[proposal] New resolver configuration#938
[proposal] New resolver configuration#938tiran wants to merge 1 commit intopython-wheel-build:mainfrom
Conversation
tiran
force-pushed
the
proposal-new-resolver-config
branch
3 times, most recently
from
ceb5b48 to
071d46c
Compare
|
@shifa-khan has been looking at one of the other tickets related to similar work. |
|
I have inspect our downstream repo. Most
|
tiran
force-pushed
the
proposal-new-resolver-config
branch
from
071d46c to
3356160
Compare
| - `resolver_dist.sdist_server_url` is replaced by `index_url` parameter. | ||
| All `pypi-*` profile support a custom index. | ||
| - `git_options.submodules` is not needed. Like pip, Fromager will always | ||
| clone all submodules. |
There was a problem hiding this comment.
In our downstream, we have packages that explicitly need submodules=False and cases where submodules point to private repos or have broken URLs. These packages when migrated to YAML would deviate from current behavior. Do we have a workaround to address such cases?
For example, adding an optional submodules parameter with true as the default?
There was a problem hiding this comment.
I'm curious, which package fails to build when submodules are enabled? Do you know? I was under the impression that engineers added submodules=False when the package did not need submodules.
Yes, we have a workaround: keep the Python plugin.
If we really need a submodules: false flag for several packages, then we could consider the feature in the future. I would like to start with a simple configuration (80/20 rules).
There was a problem hiding this comment.
Yeah, my bad - let me rephrase. We have packages (tacozip, tpu_info, vllm_neuron) that use submodules=False, which could be for optimization or maybe to avoid actual failures. There's also the future case where submodules might point to private repos or broken URLs. But as you said, keeping a plugin for those edge cases should work.
|
downstream has few plugins call |
tiran
left a comment
tiran
left a comment
There was a problem hiding this comment.
downstream has few plugins call ensure_pkg_info() in prepare_source as boilerplate. Git clones never have PKG-INFO, but build tools (like setuptools-scm) need it before building. Could the git-based providers (gitlab-git, github-git, pypi-git) handle this automatically?
Great point! Yes, this will be handled automatically in two ways:
- Fromager already adds
PKG-INFOto the final sdist. In some cases, the file is added too late for some PEP 517 hooks. - Fromager will add a
.git_archival.txtfile to clones. Lala is working on the feature in #961 / #962 . This works for all setuptools-scm cases.
I'll update the design with the PKG-INFO and .git_archival.txt details.
| - `resolver_dist.sdist_server_url` is replaced by `index_url` parameter. | ||
| All `pypi-*` profile support a custom index. | ||
| - `git_options.submodules` is not needed. Like pip, Fromager will always | ||
| clone all submodules. |
There was a problem hiding this comment.
I'm curious, which package fails to build when submodules are enabled? Do you know? I was under the impression that engineers added submodules=False when the package did not need submodules.
Yes, we have a workaround: keep the Python plugin.
If we really need a submodules: false flag for several packages, then we could consider the feature in the future. I would like to start with a simple configuration (80/20 rules).
Add design proposal for new resolver configuration. See: python-wheel-build#937 Signed-off-by: Christian Heimes <cheimes@redhat.com>
tiran
force-pushed
the
proposal-new-resolver-config
branch
from
3356160 to
6559cd3
Compare
3 participants
Add design proposal for new resolver configuration.
See: #937