On Fri, Jan 7, 2011 at 11:41 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Fri, Jan 7, 2011 at 8:20 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> Between
>>> amclusterable, amsearchnulls, and amoptionalkey, I believe that we have
>>> quite enough flags already to cover what anything else actually
>>> needs-to-know about the AM's behavior.
>
>> I've pretty much come to the conclusion that pg_am is much better at
>> providing the illusion of abstraction than it is at providing actual
>> abstraction. IIUC, the chances that a third-party AM would need to
>> patch core are nearly 100% anyway, so I'm not inclined to spend much
>> mental energy trying to figure out what flags it might hypothetically
>> need.
>
> Well, I'll grant that allowing loadable modules to emit and replay WAL
> records is an unsolved problem, but the existence of that problem
> doesn't mean that we should entirely give up on keeping AMs modular.
> I believe that they *are* pretty modular except for that one issue.
I'm not in a hurry to chuck the current system completely, but it
strikes me that the planner basically has to know everything about
what the AMs can do. I agree there's some value in encapsulating that
behind Booleans, but adding a new AM, or a new feature to an existing
AM, can be expected to result in regular rearrangements of those
Booleans. So it seems a bit porous.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
