Fly.io

On Fly.io, the server runs on one Fly Machine and the store is an R2 bucket. fly deploy can build the image on Fly's builders, so you don't need Docker locally.

Fly's own buckets come from Tigris, which no longer accepts the path-style requests Akter sends for new buckets, so this guide uses R2. Put the bucket near the app: an R2 location hint of Eastern North America suits the iad region below.

You need the src/server.ts, Dockerfile and .dockerignore from Deploying, an R2 bucket, and the fly command.

Configure the app⁠

Create fly.toml:

app = "my-app"
primary_region = "iad"
kill_signal = "SIGTERM"
kill_timeout = "15s"

[deploy]
  strategy = "rolling"

[env]
  AKTER_STORAGE_BUCKET = "my-app-store"
  AKTER_STORAGE_ENDPOINT = "https://ACCOUNT_ID.r2.cloudflarestorage.com"
  AKTER_STORAGE_REGION = "auto"
  AUTH_ISSUER = "https://login.example.com/"
  AUTH_AUDIENCE = "my-app"
  AUTH_JWKS_URL = "https://login.example.com/.well-known/jwks.json"

[http_service]
  internal_port = 8080
  force_https = true
  auto_stop_machines = "off"
  auto_start_machines = false

  [[http_service.checks]]
    interval = "10s"
    timeout = "2s"
    grace_period = "10s"
    method = "GET"
    path = "/ready"

[[vm]]
  size = "shared-cpu-1x"
  memory = "1gb"

• auto_stop_machines = "off" keeps the server running when no requests arrive.
• strategy = "rolling" replaces the machine in place: Fly stops it, then starts it on the new image.
• kill_timeout gives the server 15 seconds after SIGTERM. Fly's default is 5.
• The check sends traffic to the machine only while /ready answers 200.

Deploy⁠

fly apps create my-app
fly secrets set --app my-app AWS_ACCESS_KEY_ID=<access key id> AWS_SECRET_ACCESS_KEY=<secret access key>
fly deploy --ha=false

--ha=false creates one machine. Without it, the first deploy creates two, and the second would take the store from the first. If you ever see two, run fly scale count 1.

Check the server:

curl https://my-app.fly.dev/ready

Releases⁠

Run fly deploy again. With one machine, Fly stops it, waits for the drain, and starts it again on the new image. Clients retry through the gap. Don't switch to the bluegreen or canary strategies: they start the new machine while the old one still runs.