On 16.12.2010 15:52, Heikki Linnakangas wrote:
> On 13.12.2010 20:30, Tom Lane wrote:
>> Can we fix it so that each child page is updated, and its downlink
>> inserted, as a separate atomic action? That'd require each intermediate
>> state to be consistent and crash-safe, but I think you really need the
>> intermediate states to be consistent anyway because of concurrent scans.
>
> Here's an updated patch, using that idea.If a page split into more than
> two pages, the downlinks for the pages are inserted to the parent
> one-by-one, right-to-left, until there's only two remaining. Finally the
> downlink for the last remaining right page is inserted and the downlink
> for the original page is updated as one atomic operation.
>
> It was a pretty big rewrite again, but seems to work now. I tested
> splits that span more than two pages by rigging the btree_gist picksplit
> function to choose very bad split points, and the fix-split logic by
> adding elog(ERROR) in strategic places to sometimes leave splits
> incomplete.
One final version, with a bug fix wrt. root page split and some cleanup.
I'm planning to commit this before Christmas. It's a big patch, so
review would be much appreciated.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com