When pg_upgrade calls pg_dump, it passes in "--schema-only", which is
generally correct, except that this causes everything having to do with
large objects to be excluded. That's still usually correct, because
pg_upgrade will simply copy the pg_largeobject and
pg_largeobject_metadata tables through to the new cluster as-is
(essentially treating them as if they were user tables).
Unfortunately, those tables aren't, actually, the only places that we
store information about large objects; the general-purpose tables like
pg_seclabel and pg_description can hold information about large objects
What this means is that performing a pg_upgrade will result in any
security labels or comments on large objects being dropped. This
seems to go back at least as far as 9.2, though I found it through the
pg_dump regression testing that I've been working on.
I haven't looked at trying to fix this yet, but I'm thinking the
approach to use will probably be to modify pg_dump to still call
getBlobs() when in binary-upgrade mode (regardless of the schema-only
flag) but then have dumpBlobs(), when in binary-upgrade mode, only
output the security labels and comments. I hope that doesn't end up
causing some kind of chicken-and-egg problem.. Presumably the large
object tables are in place and correct before the dump is restored, so I
think this will work.
Just wanted to get a note out to -hackers about the issue, I'll see
about getting a fix written up for it soon.
pgsql-hackers by date
|Next:||From: Fujii Masao||Date: 2017-02-21 16:52:20|
|Subject: Re: DROP SUBSCRIPTION and ROLLBACK|
|Previous:||From: Tom Lane||Date: 2017-02-21 15:18:25|
|Subject: Re: drop support for Python 2.3|