Season 7: It Worked on My Laptop · Episode 7 of 8
Deploy Without the Dramatic Music
You play: Junior AI Developer
Included with Core programme and Certification prep
The situation
Leo's new feature for Code Nexus Live — a creator spotlight panel built on Code Nexus Assist — is ready. The old way of shipping it would be: tell everyone, take Assist down for a minute, bring the new version up, hope. Kabelo has a simpler request: ship it so nobody notices, and if it goes wrong, notice before a creator does.
What you'll learn
- Kubernetes application delivery: declarative rolling updates with maxSurge and maxUnavailable, reading rollout status and revision history, rollback with kubectl rollout undo, the difference between metrics, logs and events as observability signals, why a stuck rollout isn't fixed by scaling harder, and verifying user-facing health after recovery.
- GitOps with Argo CD and Flux, Helm charts and releases, blue-green versus canary versus rolling, kubectl exec, logs --previous and port-forward, and the HPA, VPA and Cluster Autoscaler.
Who you work with
Kabelo Mokoena
Platform and DevOps Engineer
Leo Martins
Product Manager
8-BIT
Code Nexus Internal AI Assistant
Scenes
- 1.Ship it so nobody notices
- 2.Metrics, logs, events
- 3.Read a rollout in progress
- 4.A bounded rolling update, on purpose
- 5.Ship it, then take it back
- 6.Why scaling harder doesn't help
- 7.Recover it, and prove it
- 8.Write the rollout record
- 9.Delivery and rollback check
What you leave with
A bounded rolling update — maxSurge and maxUnavailable, written explicitly for the first time — paced a normal release with zero restarts and no dip below full capacity. When the creator-spotlight build turned out to crash on start, the same bound that made the good rollout safe also made the bad one fail safely: it stalled with the working version still serving, rather than taking anything down. Scaling up under traffic pressure didn't help, because more replicas of a crashing template is not capacity, and the cluster didn't even have room for all of them. Recovery was two corrections — scale back, roll back — verified not by a green dashboard but by a real request answered correctly.