Date: Thu, 12 Nov 2020 16:36:32 +0100
From: Dmitry Dolgov <9erthalion6@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Alvaro Herrera <alvherre@alvh.no-ip.org>,
	Michael Paquier <michael@paquier.xyz>,
	James Coleman <jtc331@gmail.com>,
	Pg Hackers <pgsql-hackers@lists.postgresql.org>
Subject: Re: remove spurious CREATE INDEX CONCURRENTLY wait
Message-ID: <20201112153632.whecjbyvrxzdcjfv@localhost>
References: <20201110023115.GA18579@alvherre.pgsql>
 <2091476.1604977347@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <2091476.1604977347@sss.pgh.pa.us>
Precedence: bulk

> On Mon, Nov 09, 2020 at 10:02:27PM -0500, Tom Lane wrote:
>
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> > Yeah ... it would be much better if we can make it use atomics instead.
>
> I was thinking more like "do we need any locking at all".
>
> Assuming that a proc's vacuumFlags can be set by only the process itself,
> there's no write conflicts to worry about.  On the read side, there's a
> hazard that onlookers will not see the PROC_IN_SAFE_IC flag set; but
> that's not any different from what the outcome would be if they looked
> just before this stanza executes.  And even if they don't see it, at worst
> we lose the optimization being proposed.
>
> There is a question of whether it's important that both copies of the flag
> appear to update atomically ... but that just begs the question "why in
> heaven's name are there two copies?"

Sounds right, but after reading the thread about GetSnapshotData
scalability more thoroughly there seem to be an assumption that those
copies have to be updated at the same time under the same lock, and
claims that in some cases justification for correctness around not
taking ProcArrayLock is too complicated, at least for now.

Interesting enough, similar discussion happened about vaccumFlags before
with the same conclusion that theoretically it's fine to update without
holding the lock, but this assumption could change one day and it's
better to avoid such risks. Having said that I believe it makes sense to
continue with locking. Are there any other opinions? I'll try to
benchmark it in the meantime.