4. Keep a last-good credential cache with explicit expiry so concurrent AsyncAnthropicBedrock clients do not stampede refresh on the same thread.

Expected-vs-actual check that surfaces this class of bug early: assert that a 200ms heartbeat task still fires while a synthetic 2s sleep is injected into the credential mock. If the heartbeat stops, the loop is blocked — not "AWS is slow".

Happy to expand on the heartbeat harness if useful; this is pure free diagnostic framing, not a product pitch.

[retornam]

I'm beginning to think they don't merge contributions from people outside the company or even look at bug reports anymore looking at the most recent commits

https://github.com/anthropics/anthropic-sdk-python/commits/main/