diff --git a/.gitea/workflows/deploy.yml b/.gitea/workflows/deploy.yml index 883e789..813a42a 100644 --- a/.gitea/workflows/deploy.yml +++ b/.gitea/workflows/deploy.yml @@ -1,10 +1,21 @@ -name: Deploy to Humanitec v2 +name: Deploy to Orchestrator on: workflow_run: workflows: ["Build and Push to ACR"] types: [completed] + # Must list the branches the build actually runs on. Without standard branching + # the build runs on main, and a list of dev/staging/prod would silently match + # nothing -- the deploy would never fire and the build would look successful. branches: [ "dev", "staging", "prod" ] + # The rollback path. Editing .platform/release.yaml is how a release is chosen, and + # build-push.yml deliberately ignores .platform/, so nothing else would trigger a + # deploy for it. This is what turns an approved rollback pull request into a running + # rollback the moment it merges. + push: + branches: [ "dev", "staging", "prod" ] + paths: + - '.platform/release.yaml' workflow_dispatch: inputs: environment: @@ -12,104 +23,257 @@ on: required: true default: 'dev' type: choice - options: [dev, staging, production] + options: + - dev + - staging + - prod env: - HUMANITEC_ORG: skillful-wild-chicken-2617 - HUMANITEC_AUTH_TOKEN: ${{ secrets.HUMANITEC_TOKEN }} + PO_API_URL: https://api.dev.orchestrator.crucible.kyndemo.live + PO_ORG_ID: crucible + PO_AUTH_TOKEN: ${{ secrets.PO_AUTH_TOKEN }} + # ONE ORCHESTRATOR PROJECT PER APPLICATION. + # + # This used to be the shared `apps-cluster` project with one environment per app, which + # made every app a peer of every other: the Orchestrator tab on any component listed the + # entire estate, and an app had exactly one environment named after itself, so there was + # nowhere for dev/staging/prod to live. + # + # That shape existed to avoid a Terraform pull request against config/projects.tf for every + # scaffolded app. That constraint turned out not to be real -- `octl create project`, + # `octl create runner-rule` and `octl create environment` are all runtime operations, so + # the workflow below builds the whole thing on first deploy and needs no repository change. PROJECT_ID: demo-a-3 - DEFAULT_ENV_ID: dev - ACR_REGISTRY: bstagecjotdevacr.azurecr.io + # The runner now comes from .platform/placement.yaml, through the guard job -- see "Read the placement decision". + # The runner a project's workloads execute on. This is NOT cosmetic: the Kubernetes and + # Helm providers are ambient, so a workload lands in whichever cluster its runner lives in, + # and Terraform state is keyed per runner. A project bound to the wrong runner deploys to + # the wrong cluster, and repointing it afterwards orphans the state it already owns. + OCTL_VERSION: 1.0.0 jobs: - deploy: - name: Deploy to Humanitec v2 + guard: + name: Placement guard runs-on: ubuntu-latest - if: github.event_name == 'workflow_dispatch' || github.event.workflow_run.conclusion == 'success' + outputs: + ready: ${{ steps.check.outputs.ready }} + runner: ${{ steps.placement.outputs.runner }} steps: - uses: actions/checkout@v4 - - name: Install dependencies - run: apt-get update -qq && apt-get install -y jq - - name: Install hctl CLI + with: + # The exact commit the deploy job ships, not a branch name. On a workflow_run the + # default checkout is the DEFAULT branch, and a branch name can move between this job + # and the deploy job -- the guard must judge the very commit the deploy will ship. + ref: ${{ github.event.workflow_run.head_sha || github.sha }} + + - name: Check platform initialized + id: check run: | - HCTL_VERSION=$(curl -s https://api.github.com/repos/humanitec/hctl/releases/latest | jq -r '.tag_name') - mkdir -p /tmp/hctl-install - curl -sLo /tmp/hctl-install/hctl.tar.gz "https://github.com/humanitec/hctl/releases/download/${HCTL_VERSION}/hctl_${HCTL_VERSION#v}_linux_amd64.tar.gz" - tar -xzf /tmp/hctl-install/hctl.tar.gz -C /tmp/hctl-install - install -m 755 /tmp/hctl-install/hctl /usr/local/bin/hctl - - name: Ensure Humanitec project and environment exist - env: - HUMANITEC_AUTH_TOKEN: ${{ secrets.HUMANITEC_TOKEN }} - run: | - DISPATCH_ENV="${{ github.event.inputs.environment }}" - ENV_ID="${DISPATCH_ENV:-$DEFAULT_ENV_ID}" - # Create project if it doesn't exist (hctl exits 0 if already exists) - hctl create project "$PROJECT_ID" --set display_name="demo-a-3" 2>&1 | grep -v "already exists" || true - # Create environment if it doesn't exist - hctl create environment "$PROJECT_ID" "$ENV_ID" --set env_type_id=development --set display_name="Development" 2>&1 | grep -v "already exists" || true - echo "✓ Project $PROJECT_ID / env $ENV_ID ready" - - name: Deploy with Score - env: - HUMANITEC_AUTH_TOKEN: ${{ secrets.HUMANITEC_TOKEN }} - run: | - DISPATCH_ENV="${{ github.event.inputs.environment }}" - ENV_ID="${DISPATCH_ENV:-$DEFAULT_ENV_ID}" - DEFAULT_IMAGE="$ACR_REGISTRY/demo-a-3:latest" - # Pre-flight: wait for any deployment from a prior run to finish before calling hctl. - # hctl refuses to start a new deployment while one is still executing. - echo "Pre-flight: checking for in-progress deployments..." - MAX_PREFLIGHT=420 - PREFLIGHT_WAITED=0 - while [ $PREFLIGHT_WAITED -lt $MAX_PREFLIGHT ]; do - PREFLIGHT_STATUS=$(curl -sf -H "Authorization: Bearer $HUMANITEC_AUTH_TOKEN" "https://api.humanitec.dev/orgs/$HUMANITEC_ORG/last-deployments?env_id=$ENV_ID&project_id=$PROJECT_ID&state_change_only=true" | jq -r '.items[0].status // "none"' 2>/dev/null || echo "none") - if [ "$PREFLIGHT_STATUS" != "in progress" ] && [ "$PREFLIGHT_STATUS" != "pending" ] && [ "$PREFLIGHT_STATUS" != "executing" ]; then - echo "Pre-flight passed (status=$PREFLIGHT_STATUS). Proceeding." - break - fi - echo " Prior deployment still running ($PREFLIGHT_WAITED s elapsed, status=$PREFLIGHT_STATUS)..." - sleep 15 - PREFLIGHT_WAITED=$((PREFLIGHT_WAITED + 15)) - done - # First deploy — provisions all resources. On a brand-new Humanitec project the - # dns-k8s-ingress Terraform module runs before the K8s Service exists, so the - # ingress backend port falls back to 3000. A second deploy (below) corrects it - # once the Service is up, which is essential for Java/Python apps on port 8080. - HCTL_EXIT=0 - timeout 300 hctl score deploy "$PROJECT_ID" "$ENV_ID" score.yaml --no-prompt --default-image "$DEFAULT_IMAGE" || HCTL_EXIT=$? - if [ "$HCTL_EXIT" -eq 0 ]; then - echo "✓ First deployment complete for demo-a-3 to $ENV_ID" - elif [ "$HCTL_EXIT" -eq 124 ]; then - echo "✓ First deployment submitted (polling timed out — waiting for K8s to settle)" + if [ -f ".platform/initialized.md" ]; then + echo "ready=true" >> $GITHUB_OUTPUT else - echo "✗ hctl failed with exit code $HCTL_EXIT" - exit $HCTL_EXIT + echo "ready=false" >> $GITHUB_OUTPUT + echo "Skipping: .platform/initialized.md not found" fi - # Poll Humanitec API until the first deployment is no longer in-progress before - # re-deploying. A flat sleep is unreliable — Terraform DNS modules can take 4-6 min. - echo "Waiting for first deployment to finish (polling Humanitec API)..." - MAX_WAIT=360 - WAITED=0 - while [ $WAITED -lt $MAX_WAIT ]; do - DEPLOY_STATUS=$(curl -sf -H "Authorization: Bearer $HUMANITEC_AUTH_TOKEN" "https://api.humanitec.dev/orgs/$HUMANITEC_ORG/last-deployments?env_id=$ENV_ID&project_id=$PROJECT_ID&state_change_only=true" | jq -r '.items[0].status // "unknown"' 2>/dev/null || echo "unknown") - if [ "$DEPLOY_STATUS" != "in progress" ] && [ "$DEPLOY_STATUS" != "pending" ] && [ "$DEPLOY_STATUS" != "executing" ]; then - echo "First deployment finished with status: $DEPLOY_STATUS" - break - fi - echo " Still running ($WAITED s elapsed, status=$DEPLOY_STATUS)..." - sleep 15 - WAITED=$((WAITED + 15)) - done - if [ $WAITED -ge $MAX_WAIT ]; then - echo "Warning: first deployment still running after $MAX_WAIT s — proceeding anyway" + + # The placement decision is enforced HERE, on every deploy, not only by the form that + # wrote it. Missing, refused or unreadable: this job fails, the deploy job that needs it + # never starts, and nothing reaches the orchestrator. REFUSED lines are plain echo as + # well as ::error:: because Gitea's log view may not render the annotation. + - name: Read the placement decision + id: placement + if: steps.check.outputs.ready == 'true' + run: | + set -euo pipefail + RECORD=.platform/placement.yaml + # Top-level keys, one line each -- the same contract release.yaml keeps. + field() { sed -n "s/^$1:[[:space:]]*//p" "$RECORD" | head -1 | sed -e 's/^"//' -e 's/"$//'; } + if [ ! -f "$RECORD" ]; then + echo "REFUSED: no placement decision. $RECORD is missing, and nothing deploys without an approved record." + echo "::error::No placement decision: $RECORD is missing." + exit 1 fi - # Second deploy — dns module now reads the real K8s Service port, fixing the ingress - HCTL_EXIT2=0 - timeout 120 hctl score deploy "$PROJECT_ID" "$ENV_ID" score.yaml --no-prompt --default-image "$DEFAULT_IMAGE" || HCTL_EXIT2=$? - if [ "$HCTL_EXIT2" -eq 0 ]; then - echo "✓ Deployment finalised for demo-a-3 to $ENV_ID" - elif [ "$HCTL_EXIT2" -eq 124 ]; then - echo "✓ Second deployment submitted for demo-a-3 to $ENV_ID (polling timed out)" + DECISION=$(field decision) + REASON=$(field reason) + RUNNER=$(field runner) + ZONE=$(field landing_zone_name) + case "$DECISION" in + placed) ;; + refused) + echo "REFUSED: $REASON" + echo "::error::Placement refused: $REASON" + exit 1 ;; + *) + echo "REFUSED: $RECORD has decision '$DECISION'; expected placed or refused." + echo "::error::Unreadable placement decision '$DECISION'." + exit 1 ;; + esac + case "$RUNNER" in + ''|none) + echo "REFUSED: $RECORD is placed but names no runner." + echo "::error::Placed record names no runner." + exit 1 ;; + esac + echo "PLACED: $ZONE -- runner $RUNNER" + echo "runner=$RUNNER" >> "$GITHUB_OUTPUT" + + deploy: + name: Deploy to Orchestrator + needs: guard + # Every trigger needs a runner from the placement record. workflow_dispatch used to skip + # the guard's verdict; an empty runner now means the guard did not place this workload. + if: >- + needs.guard.outputs.runner != '' && ( + (github.event_name == 'workflow_run' && github.event.workflow_run.conclusion == 'success') || + github.event_name == 'push' || + github.event_name == 'workflow_dispatch') + runs-on: ubuntu-latest + steps: + - name: Checkout code + uses: actions/checkout@v4 + with: + # The built commit, not a branch name (see the guard). On a workflow_run the default + # checkout is the DEFAULT branch, not the one that was built. That did not matter while + # the image tag came from an event field; it matters now that the tag is read out of a file in this checkout. + ref: ${{ github.event.workflow_run.head_sha || github.sha }} + + - name: Install octl + run: | + set -euo pipefail + curl -fsSLo /tmp/octl.tar.gz \ + "https://github.com/stellwerk-labs/platform-orchestrator-cli/releases/download/v${OCTL_VERSION}/platform-orchestrator-cli_${OCTL_VERSION}_linux_amd64.tar.gz" + tar xzf /tmp/octl.tar.gz -C /tmp + install -m 755 /tmp/octl /usr/local/bin/octl + octl --version + + - name: Derive environment + run: | + # The environment is now a STAGE of this application -- dev, staging, prod -- because + # the project is the application. It used to be the component id, which was the only + # option while every app shared one project and had to be distinguishable inside it. + DISPATCH_ENV="${{ github.event.inputs.environment }}" + if [ -n "$DISPATCH_ENV" ]; then + ENV_ID="$DISPATCH_ENV" else - echo "✗ Second hctl deploy failed with exit code $HCTL_EXIT2" - exit $HCTL_EXIT2 + # On a workflow_run the branch is the triggering run's, not this job's checkout. + BRANCH="${{ github.event.workflow_run.head_branch }}" + BRANCH="${BRANCH:-${GITHUB_REF_NAME}}" + case "$BRANCH" in + staging) ENV_ID=staging ;; + prod|main|master) ENV_ID=prod ;; + *) ENV_ID=dev ;; + esac + echo "Branch '$BRANCH' maps to environment '$ENV_ID'" fi + echo "ENV_ID=$ENV_ID" >> $GITHUB_ENV + echo "Deploying $PROJECT_ID to environment: $ENV_ID" + + - name: Ensure the project and its runner binding exist + env: + # From the placement record, through the guard -- never a literal in this file. + PO_RUNNER_ID: ${{ needs.guard.outputs.runner }} + run: | + set -euo pipefail + if [ -z "$PO_RUNNER_ID" ]; then + echo "::error::No runner from the placement guard; refusing to bind." + echo "REFUSED: No runner from the placement guard." + exit 1 + fi + # Created on first deploy rather than by a pull request against config/projects.tf. + # None of this needs a repository change: project, runner rule and environment are + # all runtime objects. + # + # Neither create is idempotent, so both fall through to a read on the second run. + octl create project "$PROJECT_ID" \ + --set display_name='demo-a-3' || \ + octl get project "$PROJECT_ID" + + # The runner rule routes this project's deployments. A project already bound to a + # DIFFERENT runner is refused, never repointed: Terraform state is keyed per runner, + # so rebinding orphans what it owns. A failing read is a failure, not "no rule" -- + # guessing "absent" is how a second rule gets created. + RULES=$(octl get runner-rules -o json) + # Several rules for one project, or a rule with no runner, is a binding nobody can + # vouch for: refuse it by name rather than pick the first. unique + join stays jq 1.6-safe. + if ! BOUND=$(printf '%s' "$RULES" | jq -r --arg p "$PROJECT_ID" '[.. | objects | select(.project_id? == $p) | (.runner_id // "")] | unique | if length > 1 then error("bound to several runners: \(join(", "))") else (.[0] // empty) end' 2>&1); then + echo "REFUSED: project '$PROJECT_ID' has an ambiguous runner binding: $BOUND" + echo "::error::Runner binding is ambiguous; refusing to continue." + exit 1 + fi + if [ -z "$BOUND" ]; then + octl create runner-rule \ + --set project_id="$PROJECT_ID" \ + --set runner_id="$PO_RUNNER_ID" \ + --no-prompt + elif [ "$BOUND" != "$PO_RUNNER_ID" ]; then + echo "REFUSED: project '$PROJECT_ID' is bound to runner '$BOUND', but its placement record names '$PO_RUNNER_ID'." + echo "::error::Runner binding disagrees with the placement record; refusing to repoint." + exit 1 + else + echo "Runner rule for '$PROJECT_ID' already binds '$BOUND', as the record says." + fi + + - name: Ensure the environment exists + run: | + set -euo pipefail + # Project and environment ids are POSITIONAL; only env_type_id and display_name go + # through --set. `dev` and `stable` are the only environment TYPES that exist, so + # staging rides on the dev type -- the type governs policy, the id governs identity. + case "$ENV_ID" in + prod) ENV_TYPE=stable ;; + *) ENV_TYPE=dev ;; + esac + octl create environment "$PROJECT_ID" "$ENV_ID" \ + --set env_type_id="$ENV_TYPE" \ + --set display_name="$ENV_ID" || \ + octl get environment "$PROJECT_ID" "$ENV_ID" + + - name: Deploy the Score workload + run: | + set -euo pipefail + # `octl score deploy` is ADDITIVE — it adds or updates a workload in the manifest and + # never removes one. That is the opposite of `octl deploy`, where omission is + # deletion. Do not substitute one for the other. + # No --show-logs: octl 1.0.0 has no such flag and exits 1 with `unknown flag` + # BEFORE contacting the orchestrator, so the whole deploy dies on an argument + # typo. Its nearest relatives are --runner-logs-level (default `info`, already + # what we want) and --skip-logs (which suppresses storage). Neither streams the + # runner's logs into this job, so there is nothing to substitute -- the runner + # logs are read from the orchestrator, not from here. + # The image comes from .platform/release.yaml, the one place a deployable tag + # is written down. It used to be derived from the triggering commit, which made + # "what is running" a question you answered by reading event metadata -- and + # left a rollback with no file to edit and no way to name an older image. + RELEASE_FILE=.platform/release.yaml + if [ ! -f "$RELEASE_FILE" ]; then + echo "::error::$RELEASE_FILE is missing -- there is nothing to deploy." + exit 1 + fi + IMAGE_REF=$(sed -n 's/^image:[[:space:]]*//p' "$RELEASE_FILE" | head -1 | tr -d "\"' ") + case "$IMAGE_REF" in + */*:*) ;; + *) echo "::error::$RELEASE_FILE must set image to /:, found '$IMAGE_REF'."; exit 1 ;; + esac + echo "IMAGE_REF=$IMAGE_REF" >> $GITHUB_ENV + echo "Deploying $IMAGE_REF" + + octl score deploy "$PROJECT_ID" "$ENV_ID" score.yaml \ + --default-image "$IMAGE_REF" \ + --no-prompt + + - name: Deployment summary + if: always() + run: | + # IMAGE_REF is unset when the deploy step failed before reading the release + # file, and this step runs on failure too. + DEPLOYED_IMAGE="${IMAGE_REF:-unresolved}" + SHORT_SHA="${{ github.event.workflow_run.head_sha || github.sha }}" + SHORT_SHA="${SHORT_SHA:0:7}" + echo "## Deployment Result" >> $GITHUB_STEP_SUMMARY + echo "| Field | Value |" >> $GITHUB_STEP_SUMMARY + echo "|---|---|" >> $GITHUB_STEP_SUMMARY + echo "| Project | \`$PROJECT_ID\` |" >> $GITHUB_STEP_SUMMARY + echo "| Environment | \`$ENV_ID\` |" >> $GITHUB_STEP_SUMMARY + echo "| Image | \`$DEPLOYED_IMAGE\` |" >> $GITHUB_STEP_SUMMARY + echo "| Triggering commit | \`$SHORT_SHA\` |" >> $GITHUB_STEP_SUMMARY + echo "[View in Orchestrator Console](https://console.dev.orchestrator.crucible.kyndemo.live/orgs/$PO_ORG_ID/projects/$PROJECT_ID/environments/$ENV_ID)" >> $GITHUB_STEP_SUMMARY