Re: COPY FREEZE and setting PD_ALL_VISIBLE/visibility map bits

Hi,

I started looking at this patch again, hoping to get it committed in
this CF, but I think there's a regression in handling TOAST tables
(compared to the v3 patch submitted by Pavan in February 2019).

The test I'm running a very simple test (see test.sql):

1) start a transaction
2) create a table with a text column
3) copy freeze data into it
4) use pg_visibility to see how many blocks are all_visible both in the
main table and it's TOAST table

For v3 patch (applied on top of 278584b526 and s/hi_options/ti_options)
I get this:

pages NOT all_visible
------------------------------------------
main 637 0
toast 50001 3

There was some discussion about relcache invalidations causing a couple
TOAST pages not be marked as all_visible, etc.

However, for this patch on master I get this

pages NOT all_visible
------------------------------------------
main 637 0
toast 50001 50001

So no pages in TOAST are marked as all_visible. I haven't investigated
what's causing this, but IMO that needs fixing to make ths patch RFC.

Attached is the test script I'm using, and a v5 of the patch - rebased
on current master, with some minor tweaks to comments etc.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company