Re: GiST VACUUM - Mailing list pgsql-hackers

From Heikki Linnakangas
Subject Re: GiST VACUUM
Date
Msg-id 9cc18cc9-08d0-9838-7297-57500311d57b@iki.fi
Whole thread Raw
In response to Re: GiST VACUUM  (Andrey Borodin <x4mmm@yandex-team.ru>)
Responses Re: GiST VACUUM  (Andrey Borodin <x4mmm@yandex-team.ru>)
List pgsql-hackers
On 12/07/18 19:06, Andrey Borodin wrote:
>> 11 июля 2018 г., в 0:07, Heikki Linnakangas <hlinnaka@iki.fi>
>> написал(а):
>> 
>> This seems misplaced. This code deals with internal pages, and as
>> far as I can see, this patch never marks internal pages as deleted,
>> only leaf pages. However, we should have something like this in the
>> leaf-page branch, to deal with the case that an insertion lands on
>> a page that was concurrently deleted. Did you have any tests, where
>> an insertion runs concurrently with vacuum, that would exercise
>> this?
> 
> That bug could manifest only in case of crash between removing
> downlinks and marking pages deleted.

Hmm. The downlink is removed first, so I don't think you can see that 
situation after a crash. After a crash, you might have some empty, 
orphaned, pages that have already been unlinked from the parent, but a 
search/insert should never encounter them.

Actually, now that I think about it more, I'm not happy with leaving 
orphaned pages like that behind. Let's WAL-log the removal of the 
downlink, and marking the leaf pages as deleted, in one WAL record, to 
avoid that.

But the situation in gistdoinsert(), where you encounter a deleted leaf 
page, could happen during normal operation, if vacuum runs concurrently 
with an insert. Insertion locks only one page at a time, as it descends 
the tree, so after it has released the lock on the parent, but before it 
has locked the child, vacuum might have deleted the page. In the latest 
patch, you're checking for that just before swapping the shared lock for 
an exclusive one, but I think that's wrong; you need to check for that 
after swapping the lock, because otherwise vacuum might delete the page 
while you're not holding the lock.

> I do not know how to test this
> reliably. Internal pages are locked before leafs and locks are
> coupled. No cuncurrent backend can see downlinks to pages being
> deleted, unless crash happens.

Are you sure? At a quick glance, I don't think the locks are coupled.

We do need some way of testing this..

- Heikki


pgsql-hackers by date:

Previous
From: Claudio Freire
Date:
Subject: Re: Vacuum: allow usage of more than 1GB of work mem
Next
From: Andres Freund
Date:
Subject: Re: Internal error XX000 with enable_partition_pruning=on, pg 11beta1 on Debian