Re: planner/optimizer question - Mailing list pgsql-performance

From Tom Lane
Subject Re: planner/optimizer question
Date
Msg-id 14614.1083428488@sss.pgh.pa.us
Whole thread Raw
In response to Re: planner/optimizer question  (Jochem van Dieten <jochemd@oli.tudelft.nl>)
List pgsql-performance
Jochem van Dieten <jochemd@oli.tudelft.nl> writes:
> The moment the heap tuple is updated/deleted the visible-to-all
> flag needs to be set to false again in all indexes. This is
> critical,

Exactly.  This gets you out of the hint-bit semantics and into a ton
of interesting problems, such as race conditions.  (Process A determines
that tuple X is visible-to-all, and goes to update the index tuple.
Before it can reacquire lock on the index page, process B updates the
heap tuple and visits the index to clear the flag bit.  Once A obtains
lock it will set the flag bit.  Oops.)

Basically what you are buying into with such a thing is multiple copies
of critical state.  It may be only one bit rather than several words,
but updating it is no less painful than if it were a full copy of the
tuple's commit status.

            regards, tom lane

pgsql-performance by date:

Previous
From: "Gary Doades"
Date:
Subject: Re: planner/optimizer question
Next
From: Dave Cramer
Date:
Subject: Re: Wierd context-switching issue on Xeon