On Wed, Apr 9, 2014 at 11:24 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Maybe we should make *neither* of these the default opclass, and give
>> *neither* the name json_ops.
>
> There's definitely something to be said for that. Default opclasses are
> sensible when there's basically only one behavior that's interesting for
> most people. We can already see that that's not going to be the case
> for jsonb indexes, at least not with the currently available alternatives.
>
> Not having a default would force users to make decisions explicitly.
> Is that what we want?
I don't like the idea of having no default opclass. I think there's a
huge usability gain in being able to "just" create an index on a
column and have it do something reasonable for most use cases.
I can get behind the idea of having separate index opclasses for paths
and path-value pairs but I suspect the default should just be to index
both in the same index. If we can have one default index opclass that
supports containment and existence and then other opclasses that are
smaller but only support a subset of the operators that would seem
like the best compromise.
I'm a bit confused by Heikki's list though. I would expect path and
path-value pair to be the only useful ones. I'm not clear what an
index on keys or key-value would be -- it would index just the
top-level keys and values without recursing?
--
greg