diff --git a/.github/release-template.md b/.github/release-template.md index c67dc086..fd72746e 100644 --- a/.github/release-template.md +++ b/.github/release-template.md @@ -96,6 +96,13 @@ repositories or docs.] ## Breaking changes / Migration notes - [Delete this section if there are no breaking changes.] +- **Upgrading from a pre-PR #523 cluster**: PR #523 relocated six `bedag/raw` + helmfile releases into the `base` chart. Existing clusters must run + `bash hack/migrate-bedag-raw-to-base.sh` once before `obol stack up` to + transfer Helm ownership annotations; otherwise `helm upgrade base` fails + with `invalid ownership metadata`. See + [`docs/upgrade-from-pre-pr-523.md`](../docs/upgrade-from-pre-pr-523.md). + Fresh installs are unaffected. ## Known issues diff --git a/docs/upgrade-from-pre-pr-523.md b/docs/upgrade-from-pre-pr-523.md new file mode 100644 index 00000000..1b9b5bf7 --- /dev/null +++ b/docs/upgrade-from-pre-pr-523.md @@ -0,0 +1,82 @@ +# Upgrading clusters created before PR #523 + +PR [#523](https://github.com/ObolNetwork/obol-stack/pull/523) relocates six +`bedag/raw` helmfile releases into the `base` chart so the stack has one +source of truth for everything it ships in the `erpc`, `obol-frontend`, and +`llm` namespaces. + +**Fresh installs are unaffected.** This page only applies if you are +upgrading a cluster that was created **before** PR #523 was merged. + +## Symptom + +Running `obol stack up` on a pre-#523 cluster fails during `helm upgrade base` +with errors of the form: + +``` +Error: UPGRADE FAILED: exists and cannot be imported into the +current release: invalid ownership metadata; annotation validation error: +key "meta.helm.sh/release-name" must equal "base"; current value is +"" +``` + +Helm refuses to "adopt" resources owned by another release. About ten +resources are affected (Namespaces, HTTPRoutes, Middlewares, ConfigMaps, +PrometheusRule, PodMonitor, ClusterRole/Binding) — enough that hand-fixing +them is error prone. + +## When to run the migration script + +- **Run once**, **before** `obol stack up`, against any cluster created + before PR #523 merged. +- The script is **idempotent** — safe to re-run if `obol stack up` is + interrupted or if you migrate one cluster at a time. +- Fresh clusters (`obol stack init && obol stack up` on an empty machine) + do **not** need it. + +```bash +# Optional: point at a non-default kubeconfig +export KUBECONFIG="$HOME/.config/obol/kubeconfig.yaml" + +bash hack/migrate-bedag-raw-to-base.sh +obol stack up +``` + +## What the script does + +It re-annotates the affected resources so Helm treats them as members of +the `base` release: + +``` +meta.helm.sh/release-name=base +meta.helm.sh/release-namespace=kube-system +app.kubernetes.io/managed-by=Helm +``` + +It covers the legacy `bedag/raw` releases removed by PR #523: + +| Legacy release | Namespace | +|---|---| +| `obol-frontend-rbac` | `obol-frontend` | +| `obol-frontend-httproute` | `obol-frontend` | +| `erpc-httproute` | `erpc` | +| `erpc-x402-middleware` | `erpc` | +| `erpc-metadata` | `erpc` | +| `llm-buyer-podmonitor` | `llm` | +| `x402-verifier-podmonitor` | `x402` (partial-upgrade clusters from before PR #513 hardening) | + +It also adopts a small set of resources that may exist with no Helm +ownership at all (`namespace/erpc`, `namespace/obol-frontend`, +`prometheusrule/x402-verifier` in `x402`) so the next `helm upgrade base` +can manage them cleanly. + +## Verifying the migration + +After running the script, `obol stack up` should succeed without the +`invalid ownership metadata` errors. To spot-check a single resource: + +```bash +kubectl get httproute -n obol-frontend obol-frontend \ + -o jsonpath='{.metadata.annotations.meta\.helm\.sh/release-name}{"\n"}' +# → base +``` diff --git a/hack/migrate-bedag-raw-to-base.sh b/hack/migrate-bedag-raw-to-base.sh new file mode 100755 index 00000000..d1c7d278 --- /dev/null +++ b/hack/migrate-bedag-raw-to-base.sh @@ -0,0 +1,82 @@ +#!/usr/bin/env bash +# Migrate resources from the legacy bedag/raw helmfile releases to the +# base chart that now owns them after obol-stack PR #523. +# +# Symptom this fixes: +# Error: UPGRADE FAILED: exists and cannot be imported +# into the current release: invalid ownership metadata +# +# Run once before `obol stack up` against any cluster deployed before +# PR #523 merged. +# +# Idempotent — safe to re-run. + +set -euo pipefail + +: "${KUBECONFIG:=$HOME/.config/obol/kubeconfig.yaml}" + +ORPHAN_RELEASES=( + obol-frontend-rbac + obol-frontend-httproute + erpc-httproute + erpc-x402-middleware + erpc-metadata + llm-buyer-podmonitor + x402-verifier-podmonitor # killed by PR #513's hardening; keep in case partial-upgrade clusters still have it +) + +migrate_one() { + local target="$1" + local current + current=$(kubectl get "$target" -o jsonpath='{.metadata.annotations.meta\.helm\.sh/release-name}' 2>/dev/null || true) + if [[ "$current" == "base" ]]; then + echo " $target: already on base, skipping" + return 0 + fi + if [[ -z "$current" ]]; then + echo " $target: no Helm metadata, adopting into base" + else + echo " $target: was on '$current', migrating to base" + fi + kubectl annotate "$target" \ + meta.helm.sh/release-name=base \ + meta.helm.sh/release-namespace=kube-system --overwrite >/dev/null + kubectl label "$target" app.kubernetes.io/managed-by=Helm --overwrite >/dev/null +} + +echo "==> Scanning for resources owned by legacy bedag/raw releases..." +for release in "${ORPHAN_RELEASES[@]}"; do + echo "release: $release" + kubectl get all,clusterrole,clusterrolebinding,role,rolebinding,configmap,httproute,middleware,podmonitor,servicemonitor,prometheusrule,referencegrant,namespace \ + -A -o json 2>/dev/null \ + | jq -r --arg rel "$release" '.items[] + | select(.metadata.annotations["meta.helm.sh/release-name"] == $rel) + | "\(.kind)/\(.metadata.name)\(if .metadata.namespace then " -n " + .metadata.namespace else "" end)"' \ + | while read -r target; do + [[ -z "$target" ]] && continue + migrate_one "$target" + done +done + +# Some resources were never Helm-owned (e.g. PrometheusRule x402-verifier may have +# been created via kubectl apply somewhere). Adopt them into base too if they exist +# in the namespaces base now owns. +echo "==> Adopting unowned resources base will now claim..." +declare -a UNOWNED_TARGETS=( + "namespace/erpc" + "namespace/obol-frontend" + "prometheusrule/x402-verifier -n x402" +) +for target in "${UNOWNED_TARGETS[@]}"; do + if kubectl get $target >/dev/null 2>&1; then + owner=$(kubectl get $target -o jsonpath='{.metadata.annotations.meta\.helm\.sh/release-name}' 2>/dev/null || true) + if [[ -z "$owner" || "$owner" == "base" ]]; then + echo " $target: $([ -z "$owner" ] && echo "adopting" || echo "already base")" + kubectl annotate $target meta.helm.sh/release-name=base meta.helm.sh/release-namespace=kube-system --overwrite >/dev/null + kubectl label $target app.kubernetes.io/managed-by=Helm --overwrite >/dev/null + fi + fi +done + +echo "" +echo "✓ Migration complete. You may now run 'obol stack up'." diff --git a/internal/agentcrd/agent.go b/internal/agentcrd/agent.go index 02ca080e..c9507926 100644 --- a/internal/agentcrd/agent.go +++ b/internal/agentcrd/agent.go @@ -32,7 +32,7 @@ func Namespace(name string) string { } // HostHomePath is where the agent's .hermes data lives on the host. The -// cluster mounts this into the Hermes pod via hostPath; writing +// cluster mounts this into the Hermes pod via the data PVC; writing // SOUL.md/skills here puts them inside the pod automatically. func HostHomePath(cfg *config.Config, name string) string { desc := agentruntime.Describe(agentruntime.Hermes) diff --git a/internal/hermes/hermes.go b/internal/hermes/hermes.go index 59e8c66a..5774330b 100644 --- a/internal/hermes/hermes.go +++ b/internal/hermes/hermes.go @@ -245,10 +245,7 @@ func Sync(cfg *config.Config, id string, u *ui.UI) error { return fmt.Errorf("helmfile sync failed: %w", err) } - // Host-side chown the PVC backing dirs to the in-pod UID/GID, bypassing - // the user-namespacing that defeats the in-pod `init-hermes-perms` - // chown from #446 (see ensureHermesPVCOwnership doc comment for details). - ensureHermesPVCOwnership(cfg, id, u) + fixHermesDataPVCK3dFallback(cfg, id, u) // Publish wallet-metadata ConfigMap for the frontend (namespace now exists). applyWalletMetadataConfigMap(cfg, id, deploymentDir) @@ -775,20 +772,8 @@ func generateValues(namespace, hostname, dashboardHostname, agentBaseURL, token, runAsUser: %d runAsGroup: %d fsGroup: %d + fsGroupChangePolicy: OnRootMismatch initContainers: - - name: init-hermes-perms - image: %s - imagePullPolicy: IfNotPresent - securityContext: - runAsUser: 0 - runAsGroup: 0 - command: - - sh - - -c - - chown -R %d:%d /data - volumeMounts: - - name: data - mountPath: /data - name: init-hermes-data image: %s imagePullPolicy: IfNotPresent @@ -862,7 +847,7 @@ func generateValues(namespace, hostname, dashboardHostname, agentBaseURL, token, value: %s - name: OBOL_SKILLS_DIR value: /data/.hermes/%s - `, desc.DataPVCName, namespace, desc.ServiceName, desc.ServiceName, namespace, desc.ServiceName, desc.ServiceName, desc.ServiceName, desc.ServiceName, containerUID, containerGID, containerGID, quoteYAML(image()), containerUID, containerGID, quoteYAML(image()), desc.ServiceName, quoteYAML(image()), quoteYAML(hermesBinary), desc.DefaultPort, desc.DefaultPort, quoteYAML(primary), quoteYAML(namespace), obolSkillsDirName) + `, desc.DataPVCName, namespace, desc.ServiceName, desc.ServiceName, namespace, desc.ServiceName, desc.ServiceName, desc.ServiceName, desc.ServiceName, containerUID, containerGID, containerGID, quoteYAML(image()), desc.ServiceName, quoteYAML(image()), quoteYAML(hermesBinary), desc.DefaultPort, desc.DefaultPort, quoteYAML(primary), quoteYAML(namespace), obolSkillsDirName) if agentBaseURL != "" { fmt.Fprintf(&b, " - name: AGENT_BASE_URL\n value: %s\n", quoteYAML(agentBaseURL)) @@ -1022,7 +1007,6 @@ func syncRuntimeFiles(cfg *config.Config, id string, configData []byte, u *ui.UI if err := removeLegacyHeartbeat(targetDir); err != nil { return err } - fixRuntimeVolumeOwnership(cfg, targetDir, u) return nil } @@ -1314,94 +1298,38 @@ func fixRuntimeVolumeOwnership(cfg *config.Config, hostPath string, u *ui.UI) { } } -// hermesPVCPaths returns the host-side PVC backing directories owned by the -// Hermes pod and chowned to containerUID:containerGID. -// -// Intentionally limited to PVCs that the Hermes container itself mounts — -// `remote-signer-keystores` is excluded even though it sits in the same -// namespace because the remote-signer pod runs as runAsUser=65532 with -// fsGroup=1000 (obol/remote-signer chart) and forcing its volume to -// 10000:10000 (Hermes' UID) makes the remote-signer crash-loop on -// `failed to load keystores: Permission denied (os error 13)` against -// the read-only /data/keystores mount. The local-path-provisioner default -// of 1000:1000 already matches that pod's fsGroup contract, so leaving -// that volume untouched is the safe behavior. -func hermesPVCPaths(cfg *config.Config, id string) []string { - namespace := agentruntime.Namespace(agentruntime.Hermes, id) - return []string{ - filepath.Join(cfg.DataDir, namespace, agentruntime.Describe(agentruntime.Hermes).DataPVCName), +// fsGroup should own Hermes' data volume. This fallback only repairs legacy +// k3d/userns clusters when the init container is already visibly stuck. +func fixHermesDataPVCK3dFallback(cfg *config.Config, id string, u *ui.UI) { + backendName := "k3d" + if data, err := os.ReadFile(filepath.Join(cfg.ConfigDir, ".stack-backend")); err == nil { + backendName = strings.TrimSpace(string(data)) + } + if backendName != "k3d" { + return } -} -// ensureHermesPVCOwnership host-side chowns the Hermes PVC backing directories -// to containerUID:containerGID so the agent's init containers can write under -// /data on the first start. -// -// Why this is needed (issue #475): -// - The embedded k3d config (internal/embed/k3d-config.yaml) sets -// KubeletInUserNamespace=true. Pod "root" maps to a host subuid that -// lacks chown authority over the host bind-mount path provisioned by -// local-path-provisioner. The in-pod `init-hermes-perms` chown added in -// #446 (commit c066baa) silently no-ops in this configuration. -// - local-path-provisioner's helper-pod sets the dir to 1000:1000 (see -// internal/embed/infrastructure/base/templates/local-path.yaml). Hermes -// runs as 10000:10000, so the next init container fails on -// `mkdir /data/.hermes/home: Permission denied`. -// -// The fix is to chown from outside the user namespace: `docker exec` into the -// k3d server container runs at the host Docker daemon's authority, which is -// real root and is not subject to the kubelet's user-namespacing. -// -// Best-effort. Waits up to 60s for each PVC to be Bound (local-path uses -// WaitForFirstConsumer, so the host dir doesn't exist until the consuming -// pod is scheduled). On non-k3d backends fixRuntimeVolumeOwnership falls -// back to a plain os.Chown. -// -// If a Hermes pod is currently stuck in Init:CrashLoopBackOff because of the -// pre-fix permissions, deletes it so kubelet re-creates with the corrected -// perms immediately rather than after exponential backoff (up to ~5 min). -// Skips the delete when no pod is stuck so repeated `Sync` calls -// (e.g. `obol model sync` after `obol model prefer`) do not gratuitously -// restart a healthy agent. -func ensureHermesPVCOwnership(cfg *config.Config, id string, u *ui.UI) { namespace := agentruntime.Namespace(agentruntime.Hermes, id) - kubeconfigPath := filepath.Join(cfg.ConfigDir, "kubeconfig.yaml") - kubectlBin := filepath.Join(cfg.BinDir, "kubectl") - - // Wait only for the PVCs hermesPVCPaths chowns. remote-signer-keystores - // is intentionally NOT in this loop — see the doc comment on - // hermesPVCPaths for why. - for _, pvc := range []string{ - agentruntime.Describe(agentruntime.Hermes).DataPVCName, - } { - waitCmd := exec.Command(kubectlBin, - "wait", "--for=jsonpath={.status.phase}=Bound", - "--timeout=60s", "pvc/"+pvc, "-n", namespace) - waitCmd.Env = append(os.Environ(), "KUBECONFIG="+kubeconfigPath) - _ = waitCmd.Run() // best-effort; continue even on timeout + if !hermesInitContainerStuck(cfg, namespace) { + return } - for _, p := range hermesPVCPaths(cfg, id) { - fixRuntimeVolumeOwnership(cfg, p, u) - } + hostPath := filepath.Join(cfg.DataDir, namespace, agentruntime.Describe(agentruntime.Hermes).DataPVCName) + fixRuntimeVolumeOwnership(cfg, hostPath, u) - if hermesInitStuck(cfg, namespace) { - deleteCmd := exec.Command(kubectlBin, - "-n", namespace, "delete", "pod", - "-l", "app.kubernetes.io/name=hermes", - "--ignore-not-found", "--wait=false") - deleteCmd.Env = append(os.Environ(), "KUBECONFIG="+kubeconfigPath) - if err := deleteCmd.Run(); err == nil && u != nil { - u.Info("Restarted Hermes pod to apply fresh volume ownership") - } + kubeconfigPath := filepath.Join(cfg.ConfigDir, "kubeconfig.yaml") + kubectlBin := filepath.Join(cfg.BinDir, "kubectl") + deleteCmd := exec.Command(kubectlBin, + "-n", namespace, "delete", "pod", + "-l", "app.kubernetes.io/name=hermes", + "--ignore-not-found", "--wait=false") + deleteCmd.Env = append(os.Environ(), "KUBECONFIG="+kubeconfigPath) + if err := deleteCmd.Run(); err == nil && u != nil { + u.Info("Restarted Hermes pod after best-effort k3d PVC ownership repair") } } -// hermesInitStuck reports whether at least one Hermes pod has an init -// container in CrashLoopBackOff or an Error waiting state — the signature of -// the perm-denied symptom this fix targets. Returns false on any kubectl -// failure so that a transient API hiccup does not trigger spurious restarts. -func hermesInitStuck(cfg *config.Config, namespace string) bool { +func hermesInitContainerStuck(cfg *config.Config, namespace string) bool { kubeconfigPath := filepath.Join(cfg.ConfigDir, "kubeconfig.yaml") kubectlBin := filepath.Join(cfg.BinDir, "kubectl") cmd := exec.Command(kubectlBin, diff --git a/internal/hermes/hermes_test.go b/internal/hermes/hermes_test.go index e18499a0..2fa09d1a 100644 --- a/internal/hermes/hermes_test.go +++ b/internal/hermes/hermes_test.go @@ -142,9 +142,8 @@ func TestGenerateValues_UsesHermesNativeNames(t *testing.T) { "/data/.hermes/obol-skills", "containerPort: 8642", "containerPort: 9119", - "init-hermes-perms", + "fsGroupChangePolicy: OnRootMismatch", "init-hermes-data", - "chown -R 10000:10000 /data", `Hermes binary missing from image: /opt/hermes/.venv/bin/hermes`, `Hermes image is missing required extras: web,messaging,mcp,pty,cli,acp,google`, `import fastapi, uvicorn, telegram, mcp, ptyprocess, simple_term_menu, googleapiclient`, @@ -166,9 +165,11 @@ func TestGenerateValues_UsesHermesNativeNames(t *testing.T) { "git clone", "uv pip install", "/data/.hermes/hermes-agent", + "init-hermes-perms", + "chown -R 10000:10000 /data", } { if strings.Contains(values, banned) { - t.Fatalf("generateValues() should not rebuild Hermes inside the PVC, found %q:\n%s", banned, values) + t.Fatalf("generateValues() contains banned fragment %q:\n%s", banned, values) } } @@ -351,48 +352,3 @@ func mkdirInstance(t *testing.T, cfg *config.Config, id string) { t.Fatalf("create Hermes instance %q: %v", id, err) } } - -// TestHermesPVCPaths pins the host-side directories that -// ensureHermesPVCOwnership chowns. Renaming the Hermes data PVC or -// relocating the namespace prefix in agentruntime without updating this -// list would silently regress the #475 fix on Linux k3d, because the chown -// would land on a non-existent path while the real PVC backing dir kept -// its local-path-provisioner default ownership of 1000:1000. -func TestHermesPVCPaths(t *testing.T) { - cfg := testConfig(t) - const id = "obol-agent" - namespace := agentruntime.Namespace(agentruntime.Hermes, id) - - got := hermesPVCPaths(cfg, id) - want := []string{ - filepath.Join(cfg.DataDir, namespace, "hermes-data"), - } - if !reflect.DeepEqual(got, want) { - t.Fatalf("hermesPVCPaths(%q) = %#v; want %#v", id, got, want) - } -} - -// TestHermesPVCPaths_ExcludesRemoteSignerKeystores pins the negative half of -// the contract: the helper MUST NOT include the remote-signer-keystores PVC -// path. The first revision of this fix included it, which chowned the -// volume to 10000:10000 (Hermes' UID) and broke the remote-signer pod — -// remote-signer runs as runAsUser=65532 with fsGroup=1000, expecting the -// local-path-provisioner default ownership of 1000:1000. The result was -// a remote-signer CrashLoopBackOff with -// -// failed to load keystores: Permission denied (os error 13) -// -// against /data/keystores. This guard makes that regression impossible to -// re-introduce by accident. -func TestHermesPVCPaths_ExcludesRemoteSignerKeystores(t *testing.T) { - cfg := testConfig(t) - const id = "obol-agent" - namespace := agentruntime.Namespace(agentruntime.Hermes, id) - - keystoreVolume := filepath.Join(cfg.DataDir, namespace, "remote-signer-keystores") - for _, p := range hermesPVCPaths(cfg, id) { - if p == keystoreVolume { - t.Fatalf("hermesPVCPaths included %q — the remote-signer pod (runAsUser=65532, fsGroup=1000) crash-loops when this volume is chowned to Hermes' containerUID:containerGID", keystoreVolume) - } - } -} diff --git a/internal/serviceoffercontroller/agent_render.go b/internal/serviceoffercontroller/agent_render.go index 0619ed77..e44874cc 100644 --- a/internal/serviceoffercontroller/agent_render.go +++ b/internal/serviceoffercontroller/agent_render.go @@ -316,9 +316,10 @@ func agentPodSpec(agent *monetizeapi.Agent) map[string]any { "serviceAccountName": hermesServiceName, "automountServiceAccountToken": true, "securityContext": map[string]any{ - "runAsUser": int64(hermesContainerUID), - "runAsGroup": int64(hermesContainerGID), - "fsGroup": int64(hermesContainerGID), + "runAsUser": int64(hermesContainerUID), + "runAsGroup": int64(hermesContainerGID), + "fsGroup": int64(hermesContainerGID), + "fsGroupChangePolicy": "OnRootMismatch", }, "initContainers": []any{ buildAgentProfileInitContainer(), diff --git a/internal/serviceoffercontroller/agent_render_test.go b/internal/serviceoffercontroller/agent_render_test.go index 29735b3d..e04e1e32 100644 --- a/internal/serviceoffercontroller/agent_render_test.go +++ b/internal/serviceoffercontroller/agent_render_test.go @@ -137,6 +137,43 @@ func TestAgentManifests_DeploymentEnvCarriesContext(t *testing.T) { } } +func TestAgentManifests_DeploymentUsesFSGroup(t *testing.T) { + agent := &monetizeapi.Agent{} + agent.Name = "quant" + agent.Namespace = "agent-quant" + agent.Spec = monetizeapi.AgentSpec{Model: "qwen3.5:9b"} + + out, err := agentManifests(agent, "litellm", "api") + if err != nil { + t.Fatalf("agentManifests: %v", err) + } + var dep map[string]any + for _, m := range out { + if m.GetKind() == "Deployment" { + dep = m.UnstructuredContent() + break + } + } + if dep == nil { + t.Fatal("Deployment manifest missing") + } + + podSpec := dep["spec"].(map[string]any)["template"].(map[string]any)["spec"].(map[string]any) + securityContext := podSpec["securityContext"].(map[string]any) + if securityContext["runAsUser"] != int64(hermesContainerUID) { + t.Fatalf("runAsUser = %v, want %d", securityContext["runAsUser"], hermesContainerUID) + } + if securityContext["runAsGroup"] != int64(hermesContainerGID) { + t.Fatalf("runAsGroup = %v, want %d", securityContext["runAsGroup"], hermesContainerGID) + } + if securityContext["fsGroup"] != int64(hermesContainerGID) { + t.Fatalf("fsGroup = %v, want %d", securityContext["fsGroup"], hermesContainerGID) + } + if securityContext["fsGroupChangePolicy"] != "OnRootMismatch" { + t.Fatalf("fsGroupChangePolicy = %v, want OnRootMismatch", securityContext["fsGroupChangePolicy"]) + } +} + func TestAgentManifests_ProfileSeedInitContainer(t *testing.T) { agent := &monetizeapi.Agent{} agent.Name = "quant"