From: Emmanuel Gaillot <emmanuel.gaillot@...>
Date: 2007-07-30T00:05:11+09:00
Subject: Puzzled about Ruby's behaviour re. closed sockets

Everyone,

I'd like to write a test that ensures a Server application correctly
handles closed client connections -- more specifically: I want to set
up a TCPServer, then a client, then close the client, then attempt to
write on the client Connection from server side -- and make sure no
Error gets fired.

To my great surprise, the test doesn't behave in any predictible way
-- sometimes it passes, sometimes it fails.

I've written a piece of code that illustrates it:
=============
require 'socket'

consecutive_puts_to_attempt = 1
ok_counter = 0
error_counter = 0

10000.times do
  server = TCPServer.open '127.0.0.1', 4807
  client = TCPSocket.open '127.0.0.1', 4807
  connection = server.accept
  client.close

  begin
    consecutive_puts_to_attempt.times {connection.puts "Hello!"}
    ok_counter += 1
  rescue
    error_counter += 1
  end
  server.close
end

puts "(Unexpectedly) passed #{ok_counter} times;"
puts "(correctly) failed #{error_counter} times."
=============

Statistically, this piece of code unexpectedly passes with no raised
error about 98% of the time.  Errors are Errno::EPIPE and
Errno::ECONNRESET

When I fiddle with consecutive_puts_to_attempt value, the number of
expected raised errors changes --

puts attempts      chances of raised error
1                            2%
2                            98%
3                            98.5%
4                            99%


Here is my environment:
ruby -v
> ruby 1.8.6 (2007-03-13 patchlevel 0) [i686-darwin8.10.1]

Any idea of what's going on?
Any suggestion about how I could make this piece of code predictible?
-- Emmanuel.