Changes to the CI Infrastructure are made to the main branch in this repository. Config branches use the ci.yml workflows to workflow_call the equivalent workflow that is in model-config-tests.
Since the logic in the CI infrastructure is quite involved, it would be a good idea to read the README-DEV.md.
Each configuration has a dev-* and a release-* branch. They differ in the CI checks that are run when pull requests are made to update the branch. Any branch starting with either dev-* or release-* are protected branches. You cannot (and should not) modify them directly or create new branches starting with either dev- or release-.
The dev-* branch is where a configuration is updated. Temporary branches should be created and a pull request made to update the dev-* branch. Quality assurance (QA) CI checks are run on pull requests to dev-* branches, but not reproducibility checks. There is no requirement that the version be updated when changes are made to the dev- branch - except on the first time that you create the dev-* branch. So the dev- branch of a configuration allows for smaller changes that can be accumulated before a PR is made to the respective release-* branch.
Pull requests to the release-* branch should be made from the respective dev-* branch and are intended to create a new version of the configuration. These pull requests have CI quality assurance (QA) checks that ensure the model configuration is suitable for release. CI Model reproducibility checks are also conducted: a short test run of the configuration is checked for bitwise reproducibility. The success or otherwise of this check determines if a minor or major version bump is required.
It is expected that the version will be updated before the pull request can be merged. This in turn creates a new tag for that configuration branch. It can be confusing for users if there are a large number of versions of a configuration and it is of little benefit to them. For this reason the atomicity of updates to a released configuration should be minimised, i.e. updates should be meaningful.
Config branches are entirely separate from the main history in this repository, except for a few files in .github. Note, you may need to be an Administrator to commit to release-* or dev-* branches directly.
If you are creating a brand new configuration, and don't have the config stored in another repository, just checkout a dev-* branch from main and delete everything except .github/workflows/ci.yml, then add your config.
Create a dev-* branch by adding the config repository as a remote and checking out the config branch:
git remote add <config_repo> <config_repo_url> # ex. git remote add config git@github.com/my/configs.git
git checkout <config_repo>/<config_branch> -b dev-<config_name> # checkout config from new remote + add to branch, ex. git checkout config/main -b dev-1deg_abc_def
git checkout main -- .github/workflows/ci.yml
git add .
git commit -m "Initial commit for config branch"
git push # might require admin permissions for pushes to dev-* branchFor a brand new configuration there is no existing release-* branch, so one needs to be created. Follow the pull request process outlined below to update the dev branch so that it is passing QA checks. At this point create a release-* branch from the dev- branch and git push it to the repository:
git checkout -b release-<config_name>
git push release-<config_name>For the CI workflows to work correctly the release- branch needs to have a version set, and a reproducibility checksum committed. There is a convenience workflow for this purpose: Generate Initial Checksums. Click the "Run workflow" menu, fill in the fields and push the green "Run workflow" button.
Once the workflow is completed there should be a new commit on the release-* branch, and a tag for the specified version.
Once the release-* branch has been updated those changes need to be merged back into the dev-* branch. This step is only necessary when the release-* branch is updated independently of the dev-* branch.
- Make your changes, test them, and open a PR from a feature/change branch (or fork) to the
dev-*branch of a particular configuration. - QA checks will run to ensure the configuration meets criteria for a released configuration, and to ensure consistency of released configurations.
- Fix the problems identified in the QA checks, commit and push to the PR branch.
- Once all checks pass the pull request branch can be merged.
- Optionally, you can get the checksums for the
dev-*branch (without committing or pull requesting torelease-*) via thegenerate-initial-checksums.ymlworkflow, unchecking thecommitoption, and downloading the archive as part of the workflow run. - Optionally, consider making a PR to the equivalent
release-*branch.
Note: If this is a brand new configuration and there is no existing release-* branch you will need to create one first.
- Open a PR from the
dev-*branch of a particular configuration to the equivalentrelease-*branch - QA checks will run to ensure the configuration meets criteria for a released configuration, and to ensure consistency of released configurations.
- Checks will also run to test if changes break reproducibility with the current major version config tag on the target branch. For example, if you are opening a PR on the
release-1deg_jra55_iafbranch, and the last tagged version on this branch isrelease-1deg_jra55_iaf-1.2, the checksums between the config in your PR and the checksum in the config tag are compared. - A comment will be posted on the PR when this is completed, notifying you whether the checksums match (in this example meaning a minor bump to
*-1.3), or are different (meaning a major bump to*-2.0). - Optionally, you can now modify your PR and get more reproducibility checks. Particularly in the case where bitwise reproducibility should be retained this is an opportunity to modify the configuration to enable this.
- Bump the version using the
!bump [major|minor]command depending on the result of the reproducibility check. Additionally, if the checksums are different, the updated checksum will be automatically committed to the PR. Bumping the version in some way is a requirement before the PR will be mergable. - Merge the PR
Each configuration should have a metadata file, metadata.yaml, which uses the ACCESS-NRI experiment schema. Note that required fields of the schema are removed from file validations checks as experiment specific fields such as UUIDs should not be set for released configurations. The following fields must be set in metadata.yaml:
Use the existing release-* version. If there isn't an existing version set to null.
Acceptable values for this field must be set as a sequence:
realm:
- <acceptable value 1>
- <acceptable value 2>
# - ...Choose the appropriate value for the resolution used:
| Config resolution | Nominal Resolution |
|---|---|
| 1° | 100 km |
<Add more as applicable>
These are sourced from the CMIP6 controlled vocabulary. If your resolution differs from those listed you will need to make a pull request to add it to this documentation and the QA checks.
We have a "controlled vocabulary of keywords to prevent a proliferation of synonyms that mean the same thing, and to make it easy to populate these fields:
| Topic | Keywords (mutually exclusive) |
|---|---|
<topic> |
<keywords> |
An appropriate scientific reference for the configuration. <Give example of an appropriate scientific reference for this model>
This is the license that will apply to the model outputs for an experiment. This should be set to the SPDX identifier for CC BY 4.0 (CC-BY-4.0) to alleviate users from the burden of choosing a license, and to ensure model outputs have a permissive license for reuse to encourage open and shareable science.
This is a bit tricky. Ideally this should be a URL to the GitHub (or similar) repository of the configuration for the experiment being run. So if we include this and require it to be filled then it should either be the URL pointing at the branch being modified, or a placeholder. Either way it should include a comment that it should be updated to reference the experiment being run.
<Give acceptable model names>
This is checked to make sure a shorter run time hasn't been set during testing and forgotten to set back to the proper value. As it is difficult to create a general heuristic the values have been hard-coded to those shown below:
| Config resolution | restart_period |
|---|
<This table needs to be filled in for the model in question>
This governs how what model restart files are retained.
The requirement is simply that a date-based frequency be used so that restarts are saved in a reliable manner. Typical values are 1YS or 5YS for 0.1° models, and 5YS to 20YS for 1°. See the payu documentation for a detailed description of the meaning of the time units in restart_freq.
This should not be enabled by default. Nor should path be set to a real path. Ideally set path to null:
sync:
enable: false
path: nullUsers should enable this, and set the path themselves, as there is no safe default for this.
<any post-processing scripts should be noted here, otherwise remove this section>