Re: Proposal for Allow postgresql.conf values to be changed via SQL

Tom Lane <tgl@sss.pgh.pa.us>

From: Tom Lane <tgl@sss.pgh.pa.us>

To: Amit kapila <amit.kapila@huawei.com>

Cc: Fujii Masao <masao.fujii@gmail.com>, "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>, Dimitri Fontaine <dimitri@2ndquadrant.fr>, "cedric@2ndquadrant.com" <cedric@2ndquadrant.com>, Robert Haas <robertmhaas@gmail.com>, Greg Smith <greg@2ndquadrant.com>, Josh Berkus <josh@agliodbs.com>, Magnus Hagander <magnus@hagander.net>, Christopher Browne <cbbrowne@gmail.com>, Alvaro Herrera <alvherre@2ndquadrant.com>

Date: 2012-11-24T17:26:02Z

Lists: pgsql-hackers

Amit kapila <amit.kapila@huawei.com> writes:
> On Friday, November 23, 2012 10:10 PM Fujii Masao wrote:
>> What happens if the server crashes while SET PERSISTENT is writing the
>> setting to the file? A partial write occurs and restart of the server would fail
>> because of corrupted postgresql.auto.conf?

> This situation will not happen as SET PERSISTENT command will first write to ".lock" file and then at commit time, 
> rename it to ".auto.conf".

Yes, the right way to write the config file is to write under a
temporary name, fsync the file, and then use rename(2) to atomically
move it into place.  However, the above is contemplating some extra
complexity that I think is useless and undesirable, namely postponing
the rename until commit time.  The point of the suggestion that SET
PERSISTENT not be allowed inside a transaction block is so that you can
write the file immediately rather than have to add commit-time mechanism
to support the feature.  Aside from being extra complexity, and some
extra cycles added in *every single commit*, a post-commit write creates
another way to have post-commit failures, which we cannot cope with in
any sane way.

			regards, tom lane