repodata patch backporting pin_subpackage(..., exact=True)

[croth1]

Question

I am currently trying to figure out how to backport this change with a repodata patch:

 run_constrained:
        - {{ pin_subpackage('seaborn', exact=True) }}

Change: https://github.com/conda-forge/seaborn-feedstock/pull/35/files#diff-f3725a55bf339595bf865fec73bda8ac99f283b0810c205442021f29c06eea9aR40-R41

In index.json an exact pinning with this mechanism seems to show up as <package-name> <verision> <build> in the depends section, so my naive guess for a patch was:

if:
  name: seaborn-base
  timestamp_lt: 1713986364000
then:
  - add_constrains: seaborn ${version} ${build}

However, ${build} does not seem to be exposed in the patching yaml 🤔 https://github.com/conda-forge/conda-forge-repodata-patches-feedstock/blob/main/recipe/patch_yaml_utils.py#L10-L19

Is there currently a way to backport the exact pin, and if not, would it make sense to expose ${build} in the yaml file?

Exposing build variable and applying patch

diff --git a/recipe/patch_yaml_utils.py b/recipe/patch_yaml_utils.py
index 25d29ea..be3415d 100644
--- a/recipe/patch_yaml_utils.py
+++ b/recipe/patch_yaml_utils.py
@@ -11,6 +11,7 @@ ALLOWED_TEMPLATE_KEYS = [
     "name",
     "version",
     "build_number",
+    "build",
     "subdir",
     "next_version",
     "major_version",

After exposing build, the repodata patch would do the following:

diff --git a/before.txt b/after.txt
index 88a3601..649db17 100644
--- a/before.txt
+++ b/after.txt
@@ -1,6 +1,7 @@
 ('seaborn-base-0.12.1-pyhd8ed1ab_0.tar.bz2',
  {'build': 'pyhd8ed1ab_0',
   'build_number': 0,
+  'constrains': ['seaborn 0.12.1 pyhd8ed1ab_0'],
   'depends': ['matplotlib-base >=3.1,!=3.6.1',
               'numpy >=1.17',
               'pandas >=0.25',

This assumes that the build variable is the same for the two outputs seaborn and seaborn-base. Empirically this seems to hold, but since I don't know how the build string is generated exactly it probably would be good to sanity-check this.

For comparison, this is how the index looks like for pin_subpackage("seaborn-base", exact=True) where seaborn pins seaborn-base in its dependencies.

{
  "arch": null,
  "build": "hd8ed1ab_0",
  "build_number": 0,
  "depends": [
    "seaborn-base 0.13.2 pyhd8ed1ab_0",
    "statsmodels >=0.12"
  ],
  "license": "BSD-3-Clause",
  "license_family": "BSD",
  "name": "seaborn",
  "noarch": "python",
  "platform": null,
  "subdir": "noarch",
  "timestamp": 1706341016553,
  "version": "0.13.2"
}

[croth1]

A couple of notes:

1. It turns out {{ pin_subpackage('seaborn', exact=True) }} in run_constrained is actually broken - see run_constrained not handled properly in test with outputs and reverse pin_subpackage  conda/conda-build#4415 for reference and a workaround
2. Even if it had worked when patching with the build string one needs to be extra careful, as in my case one build string is just {hash}_{build_number} (e.g. noarch/seaborn-0.13.2-hd8ed1ab_2.conda), the other is py{hash}_{build_number} (e.g. noarch/seaborn-base-0.13.2-pyhd8ed1ab_2.conda).

Multi-output recipes are just weird 😬