Files
slp-modular-cms/.gitea/workflows/deploy-scp.yaml
T
SluijsensandClaude Sonnet 5.5 df488663bd
Continuous Integration / config-sls (pull_request) Successful in 6s
Continuous Integration / changes (pull_request) Successful in 12s
Continuous Integration / config-asf (pull_request) Successful in 5s
Continuous Integration / compile-sites (pull_request) Successful in 1m12s
Continuous Integration / frontend-prepare (pull_request) Successful in 34s
Continuous Integration / vulnerability-scan (pull_request) Successful in 1m13s
Continuous Integration / frontend-build (pull_request) Successful in 41s
Continuous Integration / backend-build (pull_request) Successful in 1m11s
Continuous Integration / frontend-lint (pull_request) Successful in 41s
Continuous Integration / frontend-test (pull_request) Successful in 2m56s
Continuous Integration / backend-test (pull_request) Successful in 3m59s
Continuous Integration / publish-sls-test (pull_request) Skipped
Continuous Integration / publish-sls-production (pull_request) Skipped
Continuous Integration / deploy-sls-test (pull_request) Skipped
Continuous Integration / deploy-sls-production (pull_request) Skipped
Continuous Integration / publish-asf-test (pull_request) Skipped
Continuous Integration / deploy-asf-test (pull_request) Skipped
Continuous Integration / smoke-sls-test (pull_request) Skipped
Continuous Integration / smoke-asf-test (pull_request) Skipped
CI: nieuwe namen voor secrets en variabelen, SSH alleen met sleutel (geen sshpass), ASF-testdeploy zonder schakelaar
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
2026-11-18 20:22:31 +01:00

259 lines
14 KiB
YAML

name: Deploy (SCP)
# Reusable deploy workflow, invoked by continuous_integration.yaml (U5) for both the automatic
# test deploy and the opt-in production deploy. One workflow, one input contract, so a second
# transport (e.g. FTPS for a future shared-hosting target) can be added later as an alternative
# without restructuring this workflow or its callers (FR-02, D-02, NFR-09).
on:
workflow_call:
inputs:
host:
description: 'Which host to deploy to: the prefix of its secrets, one of SLS_TEST, SLS_PROD, ASF_TEST (<host>_HOST_ADDRESS, _PORT, _USERNAME, _SSH_KEY)'
required: true
type: string
artifact_name:
description: 'Build artifact to download and deploy'
required: true
type: string
environment:
description: 'Environment label, used in log output and step naming'
required: true
type: string
deploy_path:
description: 'Release root on the target host for this environment'
required: true
type: string
service_name:
description: 'systemd --user unit to restart after switching releases'
required: true
type: string
health_check_url:
description: 'Public /health URL to verify after restart'
required: true
type: string
run_db_backup:
description: 'Take a database backup before deploying (production only)'
required: false
type: boolean
default: false
transport:
description: 'Transport mechanism. Only "scp" is implemented today; reserved for FTPS later'
required: false
type: string
default: scp
jobs:
deploy:
runs-on: linux
steps:
# Runner-independent: only installs what is missing, and only uses sudo when not already root and sudo exists
# (a container on the PC runner has neither sudo nor a need for it; the Pi runner may need it).
#
# Authentication is by SSH key only (no password). The deploy key of the chosen host (<host>_HOST_SSH_KEY) is
# written to a temp directory with mode 600; the wrappers `rssh` / `rscp` below use it (-i, IdentitiesOnly,
# BatchMode, StrictHostKeyChecking=accept-new with a known_hosts file of this job only) and read the address,
# port and user of the chosen host from files in the same directory, so no other step needs the secrets. The
# directory is removed at the end of the job; the key is never printed.
#
# All twelve secrets are mapped here and the right group is chosen in the shell: an expression like
# `a && secrets.X || secrets.Y` would silently fall through to another host when a secret is empty.
- name: Ensure ssh tooling and authentication
env:
TARGET: ${{ inputs.host }}
SLS_TEST_HOST_ADDRESS: ${{ secrets.SLS_TEST_HOST_ADDRESS }}
SLS_TEST_HOST_PORT: ${{ secrets.SLS_TEST_HOST_PORT }}
SLS_TEST_HOST_USERNAME: ${{ secrets.SLS_TEST_HOST_USERNAME }}
SLS_TEST_HOST_SSH_KEY: ${{ secrets.SLS_TEST_HOST_SSH_KEY }}
SLS_PROD_HOST_ADDRESS: ${{ secrets.SLS_PROD_HOST_ADDRESS }}
SLS_PROD_HOST_PORT: ${{ secrets.SLS_PROD_HOST_PORT }}
SLS_PROD_HOST_USERNAME: ${{ secrets.SLS_PROD_HOST_USERNAME }}
SLS_PROD_HOST_SSH_KEY: ${{ secrets.SLS_PROD_HOST_SSH_KEY }}
ASF_TEST_HOST_ADDRESS: ${{ secrets.ASF_TEST_HOST_ADDRESS }}
ASF_TEST_HOST_PORT: ${{ secrets.ASF_TEST_HOST_PORT }}
ASF_TEST_HOST_USERNAME: ${{ secrets.ASF_TEST_HOST_USERNAME }}
ASF_TEST_HOST_SSH_KEY: ${{ secrets.ASF_TEST_HOST_SSH_KEY }}
IN_DEPLOY_PATH: ${{ inputs.deploy_path }}
IN_SERVICE_NAME: ${{ inputs.service_name }}
IN_HEALTH_URL: ${{ inputs.health_check_url }}
run: |
# Inputs are not secret: print them so an empty one is visible, and stop before anything touches the host.
echo "Inputs: host='$TARGET' deploy_path='$IN_DEPLOY_PATH' service_name='$IN_SERVICE_NAME' health_check_url='$IN_HEALTH_URL'"
for v in IN_DEPLOY_PATH IN_SERVICE_NAME IN_HEALTH_URL; do
if [ -z "${!v}" ]; then echo "Input $v is empty; refusing to deploy."; exit 1; fi
done
case "$TARGET" in
SLS_TEST|SLS_PROD|ASF_TEST) ;;
*) echo "Unknown host '$TARGET' (expected SLS_TEST, SLS_PROD or ASF_TEST); refusing to deploy."; exit 1;;
esac
SSH_HOST_VALUE=$(printenv "${TARGET}_HOST_ADDRESS")
SSH_PORT_VALUE=$(printenv "${TARGET}_HOST_PORT")
SSH_USER_VALUE=$(printenv "${TARGET}_HOST_USERNAME")
SSH_KEY_VALUE=$(printenv "${TARGET}_HOST_SSH_KEY")
for n in HOST_ADDRESS HOST_PORT HOST_USERNAME HOST_SSH_KEY; do
if [ -z "$(printenv "${TARGET}_${n}")" ]; then echo "The secret ${TARGET}_${n} is empty or not set; refusing to deploy."; exit 1; fi
done
SUDO=""
if [ "$(id -u)" != "0" ] && command -v sudo >/dev/null 2>&1; then SUDO="sudo"; fi
if ! command -v ssh >/dev/null 2>&1 || ! command -v scp >/dev/null 2>&1; then
$SUDO apt-get update && $SUDO apt-get install -y openssh-client
fi
D=$(mktemp -d)
chmod 700 "$D"
mkdir -p "$D/bin"
echo "DEPLOY_TMP=$D" >> "$GITHUB_ENV"
echo "$D/bin" >> "$GITHUB_PATH"
printf '%s\n' "$SSH_KEY_VALUE" > "$D/key"
chmod 600 "$D/key"
printf '%s' "$SSH_HOST_VALUE" > "$D/host"
printf '%s' "$SSH_PORT_VALUE" > "$D/port"
printf '%s' "$SSH_USER_VALUE" > "$D/user"
: > "$D/known_hosts"
cat > "$D/bin/rssh" <<'WRAP'
#!/bin/sh
D="$(dirname "$0")/.."
exec ssh -i "$D/key" -o IdentitiesOnly=yes -o BatchMode=yes -o StrictHostKeyChecking=accept-new -o UserKnownHostsFile="$D/known_hosts" \
-p "$(cat "$D/port")" "$(cat "$D/user")@$(cat "$D/host")" "$@"
WRAP
cat > "$D/bin/rscp" <<'WRAP'
#!/bin/sh
# The remote side of a copy is written as @REMOTE@:<path>; it is replaced by user@host.
D="$(dirname "$0")/.."
REMOTE="$(cat "$D/user")@$(cat "$D/host")"
for a in "$@"; do shift; set -- "$@" "$(printf '%s' "$a" | sed "s|^@REMOTE@:|$REMOTE:|")"; done
exec scp -i "$D/key" -o IdentitiesOnly=yes -o BatchMode=yes -o StrictHostKeyChecking=accept-new -o UserKnownHostsFile="$D/known_hosts" \
-P "$(cat "$D/port")" "$@"
WRAP
chmod +x "$D/bin/rssh" "$D/bin/rscp"
# Connection check: fails the job here, with a clear message, instead of in the first real step.
if "$D/bin/rssh" true; then echo "Authentication: deploy key accepted by ${TARGET}"; else echo "The deploy key of ${TARGET} was rejected or the host is unreachable."; exit 1; fi
- name: Compute release timestamp
id: release
run: echo "timestamp=$(date -u +%Y%m%d%H%M%S)" >> "$GITHUB_OUTPUT"
- name: Download build artifact
uses: actions/download-artifact@v3
with:
name: ${{ inputs.artifact_name }}
path: ${{ inputs.artifact_name }}
# Production-only database backup, taken before anything on the host changes (FR-20, D-26).
# The backup script itself lives on the host (created once during Operations host setup) and
# is only invoked here — no database connection string or credential is ever known to this
# workflow, keeping D-16 ("runtime secrets live in host environment variables") intact.
# Gitea Actions does not resolve `${{ inputs.* }}` inside a step's `name:` (unlike inside
# `run:`, where it works fine — see the echo below) — kept static rather than showing the
# literal, unresolved `${{ inputs.environment }}` text in the log.
- name: Back up database
if: ${{ inputs.run_db_backup }}
run: |
echo "Environment: ${{ inputs.environment }}"
rssh \
"bash ~/scripts/backup-slpsoftware-db.sh ${{ inputs.environment }}"
# Uploads the published output to a fresh, timestamped release directory rather than
# overwriting the live one (FR-06, D-27) — the atomic switch happens in a later step, once
# this upload and the website-symlink step below have both succeeded.
#
# Plain shell step rather than a container SCP action: container-based actions fail on this
# runner with "failed to attach to container: unable to upgrade to tcp, received 409", a known
# limitation of Podman's Docker-compatible API for the attach/log-streaming that container
# actions rely on (D-05). A plain scp command needs no nested container.
# Same Gitea Actions limitation as the step above — static name, value logged via echo instead.
- name: Upload release
run: |
echo "Environment: ${{ inputs.environment }}, transport: ${{ inputs.transport }}"
RELEASE_DIR="${{ inputs.deploy_path }}/releases/${{ steps.release.outputs.timestamp }}"
rssh \
"mkdir -p $RELEASE_DIR"
rscp \
-r ${{ inputs.artifact_name }}/* \
"@REMOTE@:$RELEASE_DIR/"
# The customer's public website (wwwroot/web) must survive every CMS deploy (FR-08, ASM-01).
# It lives outside the swapped release directory in a persistent shared/ folder, and is
# symlinked into each new release. mkdir -p is idempotent, so this is also correct on the very
# first-ever deploy, before any website workspace has published anything there (U1 already
# tolerates a missing wwwroot/web at startup). The publish output's own wwwroot/web (empty, or
# containing only the placeholder page) is removed before the symlink is created, so it never
# shadows the persistent content. `mkdir -p $RELEASE_DIR/wwwroot` guards against `ln -s` failing
# with "No such file or directory": dotnet publish only emits a wwwroot/ folder at all when the
# source project has one with actual content, so a release built before any frontend content
# exists can land with no wwwroot/ directory whatsoever, not merely an empty wwwroot/web/.
- name: Link persistent website content
run: |
RELEASE_DIR="${{ inputs.deploy_path }}/releases/${{ steps.release.outputs.timestamp }}"
rssh \
"mkdir -p ${{ inputs.deploy_path }}/shared/wwwroot-web && \
mkdir -p $RELEASE_DIR/wwwroot && \
rm -rf $RELEASE_DIR/wwwroot/web && \
ln -s ../../../shared/wwwroot-web $RELEASE_DIR/wwwroot/web"
# Optional plugin modules (future modular support): a compiled SlpModularCms.Modules.*.dll
# dropped into shared/modules survives every release, the same persistence pattern as
# shared/wwwroot-web above — except there's no cross-account permission dance here, since
# shared/modules lives entirely under gitea-workflow's own tree (unlike wwwroot-web, which
# reaches into webadmin's). ModuleOrchestrator.DiscoverModules() already scans its own base
# directory on disk for matching assemblies at startup (SlpModularCms.Core/Hosting/
# ModuleOrchestrator.cs) and loads whatever it finds via reflection — no code change needed,
# only somewhere for the DLL to still be after the next deploy replaces releases/{timestamp}.
# `ls ... 2>/dev/null` piped to a `while read` (rather than a bare glob loop) keeps this
# POSIX-sh safe and correct when shared/modules is empty, which is the common case today.
- name: Link persistent modules
run: |
RELEASE_DIR="${{ inputs.deploy_path }}/releases/${{ steps.release.outputs.timestamp }}"
rssh \
"mkdir -p ${{ inputs.deploy_path }}/shared/modules && \
ls ${{ inputs.deploy_path }}/shared/modules/*.dll 2>/dev/null | while read -r f; do \
name=\$(basename \"\$f\"); \
ln -sf \"../../shared/modules/\$name\" \"$RELEASE_DIR/\$name\"; \
done"
# Atomic release switch (FR-06, D-27): `ln -sfn` replaces the `current` symlink target in a
# single filesystem operation, so there is no moment where `current` points at a half-written
# directory. The process is then restarted so it picks up the new assemblies — a running .NET
# process holds on to the ones it already loaded.
- name: Switch current release and restart service
run: |
RELEASE_DIR="${{ inputs.deploy_path }}/releases/${{ steps.release.outputs.timestamp }}"
rssh \
"ln -sfn $RELEASE_DIR ${{ inputs.deploy_path }}/current && \
systemctl --user restart ${{ inputs.service_name }}"
# Verifies the restart actually produced a healthy process before this run is allowed to
# report success. Retries absorb process-startup time, which is about 10 seconds normally but
# 30 seconds or more when the host is busy (a migration step, two services restarting at once
# on one machine): up to 40 attempts, 3 seconds apart (about 2 minutes), each request limited
# to 5 seconds so one hanging attempt cannot eat the whole window. A run that never turns
# healthy fails the job without touching `current` or pruning (see below) - no automatic
# rollback. A previous release is always still on disk to restore from manually (D-26).
- name: Verify /health
run: |
attempts=40
for attempt in $(seq 1 $attempts); do
if curl -f -s --max-time 5 -o /dev/null "${{ inputs.health_check_url }}"; then
echo "Health check passed on attempt $attempt"
exit 0
fi
echo "Health check attempt $attempt of $attempts failed, retrying..."
sleep 3
done
echo "Health check did not pass after $attempts attempts"
exit 1
# Retention: keep `current` plus exactly one previous release, so a manual rollback is always
# a re-point-and-restart away without rebuilding. Only runs after a passing health check —
# pruning after a failed check could leave the only other release as the sole survivor.
- name: Prune old releases
run: |
rssh \
"cd ${{ inputs.deploy_path }}/releases && ls -1t | tail -n +3 | xargs -r rm -rf"
- name: Remove deploy key
if: always()
run: rm -rf "${DEPLOY_TMP:-/nonexistent}"