Re: Dead code with short varlenas in toast_save_datum()

On Wed, Aug 06, 2025 at 10:30:54AM +0900, Michael Paquier wrote:
> Ah, thanks, nice one. I did not consider the trick of using two
> attributes to bypass the check when externalizing the tuple. I'll go
> add a test in strings.sql among these lines, with some TOAST slice
> scans based on substr().

Done that as of 225ebfe30a1a, with some additions:
- A check on pg_column_compression(), to make sure that the short
varlena path leads to an uncompressed on-disk Datum.
- Some substr() for the values, to cross-check slice reads.
- A query based on the TOAST relation's chunk_seq, making sure that we
have taken toast_save_datum() two times when inserting a tuple with
the two toastable attributes.

The buildfarm is OK with it, as is the latest patch for the 8-byte
TOAST values posted here:
https://www.postgresql.org/message-id/aIyCz4T858Kcm4UU@paquier.xyz

The patch posted there needs a rebase, following the changes in
varatt.h done in e035863c9a04. Will post a rebase very soon.
--
Michael