11 KiB
Reduce time to publish packages
https://blueprints.launchpad.net/fuel/+spec/reduce-time-to-publish-packages
Problem description
- As a Package Maintainer, I want my package updates to be published to official package repositories immediately after my code changes are merged in git
- At the moment, all changes to the MOS packages repositories are tested on a set of simple functional and integration tests so that they will not be merged into the stable package repository unless it passes all of the configured tests. Merged packages are then pushed to staging mirrors, then staging mirrors are validated using BVT system tests using most recent stable build of Fuel, then staging mirrors are published. Published mirrors are used to build yet another ISO image which is validated again via BVT system tests. Then the resulting mirror and ISO are passed to the swarm tests. This results in a long chain of processes maintained by different teams, and thus creates quite a long feedback loop.
- Different test pipelines for different types of commits made to MOS repositories and StackForge repositories create inconsistency in test coverage between artifacts created from these sources.
- Duplication of artifacts of the same type created in different test pipelines creates waste of computing capacity of CI infrastructure.
Proposed change
We must complete all package validation (including system tests) before package change patch set is merged, and then publish packages to the mirrors immediately after review is merged.
Proposed flow for change requests to MOS packages code and/or spec is below:
- Change request is proposed for code review to Gerrit
- Change request parameters (repository name, branch, change ID) are passed by Zuul to the Jenkins build job that fetches sources and specs from Gerrit and pushes sources and specs to the build system in a respective format (SRPM for RPM packages, and unpacked sources for DEB packages)
- On success build job triggers in series publish job, and then test jobs -install and systest. If any of is jobs failed, next jobs are not run and build job fails
- publish job moves package (or packages), built from resulting sources and specs by the build system, into a test repository. Test repository location is created as a Jenkins job artifact
- Test repository is run through install test - all built packages are installed and removed one by one
- Test repository is run through systest test - built packages are used to deploy some set of servers.
- Change request is peer-reviewed
- Change request is accepted or rejected by core team
- Change request is rebased and run through pre-merge automated checks -gate job1. Because gate job should check several change requests at once, a set of build jobs are triggered with parameters preventing tests, so each build job triggers only publish job
- If all build jobs are succeed then gate job triggers systest job using all created repositories at once to check together packages built by several jobs
- Change request is merged or rejected
- Merge event from Gerrit is picked up by Zuul to run Jenkins build job that builds final package and publish it to released repository. publish job detects if change request is merged, selects appropriate repository, puts packages to selected repository, and signes them (for RPM - a package itself, for DEB - repository metadata)
- Repository with rebuilt package is propagated to the external MOS mirrors by the means of Jenkins job that uses transactional syncronization
Test repository format
Test repository with packages should be passed to the system tests as parameter using the following conventions:
- RPM packages. Plain RPM repository accessible by HTTP.
- DEB packages. Single-component APT repository accessible by HTTP, with distribution name that equals to the current upstream OS codename.
Automated functional and integration tests
Each package built into a test repository will undergo the following tests:
- Basic syntax test
-
rpmlint and lintian, for RPM and DEB packages, respectively
- Basic package health test - install job
-
Uses plain VM with respective OS type that contains only the minimal set of packages. Package install, run basic shell script, and uninstall tests are performed.
- Full system test - systest job
-
All packages are tested using the same scenarios that are used in current BVT tests on the Product CI. In order to keep the actual test run as short and reliable as possible, the VM snapshots for Fuel master node and cluster nodes are prepared ahead of time and kept on slave servers ready for immediate use. The process of preparing the VM snapshots ahead of time reduces the network traffic and influence of an external depencencies during the run. The job that prepares the VM snapshots is executed on a daily basis or, alternatively, triggered on ISO update, depending on the frequency if ISO updates.
Depending on the package flavor, there are different artifacts that could be supplied to the integration test job.
- RPM and DEB packages for clusters. Test repository is added to the Nailgun with highest priority (in terms of a respective package manager).
- Fuel Master host OS packages. Rebuilding of a Fuel ISO with an extra test repository and redeployment of Fuel Master node must precede the system test.
- Fuel bootstrap packages. Rebuilding of bootstrap package and installing it on the Fuel Master node created from snapshot prepared ahead precedes the system test.
- Fuel Docker container packages. Rebuilding of Fuel Docker containers late-package and installing it on the Fuel Master node created from snapshot prepared ahead precedes the system test.
All tests will be running in parallel. In case if basic package health test for a given patch set is failed, the respective full system test must be aborted to prevent the waste of a CI system resources.
Full logs for a system test run must be available in the Jenkins as a job artifact. In order to support debugging of system test issues, snapshots of VMs created during the test run are stored on Jenkins slaves for 20 days.
In order to support enhanced debug cases, there should be an ability to replicate all the test parameters, including job artifacts, on a Custom CI.
After merge event in Gerrit, package built on a gate will be pushed to a respective stable repository on an internal OSCI mirror and signed in terms of a respective package manager. Right after that, this resulting repository will be synced to the external OSCI mirrors without any further verification. RPM and DEB mirrors are synced independently, list of synced packages must be included as an artifact to an appropriate Jenkins job.
Testing of concurrent requests
To ensure that no regressions are introduced by merging of concurrent requests, on merge event, every outstanding request must be speculatively retested against the tip of the package code and specs branches exactly as it is going to be merged.
To optimize handling of merge queue we are going to use Zuul as a gating tool.2
Auxiliary jobs
The following auxiliary jobs must be supported on the Jenkins as the part of this spec:
- Job to prepare environment for system tests on the Jenkins slave servers
- Jobs to support debug sessions workflow
- Job to clean expired VM snapshots on Jenkins slaves
Alternatives
As an alternative to Zuul we considered managing pre-merge checks via set of custom scripts, but it appears that we need a lot of Zuul functionality and in long-term we will benefit from reusing the existent tooling, which is supported by wide community including OpenStack Infra team.
Data model impact
None
REST API impact
None
Upgrade impact
None
Security impact
None
Notifications impact
None
Other end user impact
None
Performance Impact
None
Plugin impact
None
Other deployer impact
None
Developer impact
Due to more complex scenarios used in deploy tests configuration, time to complete all CI jobs for a single commit will increase approx. by 60%-120%. A speed of publishing of packages decreased from hours to minutes, should compensate this inconvenience, though.
Infrastructure impact
- Changes in system tests will require additional servers to be used as Jenkins slaves.
- Logs of a system tests will be stored on the Jenkins Master, therefore we will need to increase the storage capacity for this server (at least 1 TB is required). Alternatively, this requirement could be addressed by the centralized diagnostic snapshot storage feature.
- Changes described in this document will only apply to 7.0+ CI environments, no backporting for older releases is planned.
Implementation
Assignee(s)
- Primary assignee:
-
Vitaly Parakhin <vparakhin@mirantis.com>
- DevOps:
-
Andrey Nikitin <anikitin@mirantis.com> Dmitry Kaiharodtsev <dkaiharodsev@mirantis.com>
- QA:
-
<TBD>
- Mandatory Design Reviewers:
-
Roman Vyalov <rvyalov@mirantis.com> Aleksandra Fedorova <afedorova@mirantis.com>
Work Items
- Setup new Jenkins instance for packaging and tests
- Setup Zuul and Gearman instance and attach it to Jenkins
- Setup new Jenkins slaves for system test jobs3
- Implement Jenkins job for transactional syncing of mirrors4
- Implement Jenkins job that retriggers CI for outstanding requests on merge of a concurrent request
- Adapt system test Jenkins job to support different types of artifacts
- Switch packaging and tests for MOS 7.0 to new Jenkins
- Eliminate the staging mirrors phase on the MOS 7.0 Product CI
- Replace all packages sources of MOS 7.0 components in Gerrit with fully unpacked source code (get rid of orig.tar.gz etc)
Dependencies
Testing
Acceptance Criteria:
- Each package source code and build script change patch set is validated with the same level of test coverage that is currently used for package staging mirrors.
- All validation is done on patch sets before the change is merged; once the change is merged, package is propagated to official package repositories without additional validation.
- There are no binary packages in any of the MOS 7.0 package repositories that were not built by the new build system from master or 7.0 specific branches in git repositories with build scripts and fully unpacked source code (no orig.tar.gz files in git).
- Individual package and test jobs can be retriggered one at a time.
Documentation Impact
None