Overview
A typical engagement starts with a platform team that runs several Valkey 8.0 clusters on Kubernetes for sessions and idempotency keys. Valkey 9.0 is out, 8.0 is nearing the end of its maintenance window, and the team wants a written path before it touches production.
Challenge
Each Valkey minor receives fixes for a limited period, so staying on an old minor eventually means no fixes. An in-place cluster upgrade is safe only in the right order: replicas first, then a controlled failover, then the former primary, without a long period on mixed minors. Clients, scripts and monitoring that were tested on 8.0 have to be checked against 9.0, and the team has never rehearsed a rollback.
Environment
Several Valkey 8.0 Cluster deployments as StatefulSets with persistent volumes on Kubernetes, application clients in more than one language, Prometheus metrics, and snapshot persistence.
Approach
AceMQ engineers read the release notes between the versions in use and the target and list each change that touches the team's configuration, commands and clients. A staging cluster that mirrors the production topology is upgraded replica first, with a failover while clients are connected and the real command mix replayed. The plan sets the order of clusters, from lowest risk to highest, and the check that must pass before each next step.
Solution
- Release-note review from the current version to the target, mapped to the team's configuration, commands and clients
- Staging cluster that mirrors production topology, upgraded replicas first
- Controlled failover with clients connected, then the former primary
- Replay of the real command mix and a comparison of latency and error rates
- Rollback tested at each step, with a limit on time spent on mixed minors
- Order of clusters and a check to pass before each one
Outcome
The engagement delivers a release-note impact list, a rehearsed upgrade runbook per cluster type, a rollback for each step and a schedule. The team runs the production upgrades, or AceMQ engineers run the first cluster beside it. Application code changes stay with the customer's developers.