Regarding wheel ID vs. reference, I meant it might be better to pass around the client's wheel ID rather than a reference to the wheel itself. As in:

  $kernel->yield('keepalive', $io_wheel->ID, $kernel);
[download]

and

  keepalive => sub {
      my ( $io_wheel_id, $kernel, $heap ) = @_[ ARG0, KERNEL ];
      return unless exists $heap->{clients}{$io_wheel_id};
      $heap->{clients}{$io_wheel_id}->put("keepalive");
      $kernel->delay( 'keepalive' => 10, $io_wheel_id );
  },
[download]

It's more work, for you and for the program, but it stops repeating the timer after the client has disconnected. Of course, another way is to make sure that the timer is canceled wherever you are deleting the client. If you find yourself doing that in multiple places, refactor it into a subroutine like:

sub delete_client {
  my ($kernel, $heap, $io_wheel_id) = @_;
  delete $heap->{clients}{$io_wheel_id};
  $kernel->delay(keepalive => undef);
}
[download]

which would be called the usual Perlish way

        on_connect_error => sub {
            # Handle client error, including disconnect.
            my $wheel_id = $_[ARG3];
            delete_client(@_[KERNEL, HEAP], $wheel_id);
        },
[download]

It really helps when cleanup gets complex, as it often does when programs evolve and grow.