mirror of
https://github.com/kubevela/kubevela.git
synced 2026-07-30 02:41:43 +00:00
3eaf62b2a2
* Refactor(helm): extract type declarations into types.go Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract Provider, constructors, and dry-run ctx into provider.go Move the Provider singleton, NewProvider/NewProviderWithConfig constructors, DefaultCacheTTLConfig, the dry-run context helpers (WithDryRun/isDryRun), and the small log/cache key utilities out of helm.go into a dedicated provider.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract chart fetching into chart_fetch.go Move detectChartSourceType, isMutableVersion, fetchChart (cache wrapper), fetchChartWithoutCache, determineCacheTTL, and the three per-source fetchers (fetchOCIChart, fetchURLChart, fetchRepoChart) out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract values resolution into values.go Move mergeValues, loadValuesFromSource, resolveValuesFromNamespace, loadConfigMapValues, loadSecretValues, plus the valueSourceMissingError sentinel and errCrossNamespaceValuesFrom guard out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract post-renderer and ownership helpers into postrender.go Move getActionConfig, velaLabelPostRenderer (struct + Run), velaOwnerLabels, and isOwnedByVela out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract manifest parsing into manifest.go Move parseManifestResources, isClusterScopedGVK, isClusterScopedKindStaticFallback, isTestResource, cleanResource, and orderResources out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract release secret lifecycle into release_secrets.go Move validateReleaseHealth, cleanOrphanedReleaseSecrets, listReleaseSecretNames, labelReleaseSecrets, deleteReleaseSecretsDirect, and InvalidateRelease out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract dry-run render into dryrun.go Move dryRunRender (used by webhook validation) and getKubeVersion (cluster Kubernetes version capabilities lookup) out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): extract release dispatcher into release.go Move installOrUpgradeChart (dispatcher between install/upgrade with fingerprint dedup, adoption detection, and publishVersion pin short-circuit) and computeReleaseFingerprint out of helm.go. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): split install/upgrade into install.go and upgrade.go with option appliers Move freshInstall, newInstallAction, and isRetryableInstallError into install.go. Extract the inline upgrade block from installOrUpgradeChart into a new performUpgrade function in upgrade.go. Introduce applyCommonInstallOptions and applyCommonUpgradeOptions free functions that own option-to-action wiring per Helm action type. The asymmetry between the two (MaxHistory/CleanupOnFail/Recreate live only in the upgrade applier) is now visible in code and reflects the Helm SDK shape. No behavior change. All 240 unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Chore(helm): remove Uninstall provider (unused) The helmchart ComponentDefinition only calls helm.#Render. Deletion is handled by KubeVela GC and ResourceTracker, never by helm uninstall, so the Uninstall CUE template and the workflow provider registration were exercised by tests only. Touched: UninstallParams/UninstallReturns in types.go, the Uninstall Go function and its registration in helm.go, the #Uninstall block in helm.cue, the uninstall registration in pkg/workflow/providers/helm/helm.go, and the corresponding helm_test.go cases. No behavior change for any in-tree caller. All 237 remaining unit specs remain green. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Fix(helm): wire includeCRDs and install-path force; drop waitTimeout (#7151) includeCRDs was a complete no-op: the field was declared on RenderOptionsParams but never read. Wire it through SkipCRDs on both action.Install and action.Upgrade (inverted, since the user-facing field expresses inclusion). force was wired on upgrade only. action.Install has a Force field in Helm SDK v3.14.4, so wire it there too. waitTimeout had no backing Helm SDK field (only Timeout exists, covering both connection and wait). Remove the field from RenderOptionsParams and the CUE schema. timeout now documents its dual role. maxHistory, cleanupOnFail, and recreatePods stay upgrade-only per the Helm SDK shape; CUE comments now state this explicitly. Tests: 18 new specs covering applyCommonInstallOptions and applyCommonUpgradeOptions (each option, nil opts, unparseable timeout, asymmetry of upgrade-only fields). Suite: 258 of 258 passing. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Refactor(helm): split helm_test.go alongside production files helm_test.go was a single 2500-line test file. Mirror the production file layout: types_test.go, provider_test.go, chart_fetch_test.go, values_test.go, postrender_test.go, manifest_test.go, release_test.go, release_secrets_test.go, dryrun_test.go, install_test.go (extended with isRetryableInstallError/newInstallAction/freshInstall), upgrade_test.go. helm_test.go now owns only Render-entry tests (matching helm.go after the production split). No test logic change. Suite remains 258 of 258 passing. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Docs(helm): add one-line responsibility comment to each provider file Every production .go file in the helm provider now opens with a single descriptive line immediately after the package declaration, so a reader can scan the directory and know what each file owns without opening it. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * chore(helm): fix the helmchart CD Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Test(helm): drop uninstall expectations from workflow provider tests CI revealed three failing assertions in pkg/workflow/providers/helm: GetTemplate was expected to contain 'uninstall', and GetProviders was expected to expose an uninstall key with a non-nil function. The Uninstall provider was removed earlier in this branch since the native helmchart ComponentDefinition never invoked it, but the workflow-provider tests still referenced it. Update them to assert the render-only surface. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Fix(helm): address cubic AI review on PR #7154 P0: release.go installOrUpgradeChart only treats driver.ErrReleaseNotFound as a missing release; other Get errors are surfaced rather than silently falling through to freshInstall. P1: release.go guards existingRelease.Info nil before reading Status. postrender.go isOwnedByVela now matches the full velaCtx triple (AppName + AppNamespace + Name) rather than any non-empty ownership label, so two Applications targeting the same release name in the same namespace do not mis-detect each other as already owned. P2: values.go deep-clones inline values before CoalesceTables so nested maps cannot be mutated through to the caller. dryrun.go uses the full server GitVersion when present so charts with patch-level kubeVersion constraints validate correctly. install.go isRetryableInstallError now catches 'cannot reuse a name that is still in use' and the dash-hyphen variant. upgrade.go applyCommonUpgradeOptions documents that helm v3 only consults SkipCRDs on the install-mode upgrade path. postrender.go Run no longer short-circuits when only the context is nil, so direct callers still get namespace defaulting. Tests: remove three non-hermetic / no-cluster tests that did not actually assert their named behaviour. values_test now uses distinct application and release namespaces in the cross-namespace guard scenario, and adds a second positive case for the release-namespace branch. chart_fetch_test fetchOCIChart-with-auth block now captures and restores singleton.KubeClient via BeforeEach/AfterEach to match the sibling fetchURLChart and fetchRepoChart blocks. Suite stays green at 255 specs. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Fix(helm): preserve numeric types in inline-values deep clone Cubic flagged the JSON round-trip in deepCloneValues for coercing int to float64. Switch to runtime.DeepCopyJSON which walks the same nested map/slice/scalar shape but preserves Go numeric types. Templates that depend on integer type assertions (e.g. {{ .Values.replicaCount | int }}) now see the original type instead of a coerced float. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Fix(helm): deep-clone inline values without runtime.DeepCopyJSON Cubic flagged that runtime.DeepCopyJSON panics on plain int values because the helper only accepts the narrow set of types json.Unmarshal returns. CUE evaluation routinely produces Go int / int32 / int64 values, so the previous switch from json round-tripping to runtime.DeepCopyJSON traded one bug (int->float64 coercion) for another (panic on int). Replace with a hand-rolled recursive walk that recreates maps and slices and copies scalars by interface assignment. Scalars are immutable in Go, so the interface copy is sufficient. The implementation accepts every type CUE evaluation can produce. Add five focused unit tests: nil input, int preservation, int64/float64 preservation, nested-map isolation, and slice isolation. Suite at 260 specs. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Test(helm): exercise installOrUpgradeChart with helm fake KubeClient Codecov patch coverage was failing because the file split moved installOrUpgradeChart into a new file (release.go) with no direct coverage. The previous unit tests were no-cluster smokes that did not assert anything meaningful and were rightly deleted earlier in this branch on cubic feedback. Add a small seam (actionConfigFactory field on Provider, defaulting to the real cluster getter) so tests can swap in helm SDK kubefake.PrintingKubeClient plus an in-memory release storage driver. Six new specs exercise the dispatcher end-to-end: fresh install, fingerprint dedup short-circuit, chart-version upgrade, values-change upgrade, adoption of vanilla helm releases, and publishVersion pin short-circuit. Package coverage climbs from ~71% to 82.6%. installOrUpgradeChart goes from 0% to 85.2%. performUpgrade goes from 0% to 90.9%. The remaining gaps (freshInstall retry path, release_secrets cluster helpers) require a real Kubernetes API and are exercised in the e2e suite. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Test(helm): plug the second cluster-leak seam in dispatcher tests Cubic flagged that the new dispatcher tests still let real Kubernetes calls escape: validateReleaseHealth (started as a goroutine after fingerprint-dedup short-circuit) called p.getActionConfig directly, and labelReleaseSecrets / listReleaseSecretNames / deleteReleaseSecretsDirect built a clientset from p.helmClient.RESTClientGetter(). Add a kubeClientFactory field on Provider next to actionConfigFactory and route all four helpers through it. installProviderWithFake injects a clientsetfake clientset alongside the fake action.Configuration. The release_secrets tests are reworked to use the helper as well, replacing the previous 'should not panic without a real cluster' smokes with assertions against the fake clientset: list filters to vela-owned secrets, label patches missing labels but never overwrites existing ones, delete cleans only matching helm-owned secrets. Verified the suite emits zero 'Kubernetes cluster unreachable' log lines now. 266 specs green. Also strengthened the adoption test to assert all three app.oam.dev ownership labels. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Fix(helm): restore corrupt-secret recovery via fall-through-on-any-Get-error The earlier cubic-driven P0 fix narrowed the dispatcher to fall through to freshInstall only when getAction.Run returned driver.ErrReleaseNotFound. That broke the documented self-healing contract validated by test/e2e-test/helmchart_test.go ('Helmchart Self-Healing / should recover from corrupted Helm release secret'): a corrupted .data.release surfaces from the helm storage driver as a decode error (base64 / gzip / json), not as driver.ErrReleaseNotFound, so the strict check trapped those errors and skipped the orphan-state cleanup path that freshInstall + isRetryableInstallError implement. Restore the original fall-through-on-any-error behaviour with two safety improvements: (1) non-NotFound errors are now logged at warning level so operators can spot RBAC / transient API failures that the recovery path then addresses (or fails to address); (2) the fingerprint slice in the stale-cache-clear log is bounds-checked so a short cached value cannot panic. Add a unit regression that wraps the in-memory storage driver in a corruptingDriver returning a synthetic decode error for the named release. Asserts the dispatcher swallows the error, runs fresh install, and repopulates the cache with the new fingerprint. Mirrors the e2e contract at the unit-test layer so a future refactor cannot regress this without breaking both gates. Suite: 267 of 267 passing. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Test(helm): seed corruption-recovery test at the canonical helm storage key Cubic flagged that the new corrupt-secret regression seeded the in-memory storage driver at 'rel.v1' while helm SDK derives canonical keys of the form 'sh.helm.release.v1.<name>.v<version>'. The mismatch meant the seeded release never collided with helms install-path probe and the test passed without actually exercising the recovery branch. Switch the seed key to the canonical form via the same makeKey shape helm SDK uses internally. The corruptingDriver layered on top still returns a synthetic decode error for Get/Query against the named release, so the dispatcher is forced through the swallow-and-fall-through path. Tightened the comment to be explicit about test scope: this unit pins the dispatcher-layer contract, the e2e test pins the end-to-end cleanup. Also relaxed the pre-seeded cache fingerprint to a realistic-length value so the truncation-safe log path exercises cleanly. Suite: 267 of 267 passing. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * Test(helm): drop canonical-key seed from corruption-recovery test The canonical-key seed made the in-memory storage collide with the key helm SDK derives for the release, which triggered the orphan-state cleanup retry path in freshInstall. Cleanup uses the kubeClientFactory clientset (a fake clientset with no helm release secrets) so it deletes nothing; the storage-driver entry stays put; the retry install then fails with release: already exists. CI confirmed against the previous commit. Seed only the corruptingDriver this time, no pre-existing release in storage. The dispatcher swallows the synthetic decode error from Get/Query and falls through to a clean fresh install, which is the contract this unit test is meant to pin. The full storage + clientset cleanup loop remains exercised by the e2e Helmchart Self-Healing scenario. Updated the test comment to be explicit about scope: storage and clientset are intentionally not bridged in the fake setup; that bridging belongs in the e2e test, not here. Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> * ci: retrigger checks Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com> --------- Signed-off-by: Ayush Kumar <ayushshyamkumar888@gmail.com>