Re: Block level concurrency during recovery

Simon Riggs wrote:
> I'm undecided as to whether this is too much code or too little. I'm
> somewhat uncertain as to the locking requirements for the physical scan
> during a vacuum. I've worked out various options if we need to change
> this.

For the B-tree, an exclusive lock is enough to modify the contents of
the page. A cleanup lock needs to be taken on every page to ensure that
the vacuum doesn't finish and delete a heap tuple that's about to be
accessed by an index scan. That could be handled in WAL replay by
dropping the exclusive lock and re-acquiring the cleanup lock, but we
can't do that for heap vacuums.

However, we require that in b-tree vacuum, you take a cleanup lock on
*every* leaf page of the index, not only those that you modify. That's a
problem, because there's no trace of such pages in the WAL.

PS. FSM doesn't need cleanup locks.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com