On 25.09.2011 19:01, Robert Haas wrote:
> On Wed, Sep 14, 2011 at 6:52 AM, Heikki Linnakangas
> <heikki.linnakangas@enterprisedb.com> wrote:
>>> Why do you need new WAL replay routines? Can't you just use the existing
>>> XLOG_HEAP_NEWPAGE support?
>>>
>>> By any large, I think we should be avoiding special-purpose WAL entries
>>> as much as possible.
>>
>> I tried that, but most of the reduction in WAL-size melts away with that.
>> And if the page you're copying to is not empty, logging the whole page is
>> even more expensive. You'd need to fall back to retail inserts in that case
>> which complicates the logic.
>
> Where does it go? I understand why it'd be a problem for partially
> filled pages, but it seems like it ought to be efficient for pages
> that are initially empty.
A regular heap_insert record leaves out a lot of information that can be
deduced at replay time. It can leave out all the headers, including just
the null bitmap + data. In addition to that, there's just the location
of the tuple (RelFileNode+ItemPointer). At replay, xmin is taken from
the WAL record header.
For a multi-insert record, you don't even need to store the RelFileNode
and the block number for every tuple, just the offsets.
In comparison, a full-page image will include the full tuple header, and
also the line pointers. If I'm doing my math right, a full-page image
takes 25 bytes more data per tuple, than the special-purpose
multi-insert record.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com