Move an app off Heroku, step by step

The Heroku migration guide maps the concepts. This tutorial walks one concrete app through the move: a Node/Express API called acme-api with a web and a worker dyno, Heroku Postgres, Heroku Redis, two Scheduler jobs, 30 config vars and api.example.com. Adjust the commands to your stack — the language pages have the Rails, Django and Laravel equivalents.

You need: the Heroku CLI still logged in, the code on GitHub, skiffly (npm i -g skiffly && skiffly login), pg_restore and redis-cli locally. Plan for 1–2 hours including a maintenance window of a few minutes for the database copy.

0. Inventory#

heroku ps -a acme-api                       # dynos: web.1, worker.1
cat Procfile                                # web: node dist/server.js  /  worker: node dist/worker.js
heroku addons -a acme-api                   # heroku-postgresql, heroku-redis, scheduler
heroku config -s -a acme-api > heroku.env   # KEY=value, one per line
heroku domains -a acme-api                  # api.example.com

Everything in Procfile becomes a service; every add-on becomes a template; release: (if you have one) becomes part of the start command.

1. Project and databases on Skiffly#

cd acme-api
skiffly init --name acme-api
skiffly deploy --template postgres      # "Postgres", 10 GB volume — size it with the dashboard if your DB is bigger
skiffly deploy --template redis
skiffly status                          # both SUCCESS

Expected: Postgres and Redis running on the private network with DATABASE_URL and REDIS_URL variables of their own.

2. The web service#

skiffly up -y -d                        # creates service "acme-api", first build starts

The build succeeds but the app crashes without its variables — expected for now. Set them:

# drop the add-on URLs Heroku managed, keep the rest
grep -v -E '^(DATABASE_URL|REDIS_URL|REDIS_TLS_URL|HEROKU_|PORT=)' heroku.env > vars.env
skiffly variables set -s acme-api --skip-deploys $(cat vars.env | xargs)
skiffly variables set -s acme-api --skip-deploys 'DATABASE_URL=${{Postgres.DATABASE_URL}}' 'REDIS_URL=${{Redis.REDIS_URL}}'

Quoting: xargs breaks on values with spaces or #; set those individually (skiffly variables set KEY="a value") or use the dashboard's raw editor, which accepts the .env text as is.

Now the start command. Heroku ran web: node dist/server.js; Railpack would pick npm start, which is usually the same thing. If you had a release phase (release: npm run migrate), fold it in:

• Settings → Deploy → Start command: sh -c "npm run migrate && node dist/server.js"
• Healthcheck path: /healthz (add the route if the app has none; Heroku did not need it, Skiffly's zero-downtime rollouts do).

skiffly redeploy -s acme-api
skiffly logs -s acme-api -f
skiffly domain -s acme-api             # https://acme-api-x1y2.skiffly.cloud

Expected: SUCCESS, and the generated URL answers. PORT is set by Skiffly exactly as Heroku did; heroku local-style .env files are not read — everything is variables.

Differences to check now, before the data move:

3. The worker#

Create a second service from the same repository — dashboard New service → GitHub repo (same repo, name worker), or with config as code (step 7). Settings:

• Start command: sh -c 'node -e "require(\"http\").createServer((_,r)=>r.end(\"ok\")).listen(process.env.PORT)" & exec node dist/worker.js' — Skiffly probes every long-running service on its port; the one-liner keeps the rollout healthy for a worker that has no HTTP server of its own (see Workers).
• No domain.
• Variables: the same list. Use references so they are not duplicated: DATABASE_URL=${{Postgres.DATABASE_URL}}, REDIS_URL=${{Redis.REDIS_URL}}, and for app secrets either ${{acme-api.SESSION_SECRET}} or move them to shared variables of the environment (skiffly variables set --shared).

Expected: skiffly logs -s worker shows the worker connecting to Redis.

4. Scheduler jobs#

Each Heroku Scheduler entry becomes a service with a cron schedule (UTC) and the job's command as the start command; it runs one container per tick and exits. Cron services are not probed and get no domain.

Create them like the worker (same repository), or in the config file below.

5. Copy the data#

Do this in a short maintenance window: put Heroku in maintenance mode, copy, verify, then switch DNS.

heroku maintenance:on -a acme-api
heroku pg:backups:capture -a acme-api
heroku pg:backups:download -a acme-api          # latest.dump
 
skiffly connect postgres --print                 # opens a TCP proxy, prints postgresql://postgres:…@edge.skiffly.cloud:2xxxx/app
pg_restore --no-owner --no-acl --clean --if-exists -d "postgresql://postgres:…@edge.skiffly.cloud:2xxxx/app" latest.dump
psql "postgresql://postgres:…@edge.skiffly.cloud:2xxxx/app" -c "select count(*) from users"
skiffly proxy delete <port>                      # close the public proxy

pg_restore warnings about heroku_ext schema or extensions owned by Heroku roles are normal; errors about missing extensions mean you need the postgis/pgvector template instead of plain postgres (deploy that one and repeat).

Redis on Heroku is usually cache and queues: let it start empty. If you must move keys, redis-cli --rdb on the Heroku side and redis-cli -u "$(skiffly connect redis --print)" --pipe after DEBUG RELOAD is the manual route.

Restart the app so pools reconnect: skiffly restart -s acme-api.

6. Domain cutover#

skiffly domain -s acme-api api.example.com
# add the two records at your DNS provider:
#   api.example.com.            CNAME  edge.skiffly.cloud.
#   _skiffly.api.example.com.   TXT    "<token>"
skiffly domain status api.example.com            # wait for "verified", certificate issued within a minute

Lower the TTL of api.example.com a day before if you can. Until the CNAME propagates, Heroku keeps answering (it is in maintenance mode, so users see the maintenance page for a few minutes at most). Once curl -sI https://api.example.com returns headers from Skiffly, you are live.

Expected: SKIFFLY_PUBLIC_DOMAIN on the service now shows the custom domain; update any variable that held the Heroku hostname (APP_URL, OAuth callbacks, CORS origins).

7. Config as code (recommended before you delete anything)#

skiffly config init                     # writes .skiffly/skiffly.ts with everything created above
skiffly config plan                     # no changes

import { defineSkiffly, github, postgres, preserve, project, redis, service, shared } from "@skiffly/config";
 
export default defineSkiffly(() => {
  const db = postgres("Postgres");
  const cache = redis("Redis");
  const env = { NODE_ENV: "production", DATABASE_URL: db.env.DATABASE_URL, REDIS_URL: cache.env.REDIS_URL, SESSION_SECRET: shared("SESSION_SECRET") };
  const src = github("acme/acme-api");
  const web = service("acme-api", { source: src, start: 'sh -c "npm run migrate && node dist/server.js"', healthcheck: "/healthz", env, domains: ["api.example.com"], domain: true });
  const worker = service("worker", { source: src, start: "sh -c 'node -e \"require(\\\"http\\\").createServer((_,r)=>r.end(\\\"ok\\\")).listen(process.env.PORT)\" & exec node dist/worker.js'", env });
  const digest = service("digest", { source: src, start: "node dist/jobs/digest.js", cron: "0 3 * * *", env });
  return project("acme-api", { resources: [db, cache, web, worker, digest] });
});

Commit it. skiffly config apply --yes in CI (with a SKIFFLY_TOKEN secret) now replaces the Heroku pipeline.

8. Decommission Heroku#

After a few days of traffic on Skiffly with no surprises in skiffly logs and Project → Observability:

heroku pg:backups:capture -a acme-api        # one last backup, download and keep it
heroku apps:destroy -a acme-api

What is different day to day#

• Deploys: every push to main builds (or skiffly up from your machine). PR previews replace review apps.
• Logs: last 5 000 lines per deployment in skiffly logs; forward to your log service if you kept Papertrail-style history.
• Scaling: replicas (numReplicas) and CPU/memory limits per service instead of dyno types; a service with a volume stays at one replica.
• Billing: a prepaid balance instead of a monthly invoice — top up before the balance hits zero (3-day grace, then services pause).
• Backups: daily volume snapshots plus a logical dump for Postgres, restored by support; keep your own pg_dump cron as well.