Re: Bug in nbtree optimization to skip > operator comparisons (or < comparisons in backwards scans)

Hi, Anton!
Looks like an oversight when refactoring BTScanOpaqueData.firstPage into
using function argument in 06b10f80ba4.

@@ -2487,14 +2486,13 @@ _bt_endpoint(IndexScanDesc scan, ScanDirection dir)
104
105 /* remember which buffer we have pinned */
106 so->currPos.buf = buf;
107 - so->firstPage = true;
108
109 _bt_initialize_more_data(so, dir);
110
111 /*
112 * Now load data from the first page of the scan.
113 */
114 - if (!_bt_readpage(scan, dir, start))
115 + if (!_bt_readpage(scan, dir, start, false))

Attached is a fix.
Thank you!

Regards,
Pavel

On Fri, 22 Mar 2024 at 11:29, Anton A. Melnikov <a(dot)melnikov(at)postgrespro(dot)ru>
wrote:

> Hi!
>
> Maybe _bt_readpage(scan, dir, start, true) needed at this line:
>
>
> https://github.com/postgres/postgres/blob/b4080fa3dcf6c6359e542169e0e81a0662c53ba8/src/backend/access/nbtree/nbtsearch.c#L2501
>
> ?
>
> Do we really need to try prechecking the continuescan flag here?
>
> And the current "false" in the last arg does not match the previous code
> before 06b10f80ba
> and the current comment above.
>
> Would be very grateful for clarification.
>
> With the best regards!
>
> --
> Anton A. Melnikov
> Postgres Professional: http://www.postgrespro.com
> The Russian Postgres Company
>