9.1 KiB
HAN Chat production-like deployment runbook
This is the executable checklist for the single-VM contour. PostgreSQL and S3
are managed external services. Never use docker compose down -v, an Alembic
downgrade, or a mutable image tag during deployment.
Gate 0 — decisions and ownership
- Release SHA/digests, maintenance window, on-call and rollback owner recorded.
- RPO/RTO accepted; initial targets are PG RPO <=15 minutes and RTO <=4 hours.
- Remote OTLP backend selected, or debug-only acceptance limitation accepted.
- Mock OTP, Safety stub and bitrix-sync stub risks explicitly accepted.
Gate 1 — VPC, DNS and security groups
- Managed PostgreSQL has only a private endpoint and accepts traffic from VM SG.
- Internet can reach only VM TCP 80/443; SSH is restricted to VPN/ops CIDR.
- Ports 6379, 4317/4318, 8000, 8080 and 9000 are denied externally.
- DNS
AforPUBLIC_HOSTpoints at the VM and outbound HTTPS is available.
Gate 2 — VM hardening
On a fresh Ubuntu 24.04 VM, run:
sudo deployment/scripts/setup-vm.sh
Before setting HARDEN_SSH=true, verify key-based access in a separate SSH
session. The script header documents its parameters and safe defaults.
- Ubuntu 24.04, NTP, unattended security updates and disk alerts are active.
- Key-only deploy account works in a second session; root/password SSH is off.
- UFW/cloud SG and
DOCKER-USERpolicy survive reboot. - Docker Engine and Compose support
includeand long-formenv_file.
Gate 3 — managed PostgreSQL
- Daily backup, PITR, deletion protection, encryption and alerts are enabled.
- Provider CA is installed at
PG_CA_HOST_PATH; all DSNs useverify-full. - Schemas
han_app,bitrix_local,bitrix_sync,message_safety,keycloakhave separate migration/runtime roles with tested negative grants. - Migration tested against an empty DB and a clone of the previous release.
Gate 4 — Selectel S3
- Quarantine, attachments and documents buckets are private and encrypted.
- API credentials are prefix-scoped; Safety credentials are quarantine read-only.
- Browser CORS permits exact HTTPS origin and PUT headers only.
- Quarantine lifecycle exceeds Safety poll/recovery; data retention is approved.
Gate 5 — immutable release
- Checkout is detached at the approved SHA and working tree is clean.
- Service images are immutable and scanned; no unresolved critical/high issue.
- Root
docker-compose.ymlis the only deployment entry point.
Gate 6 — environment and secrets
umask 077
cp .env.example .env
chmod 600 .env
# Replace placeholders using a protected editor/secret manager.
./scripts/validate-env .env
docker compose --env-file .env config --quiet
- Token pairs match, PG verifies TLS, public URLs are HTTPS.
- Mock OTP risk is accepted and all secrets are unique >=128-bit values.
FRONTEND_DEV_PROXY_ENABLED=falseand Safety/nginx timeout budgets match.
Gate 7 — images and static frontend
docker compose --env-file .env pull
docker compose --env-file .env build --pull frontend-static nginx redis
docker compose --env-file .env run --rm frontend-static
- Frontend export was tested/scanned and copied by
frontend-staticinto its named volume. - Build artifacts contain no secrets or unintended source maps.
- At least 30% VM disk remains free.
Gate 8 — topology
docker compose --env-file .env config --services
python3 -m unittest discover -s tests -v
- Exactly nginx publishes
80:80and443:443; no PostgreSQL service exists. - Redis AOF/RDB/ACL and OTEL persistent queue volumes are present.
backendandobservabilityare internal networks.
Gate 9 — ACME/TLS bootstrap
Set NGINX_TLS_ENABLED=false only for this bootstrap command:
NGINX_TLS_ENABLED=false docker compose --env-file .env up -d nginx
docker compose --profile certbot run --rm certbot certonly \
--webroot -w /var/www/certbot -d "$PUBLIC_HOST" \
--cert-name "$PUBLIC_HOST" --email "$ACME_EMAIL" \
--agree-tos --no-eff-email --non-interactive
docker compose --env-file .env up -d --force-recreate nginx
docker compose exec -T nginx nginx -t -c /tmp/nginx.conf
First rehearse with Certbot --staging. Install a twice-daily systemd timer for
deployment/scripts/ssl-renew.sh; test certbot renew --dry-run. Enable HSTS
only after chain, hostname, redirect and TLS 1.2/1.3 checks pass.
Gate 10 — migrations and seed
Create a provider PITR marker, then:
PITR_MARKER_CONFIRMED=true deployment/scripts/migrate.sh
deployment/scripts/seed.sh
- Expected Alembic revisions are active and runtime users did not perform DDL.
- Seed succeeds twice and mandatory settings contain no secret.
- Schema remains backward-compatible with the previous images.
Gate 11 — Keycloak
docker compose up -d keycloak
docker compose ps keycloak
- Discovery/JWKS issuer is the exact public
/authHTTPS URL. - Frontend client is public PKCE S256; implicit/password/social flows are off.
- Wrong/replayed OTP and limits fail safely; settings bridge is fail-closed.
- Bootstrap admin was removed/rotated and named admin MFA is enabled.
Gate 12 — ordered startup and readiness
docker compose up -d redis
docker compose up -d keycloak otel-collector
docker compose up -d message-safety
docker compose up -d api-backend
docker compose up -d bitrix-local-app bitrix-sync
docker compose up -d nginx
docker compose ps
- No restart loop/OOM; critical readiness is green.
- Only documented Bitrix not-installed/sync-stub degradation remains.
- External
/internal/*is 404 and OTEL accepts telemetry.
Gate 13 — Bitrix24
- Install, handler and placement URLs use the exact public HTTPS paths.
- Connector
han_mobile_appis active on Open Line 8; events are bound once. - OAuth is encrypted; callback/application/service tokens never enter logs.
- Outbound and operator reply paths are idempotent; internal status is private.
Gate 14 — smoke and E2E
deployment/scripts/smoke.sh
- Guest, OTP/PKCE/bootstrap/session, refresh and logout paths pass.
- Safety allow/deny/pending/timeout and one concurrent slow poll pass.
- File quarantine/promote/delete, owner-only download and audit pass.
- WS reconnect plus REST reconciliation, ownership 404, idempotency and 429 pass.
- Logs contain no PII, message body, token or presigned query.
Gate 15 — observability
- Known request ID links nginx, API and downstream trace; UX ID is not a label.
- Three signals reach the selected backend; SLO queries and alerts are tested.
- Remote outage fills/drains the bounded persistent queue without business outage.
- Secret/PII canary is absent. Collector restart/drop/refused metrics are checked.
For local acceptance only, start the redacted debug collector with:
docker compose --profile observability-local up -d otel-collector-local.
Gate 16 — open traffic
- Gates 0–15 are signed; fresh backup/PITR evidence and previous images exist.
- HSTS is enabled, release digests/schema/realm versions are recorded.
- No active page; on-call and product owner accept stub limitations.
- Observe 5xx/auth/delivery/DB/Redis/OOM/OTEL queue/Bitrix for 60 minutes.
Backup and restore
Provider backup/PITR is authoritative. A supplemental verified logical dump:
PG_BACKUP_DSN='postgresql://...?...sslmode=verify-full&sslrootcert=...' \
deployment/scripts/backup.sh /opt/han-chat/backups
Quarterly, restore PG and S3 into an isolated VPC, deploy the same image digests, do not route production DNS/Bitrix callbacks, run smoke, and record measured RPO/RTO. Redis may be restored empty; its AOF/RDB is not a business backup.
Rollback
Only roll back to images compatible with the current schema:
SCHEMA_BACKWARD_COMPATIBLE_CONFIRMED=true \
deployment/scripts/rollback.sh /secure/path/previous-release.env
ENV_FILE=/secure/path/previous-release.env deployment/scripts/smoke.sh
Real SMS rollout addendum
This runbook remains mock-only until module-11 artifacts exist. An SMS release requires schema/role sms, versioned migrations and an active approved auth_otp seed, sms-service/worker, the exact callback route, paired service tokens, Direct TOKEN_1, approved sender/template, separate callback credentials, a reconfirmed callback source IP, and a static worker egress IP.
Order: App DB OTP seed → SMS schema/migrations/seed → mock Direct tests → production SMS deployment while Keycloak remains in mock mode → Keycloak expand migration/SPI → controlled provider smoke plus callback/redaction evidence → real mode. Roll back by restoring mock mode without deleting the journal/schema; stop new real orders and drain or record in-flight/uncertain rows. Downgrade only with proven schema compatibility.
Never run Alembic downgrade. After a backward-incompatible migration choose a forward fix or coordinated PITR/S3/Bitrix reconciliation under maintenance. Always verify outbox/inbox/recovery so an ambiguous message is not sent twice.