Set write buffer limit to zero on client and server

[asvetlov]

#2698

[fafhrd91]

You should run some benchmarks

[codecov-io]

Codecov Report

Merging #2869 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #2869      +/-   ##
==========================================
+ Coverage   97.97%   97.97%   +<.01%     
==========================================
  Files          39       39              
  Lines        7464     7466       +2     
  Branches     1310     1310              
==========================================
+ Hits         7313     7315       +2     
  Misses         48       48              
  Partials      103      103

Impacted Files	Coverage Δ
aiohttp/client_proto.py	94.44% <100%> (+0.04%)	⬆️
aiohttp/web_protocol.py	91.83% <100%> (+0.02%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 39831ca...98e96f7. Read the comment docs.

[asvetlov]

Sure will do

[pfreixes]

@asvetlov just some general considerations, I would like to put myself int he same page as you because sure I'm missing something.

The StreamWriter.drain method uses the Asyncio method _drain_helper which checks if the stream is paused, in case that is paused will wait till the stream is resumed. Having a new limit of 0 for the high water mark - that is what this MR is doing - shouldn't we expect that the stream is never paused? Making useless the call to the drain?

At the same time, the call to the StreamWriter.drain made by the StreamWriter.write it's at least a bit confusing IMHO - sure I'm missing something again. It calls explicitly the drain if the size of the buffer sent till that specific moment is greater than 0x10000, my questions here are:

1. Why 16 bytes?
2. Why call the drain when we have sent 16 bytes?
3. Why we are not just calling the Asyncio._drain_helper setting the high watermark to the desired value?

[asvetlov]

Good questions. The initial idea was not using aiohttp/asyncio buffering but only OS kernel write buffer for a socket.
On benchmarking I see ~7% slowdown, need to investigate more.