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

Bruce Momjian wrote:

> It appears this item is on hold until Alvaro can comment. I can
> probably find the impact and correct the problem for previous upgrades,
> but it might be 1-2 months until I can get to it. The fix for future
> upgrades is simple.

TBH I think future upgrades is what we need to fix now, so that people
don't do broken upgrades from 9.2 to 9.4 once the latter is out. (I
would assume that upgrades from 9.3 will not have a problem, unless the
9.3 setup itself is already broken due to a prior pg_upgrade). Also, we
need to fix pg_upgrade to protect people upgrading to 9.3.5 from <= 9.2.

As for detecting the case in existing installs, I think we can supply a
query in the next minor release notes that uses pg_ls_dir() to see
whether there is a 0000 file outside the working range of
multixact/offset, whose result if nonempty would indicate the need to
delete (or more likely rename) a file.

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services