case study / Enfaq (freelance) · 2025

A fintech platform to production in three weeks

Java microservices moved from Docker Compose to private GKE while the team kept shipping, plus a controlled Saudi network path for a bank integration.

Client
Enfaq (freelance)
Period
2025
Role
DevOps Engineer
Result
live in 3 weeks
Stack
GKE, Terraform, Helm, Argo CD, Cloud Build, Tailscale
DockerCompose3 weeksprivate GKETerraform · Helm · Argo CDbank

In short

  • Took a Saudi corporate-finance platform from Docker Compose to production on private GKE in three weeks.
  • Built the new platform next to the old one, so developers were never blocked.
  • Solved a bank integration that only accepted traffic from a Saudi ISP, first with OpenVPN, then with a Tailscale mesh and selective routing.
  • Worked directly with the CTO on pipelines, GitOps and the bank requirements.

Context

Enfaq is a Saudi fintech company. The product was a set of Java microservices with Redis and SQL dependencies, running on Docker Compose. The goal was a production-ready operating model in three weeks, while engineering kept developing against the existing environment.

Constraints

  • Time: three weeks to production readiness.
  • No downtime for the team: developers had to keep shipping features during the move.
  • Network: a partner bank’s endpoints could only be reached from approved Saudi network paths. Ordinary GCP public addresses were not accepted, because the bank recognised specific ISP ranges.
  • Access: no unrestricted remote access for every developer.

Decisions

Build in parallel, not in place

Replacing the environment in place would have let any network, container or database problem block the developers. So the new path was built next to the old one: services were containerized and validated one at a time, cloud infrastructure was provisioned separately, deployment patterns were tested before cutover, and data and external dependencies were mapped before traffic moved. Rollback stayed possible the whole time.

A platform in layers

  • Foundation (Terraform): separate pre-production and production, private GKE networking, administration only through IAP, Cloud SQL and Redis connectivity.
  • Identity and secrets: Workload Identity from Kubernetes to GCP, Google Secret Manager synced with External Secrets.
  • Delivery: Cloud Build and Artifact Registry for images, Helm charts for workloads, Argo CD reconciling the cluster state.
  • Edge and operations: ingress behind a WAF, health checks, autoscaling, logging, metrics and alerting, plus GKE scaling workflows that lowered non-production cost.

The Java services shared one Helm chart pattern: standard health probes, resource settings, OpenTelemetry options, secrets and service definitions. Onboarding a service became a values file, and production differences were visible instead of hidden in copied manifests.

Readiness is more than a running pod

Under a short deadline it is tempting to treat a running pod as done. Each service went through the same checklist: immutable image tags, secret sync, database and Redis connectivity, service discovery, probes, resource limits, external integrations, logs and alerts, and a deploy and rollback through the pipeline. That separated application bugs from platform bugs and made the daily conversations with the CTO short and concrete.

A Saudi network path for the bank

The application needed a Saudi egress path without moving the platform or opening wide remote access. It evolved in two steps:

  1. OpenVPN through a private server in Dammam. A Cloudflare private tunnel gave controlled administrative access, and OpenVPN carried development and integration traffic through the approved Saudi ISP. It unblocked testing quickly, but needed client setup and care around routing and availability.
  2. A private Tailscale mesh. Developer devices and the Dammam server joined one private network. Only traffic for the bank endpoints went through a controlled proxy on the server; everything else used each developer’s own connection. A full-tunnel VPN would have added latency and made all developer traffic depend on one server.

The mesh gave identity-based membership, easy onboarding and revocation, a stable path through the approved ISP, and no exposed administration ports.

One source of truth during cutover

Running two environments side by side avoids downtime but invites drift between them. The new repositories were the single source of truth: Terraform for cloud resources, Helm for application configuration, Cloud Build for artifacts, Argo CD for the cluster state, and Secret Manager for runtime secrets. Every fix during the migration went into that model, not into a one-off change.

Results

  • Production-ready platform on private GKE in three weeks, with separate pre-production and production.
  • Feature development never stopped during the migration.
  • Repeatable releases through Cloud Build and Argo CD, with rollback.
  • A working, controlled path for partner-bank integration testing from Saudi Arabia.
  • Lower non-production cost through scheduled GKE scaling.
← All case studies