Date: Fri, 10 Jul 2020 16:45:21 -0500
From: Justin Pryzby <pryzby@telsasoft.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: David Gilman <davidgilman1@gmail.com>,
	pgsql-hackers@lists.postgresql.org
Subject: Re: "tuple concurrently updated" in pg_restore --jobs
Message-ID: <20200710214521.GS26220@telsasoft.com>
References: <20200706050129.GW4107@telsasoft.com>
 <2116463.1594414480@sss.pgh.pa.us>
 <20200710210607.GR26220@telsasoft.com>
 <2164430.1594416988@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <2164430.1594416988@sss.pgh.pa.us>
User-Agent: Mutt/1.9.4 (2018-02-28)
Precedence: bulk

On Fri, Jul 10, 2020 at 05:36:28PM -0400, Tom Lane wrote:
> Justin Pryzby <pryzby@telsasoft.com> writes:
> > On Fri, Jul 10, 2020 at 04:54:40PM -0400, Tom Lane wrote:
> >> This works about 99% of the time, in fact.  It falls down in the --clean
> 
> > Note that this fails for me (sometimes) even without --clean.
> 
> Oh, I was thinking that REVOKE would only be issued in the --clean
> case, but apparently that's not so.  Doesn't really affect the fix
> proposal though.  I just finished a patch for HEAD, as attached.
> 
> (I flushed the "CatalogId objCatId" argument of dumpACL, which was
> not used.)
> 
> I'm not sure how far to back-patch it -- I think the parallel restore
> of ACLs behavior is not very old, but we might want to teach older
> pg_dump versions to insert the extra dependency anyway, for safety.

Yes, and the test case in David's patch on other thread [0] can't be
backpatched further than this patch is.  A variant on his test case could just
as well be included in this patch (with pg_dump writing to a seekable FD) and
then amended later to also test writing to an unseekable FD.

[0] https://commitfest.postgresql.org/28/2568/