On 10/22/22 11:20, Adrian Klaver wrote:
> On 10/20/22 14:34, Ron wrote:
>> On 10/20/22 10:02, Adrian Klaver wrote:
>>> On 10/20/22 06:20, Ron wrote:
>>>> On 10/20/22 00:12, Tom Lane wrote:
>
>>>> I ran "pg_dumpall --globals-only --no-role-passwords" on the source
>>>> instance, and applied it to the new instance before doing the
>>>> pg_restore. If I hadn't done that, pg_restore would have thrown errors
>>>> on all the GRANT and ALTER TABLE ... OWNER TO statements embedded in
>>>> the backup.
>>>>
>>>
>>> Some questions:
>>>
>>> 1) The backup was from a Postgres 12.x database using a version 12 or
>>> higher instance of pg_backup?
>>
>> pg_dump and pg_restore are 12.11 from RHEL8.
>>
>>> 3) What if you run without --jobs?
>>
>> It runs without error. Add "--jobs=2" and the errors appear.
>>
>
> Hmm, that is beyond me.
>
> 1) I did notice that the pg_restore errors all where; ERROR: permission
> denied for schema strans
>
> 2) They all occurred during CREATE INDEX or COPY, which would be the part
> where --jobs kicks in.
>
> To me it looks like out of order execution where the jobs starting on
> their tasks before the main task got done granting permissions. I just
> have no idea how that could happen.
I was afraid you were going to say that.
The work-around is to:
pg_dump $SRCDB --schema-only | grep -e '^\(GRANT|REVOKE\)' > all_GRANT.sql
pg_dump $SRCDB --schema-only | grep OWNER > all_OWNER.sql
pg_restore --jobs=X --no-owner $NEWDB
psql $NEWDB -f all_OWNER.sql
psql $NEWDB -f all_GRANT.sql
This is, of course, why we need to test the backup/restore process.
--
Angular momentum makes the world go 'round.