2011/7/9 Robert Haas <robertmhaas@gmail.com>:
> On Fri, Jul 8, 2011 at 4:57 PM, Noah Misch <noah@2ndquadrant.com> wrote:
>> Note that it does not matter whether we're actually doing an index scan -- a seq
>> scan with a filter using only leakproof operators is equally acceptable. What I
>> had in mind was to enumerate all operators in operator classes of indexes below
>> each security view. Those become the leak-free operators for that security
>> view. If the operator for an OpExpr is considered leak-free by all sources of
>> its operands, then we may push it down. That's purely a high-level sketch: I
>> haven't considered implementation concerns in any detail. The resulting
>> behavior could be surprising: adding an index may change a plan without the new
>> plan actually using the index.
>>
>> I lean toward favoring the pg_proc flag. Functions like "texteq" will be taken
>> as leakproof even if no involved table has an index on a text column. It works
>> for functions that will never take a place in an operator class, like
>> length(text). When a user reports a qualifier not getting pushed down, the
>> answer is much more satisfying: "Run 'CREATE OR REPLACE FUNCTION
>> ... I_DONT_LEAK' as a superuser." Compare to "Define an operator class that
>> includes the function, if needed, and create an otherwise-useless index." The
>> main disadvantage I see is the loss of policy locality. Only a superuser (or
>> maybe database owner?) can create or modify declared-leakproof functions, and
>> that decision applies throughout the database. However, I think the other
>> advantages clearly outweigh that loss.
>
> This strikes me as a fairly compelling refutation of Heikki's proposed approach.
>
OK, I'll try to modify the patch according to the flag of pg_proc design.
As long as the default of user-defined function is off, and we provide
built-in functions
with appropriate configurations, it seems to me the burden of DBA is
quite limited.
Thanks,
--
KaiGai Kohei <kaigai@kaigai.gr.jp>