jonathan rochkind <jonathan / dnil.net> wrote:
> 1. Huge performance problem of the way timeouts are implemented. 

I would like to split the HTTP open_timeout into two components:
  connect_timeout and dns_timeout

connect_timeout would be trivial to implement

> 2.  global-interpreter-blocking nature of DNS lookups, instead of a
> non-blocking select DNS lookup. 

Like akr said, it's not an issue now.  The performance issue is that
getaddrinfo() in standard C libraries doesn't provide configurable
timeouts so we have to use the nasty timeout library that spawns
a thread every time we need to do a DNS lookup...

-- 
Eric Wong