On Thursday 11 February 2010 11:10:32 Simon Riggs wrote:
> On Mon, 2010-02-08 at 04:33 +0000, Tom Lane wrote:
> > We still have to retain all code that copes with finding
> > HEAP_MOVED_OFF and HEAP_MOVED_IN flag bits on existing tuples. This
> > can't be removed as long as we want to support in-place update from
> > pre-9.0 databases.
>
> This doesn't seem to be a great reason. Letting weird states exist is
> not a feature, its a risk. Let me explain.
>
> This would only happen if a VACUUM FULL had been run on the pre-9.0
> database and it had failed part way through. Re-VACUUMing would remove
> those settings.
>
> ISTM that that the upgrade process should cover this, not force the
> server to cope with rare and legacy situations. If we do not do this,
> then we might argue it should *never* be removed because this same rare
> situation can persist into 9.1 etc..
>
> There were data loss situations possible in early 8.4 and these
> persisted into later releases *because* the minor release upgrade
> process did not contain a scan to detect and remove the earlier
> problems. If we allow tuples to be in strange legacy states we greatly
> increase the difficulty of diagnosing and fixing problems. People will
> say "moved in/off can be ignored now" and mistakes will happen.
>
> We should remove the moved in/off flag bits and make it a part of the
> upgrade process to ensure the absence of those states.
Essentially requiring a successfull VACUUM FULL or CLUSTER on all tables is
imho in the same ballpark as requiring a dump+restore timewise on bigger
databases.
Andres