Re: preserving forensic information when we freeze

On 2013-12-23 10:56:07 -0500, Robert Haas wrote:
> > Well, it really depends on the usecase. Reading
> > SELECT header.xmin, header.xmax
> > FROM sometable tbl,
> > pg_tuple_header(tbl.tableoid, tbl.ctid) header;
> > is surely harder to understand than
> > SELECT tbl.xmin, tbl.xmax
> > FROM sometable tbl;
> > and the latter is noticeably cheaper since we're not re-fetching the
> > tuple, especially when the tuple is the result of a more complicated
> > query including a seqscan, we might often need to re-fetch the tuple
> > from disk, thanks to the ringbuffer.
> >
> > That really makes me less than convinced that the function based
> > approach is the right tool for everything.

> Meh. Who are all of these people who are fetching xmin, xmax, cmin,
> and cmax in complex queries, and why are they doing that?

I have seen it done to avoid ABA problems in cache invalidation
mechanisms by simply checking ctid, xmin, xmax to stay the same.

> If those
> columns are just for forensic purposes, then whatever performance
> disadvantages the function-based approach has don't really matter. If
> people are using them as integral parts of their application, then
> that's more of a concern, but how are those people not getting broken
> every release or three anyway? We keep inventing things like
> combo-cids and multi-xacts and multi-xacts that also contain an update
> and now freezing without changing xmin, and if you're actually relying
> on those columns for anything but forensics your application is
> presumably going to break when we change whatever part it depends on.

Well, all of the fundamental changes (combocids, the initial multixact
introduction) have been quite some time ago. I think backward compat has
a much higher value these days (I also don't see much point in looking
at cmin/cmax for anything but diagnostic purposes). I don't think any of
the usecases I've seen would be broken by either fk-locks (multixacts
have been there before, doesn't matter much that they now contain
updates) nor by forensic freezing. The latter because they really only
checked that xmin/xmax were the same, and we hopefully haven't broken
that...

But part of my point really was the usability, not only the
performance. Requiring LATERAL queries to be usable sensibly causes a
"Meh" from my side ;)

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services