Re: crash-safe visibility map, take three

On 01.12.2010 15:39, Bruce Momjian wrote:
> Heikki Linnakangas wrote:
>> On 01.12.2010 03:35, Bruce Momjian wrote:
>>> Heikki Linnakangas wrote:
>>>> Let's recap what happens when a VM bit is set: You set the
>>>> PD_ALL_VISIBLE flag on the heap page (assuming it's not set already, it
>>>> usually isn't), and then set the bit in the VM while keeping the heap
>>>> page locked.
>>>
>>> What if we set PD_ALL_VISIBLE on the heap page, wait for a checkpoint to
>>> happen so the heap page is guaranteed to be on disk, then on next read,
>>> if PD_ALL_VISIBLE is set and the VM all-visible bit is not set, set the
>>> VM bit.
>>
>> Hmm, you'd somehow have to know if a checkpoint has happened since the
>> flag was set. And it might be a long wait, which makes it less
>
> Well, doesn't the page LSN tell you that already?

If we update the LSN when we set the flag, then we have to write the
full-page-image. That's very expensive. If we don't update the LSN, then
the LSN says nothing about when the flag was set.

>> attractive for index-only scans.
>
> My assumption is that this page will remain read-only for a while, so I
> don't see the big downside in a delay of max 5-10 minutes. For sites
> where pages go frequently in and out of read-only status, I don't think
> index-only scans are every going to be a big win. Long-running queries
> alone are going to delay how quickly we can set PD_ALL_VISIBLE, so I
> don't see an additional 5-10 minutes as a big problem.

Perhaps we could live with it, but it would be annoying to have to
checkpoint after a data load, before index-only scans kick in.

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