> 1. After a certain point, consecutive GIN index splits cause a problem.
> The new RHS block numbers are consecutive from 111780+
That's newly created page. Splitted page might have any number
>
> 2. The incomplete splits stay around indefinitely after creation and we
> aren't trying to remove the wrong split at any point. We're either never
> creating an xlog record, or we are ignoring it in recovery, or we are
> somehow making multiple entries then not removing all of them.
Agreed
> 3. The root seems to move, which isn't what I personally was expecting
> to see. It seems root refers to the highest parent involved in the
> split.
root in this context means parent of splitted page. Actually, there is a lot of
B-tree in GIN, see http://www.sigaev.ru/gin/GinStructure.pdf
>
> 4. We're writing lots of redo in between failed page splits. So *almost*
> everything is working correctly.
>
> 5. This starts to happen when we have very large indexes. This may be
> coincidental but the first relation file is fairly full (900+ MB).
Yes. It seems to me that conditions of error are very rare and B-tree over
ItemPointers (second level of GIN) has a big capacity, 1000+ items per page. So,
splits occur rather rare.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/
