On Thu, May 3, 2018 at 3:18 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Peter Geoghegan <pg@bowt.ie> writes:
>> Why should the drop cascade to 63 objects rather than 62 because I've
>> set ignore_system_indexes=on?
>
> Indeed, that seems weird. Maybe tweak the test scripts so you can see
> all the objects cascaded to, and then find out what the additional
> object is? (I think also you could look into the postmaster log,
> without changing the test.)
I wrote a patch that makes ignore_system_indexes=off agree with
ignore_system_indexes=on in the two specific places I see problems
(the create_view and collate tests). Actually, I started with the
patch, and only ended up finding this issue later, while debugging my
own work.
The goal of this patch of mine is to have nbtree use heap TID as a
unique-ifier -- an implicit "final column" that is a first class part
of the key space (we sometimes have to store an "extra" heap TID
attribute in internal page tuples to make this work across the entire
tree structure). Vadim wanted to do something like this about 20 years
ago, and I think he had the right idea -- it's more or less what
Lehman and Yao intended.
I don't actually know what the problem is just yet, but I suspect some
subtle nbtree issue. I haven't actually debugged the underlying issue
at all just yet, but I'll start on that now. I don't want to post my
patch, since it's very much still a work in progress, and I'm not yet
100% sure that it actually truly fixes the issue.
--
Peter Geoghegan