On Thu, Jul 28, 2016 at 1:57 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Aleksander Alekseev <a.alekseev@postgrespro.ru> writes:
>>> Can you explain use case where you need it?
>
>> ... Or maybe you have different objects, e.g. IndexScanDesc's, that should
>> iterate over some tree's independently somewhere in indexam.c
>> procedures. Exact order may depend on user's query so you don't even
>> control it.
>
> It seems clear to me that the existing arrangement is hazardous for any
> RBTree that hasn't got exactly one consumer. I think Aleksander's plan to
> decouple the iteration state is probably a good one (NB: I've not read the
> patch, so this is not an endorsement of details). I'd go so far as to say
> that we should remove the old API as being dangerous, rather than preserve
> it on backwards-compatibility grounds. We make bigger changes than that
> in internal APIs all the time.
>
> Having said that, though: if the iteration state is not part of the
> object, it's not very clear whether we can behave sanely if someone
> changes the tree while an iteration is open. It will need careful
> thought as to what sort of guarantees we can make about that. If it's
> too weak, then a separated-state version would have enough hazards of
> its own that it's not necessarily any safer.
+1 to all of that.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
