.. Copyright 2013-2023 Lawrence Livermore National Security, LLC and other Spack Project Developers. See the top-level COPYRIGHT file for details. SPDX-License-Identifier: (Apache-2.0 OR MIT) .. _pipelines: ============ CI Pipelines ============ Spack provides commands that support generating and running automated build pipelines in CI instances. At the highest level it works like this: provide a spack environment describing the set of packages you care about, and include a description of how those packages should be mapped to Gitlab runners. Spack can then generate a ``.gitlab-ci.yml`` file containing job descriptions for all your packages that can be run by a properly configured CI instance. When run, the generated pipeline will build and deploy binaries, and it can optionally report to a CDash instance regarding the health of the builds as they evolve over time. ------------------------------ Getting started with pipelines ------------------------------ To get started with automated build pipelines a Gitlab instance with version ``>= 12.9`` (more about Gitlab CI `here `_) with at least one `runner `_ configured is required. This can be done quickly by setting up a local Gitlab instance. It is possible to set up pipelines on gitlab.com, but the builds there are limited to 60 minutes and generic hardware. It is possible to `hook up `_ Gitlab to Google Kubernetes Engine (`GKE `_) or Amazon Elastic Kubernetes Service (`EKS `_), though those topics are outside the scope of this document. After setting up a Gitlab instance for running CI, the basic steps for setting up a build pipeline are as follows: #. Create a repository in the Gitlab instance with CI and a runner enabled. #. Add a ``spack.yaml`` at the root containing your pipeline environment #. Add a ``.gitlab-ci.yml`` at the root containing two jobs (one to generate the pipeline dynamically, and one to run the generated jobs). #. Push a commit containing the ``spack.yaml`` and ``.gitlab-ci.yml`` mentioned above to the gitlab repository See the :ref:`functional_example` section for a minimal working example. See also the :ref:`custom_Workflow` section for a link to an example of a custom workflow based on spack pipelines. Spack's pipelines are now making use of the `trigger `_ syntax to run dynamically generated `child pipelines `_. Note that the use of dynamic child pipelines requires running Gitlab version ``>= 12.9``. .. _functional_example: ------------------ Functional Example ------------------ The simplest fully functional standalone example of a working pipeline can be examined live at this example `project `_ on gitlab.com. Here's the ``.gitlab-ci.yml`` file from that example that builds and runs the pipeline: .. code-block:: yaml stages: [generate, build] variables: SPACK_REPO: https://github.com/scottwittenburg/spack.git SPACK_REF: pipelines-reproducible-builds generate-pipeline: stage: generate tags: - docker image: name: ghcr.io/scottwittenburg/ecpe4s-ubuntu18.04-runner-x86_64:2020-09-01 entrypoint: [""] before_script: - git clone ${SPACK_REPO} - pushd spack && git checkout ${SPACK_REF} && popd - . "./spack/share/spack/setup-env.sh" script: - spack env activate --without-view . - spack -d ci generate --artifacts-root "${CI_PROJECT_DIR}/jobs_scratch_dir" --output-file "${CI_PROJECT_DIR}/jobs_scratch_dir/pipeline.yml" artifacts: paths: - "${CI_PROJECT_DIR}/jobs_scratch_dir" build-jobs: stage: build trigger: include: - artifact: "jobs_scratch_dir/pipeline.yml" job: generate-pipeline strategy: depend The key thing to note above is that there are two jobs: The first job to run, ``generate-pipeline``, runs the ``spack ci generate`` command to generate a dynamic child pipeline and write it to a yaml file, which is then picked up by the second job, ``build-jobs``, and used to trigger the downstream pipeline. And here's the spack environment built by the pipeline represented as a ``spack.yaml`` file: .. code-block:: yaml spack: view: false concretizer: unify: false definitions: - pkgs: - zlib - bzip2 - arch: - '%gcc@7.5.0 arch=linux-ubuntu18.04-x86_64' specs: - matrix: - - $pkgs - - $arch mirrors: { "mirror": "s3://spack-public/mirror" } ci: enable-artifacts-buildcache: True rebuild-index: False pipeline-gen: - any-job: before_script: - git clone ${SPACK_REPO} - pushd spack && git checkout ${SPACK_CHECKOUT_VERSION} && popd - . "./spack/share/spack/setup-env.sh" - build-job: tags: [docker] image: name: ghcr.io/scottwittenburg/ecpe4s-ubuntu18.04-runner-x86_64:2020-09-01 entrypoint: [""] The elements of this file important to spack ci pipelines are described in more detail below, but there are a couple of things to note about the above working example: .. note:: There is no ``script`` attribute specified for here. The reason for this is Spack CI will automatically generate reasonable default scripts. More detail on what is in these scripts can be found below. Also notice the ``before_script`` section. It is required when using any of the default scripts to source the ``setup-env.sh`` script in order to inform the default scripts where to find the ``spack`` executable. Normally ``enable-artifacts-buildcache`` is not recommended in production as it results in large binary artifacts getting transferred back and forth between gitlab and the runners. But in this example on gitlab.com where there is no shared, persistent file system, and where no secrets are stored for giving permission to write to an S3 bucket, ``enabled-buildcache-artifacts`` is the only way to propagate binaries from jobs to their dependents. Also, it is usually a good idea to let the pipeline generate a final "rebuild the buildcache index" job, so that subsequent pipeline generation can quickly determine which specs are up to date and which need to be rebuilt (it's a good idea for other reasons as well, but those are out of scope for this discussion). In this case we have disabled it (using ``rebuild-index: False``) because the index would only be generated in the artifacts mirror anyway, and consequently would not be available during subsequent pipeline runs. .. note:: With the addition of reproducible builds (#22887) a previously working pipeline will require some changes: * In the build-jobs, the environment location changed. This will typically show as a ``KeyError`` in the failing job. Be sure to point to ``${SPACK_CONCRETE_ENV_DIR}``. * When using ``include`` in your environment, be sure to make the included files available in the build jobs. This means adding those files to the artifact directory. Those files will also be missing in the reproducibility artifact. * Because the location of the environment changed, including files with relative path may have to be adapted to work both in the project context (generation job) and in the concrete env dir context (build job). ----------------------------------- Spack commands supporting pipelines ----------------------------------- Spack provides a ``ci`` command with a few sub-commands supporting spack ci pipelines. These commands are covered in more detail in this section. .. _cmd-spack-ci: ^^^^^^^^^^^^ ``spack ci`` ^^^^^^^^^^^^ Super-command for functionality related to generating pipelines and executing pipeline jobs. .. _cmd-spack-ci-generate: ^^^^^^^^^^^^^^^^^^^^^ ``spack ci generate`` ^^^^^^^^^^^^^^^^^^^^^ Throughout this documentation, references to the "mirror" mean the target mirror which is checked for the presence of up-to-date specs, and where any scheduled jobs should push built binary packages. In the past, this defaulted to the mirror at index 0 in the mirror configs, and could be overridden using the ``--buildcache-destination`` argument. Starting with Spack 0.23, ``spack ci generate`` will require you to identify this mirror by the name "buildcache-destination". While you can configure any number of mirrors as sources for your pipelines, you will need to identify the destination mirror by name. Concretizes the specs in the active environment, stages them (as described in :ref:`staging_algorithm`), and writes the resulting ``.gitlab-ci.yml`` to disk. During concretization of the environment, ``spack ci generate`` also writes a ``spack.lock`` file which is then provided to generated child jobs and made available in all generated job artifacts to aid in reproducing failed builds in a local environment. This means there are two artifacts that need to be exported in your pipeline generation job (defined in your ``.gitlab-ci.yml``). The first is the output yaml file of ``spack ci generate``, and the other is the directory containing the concrete environment files. In the :ref:`functional_example` section, we only mentioned one path in the ``artifacts`` ``paths`` list because we used ``--artifacts-root`` as the top level directory containing both the generated pipeline yaml and the concrete environment. Using ``--prune-dag`` or ``--no-prune-dag`` configures whether or not jobs are generated for specs that are already up to date on the mirror. If enabling DAG pruning using ``--prune-dag``, more information may be required in your ``spack.yaml`` file, see the :ref:`noop_jobs` section below regarding ``noop-job``. The optional ``--check-index-only`` argument can be used to speed up pipeline generation by telling spack to consider only remote buildcache indices when checking the remote mirror to determine if each spec in the DAG is up to date or not. The default behavior is for spack to fetch the index and check it, but if the spec is not found in the index, to also perform a direct check for the spec on the mirror. If the remote buildcache index is out of date, which can easily happen if it is not updated frequently, this behavior ensures that spack has a way to know for certain about the status of any concrete spec on the remote mirror, but can slow down pipeline generation significantly. The ``--optimize`` argument is experimental and runs the generated pipeline document through a series of optimization passes designed to reduce the size of the generated file. The ``--dependencies`` is also experimental and disables what in Gitlab is referred to as DAG scheduling, internally using the ``dependencies`` keyword rather than ``needs`` to list dependency jobs. The drawback of using this option is that before any job can begin, all jobs in previous stages must first complete. The benefit is that Gitlab allows more dependencies to be listed when using ``dependencies`` instead of ``needs``. The optional ``--output-file`` argument should be an absolute path (including file name) to the generated pipeline, and if not given, the default is ``./.gitlab-ci.yml``. While optional, the ``--artifacts-root`` argument is used to determine where the concretized environment directory should be located. This directory will be created by ``spack ci generate`` and will contain the ``spack.yaml`` and generated ``spack.lock`` which are then passed to all child jobs as an artifact. This directory will also be the root directory for all artifacts generated by jobs in the pipeline. .. _cmd-spack-ci-rebuild: ^^^^^^^^^^^^^^^^^^^^ ``spack ci rebuild`` ^^^^^^^^^^^^^^^^^^^^ The purpose of ``spack ci rebuild`` is to take an assigned spec and ensure a binary of a successful build exists on the target mirror. If the binary does not already exist, it is built from source and pushed to the mirror. The associated stand-alone tests are optionally run against the new build. Additionally, files for reproducing the build outside of the CI environment are created to facilitate debugging. If a binary for the spec does not exist on the target mirror, an install shell script, ``install.sh``, is created and saved in the current working directory. The script is run in a job to install the spec from source. The resulting binary package is pushed to the mirror. If ``cdash`` is configured for the environment, then the build results will be uploaded to the site. Environment variables and values in the ``ci::pipeline-gen`` section of the ``spack.yaml`` environment file provide inputs to this process. The two main sources of environment variables are variables written into ``.gitlab-ci.yml`` by ``spack ci generate`` and the GitLab CI runtime. Several key CI pipeline variables are described in :ref:`ci_environment_variables`. If the ``--tests`` option is provided, stand-alone tests are performed but only if the build was successful *and* the package does not appear in the list of ``broken-tests-packages``. A shell script, ``test.sh``, is created and run to perform the tests. On completion, test logs are exported as job artifacts for review and to facilitate debugging. If `cdash` is configured, test results are also uploaded to the site. A snippet from an example ``spack.yaml`` file illustrating use of this option *and* specification of a package with broken tests is given below. The inclusion of a spec for building ``gptune`` is not shown here. Note that ``--tests`` is passed to ``spack ci rebuild`` as part of the ``build-job`` script. .. code-block:: yaml ci: pipeline-gen: - build-job script: - . "./share/spack/setup-env.sh" - spack --version - cd ${SPACK_CONCRETE_ENV_DIR} - spack env activate --without-view . - spack config add "config:install_tree:projections:${SPACK_JOB_SPEC_PKG_NAME}:'morepadding/{architecture}/{compiler.name}-{compiler.version}/{name}-{version}-{hash}'" - mkdir -p ${SPACK_ARTIFACTS_ROOT}/user_data - if [[ -r /mnt/key/intermediate_ci_signing_key.gpg ]]; then spack gpg trust /mnt/key/intermediate_ci_signing_key.gpg; fi - if [[ -r /mnt/key/spack_public_key.gpg ]]; then spack gpg trust /mnt/key/spack_public_key.gpg; fi - spack -d ci rebuild --tests > >(tee ${SPACK_ARTIFACTS_ROOT}/user_data/pipeline_out.txt) 2> >(tee ${SPACK_ARTIFACTS_ROOT}/user_data/pipeline_err.txt >&2) broken-tests-packages: - gptune In this case, even if ``gptune`` is successfully built from source, the pipeline will *not* run its stand-alone tests since the package is listed under ``broken-tests-packages``. Spack's cloud pipelines provide actual, up-to-date examples of the CI/CD configuration and environment files used by Spack. You can find them under Spack's `stacks `_ repository directory. .. _cmd-spack-ci-rebuild-index: ^^^^^^^^^^^^^^^^^^^^^^^^^^ ``spack ci rebuild-index`` ^^^^^^^^^^^^^^^^^^^^^^^^^^ This is a convenience command to rebuild the buildcache index associated with the mirror in the active, gitlab-enabled environment (specifying the mirror url or name is not required). .. _cmd-spack-ci-reproduce-build: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ``spack ci reproduce-build`` ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Given the url to a gitlab pipeline rebuild job, downloads and unzips the artifacts into a local directory (which can be specified with the optional ``--working-dir`` argument), then finds the target job in the generated pipeline to extract details about how it was run. Assuming the job used a docker image, the command prints a ``docker run`` command line and some basic instructions on how to reproduce the build locally. Note that jobs failing in the pipeline will print messages giving the arguments you can pass to ``spack ci reproduce-build`` in order to reproduce a particular build locally. ------------------------------------ Job Types ------------------------------------ ^^^^^^^^^^^^^^^ Rebuild (build) ^^^^^^^^^^^^^^^ Rebuild jobs, denoted as ``build-job``'s in the ``pipeline-gen`` list, are jobs associated with concrete specs that have been marked for rebuild. By default a simple script for doing rebuild is generated, but may be modified as needed. The default script does three main steps, change directories to the pipelines concrete environment, activate the concrete environment, and run the ``spack ci rebuild`` command: .. code-block:: bash cd ${concrete_environment_dir} spack env activate --without-view . spack ci rebuild .. _rebuild_index: ^^^^^^^^^^^^^^^^^^^^^^ Update Index (reindex) ^^^^^^^^^^^^^^^^^^^^^^ By default, while a pipeline job may rebuild a package, create a buildcache entry, and push it to the mirror, it does not automatically re-generate the mirror's buildcache index afterward. Because the index is not needed by the default rebuild jobs in the pipeline, not updating the index at the end of each job avoids possible race conditions between simultaneous jobs, and it avoids the computational expense of regenerating the index. This potentially saves minutes per job, depending on the number of binary packages in the mirror. As a result, the default is that the mirror's buildcache index may not correctly reflect the mirror's contents at the end of a pipeline. To make sure the buildcache index is up to date at the end of your pipeline, spack generates a job to update the buildcache index of the target mirror at the end of each pipeline by default. You can disable this behavior by adding ``rebuild-index: False`` inside the ``ci`` section of your spack environment. Reindex jobs do not allow modifying the ``script`` attribute since it is automatically generated using the target mirror listed in the ``mirrors::mirror`` configuration. ^^^^^^^^^^^^^^^^^ Signing (signing) ^^^^^^^^^^^^^^^^^ This job is run after all of the rebuild jobs are completed and is intended to be used to sign the package binaries built by a protected CI run. Signing jobs are generated only if a signing job ``script`` is specified and the spack CI job type is protected. Note, if an ``any-job`` section contains a script, this will not implicitly create a ``signing`` job, a signing job may only exist if it is explicitly specified in the configuration with a ``script`` attribute. Specifying a signing job without a script does not create a signing job and the job configuration attributes will be ignored. Signing jobs are always assigned the runner tags ``aws``, ``protected``, and ``notary``. ^^^^^^^^^^^^^^^^^ Cleanup (cleanup) ^^^^^^^^^^^^^^^^^ When using ``temporary-storage-url-prefix`` the cleanup job will destroy the mirror created for the associated Gitlab pipeline. Cleanup jobs do not allow modifying the script, but do expect that the spack command is in the path and require a ``before_script`` to be specified that sources the ``setup-env.sh`` script. .. _noop_jobs: ^^^^^^^^^^^^ No Op (noop) ^^^^^^^^^^^^ If no specs in an environment need to be rebuilt during a given pipeline run (meaning all are already up to date on the mirror), a single successful job (a NO-OP) is still generated to avoid an empty pipeline (which GitLab considers to be an error). The ``noop-job*`` sections can be added to your ``spack.yaml`` where you can provide ``tags`` and ``image`` or ``variables`` for the generated NO-OP job. This section also supports providing ``before_script``, ``script``, and ``after_script``, in case you want to take some custom actions in the case of any empty pipeline. Following is an example of this section added to a ``spack.yaml``: .. code-block:: yaml spack: ci: pipeline-gen: - noop-job: tags: ['custom', 'tag'] image: name: 'some.image.registry/custom-image:latest' entrypoint: ['/bin/bash'] script:: - echo "Custom message in a custom script" The example above illustrates how you can provide the attributes used to run the NO-OP job in the case of an empty pipeline. The only field for the NO-OP job that might be generated for you is ``script``, but that will only happen if you do not provide one yourself. Notice in this example the ``script`` uses the ``::`` notation to prescribe override behavior. Without this, the ``echo`` command would have been prepended to the automatically generated script rather than replacing it. ------------------------------------ ci.yaml ------------------------------------ Here's an example of a spack configuration file describing a build pipeline: .. code-block:: yaml ci: target: gitlab rebuild_index: True broken-specs-url: https://broken.specs.url broken-tests-packages: - gptune pipeline-gen: - submapping: - match: - os=ubuntu18.04 build-job: tags: - spack-kube image: spack/ubuntu-bionic - match: - os=centos7 build-job: tags: - spack-kube image: spack/centos7 cdash: build-group: Release Testing url: https://cdash.spack.io project: Spack site: Spack AWS Gitlab Instance The ``ci`` config section is used to configure how the pipeline workload should be generated, mainly how the jobs for building specs should be assigned to the configured runners on your instance. The main section for configuring pipelines is ``pipeline-gen``, which is a list of job attribute sections that are merged, using the same rules as Spack configs (:ref:`config-scope-precedence`), from the bottom up. The order sections are applied is to be consistent with how spack orders scope precedence when merging lists. There are two main section types, ``-job`` sections and ``submapping`` sections. ^^^^^^^^^^^^^^^^^^^^^^ Job Attribute Sections ^^^^^^^^^^^^^^^^^^^^^^ Each type of job may have attributes added or removed via sections in the ``pipeline-gen`` list. Job type specific attributes may be specified using the keys ``-job`` to add attributes to all jobs of type ```` or ``-job-remove`` to remove attributes of type ````. Each section may only contain one type of job attribute specification, ie. , ``build-job`` and ``noop-job`` may not coexist but ``build-job`` and ``build-job-remove`` may. .. note:: The ``*-remove`` specifications are applied before the additive attribute specification. For example, in the case where both ``build-job`` and ``build-job-remove`` are listed in the same ``pipeline-gen`` section, the value will still exist in the merged build-job after applying the section. All of the attributes specified are forwarded to the generated CI jobs, however special treatment is applied to the attributes ``tags``, ``image``, ``variables``, ``script``, ``before_script``, and ``after_script`` as they are components recognized explicitly by the Spack CI generator. For the ``tags`` attribute, Spack will remove reserved tags (:ref:`reserved_tags`) from all jobs specified in the config. In some cases, such as for ``signing`` jobs, reserved tags will be added back based on the type of CI that is being run. Once a runner has been chosen to build a release spec, the ``build-job*`` sections provide information determining details of the job in the context of the runner. At lease one of the ``build-job*`` sections must contain a ``tags`` key, which is a list containing at least one tag used to select the runner from among the runners known to the gitlab instance. For Docker executor type runners, the ``image`` key is used to specify the Docker image used to build the release spec (and could also appear as a dictionary with a ``name`` specifying the image name, as well as an ``entrypoint`` to override whatever the default for that image is). For other types of runners the ``variables`` key will be useful to pass any information on to the runner that it needs to do its work (e.g. scheduler parameters, etc.). Any ``variables`` provided here will be added, verbatim, to each job. The ``build-job`` section also allows users to supply custom ``script``, ``before_script``, and ``after_script`` sections to be applied to every job scheduled on that runner. This allows users to do any custom preparation or cleanup tasks that fit their particular workflow, as well as completely customize the rebuilding of a spec if they so choose. Spack will not generate a ``before_script`` or ``after_script`` for jobs, but if you do not provide a custom ``script``, spack will generate one for you that assumes the concrete environment directory is located within your ``--artifacts_root`` (or if not provided, within your ``$CI_PROJECT_DIR``), activates that environment for you, and invokes ``spack ci rebuild``. Sections that specify scripts (``script``, ``before_script``, ``after_script``) are all read as lists of commands or lists of lists of commands. It is recommended to write scripts as lists of lists if scripts will be composed via merging. The default behavior of merging lists will remove duplicate commands and potentially apply unwanted reordering, whereas merging lists of lists will preserve the local ordering and never removes duplicate commands. When writing commands to the CI target script, all lists are expanded and flattened into a single list. ^^^^^^^^^^^^^^^^^^^ Submapping Sections ^^^^^^^^^^^^^^^^^^^ A special case of attribute specification is the ``submapping`` section which may be used to apply job attributes to build jobs based on the package spec associated with the rebuild job. Submapping is specified as a list of spec ``match`` lists associated with ``build-job``/``build-job-remove`` sections. There are two options for ``match_behavior``, either ``first`` or ``merge`` may be specified. In either case, the ``submapping`` list is processed from the bottom up, and then each ``match`` list is searched for a string that satisfies the check ``spec.satisfies({match_item})`` for each concrete spec. The the case of ``match_behavior: first``, the first ``match`` section in the list of ``submappings`` that contains a string that satisfies the spec will apply it's ``build-job*`` attributes to the rebuild job associated with that spec. This is the default behavior and will be the method if no ``match_behavior`` is specified. The the case of ``merge`` match, all of the ``match`` sections in the list of ``submappings`` that contain a string that satisfies the spec will have the associated ``build-job*`` attributes applied to the rebuild job associated with that spec. Again, the attributes will be merged starting from the bottom match going up to the top match. In the case that no match is found in a submapping section, no additional attributes will be applied. ^^^^^^^^^^^^^ Bootstrapping ^^^^^^^^^^^^^ The ``bootstrap`` section allows you to specify lists of specs from your ``definitions`` that should be staged ahead of the environment's ``specs``. At the moment the only viable use-case for bootstrapping is to install compilers. Here's an example of what bootstrapping some compilers might look like: .. code-block:: yaml spack: definitions: - compiler-pkgs: - 'llvm+clang@6.0.1 os=centos7' - 'gcc@6.5.0 os=centos7' - 'llvm+clang@6.0.1 os=ubuntu18.04' - 'gcc@6.5.0 os=ubuntu18.04' - pkgs: - readline@7.0 - compilers: - '%gcc@5.5.0' - '%gcc@6.5.0' - '%gcc@7.3.0' - '%clang@6.0.0' - '%clang@6.0.1' - oses: - os=ubuntu18.04 - os=centos7 specs: - matrix: - [$pkgs] - [$compilers] - [$oses] exclude: - '%gcc@7.3.0 os=centos7' - '%gcc@5.5.0 os=ubuntu18.04' ci: bootstrap: - name: compiler-pkgs compiler-agnostic: true pipeline-gen: # similar to the example higher up in this description ... The example above adds a list to the ``definitions`` called ``compiler-pkgs`` (you can add any number of these), which lists compiler packages that should be staged ahead of the full matrix of release specs (in this example, only readline). Then within the ``ci`` section, note the addition of a ``bootstrap`` section, which can contain a list of items, each referring to a list in the ``definitions`` section. These items can either be a dictionary or a string. If you supply a dictionary, it must have a name key whose value must match one of the lists in definitions and it can have a ``compiler-agnostic`` key whose value is a boolean. If you supply a string, then it needs to match one of the lists provided in ``definitions``. You can think of the bootstrap list as an ordered list of pipeline "phases" that will be staged before your actual release specs. While this introduces another layer of bottleneck in the pipeline (all jobs in all stages of one phase must complete before any jobs in the next phase can begin), it also means you are guaranteed your bootstrapped compilers will be available when you need them. The ``compiler-agnostic`` key can be provided with each item in the bootstrap list. It tells the ``spack ci generate`` command that any jobs staged from that particular list should have the compiler removed from the spec, so that any compiler available on the runner where the job is run can be used to build the package. When including a bootstrapping phase as in the example above, the result is that the bootstrapped compiler packages will be pushed to the binary mirror (and the local artifacts mirror) before the actual release specs are built. In this case, the jobs corresponding to subsequent release specs are configured to ``install_missing_compilers``, so that if spack is asked to install a package with a compiler it doesn't know about, it can be quickly installed from the binary mirror first. Since bootstrapping compilers is optional, those items can be left out of the environment/stack file, and in that case no bootstrapping will be done (only the specs will be staged for building) and the runners will be expected to already have all needed compilers installed and configured for spack to use. ^^^^^^^^^^^^^^^^^^^ Pipeline Buildcache ^^^^^^^^^^^^^^^^^^^ The ``enable-artifacts-buildcache`` key takes a boolean and determines whether the pipeline uses artifacts to store and pass along the buildcaches from one stage to the next (the default if you don't provide this option is ``False``). ^^^^^^^^^^^^^^^^ Broken Specs URL ^^^^^^^^^^^^^^^^ The optional ``broken-specs-url`` key tells Spack to check against a list of specs that are known to be currently broken in ``develop``. If any such specs are found, the ``spack ci generate`` command will fail with an error message informing the user what broken specs were encountered. This allows the pipeline to fail early and avoid wasting compute resources attempting to build packages that will not succeed. ^^^^^ CDash ^^^^^ The optional ``cdash`` section provides information that will be used by the ``spack ci generate`` command (invoked by ``spack ci start``) for reporting to CDash. All the jobs generated from this environment will belong to a "build group" within CDash that can be tracked over time. As the release progresses, this build group may have jobs added or removed. The url, project, and site are used to specify the CDash instance to which build results should be reported. Take a look at the `schema `_ for the ci section of the spack environment file, to see precisely what syntax is allowed there. .. _reserved_tags: ^^^^^^^^^^^^^ Reserved Tags ^^^^^^^^^^^^^ Spack has a subset of tags (``public``, ``protected``, and ``notary``) that it reserves for classifying runners that may require special permissions or access. The tags ``public`` and ``protected`` are used to distinguish between runners that use public permissions and runners with protected permissions. The ``notary`` tag is a special tag that is used to indicate runners that have access to the highly protected information used for signing binaries using the ``signing`` job. .. _staging_algorithm: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Summary of ``.gitlab-ci.yml`` generation algorithm ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ All specs yielded by the matrix (or all the specs in the environment) have their dependencies computed, and the entire resulting set of specs are staged together before being run through the ``ci/pipeline-gen`` entries, where each staged spec is assigned a runner. "Staging" is the name given to the process of figuring out in what order the specs should be built, taking into consideration Gitlab CI rules about jobs/stages. In the staging process the goal is to maximize the number of jobs in any stage of the pipeline, while ensuring that the jobs in any stage only depend on jobs in previous stages (since those jobs are guaranteed to have completed already). As a runner is determined for a job, the information in the merged ``any-job*`` and ``build-job*`` sections is used to populate various parts of the job description that will be used by the target CI pipelines. Once all the jobs have been assigned a runner, the ``.gitlab-ci.yml`` is written to disk. The short example provided above would result in the ``readline``, ``ncurses``, and ``pkgconf`` packages getting staged and built on the runner chosen by the ``spack-k8s`` tag. In this example, spack assumes the runner is a Docker executor type runner, and thus certain jobs will be run in the ``centos7`` container, and others in the ``ubuntu-18.04`` container. The resulting ``.gitlab-ci.yml`` will contain 6 jobs in three stages. Once the jobs have been generated, the presence of a ``SPACK_CDASH_AUTH_TOKEN`` environment variable during the ``spack ci generate`` command would result in all of the jobs being put in a build group on CDash called "Release Testing" (that group will be created if it didn't already exist). ------------------------------------- Using a custom spack in your pipeline ------------------------------------- If your runners will not have a version of spack ready to invoke, or if for some other reason you want to use a custom version of spack to run your pipelines, this section provides an example of how you could take advantage of user-provided pipeline scripts to accomplish this fairly simply. First, consider specifying the source and version of spack you want to use with variables, either written directly into your ``.gitlab-ci.yml``, or provided by CI variables defined in the gitlab UI or from some upstream pipeline. Let's say you choose the variable names ``SPACK_REPO`` and ``SPACK_REF`` to refer to the particular fork of spack and branch you want for running your pipeline. You can then refer to those in a custom shell script invoked both from your pipeline generation job and your rebuild jobs. Here's the ``generate-pipeline`` job from the top of this document, updated to clone and source a custom spack: .. code-block:: yaml generate-pipeline: tags: - before_script: - git clone ${SPACK_REPO} - pushd spack && git checkout ${SPACK_REF} && popd - . "./spack/share/spack/setup-env.sh" script: - spack env activate --without-view . - spack ci generate --check-index-only --artifacts-root "${CI_PROJECT_DIR}/jobs_scratch_dir" --output-file "${CI_PROJECT_DIR}/jobs_scratch_dir/pipeline.yml" after_script: - rm -rf ./spack artifacts: paths: - "${CI_PROJECT_DIR}/jobs_scratch_dir" That takes care of getting the desired version of spack when your pipeline is generated by ``spack ci generate``. You also want your generated rebuild jobs (all of them) to clone that version of spack, so next you would update your ``spack.yaml`` from above as follows: .. code-block:: yaml spack: ... ci: pipeline-gen: - build-job: tags: - spack-kube image: spack/ubuntu-bionic before_script: - git clone ${SPACK_REPO} - pushd spack && git checkout ${SPACK_REF} && popd - . "./spack/share/spack/setup-env.sh" script: - spack env activate --without-view ${SPACK_CONCRETE_ENV_DIR} - spack -d ci rebuild after_script: - rm -rf ./spack Now all of the generated rebuild jobs will use the same shell script to clone spack before running their actual workload. Now imagine you have long pipelines with many specs to be built, and you are pointing to a spack repository and branch that has a tendency to change frequently, such as the main repo and its ``develop`` branch. If each child job checks out the ``develop`` branch, that could result in some jobs running with one SHA of spack, while later jobs run with another. To help avoid this issue, the pipeline generation process saves global variables called ``SPACK_VERSION`` and ``SPACK_CHECKOUT_VERSION`` that capture the version of spack used to generate the pipeline. While the ``SPACK_VERSION`` variable simply contains the human-readable value produced by ``spack -V`` at pipeline generation time, the ``SPACK_CHECKOUT_VERSION`` variable can be used in a ``git checkout`` command to make sure all child jobs checkout the same version of spack used to generate the pipeline. To take advantage of this, you could simply replace ``git checkout ${SPACK_REF}`` in the example ``spack.yaml`` above with ``git checkout ${SPACK_CHECKOUT_VERSION}``. On the other hand, if you're pointing to a spack repository and branch under your control, there may be no benefit in using the captured ``SPACK_CHECKOUT_VERSION``, and you can instead just clone using the variables you define (``SPACK_REPO`` and ``SPACK_REF`` in the example above). .. _custom_workflow: --------------- Custom Workflow --------------- There are many ways to take advantage of spack CI pipelines to achieve custom workflows for building packages or other resources. One example of a custom pipelines workflow is the spack tutorial container `repo `_. This project uses GitHub (for source control), GitLab (for automated spack ci pipelines), and DockerHub automated builds to build Docker images (complete with fully populate binary mirror) used by instructors and participants of a spack tutorial. Take a look a the repo to see how it is accomplished using spack CI pipelines, and see the following markdown files at the root of the repository for descriptions and documentation describing the workflow: ``DESCRIPTION.md``, ``DOCKERHUB_SETUP.md``, ``GITLAB_SETUP.md``, and ``UPDATING.md``. .. _ci_environment_variables: -------------------------------------------------- Environment variables affecting pipeline operation -------------------------------------------------- Certain secrets and some other information should be provided to the pipeline infrastructure via environment variables, usually for reasons of security, but in some cases to support other pipeline use cases such as PR testing. The environment variables used by the pipeline infrastructure are described here. ^^^^^^^^^^^^^^^^^ AWS_ACCESS_KEY_ID ^^^^^^^^^^^^^^^^^ Optional. Only needed when binary mirror is an S3 bucket. ^^^^^^^^^^^^^^^^^^^^^ AWS_SECRET_ACCESS_KEY ^^^^^^^^^^^^^^^^^^^^^ Optional. Only needed when binary mirror is an S3 bucket. ^^^^^^^^^^^^^^^ S3_ENDPOINT_URL ^^^^^^^^^^^^^^^ Optional. Only needed when binary mirror is an S3 bucket that is *not* on AWS. ^^^^^^^^^^^^^^^^^ CDASH_AUTH_TOKEN ^^^^^^^^^^^^^^^^^ Optional. Only needed in order to report build groups to CDash. ^^^^^^^^^^^^^^^^^ SPACK_SIGNING_KEY ^^^^^^^^^^^^^^^^^ Optional. Only needed if you want ``spack ci rebuild`` to trust the key you store in this variable, in which case, it will subsequently be used to sign and verify binary packages (when installing or creating buildcaches). You could also have already trusted a key spack know about, or if no key is present anywhere, spack will install specs using ``--no-check-signature`` and create buildcaches using ``-u`` (for unsigned binaries).