Re: Partial match fix for fast scan

On 04/10/2014 10:00 PM, Alexander Korotkov wrote:
> On Thu, Apr 10, 2014 at 8:22 PM, Fabrízio de Royes Mello <
> fabriziomello(at)gmail(dot)com> wrote:
>
>> On Thu, Apr 10, 2014 at 11:09 AM, Alexander Korotkov <aekorotkov(at)gmail(dot)com
>>> wrote:
>>
>>> GIN partial match appears to be broken after fast scan. Following simple
>>> test case raises assertion failure.
>>>
>>> create extension btree_gin;
>>> create table test as (select id, random() as val from
>>> generate_series(1,1000000) id);
>>> create index test_idx on test using gin (val);
>>> vacuum test;
>>> select * from test where val between 0.1 and 0.9;
>>>
>>> Attached patch fixes bugs in entryGetItem function.
>>> I would especially point that "continue;" checks "while" condition even
>>> if it's postfix "while". That's why I surrounded tbm_iterate with another
>>> "while".
>>>
>>>
>> Interesting... to me (using current master) your test case doesn't fail...
>>
>> fabrizio=# select * from test where val between 0.1 and 0.9;
>> id | val
>> --------+-------------------
>> 1 | 0.554413774050772
>> 2 | 0.767866868525743
>> 3 | 0.601187175605446
>> ...
>>
>>
>> But fail if I change the values of between clause:
>>
>> fabrizio=# select * from test where val between 0.1 and 0.19;
>> ERROR: tuple offset out of range: 8080
>>
>
>
> It must be compiled with --enable-cassert to fail on assertion.

I'm actually getting the "tuple offset out of range" with Fabrizio's
query, even after your fix (not every time, run it a few times,
launching a new connection each time). So there's another bug lurking
there. The problem seems to be that even though we've checked in the
innermost loop that not all of the items on the page are <= advancePast,
that situation can change later if advancePast is advanced. So on next
invocation entryGetItem, the loop to skip past advancePast might read
bogus offsets in the array.

I pushed a patch, which includes Alexander's fix, and also fixes the
second issue.

- Heikki