=?UTF-8?Q?Filip_Rembia=C5=82kowski?= <filip.rembialkowski@gmail.com> writes:
> On Wed, Apr 25, 2012 at 8:56 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Hmm. A look at the code in pg_dumpall suggests that the problem is
>> unmatched entries in pg_auth_members, ie this query:
>>
>> SELECT ur.rolname AS roleid
>> FROM pg_auth_members a LEFT JOIN pg_authid ur on ur.oid = a.roleid
>>
>> is returning some null results.
> Yes that is the case:
Ah-ha. How about the member and grantor OIDs in those rows --- do
they correspond to still-existing roles? (I believe "10" would be
the bootstrap superuser, so that should certainly still exist, but
those other numbers are for user-made roles.)
> Could it be (theoretically) caused by human-made insertions into
> pg_auth_members?
Well, perhaps ... are you in the habit of hacking that catalog directly?
Aside from the possibility of inserting a just-plain-wrong OID, there's
the possibility of inserting a valid row but forgetting to add a
pg_shdepend entry that would protect the row against the underlying
role being dropped.
> Maybe you remember some bug which could have caused this in the past?
Doesn't ring a bell offhand.
regards, tom lane