Re: Parallel Index Scan vs BTP_DELETED and BTP_HALF_DEAD

On Wed, Dec 13, 2017 at 4:20 PM, Thomas Munro
<thomas(dot)munro(at)enterprisedb(dot)com> wrote:
> On Wed, Dec 13, 2017 at 3:41 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>> Good. I hope that the patch I have posted above is able to resolve
>> this problem. I am asking as you haven't explicitly mentioned that.
>
> I can confirm that your patch fixes the problem for forward scans.
> That is, I can see it reaching the BTP_DELETED case via an extra LOG
> statement I added, and it worked correctly. Good.
>
> I don't know how to make it hit the backwards scan case. I can get a
> backward scan in a worker by changing the query to "select count(*)
> from (select * from jobs where id + 1 > id order by status desc) ss"
> but I suspect that _bt_walk_left() may be hiding deleted pages from us
> so the condition may not be reachable with this technique.

Hmm, no that's not right: clearly it can return half-dead or deleted
pages to the caller. So I don't know why I never seem to encounter
any, despite concurrent vacuums producing them; maybe something to do
with the interlocking you get with vacuum when you traverse the btree
by walking left -- my btree-fu is not yet strong enough.

--
Thomas Munro
http://www.enterprisedb.com