Re: Hot Standby b-tree delete records review

On Tue, 2010-11-09 at 13:34 +0200, Heikki Linnakangas wrote:
> (cleaning up my inbox, and bumped into this..)
> 
> On 22.04.2010 12:31, Simon Riggs wrote:
> > On Thu, 2010-04-22 at 12:18 +0300, Heikki Linnakangas wrote:
> >> Simon Riggs wrote:
> >>> On Thu, 2010-04-22 at 11:56 +0300, Heikki Linnakangas wrote:
> >>>
> >>>>>>>> If none of the removed heap tuples were present anymore, we currently
> >>>>>>>> return InvalidTransactionId, which kills/waits out all read-only
> >>>>>>>> queries. But if none of the tuples were present anymore, the read-only
> >>>>>>>> queries wouldn't have seen them anyway, so ISTM that we should treat
> >>>>>>>> InvalidTransactionId return value as "we don't need to kill anyone".
> >>>>>>> That's not the point. The tuples were not themselves the sole focus,
> >>>>>> Yes, they were. We're replaying a b-tree deletion record, which removes
> >>>>>> pointers to some heap tuples, making them unreachable to any read-only
> >>>>>> queries. If any of them still need to be visible to read-only queries,
> >>>>>> we have a conflict. But if all of the heap tuples are gone already,
> >>>>>> removing the index pointers to them can'ẗ change the situation for any
> >>>>>> query. If any of them should've been visible to a query, the damage was
> >>>>>> done already by whoever pruned the heap tuples leaving just the
> >>>>>> tombstone LP_DEAD item pointers (in the heap) behind.
> >>>>> You're missing my point. Those tuples are indicators of what may lie
> >>>>> elsewhere in the database, completely unreferenced by this WAL record.
> >>>>> Just because these referenced tuples are gone doesn't imply that all
> >>>>> tuple versions written by the as yet-unknown-xids are also gone. We
> >>>>> can't infer anything about the whole database just from one small group
> >>>>> of records.
> >>>> Have you got an example of that?
> >>>
> >>> I don't need one, I have suggested the safe route. In order to infer
> >>> anything, and thereby further optimise things, we would need proof that
> >>> no cases can exist, which I don't have. Perhaps we can add "yet", not
> >>> sure about that either.
> >>
> >> It's good to be safe rather than sorry, but I'd still like to know
> >> because I'm quite surprised by that, and got me worried that I don't
> >> understand how hot standby works as well as I thought I did. I thought
> >> the point of stopping replay/killing queries at a b-tree deletion record
> >> is precisely that it makes some heap tuples invisible to running
> >> read-only queries. If it doesn't make any tuples invisible, why do any
> >> queries need to be killed? And why was it OK for them to be running just
> >> before replaying the b-tree deletion record?
> >
> > I'm sorry but I'm too busy to talk further on this today. Since we are
> > discussing a further optimisation rather than a bug, I hope it is OK to
> > come back to this again later.
> 
> Would now be a good time to revisit this? I still don't see why a b-tree 
> deletion record should conflict with anything, if all the removed index 
> tuples point to just LP_DEAD tombstones in the heap.

I want what you say to be true. The question is: is it? We just need to
explain why that will never be a problem.

-- Simon Riggs           http://www.2ndQuadrant.com/books/PostgreSQL Development, 24x7 Support, Training and Services