rack.input already read on first upload to a Rails Metal in production

rack.input is being read or seeked through improperly before I touch it. It's only upon the first upload to each of my mongrel cluster's servers. The mongrel server is fronted by Apache mod_proxy_balancer.

Steps to reproduce:

1) Set up a basic Rails app, with a mongrel_cluster and Apache mod_proxy_balancer front end.

2) Create a new Metal, app/metal/test.rb with:

require(File.dirname(__FILE__) + "/../../config/environment") unless defined?(Rails)

class Test
  def self.call(env)
    if env["PATH_INFO"] =~ /^\/test/i
      req = Rack::Request.new(env)
      if req.post?
        return [200, {"Content-Type" => "text/html"}, [env['rack.input'].pos.to_s]]
      else
        return [200, {"Content-Type" => "text/html"}, [ %Q|
<form enctype="multipart/form-data" id="post_form" method="post">
  <input id="file" name="file" size="30" type="file">
  <input id="submit" name="commit" value="submit" type="submit">
</form>
| ]]
      end
    else
      [404, {"Content-Type" => "text/html"}, ["Not Found"]]
    end
  end
end 

3) Run the app, and navigate to /test.

4) Upload a file. Note the seek position reported. It is not zero. Instead, in my tests, it's the end of the StringIO.

5) Repeat the upload N times, where N is the number of mongrel servers in your cluster. (Assuming a round-robin or load-based proxy. Other proxy methods might require less.)

6) On the Nth try, the seek position has gone to 0. I believe this is the proper value.

Note that this does not happen in development mode, under an unproxied Rails rackup server, nor an unproxied mongrel server. I have not tried a rackup with just this Metal (not familiar enough with it).

The workaround is simple: call env['rack.input'].rewind before doing anything else in the Metal.

Versions: - Rails 2.3.2 - Rack 0.9.1 - Ruby 1.8.6 - Mongrel 1.1.5 - Apache 2.2.9 - Ubuntu 8.10

I'm not sure which subsystem is doing this. Please let me know if this is mongrel's or rack's problem.