If the file exists we can still open it with "wb" mode.If we open it with "r+b" mode, then we are not overwriting the whole file with new content but just changing the file starting from the beginning of the file. I doubt that is what we want.

Imagine file with content b"abc"
and you want to change it

with open(path, 'r+b') as f:
    f.write(b"x") 

-> b"xbc"

with open(path, 'wb') as f:
    f.write(b"x") 

-> b"x"

Maybe I have missed the point here, but so far it does not seem right to me.

Yes, I suppose it is because of wb + the fact that we work with files asynchronously that we have a situation where empty files are created.

We open the file in wb mode, deleting the contents and switching the asynchronous context. If the crawler is interrupted at this point, the file remains empty.

I settled on r+b mode because we always write formatted json files. These are also metadata files, so the fields in them are not changed. So I think it should work, since we will be overwriting the same number of lines each time.

Thanks for explanation. I think I am getting it now. But for sure this would be great to have in well commented test as it is by no means self-explanatory.

After testing and thinking about it, I decided to go back to wb. It's quite a rare bug and it doesn't have a critical impact. But if we encounter some artifacts due to r+b, it can be much more painful to find the cause.

I added a test to reproduce the case of creating an empty file. Maybe we can use it if we try to find a more stable solution for overwriting files.