Re: [PERFORM] Hash Anti Join performance degradation

2011/6/1 Tom Lane <tgl@sss.pgh.pa.us>:
> We do need to look into putting a CHECK_FOR_INTERRUPTS call in here
> somewhere, though.  I'm inclined to think that right before the
> ExecScanHashBucket is the best place.  The reason that nest and merge
> joins don't show a comparable non-responsiveness to cancels is that they
> always call a child plan node at the equivalent place, and ExecProcNode
> has got a CHECK_FOR_INTERRUPTS.  So we ought to check for interrupts
> at the point of "fetching a tuple from the inner child plan", and
> ExecScanHashBucket is the equivalent thing in this logic.  Cedric's
> suggestion of putting it before the switch would get the job done, but
> it would result in wasting cycles during unimportant transitions from
> one state machine state to another.


exact, thanks to your last email I read more the code and get the same
conclusion and put it in a more appropriate place : before
ExecScanHashBucket.

I was about sending it, so it is attached.



>
>                        regards, tom lane
>



--
Cédric Villemain               2ndQuadrant
http://2ndQuadrant.fr/     PostgreSQL : Expertise, Formation et Support