Re: pg_upgrade < 9.3 -> >=9.3 misses a step around multixacts

On Wed, Jul  2, 2014 at 11:56:34AM -0400, Alvaro Herrera wrote:
> > How are we not creating a gap in members files by setting the members
> > value?  Why will that not cause problems?
>
> We're not setting the offset and member value here; we're setting the
> "nextMulti" and "oldestMulti" values.  Both affect the offsets counter,
> not the members counter.  The members counter is kept at zero, so the
> next multi to be allocated will create members starting from the first
> members position in page zero.  initdb of the new cluster created the
> first members page, which corresponds to the first members value that
> will be used.
>
> Note pg_resetxlog --help says:
>
>   -m MXID,MXID     set next and oldest multitransaction ID
>
> I think you're confusing the fact that we pass two values here (next and
> oldest, both apply to offset counters) with offsets vs. members.
>
> To affect the members counter you would use this other pg_resetxlog switch:
>   -O OFFSET        set next multitransaction offset
> which, AFAICS, we only use when upgrading from a 9.3+ cluster to a newer
> 9.3+ cluster (and we also copy the files from old cluster to the new
> one).

OK, thanks for the analysis.  Attached patch applied back through 9.3.

--
  Bruce Momjian  <bruce@momjian.us>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

  + Everyone has their own god. +