Tom Lane wrote:
"Jim C. Nasby" <jnasby@xxxxxxxxxxxxx> writes:
I'd actually been thinking about this recently, and had come up with the
following half-baked ideas:
Allow a transaction to specify exactly what tables it will be touching,
perhaps as an extension to BEGIN. Should any action that transaction
takes attempt to access a table not specified, throw an error.
A possible variant on that would be to automatically determine at
transaction start all the tables that would be accessed by that
transaction.
Once that list is available, vacuum should be able to use it to ignore
any transactions that have promised not to touch whatever table it's
vacuuming.
No, you missed my point entirely. The above would help not at all,
unless the restrictions were somehow propagated through XMIN
calculations, which seems impracticable. (Every backend calculate a
separate XMIN with respect to every table that's being mentioned by any
other backend? I don't think so...)
So basically the problem is that calculating a cluster-wide XMIN is fine from
a performance point of view, but to coarse from a vacuum point of view. OTOH
calculating a table-specific XMIN is fine enough for vacuum to vacuum queue-like
tables efficiently, but bad from a performance point of view.
What about doing the xmin-calculations and the xmin exporting per tablespace instead?
This would be more generic than the local/global split, because all global objects
are in the pg_global tablespace I believe. OTOH most people probably don't have more
than a handfull of tablespaces, which seems a manageable amount...
greetings, Florian Pflug
