unpackaged upgrade from PostGIS 3.2 failing on ST_AsGeoJSON

[robe]

As a result of #5596 we are getting this regression on woodie. I can't replicate it locally, perhaps cause I am running PG16 and only have 3.3 and 3.4 installed or because I can't get the unpackaged path to kick in.

ALTER TABLE
ALTER TABLE
SELECT 8500
DELETE 8500
SET
SET
SET
SET
SET
 set_config 
------------
 
(1 row)

SET
SET
SET
SET
ERROR:  function "st_asgeojson" already exists with same argument types
-----------------------------------------------------------------------------
make: *** [/woodpecker/src/git.osgeo.org/gitea/postgis/postgis/regress/runtest.mk:24: check-regress] Error 3
make: Leaving directory '/woodpecker/src/git.osgeo.org/gitea/postgis/postgis/build/pg15/regress'
FAIL: postgis script hard upgrade unpackaged3.2--:auto
Cleaning up

This doesn't seem to affect extension upgrades.

I see the new signature in postgis_restore_data.generated

Does the old signature need to be put somewhere besides the postgis_drop_before.sql?

[strk]

I think it should be ok to use _postgis_drop_function_by_identity in postgis_drop_before as you did, although I wonder why you didn't use a Replaces tag instead, which would have been quicker to type.

I'll see to reproduce locally to fix it

[strk]

I could reproduce, locally, with:

regress/run_test.pl -v --dumprestore --upgrade-path unpackaged3.2--:auto regress/core/regress
ERROR:  function "st_asgeojson" already exists with same argument types

The upgrade script is clearly not finding a match for the identity passed to _postgis_drop_function_by_identity

[strk]

To be noted that soft-upgrade works, so this is related to the postgis_restore.pl script failing to recognize the identity of the function in the dump

[strk]

I confirm using Replaces fixes this problem. This is because Replaces will use signature, not identity, and we only care about signatures (not identities) for the skip-list of postgis_restore.pl

[Sandro Santilli <strk@…>]

In a6f4fa8/git:

Fix hard-upgrades failing to skip old st_asgeojson signature

Uses "Replaces" comment instead of _postgis_drop_function_by_identity
function. The latter is only required when the function signature
remains the same but argument names change. These cases are of no
interest for postgis_restore because only the signature (and not the
identity) can be found in PostgreSQL dumps, and so we didn't bother
to implement support for _postgis_drop_function_by_identity in the
script that extracts the skip-list for postgis_restore.

Closes #5623