Re: BUG #17586: Look like there are an another bug in REINDEX INDEX CONCURRENTLY of pg_toast indexes

Hi,

On 2022-08-15 11:36:52 +0000, PG Bug reporting form wrote:
> Bug reference:      17586
> Logged by:          Maxim Boguk
> Email address:      maxim.boguk@gmail.com
> PostgreSQL version: 14.4
> Operating system:   Ubuntu Linux
> Description:
>
> Details about the data structure
> pg_toast.pg_toast_2502234687_index is index toast of exports_new_2022_32w
> partition of exports_new table (Partition key: RANGE (epoch))
>
> During the reindex job there were quite a lot new inserts into
> exports_new_2022_32w table via partition head.
>

> 2022-08-11 02:46:16.972 UTC 1175180 ***_push@***_exports from [local]
> [vxid:12/63545267 txid:354388805] [INSERT] LOG:  duration: 2711.995 ms
>  statement: INSERT INTO exports_new(field_board_id, field_id_mod_256, epoch,
> field_ids)
>                 (SELECT field_board_id, (field_id % 256)::int
> field_id_mod_256, epoch, ARRAY_AGG(field_id) AS field_ids FROM (SELECT
> field_board_id, epoch, unnest(field_ids) AS field_id FROM temp_export_ids)
> t
>                 GROUP BY 1,2,3)
>
>
> 2022-08-11 02:47:28.927 UTC 1187482 postgres@***_exports from [local]
> [vxid:14/66875173 txid:0] [REINDEX] LOG:  temporary file: path
> "base/pgsql_tmp/pgsql_tmp1187482.1.sharedfileset/2.0", size 8192



> [vxid:10/63478386 txid:0] [SELECT] ERROR:  missing chunk number 0 for toast
> value 2546878653 in pg_toast_2502234687
> 2022-08-11 02:48:49.512 UTC 1184504 ***_field_sync@***_exports from [local]
> [vxid:10/63478386 txid:0] [SELECT] STATEMENT:  SELECT field_ids FROM
> exports_new WHERE field_board_id = $1 AND field_id_mod_256 = $2 AND epoch =
> $3

Hm. What is the setup of this system? Do you have a hot_standby_feedback
enabled replica? Does it use replication slots? Did it reconnect at the time?

This issue could be due to the xmin horizon going backward, rather than a bug
in reindex concurrently itself.


Too bad we don't have an errcontext set up for this, it'd be quite helpful to
know what row the missing toast value is in, so we could check its visibility.