Optimize request.pipeline writing

[ronag]

It should be possible to have the returned Duplex write directly to the socket. Instead of providing a Readable as request body we can provide a function, i.e. the reverse of what we do on the response side in request.enqueue.

This would allow us to entirely bypass the intermediate PassThrough or Readable between the Duplex and the socket.

[ronag]

Might also make use of WritableBase which is being worked on here. Could extract that out into this library in order to use of it across Node versions.

[mcollina]

go for it

[szmarczak]

I ran the benchmarks on master again. I got:

http - keepalive - pipe x 9,623 ops/sec ±3.24% (76 runs sampled)
undici - request - pipe x 16,881 ops/sec ±0.45% (86 runs sampled)
undici - pipeline - pipe x 11,107 ops/sec ±0.81% (87 runs sampled)
undici - stream - pipe x 17,350 ops/sec ±0.61% (86 runs sampled)

Indeed it's still faster than the native http2 module 🎉

Also I fixed the h2o.conf in http2-wrapper so when you clone the repo you can easily start the server just by running
h2o -c h2o.conf (no need to clone the examples in the h2o repository).

[ronag]

Let's stick to normal streams and not go to crazy here. pipeline has reasonable perf now.