Jeff Davis wrote:
> On Mon, 2007-11-05 at 22:45 +0000, Heikki Linnakangas wrote:
>>> 1) Do as you say above. What are some of the cost trade-offs here? It
>>> seems that frequent VACUUM FREEZE runs would keep the visibility map
>>> mostly full, but will also cause more writing. I suppose the worst case
>>> is that every tuple write needs results in two data page writes, one
>>> normal write and another to freeze it later, which sounds bad. Maybe
>>> there's a way to try to freeze the tuples on a page before it's written
>>> out?
>> It would also create more WAL traffic, because freezing tuples needs to
>> be WAL-logged.
>
> The thought crossed my mind, but I couldn't think of any reason that
> would need to be logged. Of course you're right, and the comments
> explain it well.
>
>> 5) Have a more fine-grain equivalent of relfrozenxid. For example one
>> frozenxid per visibility map page, so that whenever you update the
>> visibility map, you also update the frozenxid. To advance the
>> relfrozenxid in pg_class, you scan the visibility map and set
>> relfrozenxid to the smallest frozenxid. Unlike relfrozenxid, it could be
>> set to FrozenXid if the group of pages are totally frozen.
>>
>
> Wouldn't that still require WAL traffic? Otherwise how can you guarantee
> that the FrozenXid hits disk before TruncateCLOG truncates the old xmin
> away?
Updating the fine-grain frozenxid would still need to be WAL-logged. But
it would be lot less frequent than aggressively freezing tuples.
Compared to the idea of having a separate bitmap or two bits per tuple
in one data structure, you wouldn't necessarily have to freeze tuples to
advance it, you could just observe what the smallest xid on a group of
pages is. Like regular lazy vacuum does right now for relfrozenxid, just
more fine-grained.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com
