Re: Yet another fast GiST build

On 18/01/2021 00:35, Peter Geoghegan wrote:
> On Sun, Jan 17, 2021 at 12:50 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I noticed that gist_page_items() thinks it can hold inter_call_data->rel
>> open across a series of calls.  That's completely unsafe: the executor
>> might not run the call series to completion (see LIMIT), resulting in
>> relcache leak complaints.

Fixed, thanks! I changed it to return a tuplestore.

> It also has the potential to run into big problems should the user
> input a raw page image with an regclass-argument-incompatible tuple
> descriptor. Maybe that's okay (this is a tool for experts), but it
> certainly is a consideration.

I'm not sure I understand. It's true that the raw page image can contain 
data from a different index, or any garbage really. And the function 
will behave badly if you do that. That's an accepted risk with 
pageinspect functions, that's why they're superuser-only, although some 
of them are more tolerant of corrupt pages than others. The 
gist_page_items_bytea() variant doesn't try to parse the key data and is 
less likely to crash on bad input.

- Heikki