Andres Freund <andres@2ndquadrant.com> writes:
> On 2014-01-04 11:54:46 -0500, Robert Haas wrote:
>> Well, as I said before, somebody can make their own configuration
>> table and store stuff there, rather than using pg_class.reloptions.
>> As I recall, the only response to that proposal was "well, they might
>> not want to do it that way", which does not strike me as a sufficient
>> reason.
> Well, there's some things you get by that integration:
> * Proper dependency tracking when relations are dropped & renamed
> * Sensible integration into pg_dump, with only the relevant options
> being dumped/restored on partial dump/restores
> * Caching of values, with proper cache invalidation
If you have some settings that need to be table-specific, then
I agree that the reloptions infrastructure is a nice place to track them.
What's actually missing here is some compelling examples of such settings
for plausible extensions. (The original example was pure handwaving, as
I don't believe it's possible to build a "replication extension" with no
core-code changes. As long as you need some of those, patching in a few
more standard reloptions is hardly a deal-breaker.)
Assuming that such examples are forthcoming, though, I think my main
objection to this proposal is the "ext." prefix, which seems precisely
100% useless, not to mention inconsistent with the naming of custom GUCs,
which the same extension might well have some of. I think that custom
reloptions should just have names like "extension_name.option_name", the
same as custom GUCs do. We have enough experience now with custom GUCs
that I don't think it's unreasonable to model the behavior of custom
reloptions on them as closely as possible.
I would suggest addressing Robert's concern about lack of error checking
by refusing to allow a custom reloption to be set unless the relevant
extension is loaded and checks it. Unlike the postgresql.conf problem,
I don't see any very good use-case for allowing an unchecked ALTER TABLE
to occur.
regards, tom lane