Dan found a few low-frequency assertions when he let the patch cook
in DBT-2 for extended periods. For the most part they were
over-zealous assertions, although he made a couple minor adjustments
related to their causes in tracking the process local predicate lock
information used to make decisions about lock granularity promotion.
There is still the possibility of some drift in that data from the
"official" copy in shared memory due to index page splits and
combines from other processes, but that doesn't affect correctness,
just granularity promotion; so it doesn't seem like it's worth
creating LW lock contention to deal with it.
So, as of this moment the code has been through some significant
stress testing which should have shaken out most or all race
conditions. All known defects are fixed, except for a couple areas
on which we're doing some last-minute work. The portions which I
think should be considered mandatory which aren't complete are
prepared transactions, which Dan is currently working on, and
documentation work for the Concurrency Control chapter and a README
file, which I guess I'm going to be working on this weekend (the guy
who had volunteered apparently having fallen off the face of the
earth). Lack of support for prepared transactions will not affect
anyone who isn't trying to use them.
The index types other than btree don't have fine-grained support,
which I don't think is a fatal defect, but it would be nice to
implement. I may be able to get GiST working again this weekend in
addition to the documentation work. The others might not get done
for 9.1 unless someone who knows their way around the guts of those
AMs can give us some advice soon.
-Kevin
