“CPU Costing” and “System statistics” are synonyms for the same mechanism. Because of the version-dependency in the rules for using system statistics, I call dbms_stats.delete_system_stats and set _optimizer_cost_model to “io” to ensure that I’ve disabled system statistics.

As far as repeatable results go – the use of system statistics will change the cost of a tablescan relative to an index access path in ways that are dependent on the actual numbers in the sys.aux_stats$ table, getting rid of them will ensure that people don’t see different plans because they have different stats.

There are cases though (distributed queries come to mind) where certain features cannot appear unless system statistics are enabled, so sometimes I will explicitly set some values for the stats using dbms_stats.set_system_stats().

Despite appearing in recent code released by Oracle Corp. the old first_rows hint was originally labelled in 9i as “available only for backwards compatibility”. As Pavol indicated, some of its paths (or avoidance of pathcs) were dictated by simple rules – so it’s fairly easy to find cases where first_rows and first_rows(n) – even with n = 1 – differ dramatically in their outcome.

Regards

Pavol Babel

SELECT *
FROM (SELECT /*+ FIRST_ROWS */
ROWNUM AS rnum, cust_last_name
FROM ( SELECT cust_last_name
FROM customers
WHERE CUST_LAST_NAME BETWEEN ‘A%’ AND ‘D%’
ORDER BY cust_last_name) a
WHERE ROWNUM = :1; — set to 10000

Could you please provide the full test case for the explain plan that has got you the problem with first_rows hint.?