Re: [HACKERS] Database replication... - Mission Critica

"Christopher Kings-Lynne" <chriskl(at)familyhealth(dot)com(dot)au> writes:
> Also, what about if the two servers get the 'begin' command at marginally
> different times, then even:
>
> INSERT INTO foo VALUES (CURRENT_TIMESTAMP);
>
> Will be different on each different machine.

Yeah -- and gettimeofday() will be inconsistent for a slightly
different reason. In fact, if a function depends on anything other
than its arguments, you're probably going to have problems replicating
it in a sane manner if you send the query string alone.

One possible solution would be to dump the idea of sending query
strings, and just send tuple-level changes. The basic idea is to
execute a data-modifying txn locally, determine the tuple-level
changes that it makes, rollback the txn locally, and then broadcast
the tuple modifications to all the nodes in the cluster. Of course,
it's a bit more complex than that (read the Postgres-R paper for more
details), but that's the gist of it.

... which makes me think that rather than wasting our time discussing
a more-or-less solved problem, we'd be better off helping to implement
Postgres-R.

> Even nastier, what about if the different postgres servers in the cluster
> run on different architectures! That way you'd get different floating point
> results on each machine...

IMHO, that falls into the category of: "Dr., it hurts when I do this!"
-- "Well, don't do it then" :-)

Cheers,

Neil

--
Neil Conway <neilc(at)samurai(dot)com> || PGP Key ID: DB3C29FC