On 08/22/2012 08:36 AM, Denis Kolesnik wrote:
> I have now VERY strong argument to consider it is as a bug:
>
> if there a understandable for SQL language sequence which sorts
> in other fashion when adding "LIMIT".
Underspecified sorts can have unstable results, that's allowed by the
spec and is a reasonable performance optimisation. If you want stable
sorts you have to provide a fully defined ordering.
To say this is a bug is a bit like saying that undefined behaviour in C
is a bug. It isn't, the language and compiler are giving you what
they're supposed to, just not what you want.
http://en.wikipedia.org/wiki/Undefined_behavior
Unstable sorts are a common optimisation
http://en.wikipedia.org/wiki/Sorting_algorithm#Stability
and a very useful one. Not allowing Pg to use unstable sorts, or use
different sorts for different query plans, would slow correct queries
down to make an incorrect query like yours run how you expect.
--
Craig Ringer
