Process hangs until SQLiteBackend instance is closed

[aaraney]

The problem

Process will hang until instance of SQLiteBackend is closed. More specifically, the process will hang until the SQLiteBackend's responses: SQLitePickleCache and redirects: SQLiteCache instance variables are closed. Note, Ive not tested this on other backends.

Expected behavior

Process will exit without explicit close of SQLiteBackend.

Steps to reproduce the behavior

import asyncio
from aiohttp_client_cache import CachedSession, SQLiteBackend

async def hangs_forever():
    cache = SQLiteBackend("test_cache")
    async with CachedSession(cache=cache) as session:
        await session.get("http://httpbin.org/delay/1")
    # await cache.close()


if __name__ == "__main__":
    asyncio.run(hangs_forever())

Workarounds

The only workaround ive found thus far is to explicitly close the SQLiteBackend cache instance (e.g. await cache.close()).

Environment

• aiohttp-client-cache version: [e.g. 0.8.2]
• Python version: [e.g. 3.9]
• Platform: [e.g. macOS Ventura 13.4.1 arm64]

[aaraney]

This looks like an issue stemming from aiosqlite.Connection which inherits from threading.Thread.

[JWCook]

Could you give the changes in main a try for me? I believe that will fix the issue. I'll get a patch release out for that soon.

[aaraney]

@JWCook, certainly! Just tested at ae63533 and the process exited as expected.

[aaraney]

For those interested, the fix was introduced in a33331f.

[aaraney]

Fixed in 0.9.0

[aaraney]

Related to #147.

[aaraney]

I am reopening this because it seems that the switch to aiosqlite has altered the supported semantics for using aiohttp_client_cache.CachedSession. The following code hangs:

import asyncio
from aiohttp_client_cache import SQLiteBackend, CachedSession


async def main():
    cache = SQLiteBackend()
    client = CachedSession(cache=cache)

    req = await client.get("http://httpbin.org/get")
    resp = await req.text()
    req.close()
    print(resp)

    # does not hang if properly close (identical behavior to async context manager)
    # await client.close()


if __name__ == "__main__":
    asyncio.run(main())

TLDR; the above code hangs because client.close() is never called and awaited.

It is required that client.close() is called and awaited because SQLiteBackend creates and uses a aiosqlite.Connection that runs in its own thread. The aiosqlite.Connection acts like an event loop (see its run() method). It accepts work (coroutines) over a queue and then reschedules coroutines on the main thread's event loop. This is the problem, in the above code, the event loop runs for a shorter amount of time than the aiosqlite.Connection thread (as soon as main() returns, the event loop starts its shutdown). This makes it impossible to close unless done implicitly or explicitly at the call site. In some cases adding a finalizer (__del__) to SQLiteBackend that schedules a close() task on the main event loop to could remedy this, but there is no guarantee that the event loop will be open long enough for that task to be scheduled and run.

The unfortunate side-affect of this behavior is that all code that uses a SQLiteBackend passes the responsibility of handling the lifetime of the aiosqlite.Connection's thread to the end user. Meaning users of that code need to either explicitly call a shutdown method (i.e. close()) or use the object via a context manager. For example:

import asyncio
from aiohttp_client_cache import SQLiteBackend, CachedSession


class Client:
    def __init__(self):
        self._session = CachedSession(cache=SQLiteBackend())

    async def get(self) -> str:
        async with self._session.get("http://httpbin.org/delay/1") as req:
            return await req.text()

    async def close(self) -> None:
        # failing to call close will result in the program hanging
        await self._session.close()

    async def __aenter__(self) -> "Client":
        return self

    async def __aexit__(self, exc_type, exc_val, exc_tb) -> None:
        await self.close()


async def main():
    client = Client()
    res = await client.get()
    print(res)
    # this hangs because client.close() is never called


if __name__ == "__main__":
    asyncio.run(main())

[JWCook]

Are you still seeing this behavior in the latest release (0.9.1)? SQLite connections are now closed automatically on contextmanager exit, and your code snippet appears to work fine for me locally and in CI. Note that this behavior can also be changed with the autoclose parameter (see #162).

[netomi]

If the CachedSession is not properly closed this can happen, however, close is not called magically without using a contextmanager.

You have to do:

  async with CachedSession(cache=SQLiteBackend()) as client:
      # use client

after the with block finished, the session and the associated cache will be properly closed and no dangling sqlite connection will be running.