Re: BUG #5946: Long exclusive lock taken by vacuum (not full)

On Fri, Mar 25, 2011 at 5:56 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Christopher Browne <cbbrowne(at)gmail(dot)com> writes:
>> What seems natural-ish to me might include:
>> - Stomping a bit on the FSM replacement to make sure nobody's going to
>> be writing to the later extensions;
>> - Watching free space during the process so the "first" extension gets
>> re-opened up if the free space in the much earlier parts of the table
>> (e.g. - that are not planned to be dropped off) is running out.
>
> You seem to be thinking only about the possibility that somebody would
> try to write a new tuple into the space-to-be-freed.  The problem that
> necessitates use of AccessExclusiveLock is that somebody could be doing
> a seqscan that tries to *read* the blocks that are about to be truncated
> away.  We can't really improve matters much here unless we think of a
> way to fix that.  It would be okay if the scan just ignored blocks it
> failed to read, but how do you distinguish the case from a filesystem
> error that really should be reported?

It's struck me a number of times that it would make some things
simpler if we were able to maintain some state in shared memory about
the tables people were using - for example, in this case, we could
cache the table size, or the fact that vacuum has just truncated away
N blocks, or, uh, something. *waves hands* But it's hard to know how
such an area could reasonably be sized or managed.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company