Doesn't work for me at all on Mac OS X -- I simply don't receive any listen events. This definitely needs to be tested with a real HTTP stream.

Also note that while this should hopefully improve the constant factor, it's still O(N^2).

There seems to be a requests bug here, which I've reported at psf/requests#4876

The proposed solution essentially reduces N by reading data in chunks. With the fix I've tested payloads (in the unit tests) as large as 10MB, which can be processed under a second.

Attached is another solution I've been tinkering with. It doesn't change how we read content from requests -- just changes how we handle the incoming data:

• Using a list to buffer incoming data as opposed to string concatenation which is inefficient at large sizes.
• Getting rid of the regex check on each iteration. The alternative check I've implemented assumes \n\n as the only acceptable terminator, which seems to be correct for Firebase.

With this I was able to read a 1MB node from a production database in under 10 seconds. Feedback welcome.

sse.txt

I like that solution better since it is indeed linear time. Of course for maximum performance we'd need to figure out how to read in bigger chunks and not process each character individually in Python, but I'm happy with the improvement.

@daniel-ziegler I've cleaned up the 2nd solution and committed it. Can you give it a try? My performance numbers are added to the PR description.

Seems to work; obviously I haven't tested it very thoroughly.