How Northstar Systems cut deploy time by 80%

The challenge

Northstar Systems processes card payments for online merchants across Europe. Its platform had grown to 40 services running on virtual machines that a two-person ops team configured by hand.

Every release went out on a Tuesday or Thursday train. A deploy took most of an afternoon, needed a maintenance window for database changes, and blocked other teams while it ran. Engineers batched changes to avoid the overhead, so each release carried more risk than the last.

The solution

The team described its services, private network, and environment variables in one cloud.yaml file and linked the repository with the CLI. A staging copy of the stack ran in parallel for two weeks while they compared latency and error rates.

services:
  - name: payments-api
    type: web
    healthCheckPath: /healthz
    scaling: { minInstances: 3, maxInstances: 12 }
  - name: ledger-worker
    type: worker

On cut-over weekend they switched DNS service by service. Zero-downtime deploys meant the payments API never dropped a request, and private networking kept the ledger database off the public internet.

The results

Releases now go out on every merge to main. Health checks gate each deploy, and a failed check rolls back automatically, so engineers stopped batching work. The ops team moved on to fraud tooling instead of patching servers.