It seems that for clustered caches, the ezcache cli script does not do its job, unless it is called with the purge option.

To test this, create an rss feed, access it, check that an rss cache file is created on both cluster db and nfsvar, then run

php bin/php/ezcache.php --clear-id=rss_cache -d --logfiles

Then check again: the rss-cache is still on nfsvar and in db, tagged as valid.

The snafu seems to come from the fact that the eZDFSFileHandler::delete() call, which has in its javadoc a phrase stating that it works for both files and directories, does in fact make a direct call to
self::$dbbackend->_delete

and neither the Mysql nor the Oracle dfs-filehandler-backends support deleting directories via _delete(). They explicitly mention that _deletebylike should be used instead.

As proof of the reasoning: eZFSFileHandler::delete() has in its body a check which allows it to delete directories and files using different code.