Tom Lane wrote:
> Heikki Linnakangas <heikki@enterprisedb.com> writes:
>> We wouldn't clean up tuples that are visible to a transaction, but if
>> you have one long-running transaction like pg_dump in a database with
>> otherwise short transaction, you'll have a lot of tuples that are not
>> vacuumable because of the long-running process, but are not in fact
>> visible to any transaction.
>
> It sounds to me like you are proposing to remove the middles of update
> chains, which would break READ-COMMITTED updates initiated by the older
> transactions. Now admittedly pg_dump isn't going to issue any such
> updates, but VACUUM doesn't know that.
I was thinking of inserts+deletes. Updates are harder, you'd need to
change the ctid of the old version to skip the middle part of the chain,
atomically, but I suppose they could be handled as well.
Isolation level doesn't really matter. We just need a global view of
in-use *snapshots* in the system, serializable or not. Not that that's
an easy thing to do...
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com