Although the table creation itself does not use the transaction, we publish the table as part of committing the transaction and so skip publishing if there is no transactions. In practice, this only effected one thing: initializing replicated fastcount on secondaries. In all other cases when we create a table we also perform a write to store that table's ident somewhere (either in the catalog or index build colection), and only replicated fastcount doesn't need to do that. On primaries we write the oplog entry, so it could only happen on secondaries.