Re: Error during restore - dump taken with pg_dumpall -c option

On Thu, May 12, 2016 at 9:48 PM, Fabrízio de Royes Mello
<fabriziomello(at)gmail(dot)com> wrote:
>
>
> Em quinta-feira, 12 de maio de 2016, Rushabh Lathia
> <rushabh(dot)lathia(at)gmail(dot)com> escreveu:
>>
>>
>> On master branch when we do pg_dumpall with -c option, I can see that
>> it also dumping the "DROP ROLE pg_signal_backend", which seems wrong.
>> Because when you restore the dump, its throwing an error
>> "ERROR: cannot drop role pg_signal_backend because it is required by the
>> database system".
>>
>>
>> dumpRoles()::pg_dumpall.c does have logic to not dump "CREATE ROLE" if
>> the
>> rolename starts with "pg_", but similar check is missing into dropRoles()
>> function.
>>
>> PFA patch, to fix the problem in the similar way its been handled into
>> dumpRoles().
>>
>
> Shouldn't this logic be applied just to version >= 9.6? I ask it because you
> write a special query filtering rolname !~ '^pg_' and again check it using
> strcmp before the drop role output. Is this the expected behavior?

The patch looks correct to me: as far as I recall we give no guarantee
that a dump generated by pg_dump based on a given version would load
data correctly in an older version of the backend. So, when with 9.6's
pg_dump, dumping from a < 9.6 backend, bypassing the role names
beginning by "pg_" and letting the user know about their existence
without dumping them looks fine.
--
Michael