Re: WIP: Rework access method interface

On Mon, Aug 10, 2015 at 5:48 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Alexander Korotkov <a(dot)korotkov(at)postgrespro(dot)ru> writes:
> > On Mon, Aug 10, 2015 at 1:12 PM, Petr Jelinek <petr(at)2ndquadrant(dot)com>
> wrote:
> >> I don't understand this, there is already AmRoutine in RelationData, why
> >> the need for additional field for just amsupport?
>
> > We need amsupport in load_relcache_init_file() which reads
> > "pg_internal.init". I'm not sure this is correct place to call
> am_handler.
> > It should work in the case of built-in AM. But if AM is defined in the
> > extension then we wouldn't be able to do catalog lookup for am_handler on
> > this stage of initialization.
>
> This is an issue we'll have to face before there's much hope of having
> index AMs as extensions: how would you locate any extension function
> without catalog access? Storing raw function pointers in pg_internal.init
> is not an answer in an ASLR world.
>
> I think we can dodge the issue so far as pg_internal.init is concerned by
> decreeing that system catalogs can only have indexes with built-in AMs.
> Calling a built-in function doesn't require catalog access, so there
> should be no problem with re-calling the handler function by OID during
> load_relcache_init_file().
>

That should work, thanks! Also we can have SQL-visible functions to get
amsupport and amstrategies and use them in the regression tests.

> We could also have problems with WAL replay, though I think the consensus
> there is that extension AMs have to use generic WAL records that don't
> require any index-specific replay code.
>

Yes, I've already showed one version of generic WAL records. The main
objecting against them was it's hard insure that composed WAL-record is
correct.
Now I'm working on new version which would be much easier and safe to use.

------
Alexander Korotkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company