A program doesn't have to know if it's writing to a FIFO or not, unless you're dealing with networked files, i.e. files on some NFS or SMB like share. The only issue I can see here is that opening a FIFO for writing usually requires a listener, so if your filter application crashes (or you don't start it) your program is likely to have issues (some SIGPIPE in the first case, hanging in the second). Another possible issue deals with the limited buffer between these applications, so you must ensure that your filter program doesn't lose time possibly blocking the log producer.

The FIFO solution leaves the duty to write the logs to the Perl filter, not to the original application, and the filter is likely to write regular files - something that a logrotate program should not be upset with.

As for the redirection, I don't really understand how it should work. Redirecting STDOUT to STDERR means that you basically lose them all in the listener application. The pipeline

producer | consumer
[download]

links producer's STDOUT to consumer's STDIN, so the suggested redirection leaves you with an empty STDIN and nowhere to read log lines from.