On Tue, Jan 15, 2013 at 1:48 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Tue, Jan 15, 2013 at 1:10 PM, Heikki Linnakangas
>> <hlinnakangas@vmware.com> wrote:
>>> Could we stash the counter e.g. in the root page of the index?
>
>> That would require maintaining a counter per table rather than a
>> single global counter, which would be bad because then we'd need to
>> store one counter in shared memory for every table, rather than just
>> one, period, which runs up against the fixed sizing of shared memory.
>
> I think what Heikki had in mind was that the copy in the index would be
> the authoritative one, not some image in shared memory. This'd imply
> dirtying the root page on every insert, as well as increased contention
> for the root page, so it might have performance problems.
>
> I think a bigger issue is where we'd find any space for it. There's no
> easily-spare space in a GIST page. This reminds me again that the lack
> of a metapage in GIST was a serious design error, which we should
> correct if we ever break on-disk compatibility again.
>
> I concur that adding such a counter to pg_control is a nonstarter,
> though.
>
> Given that we don't need crash recovery for an unlogged table, could
> we get away with some variant of NSN that has weaker semantics than
> XLOG LSNs?
It needs to be strictly ascending and survive clean shutdowns. Is
there some place we could preserve it other than the control file?
I was assuming we wanted a single sequence shared across all relations
rather than a sequence per relation, but I don't know of any reason
why that's actually required.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company