Learning Platform Cloud Migration

Containerising a legacy application starts with untangling it. We separated the web front end, the API, the transcoder and the background workers into their own Docker images, each with explicit configuration and health checks, so they can be deployed, scaled and restarted independently.

Infrastructure is described in Terraform, which means environments can be recreated from code rather than from memory — the exact problem behind the earlier outages.

• Web, API, transcoder and workers in separate containers
• Health checks on every service
• Infrastructure recreated from Terraform, not notes

Every change now goes through a GitHub Actions pipeline that runs tests, builds images, deploys to staging and then to production. If health checks fail after a production deploy, the pipeline rolls back automatically to the previous version.

Because releases are small, tested and reversible, the team ships during working hours instead of waiting for the evening window. Deployments take about four minutes.

• Build, staging and production pipelines
• Automatic rollback when health checks fail
• Releases during the working day

On the old server, a burst of video uploads competed with page requests for the same CPU, and students felt it. Transcoding now runs on its own scaled workers, fed by a queue, so a spike in uploads can no longer affect lesson pages.

Separating the heavy workload also made right-sizing possible: the web and API services run on smaller machines, while transcoding capacity scales with demand. Together with object storage for media, that is where the 58% hosting reduction came from.

• Queue-fed transcoding workers
• Page serving isolated from upload spikes
• Right-sized services and object storage for media

Prometheus and Grafana provide metrics and dashboards, with alerting on errors, latency and resource use, and logs are aggregated so problems can be traced across services. The team now hears about issues before students report them.

Backups are no longer assumed to work. A scheduled job restores the latest PostgreSQL backup into a scratch environment and checks it, so the platform has evidence — not hope — that it can recover from a failure.

• Metrics, dashboards and alerting with Prometheus and Grafana
• Aggregated logs across services
• Scheduled, verified restore tests