答复: GiST API Adancement

Hi Andrey,

Thank you for your suggestion.

I should still maintain original API of GiST after modification.

--

Dong

________________________________
发件人: Andrew Borodin <borodin(at)octonica(dot)com>
发送时间: 2017年6月16日 6:24:03
收件人: Yuan Dong
抄送: pgsql-hackers(at)postgresql(dot)org
主题: Re: GiST API Adancement

Hi, Dong!

2017-06-15 21:19 GMT+05:00 Yuan Dong <Doffery(at)hotmail(dot)com>:
> I'm going to hack on my own. With the help of Andrew Borodin, I want to
> start the project with adding a third state to collision check. The third
> state is that:
> subtree is totally within the query. In this case, GiST scan can scan down
> subtree without key checks.

That's fine!

> After reading some code, I get this plan to modify the code:
>
> 1. Modify the consistent function of datatypes supported by GiST.
>
> 1.1 Start with cube, g_cube_consistent should return a third state when
> calling g_cube_internal_consistent. Inspired by Andrew Borodin, I have two
> solutions, 1)modify return value, 2) modify a reference type parameter.
>
> 2. Modify the gistindex_keytest in gistget.c to return multiple states
>
> 2.1 Need declare a new enum type(or define values) in gist_private.h or
> somewhere
>
> 3. Modify the gitsScanPage in gistget.c to deal with the extra state
>
> 4. Add a state to mark the nodes under this GISTSearchItem are all with in
> the query
>
> 4.1 Two ways to do this: 1) add a state to GISTSearchItem (prefered) 2)
> Somewhere else to record all this kind of items
>
> 4.2 We may use the block number as the key
>
> 4.3 Next time when the gistScanPage met this item, just find all the leaves
> directly(need a new function)
>
> After this, I shall start to benchmark the improvement and edit the code of
> other datatypes.
>
> Hope you hackers can give me some suggestions~

I think there is one more aspect of development: backward
compatibility: it's impossible to update all existing extensions. This
is not that major feature to ignore them.

Though for benchmarking purposes backward compatibility can be omitted.

Best regards, Andrey Borodin