The example given for selective columns on the left is bogus, though; of course a composite index can't be used if the first column in the index isn't used by the predicate - and this has nothing to do with the selectivity of the column.
Selectivity of the column in principle matters to reduce the time spent doing comparisons to find the right slot in each btree page; if more selective columns are at the end, the comparison function will have to scan through more of the index data to determine ordering.
Much more important, IMO, is reuse of the index. If you frequently use predicates referencing N-1 columns and have an index on N columns, you're better off if the least used column is in the Nth position. I think this is what the article was trying to get at, if somewhat hamfistedly.
Selectivity of the column in principle matters to reduce the time spent doing comparisons to find the right slot in each btree page; if more selective columns are at the end, the comparison function will have to scan through more of the index data to determine ordering.
Much more important, IMO, is reuse of the index. If you frequently use predicates referencing N-1 columns and have an index on N columns, you're better off if the least used column is in the Nth position. I think this is what the article was trying to get at, if somewhat hamfistedly.