aiohttp streaming fails on initial cassette recording

When using the aiohttp Streaming API and recording a new cassette the actual data read will fail, although the cassette is successfully recorded and subsequent tests pass.

Minimal example:

    # Will fail on initial record, but succeed after that
    with vcr.use_cassette('./vcrpytest1.yaml'):
        async with aiohttp.ClientSession() as session:
            async with session.get('https://httpbin.org/range/50') as response:
                byts = b''
                async for data in response.content.iter_chunked(1024*1024):
                    byts += data
                assert len(byts) == 50

    # Fails everytime
    with vcr.use_cassette('./vcrpytest1.yaml', record_mode='all'):
        async with aiohttp.ClientSession() as session:
            async with session.get('https://httpbin.org/range/50') as response:
                byts = b''
                async for data in response.content.iter_chunked(1024*1024):
                    byts += data
                assert len(byts) == 50

Run with Python 3.7.4: aiohttp==3.6.2 vcrpy==3.0.0

I also created a gist that runs through a few different scenarios: https://gist.github.com/mikemoritz/01f7f222f9353846db16a6d07d98d806#file-vcrpytest-py

[nickdirienzo]

Ah! Nice find! And thanks for the detailed investigation @mikemoritz 🙏

Dang, there really isn’t a way to unread. That API is being deprecated in the future so we can’t rely on that…

Thinking aloud here: I’m a tad worried about returning the mock response initially because I don’t think the vcrpy mock response maps directly to the aiohttp response object. However, I think it means that we should be moving towards ensuring they’re as close as possible. I can’t remember if the other clients behave this way, so this might be alright.

I know that play_responses also marks the response as played, so as long as that doesn’t interfere with being able to make requests multiple times in the same test run, then that seems like a good approach. It’d be really cool if we could unread but 🤷‍♂

I’m using aiohttp in my day to day, so this is a neat way to continue learning about it. I don’t have time to work on this library during the day; it’s more of a nights and weekends kinda deal for me right now. If that’s okay with y’all, I’m happy to help review and help here best I can!

[mikemoritz]

Sure thing. When I have some spare minutes there were a few other avenues I wanted to investigate first since I appreciate that returning the mock response initially is a fundamental change for all users to solve a lower percentage use case (probably).