Date: Sat, 9 Apr 2022 21:24:42 +0800
From: Julien Rouhaud <rjuju123@gmail.com>
To: Nathan Bossart <nathandbossart@gmail.com>
Cc: Robert Haas <robertmhaas@gmail.com>, Andres Freund <andres@anarazel.de>,
	Michael Paquier <michael@paquier.xyz>,
	"Bossart, Nathan" <bossartn@amazon.com>,
	Fujii Masao <masao.fujii@oss.nttdata.com>,
	"wangsh.fnst@fujitsu.com" <wangsh.fnst@fujitsu.com>,
	Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>,
	Greg Sabino Mullane <htamfids@gmail.com>,
	Tom Lane <tgl@sss.pgh.pa.us>,
	"pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>
Subject: Re: make MaxBackends available in _PG_init
Message-ID: <YlGJGiofZiWN3elx@jrouhaud>
References: 
 <CA+TgmoaonHdak_8utKJCirZNA==62Om_xSnMrm7j2zZox-U8WA@mail.gmail.com>
 <20220325023951.purgu4gb5jyrtc7w@jrouhaud>
 <20220325031146.cd23gaak5qlzdy6l@jrouhaud>
 <Yj1OOR/oQOrEt7OI@paquier.xyz>
 <20220325063542.n6die7fmawluqkfp@jrouhaud>
 <20220325222317.dftrmzuyrb3iv62o@alap3.anarazel.de>
 <20220326072203.asge6bua7dxczf7m@jrouhaud>
 <20220326172316.obspcdmjqerislri@alap3.anarazel.de>
 <CA+Tgmoawc-4i3uFkNdwp8wUR9-jmfV0xjKDyJvGR9t2k2f=1EQ@mail.gmail.com>
 <20220330163051.GA784643@nathanxps13>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220330163051.GA784643@nathanxps13>
Archived-At: 
 <https://www.postgresql.org/message-id/YlGJGiofZiWN3elx%40jrouhaud>
Precedence: bulk

On Wed, Mar 30, 2022 at 09:30:51AM -0700, Nathan Bossart wrote:
> On Tue, Mar 29, 2022 at 12:22:21PM -0400, Robert Haas wrote:
> > It's not, though, because the original proposal was to change things
> > around so that the value of MaxBackends would have been reliable in
> > _PG_init(). If we'd done that, then extensions that are using it in
> > _PG_init() would have gone from being buggy to being not-buggy. But
> > since you advocated against that change, we instead committed
> > something that caused them to go from being buggy to failing outright.
> > That's pretty painful for people with such extensions. And IMHO, it's
> > *much* more legitimate to want to size a data structure based on the
> > value of MaxBackends than it is for extensions to override GUC values.
> > If we can make the latter use case work in a sane way, OK, although I
> > have my doubts about how sane it really is, but it can't be at the
> > expense of telling extensions that have been (incorrectly) using
> > MaxBackends in _PG_init() that we're throwing them under the bus.
> > 
> > IMHO, the proper thing to do if certain GUC values are required for an
> > extension to work is to put that information in the documentation and
> > error out at an appropriate point if the user does not follow the
> > directions. Then this issue does not arise. But there's no reasonable
> > workaround for being unable to size data structures based on
> > MaxBackends.
> 
> FWIW I would be on board with reverting all the GetMaxBackends() stuff if
> we made the value available in _PG_init() and stopped supporting GUC
> overrides by extensions (e.g., ERROR-ing in SetConfigOption() when
> process_shared_preload_libraries_in_progress is true).

Yeah I would prefer this approach too, although it couldn't prevent extension
from directly modifying the underlying variables so I don't know how effective
it would be.

On the bright side, I see that citus is using SetConfigOption() to increase
max_prepared_transactions [1].  That's the only extension mentioned in that
thread that does modify some GUC, and this GUC was already marked as
PGDLLIMPORT since 2017 so it probably wasn't done that way for Windows
compatibility reason.

In the meantime, should we add an open item?

[1] https://github.com/citusdata/citus/blob/master/src/backend/distributed/transaction/transaction_management.c#L656-L657