Re: GiST VACUUM

On 25/03/2019 15:20, Heikki Linnakangas wrote:
> On 24/03/2019 18:50, Andrey Borodin wrote:
>> I was working on new version of gist check in amcheck and understand one more thing:
>>
>> /* Can this page be recycled yet? */
>> bool
>> gistPageRecyclable(Page page)
>> {
>> return PageIsNew(page) ||
>> (GistPageIsDeleted(page) &&
>> TransactionIdPrecedes(GistPageGetDeleteXid(page), RecentGlobalXmin));
>> }
>>
>> Here RecentGlobalXmin can wraparound and page will become unrecyclable for half of xid cycle. Can we prevent it by resetting PageDeleteXid to InvalidTransactionId before doing RecordFreeIndexPage()?
>> (Seems like same applies to GIN...)
>
> True, and B-tree has the same issue. I thought I saw a comment somewhere
> in the B-tree code about that earlier, but now I can't find it. I
> must've imagined it.
>
> We could reset it, but that would require dirtying the page. That would
> be just extra I/O overhead, if the page gets reused before XID
> wraparound. We could avoid that if we stored the full XID+epoch, not
> just XID. I think we should do that in GiST, at least, where this is
> new. In the B-tree, it would require some extra code to deal with
> backwards-compatibility, but maybe it would be worth it even there.

I suggest that we do the attached. It fixes this for GiST. The patch
changes expands the "deletion XID" to 64-bits, and changes where it's
stored. Instead of storing it pd_prune_xid, it's stored in the page
contents. Luckily, a deleted page has no real content.

I think we should fix this in a similar manner in B-tree, too, but that
can be done separately. For B-tree, we need to worry about
backwards-compatibility, but that seems simple enough; we just need to
continue to understand old deleted pages, where the deletion XID is
stored in the page opaque field.

- Heikki