Обсуждение: pg_authid has duplicated rows

Hi,

On 2018-07-23 21:03:32 +0300, Mariel Cherkassky wrote:
> Today I faced a situation where I saw a postgresql 9.2 database that had
> duplicated rows in pg_autid. I saw 6 roles while 3 of them were duplicated
> (3 uniques rolenames in total). I checked and the oid of the duplicates
> were the same oid as the originals. The solution was running vacuum on the
> table but I wanted to understand how this situation was created ? Why even
> with unique index on the rolname I saw duplicated values for rolname ?

That suggests some form of corruption. Which version of 9.2 did this
occur on? Did you have any OS level crashes? Ran with fsync=off?

Please note that 9.2 is not supported anymore.

Greetings,

Andres Freund

On 2018-Jul-23, Andres Freund wrote:

> Hi,
> 
> On 2018-07-23 21:03:32 +0300, Mariel Cherkassky wrote:
> > Today I faced a situation where I saw a postgresql 9.2 database that had
> > duplicated rows in pg_autid. I saw 6 roles while 3 of them were duplicated
> > (3 uniques rolenames in total). I checked and the oid of the duplicates
> > were the same oid as the originals. The solution was running vacuum on the
> > table but I wanted to understand how this situation was created ? Why even
> > with unique index on the rolname I saw duplicated values for rolname ?
> 
> That suggests some form of corruption. Which version of 9.2 did this
> occur on? Did you have any OS level crashes? Ran with fsync=off?

I have vague memories of seeing this recently.  Can you please show 

select xmin, xmax, ctid, oid, rolname from pg_authid

And the output of pg_controldata.

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On July 23, 2018 1:33:11 PM PDT, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>On 2018-Jul-23, Andres Freund wrote:
>
>> Hi,
>>
>> On 2018-07-23 21:03:32 +0300, Mariel Cherkassky wrote:
>> > Today I faced a situation where I saw a postgresql 9.2 database
>that had
>> > duplicated rows in pg_autid. I saw 6 roles while 3 of them were
>duplicated
>> > (3 uniques rolenames in total). I checked and the oid of the
>duplicates
>> > were the same oid as the originals. The solution was running vacuum
>on the
>> > table but I wanted to understand how this situation was created ?
>Why even
>> > with unique index on the rolname I saw duplicated values for
>rolname ?
>>
>> That suggests some form of corruption. Which version of 9.2 did this
>> occur on? Did you have any OS level crashes? Ran with fsync=off?
>
>I have vague memories of seeing this recently.  Can you please show

It's at least one of the potential effects of the out-of-date shared relcache bug I recently fixed (but wasn't
backpatchedto 9.2, as it's unsupported). 

Andres

--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

Mariel Cherkassky <mariel.cherkassky@gmail.com> writes:
> I checked it also in version 9.6, and in this version I can edit the
> pg_authid table. I'm suppose to get an error that the table is a system
> catalog or am I wrong ?

No, the "system catalog modification" restriction only applies to DDL
on the system catalogs --- for instance, adding/removing a column, which
our C code couldn't cope with.  You're free to change their contents by
hand if you're a superuser.  (And if you break something that way, you
get to keep both pieces.  But sometimes it's necessary.)

            regards, tom lane