Deadlock when using GoogleAzureOCR with parallel threads

[phschoepf]

When using GoogleAzureOCR and calling ocr from multiple concurrent threads (e.g. via asyncio.to_thread), the program will sometimes hang on the creation of a new GoogleOCR instance.

Setting breakpoints before creating the GoogleOCR instance resolves the issue - it's a heisenbug.

• Python 3.9.19
• ocr_wrapper 0.0.24

MWE

(the ocr function is the relevant part of GoogleAzureOCR.ocr)

import asyncio
import time

from ocr_wrapper import GoogleOCR, AzureOCR

# Requirements to reproduce the issue:
# - Async function running in asyncio.run()
# - In that function, asyncio.gather() awaitables under a semaphore (>1)
# - The awaitables are synchronous functions wrapped in asyncio.to_thread()


async def preprocess():
    semaphore = asyncio.Semaphore(4)

    async def _sem_wrapper():
        async with semaphore:
            return await asyncio.to_thread(ocr)

    awaitables = [_sem_wrapper() for _ in range(2)]
    page_texts = await asyncio.gather(*awaitables)
    return page_texts


def ocr():
    google_ocr = GoogleOCR(
        auto_rotate=True,
        correct_tilt=False,
        ocr_samples=1,
        max_size=4096,
        verbose=False,
    )
    azure_ocr = AzureOCR(
        auto_rotate=False,
        correct_tilt=False,
        ocr_samples=1,
        max_size=4096,
        verbose=False,
    )
    return (google_ocr, azure_ocr)


if __name__ == "__main__":
    start = time.perf_counter()
    res = asyncio.run(preprocess())

    print(res)
    print(f"Time taken: {time.perf_counter() - start:.2f}s")

Expected behavior

Runs and exits without error in <<1s

Actual behavior

Hangs forever, process cannot be killed with Ctr+C

[phschoepf]

Solved by upgrading grpcio (dependency of google-cloud-vision) manually to 1.62.2.

Did a quick grid search, it seems that all versions grpcio>=1.59,<=1.62.1 have the bug.

[phschoepf]

Relevant gRPC issue: grpc/grpc#36376

[phschoepf]

fixed in #18