| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
chore: consolidate initial sandbox follow-up fixes Signed-off-by: mhsong2 <songminhui2@huawei.com> | 13 天前 | |
ci: run sandbox release through K8S smoke The webhook-triggered sandbox pipeline needed to exercise the full release path instead of stopping at package/image creation. This keeps macOS and arm builds enabled by default, builds the sandbox images from the packaged wheel runtime contents, deploys to the configured K8S target, and runs the off-cluster smoke suite after deployment. The squashed change also keeps the release path practical for repeated CI runs: K8S image/test steps download only the wheel artifacts they need, package assembly reuses staged trees, cache paths are reusable and aged, and smoke coverage is bounded so scheduler pressure does not turn the validation step into a long hang. Constraint: The K8S runtime launcher is now supplied by the openyuanrong wheel and deploy defaults, so the node image must not extract or pass a standalone runtime-launcher socket path. Constraint: Buildkite webhook runs need macOS and arm enabled by default while still publishing only the required release artifacts for downstream K8S work. Constraint: Test K8S deploys use the fixed kubeconfig path supplied by the agent-stack podSpec patch under yuanrong-ci chart ownership. Constraint: TestPyPI publishing must be limited to openyuanrong_sdk wheels and must read its token from a Kubernetes Secret, never from checked-in values. Rejected: Port the broader historical PR601 build-script refactor | it introduced more variables and churn than needed before this pipeline was green. Rejected: Keep the six-ref wait exception smoke | it tested scheduler pressure and hit the outer 600s timeout instead of proving wait exception semantics. Confidence: high Scope-risk: moderate Directive: Do not reintroduce standalone runtime-launcher extraction into the node image without first checking the wheel packaging and deploy/process launcher contract. Tested: Buildkite #323 passed on 4969d982d56a; Test K8S passed with 25 passed and 4 skipped in 190.59s; Build X86 + Build Image improved from 46m40s in #316 to 35m42s in #323. Tested: python3 -m py_compile test/st/python/test_off_cluster.py Tested: bash -n test/st/run_off_cluster_test.sh .buildkite/test_sandbox_k8s.sh deploy/sandbox/k8s/bin/start-node.sh deploy/sandbox/k8s/bin/supervisord-node-entrypoint.sh .buildkite/package_sandbox_release.sh deploy/sandbox/k8s/build-images.sh scripts/package_yuanrong.sh Tested: python3 -m unittest deploy.sandbox.k8s.tests.test_yr_k8s_layout Tested: helm template yuanrong-ci deploy/helm/yuanrong-ci with and without secrets.targetKubeconfig.create=true Tested: bash .buildkite/pipeline.dynamic.yml rendered Test K8S without kubeconfig secret injection and with .buildkite/test_sandbox_k8s.sh Tested: bash .buildkite/pipeline.dynamic.yml rendered release artifact collection from output/*.whl only, without datasystem/functionsystem source whl recopy. Tested: rendered dynamic pipeline uploads only artifacts/release/openyuanrong_sdk*.whl to https://test.pypi.org/legacy/ from Build X86 when TEST_PYPI_API_TOKEN is present. Not-tested: Local direct kubectl access to guiyang target cluster from this shell, which timed out on TLS handshake after the pipeline had already passed. Signed-off-by: YuchaoWang <wyc984872769@hotmail.com> Signed-off-by: yuchaow <wangyuchao12@huawei.com> | 1 个月前 | |
ci: run sandbox release through K8S smoke The webhook-triggered sandbox pipeline needed to exercise the full release path instead of stopping at package/image creation. This keeps macOS and arm builds enabled by default, builds the sandbox images from the packaged wheel runtime contents, deploys to the configured K8S target, and runs the off-cluster smoke suite after deployment. The squashed change also keeps the release path practical for repeated CI runs: K8S image/test steps download only the wheel artifacts they need, package assembly reuses staged trees, cache paths are reusable and aged, and smoke coverage is bounded so scheduler pressure does not turn the validation step into a long hang. Constraint: The K8S runtime launcher is now supplied by the openyuanrong wheel and deploy defaults, so the node image must not extract or pass a standalone runtime-launcher socket path. Constraint: Buildkite webhook runs need macOS and arm enabled by default while still publishing only the required release artifacts for downstream K8S work. Constraint: Test K8S deploys use the fixed kubeconfig path supplied by the agent-stack podSpec patch under yuanrong-ci chart ownership. Constraint: TestPyPI publishing must be limited to openyuanrong_sdk wheels and must read its token from a Kubernetes Secret, never from checked-in values. Rejected: Port the broader historical PR601 build-script refactor | it introduced more variables and churn than needed before this pipeline was green. Rejected: Keep the six-ref wait exception smoke | it tested scheduler pressure and hit the outer 600s timeout instead of proving wait exception semantics. Confidence: high Scope-risk: moderate Directive: Do not reintroduce standalone runtime-launcher extraction into the node image without first checking the wheel packaging and deploy/process launcher contract. Tested: Buildkite #323 passed on 4969d982d56a; Test K8S passed with 25 passed and 4 skipped in 190.59s; Build X86 + Build Image improved from 46m40s in #316 to 35m42s in #323. Tested: python3 -m py_compile test/st/python/test_off_cluster.py Tested: bash -n test/st/run_off_cluster_test.sh .buildkite/test_sandbox_k8s.sh deploy/sandbox/k8s/bin/start-node.sh deploy/sandbox/k8s/bin/supervisord-node-entrypoint.sh .buildkite/package_sandbox_release.sh deploy/sandbox/k8s/build-images.sh scripts/package_yuanrong.sh Tested: python3 -m unittest deploy.sandbox.k8s.tests.test_yr_k8s_layout Tested: helm template yuanrong-ci deploy/helm/yuanrong-ci with and without secrets.targetKubeconfig.create=true Tested: bash .buildkite/pipeline.dynamic.yml rendered Test K8S without kubeconfig secret injection and with .buildkite/test_sandbox_k8s.sh Tested: bash .buildkite/pipeline.dynamic.yml rendered release artifact collection from output/*.whl only, without datasystem/functionsystem source whl recopy. Tested: rendered dynamic pipeline uploads only artifacts/release/openyuanrong_sdk*.whl to https://test.pypi.org/legacy/ from Build X86 when TEST_PYPI_API_TOKEN is present. Not-tested: Local direct kubectl access to guiyang target cluster from this shell, which timed out on TLS handshake after the pipeline had already passed. Signed-off-by: YuchaoWang <wyc984872769@hotmail.com> Signed-off-by: yuchaow <wangyuchao12@huawei.com> | 1 个月前 | |
test: cover pipeline smoke paths Expand the off-cluster smoke coverage for sandbox and yrcli flows while keeping CI defaults explicit enough for the current pipeline rollout. The yrcli deploy path now uses the configured tenant when uploading packages so the smoke cases exercise the same tenant namespace as runtime invocation. Constraint: Pipeline stability is still being validated, so build parallelism stays as simple explicit defaults instead of adaptive cgroup logic. Rejected: Keep adaptive build job detection | too much behavior surface while the pipeline is still being brought up. Confidence: medium Scope-risk: moderate Directive: Keep functionsystem parallelism explicit and environment-overridable unless the pipeline has enough evidence to justify auto-tuning. Tested: bash -n .buildkite/pipeline.dynamic.yml ci/ubuntu/compile-make-all.sh; git diff --check; bazel test //test:libruntime_test --test_filter=LibruntimeTest.InitAppliesOffClusterTenantIdToObjectStore --test_output=errors Not-tested: Full Buildkite pipeline after the final squash Signed-off-by: yuchaow <wangyuchao12@huawei.com> | 1 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 13 天前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 |