On 2015-12-09 20:23:06 -0500, Noah Misch wrote:
> By the way, it occurs to me that I should also make pg_upgrade blacklist the
> range of catversions that might have data loss. No sense in putting ourselves
> in the position of asking whether data files of a 9.9.3 cluster spent time in
> a 9.5beta2 cluster.
I can't see any benefit in that. We're talking about a bug that, afaics,
needs another unknown bug to trigger (so find_multixact_start() fails),
and then very likely needs significant amounts of new multixacts
consumed, without a restart and without find_multixact_start()
succeeding later.
What I think would actually help for questions like this, is to add, as
discussed in some other threads, the following:
1) 'creating version' to pg_control
2) 'creating version' to each pg_class entry
3) 'last relation rewrite in version' to each pg_class entry
4) 'last full vacuum in version' to each pg_class entry
I think for this purpose 'version' should be something akin to
$catversion||$numericversion (int64 probably?) - that way development
branches and release branches are handled somewhat usefully.
I think that'd be useful, both from an investigatory perspective, as
from a tooling perspective, because it'd allow reusing things like hint
bits.
> > Ripping it out and replacing it monolithically will not
> > change that; it will only make the detailed history harder to
> > reconstruct, and I *will* want to reconstruct it.
>
> What's something that might happen six months from now and lead you to inspect
> master or 9.5 multixact.c between 4f627f8 and its revert?
"Hey, what has happened to multixact.c lately? I'm investigating a bug,
and I wonder if it already has been fixed?", "Uh, what was the problem
with that earlier large commit?", "Hey, what has changed between beta2
and the final release?"...