Re: Re: [COMMITTERS] pgsql: Perform only one ReadControlFile() during startup.

On 2017-09-16 14:30:21 -0400, Tom Lane wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
> > Looking into it.
>
> I wonder whether we shouldn't just revert this patch altogether.
> Certainly, extra reads of pg_control are not a noticeable performance
> problem.

The problem is that the patch that makes the segment size configurable
also adds a bunch more ordering constraints due to the fact that the
contents of the control file influence how much shared buffers are
needed (via wal_buffers = -1, which requires the segment size, which is
read from the control file). Reading things in the wrong order leads to
bad results too.

> I'm now quite worried about whether we aren't introducing
> hazards of using stale values from the file; if a system crash isn't
> enough to get it to flush its cache, then what is?

I don't think the problem here is stale values, it's "just" a stale
pointer pointing into shared memory that gets reiniitalized?

Greetings,

Andres Freund