I love learning things, and this was no exception.

First of all, a minor clarification:

perl is smarter than me, in fact much to my consternation it's sometimes just darn clever (didn't it read Damian's book?). IAC, File::Glob does not get loaded at runtime. The compiler apparently notices you used <*> or glob() and loads File::Glob at compile time. So my solution of wrapping our glob() calls in an eval{} woulod not work.

So I had to fight clever with clever. I'm sure this can be improved, so I'd love commentary. Here's the setup:

Module A uses Module B
Module B has glob() calls
[download]

So I went into Module A and removed the use Module B; and replaced it with this:

BEGIN {
   # Module B was failing to compile on a very random schedule.  The
   # failure was actually on Windows only, and had to do with the 
   # perl compiler seeing use of the glob() function, and 
   # proceeding to load the File::Glob module, including the 
   # File::Glob.dll.  In some very rare occurrances, it
   # would fail, claiming the .dll was not valid.
   #
   # Since this was a compile time failure, we needed to find a 
   # clever way to catch it unfortunately.  This might just work.
   #
   # Try 5 times to load the module
   my $retry   = 5;
   my $succeed = 0;

   while ( $retry-- > 0 ) {

      eval {
         require ModuleB;
         import ModuleB;
      };

      if ( $@ ) {
         # Print an error it if failed
         print "INTERNAL ERROR: ModuleB failed to compile. Retrying\n"
+;

         # Delete the key from the %INC hash, otherwise it 
         # will refuse to try to reload it again
         delete $INC{ 'ModuleB.pm' };

         # Undef the whole darn namespace, to avoid any errors 
         # about redefining things in the namespace
         undef %ModuleB:: ;
      } else {
         $retry   = 0;
         $succeed = 1;
      }
   }

   # If it never loaded, bail for good.
   if ( $succeed == 0 ) {
      require Carp;
      Carp::croak("Failed to load ModuleB: $@\n");
   }
}
[download]

Some limited testing done by forcing a croak() in ModuleB on the first 4 attempts shows it seems to work. We'll see in production use!