[albertompe]

Summary

Closes #77.

Robin no longer forces appendonly yes on persistent (ephemeral: false) clusters. Each persistence directive is now written by exactly one layer:

Mode	appendonly / save written by	If not set
ephemeral: true	Robin (appendonly no + save ""), unchanged	—
ephemeral: false	The user, via spec.config	Redis defaults: AOF off, RDB save 3600 1 300 100 60 10000

The AOF/RDB decision for persistent clusters goes back to the user. The generated redis.conf also stops repeating directives that clients such as the Helm chart already set explicitly.

Changes

• internal/kubernetes/objects.go: buildRedisConf drops the else { appendonly yes } branch and documents the ownership.
• internal/kubernetes/objects_test.go:
  • New table-driven TestBuildRedisConf_PersistenceOwnership over {ephemeral, persistent} × {cluster, standalone} × {no user persistence, user enables, user disables}. It counts directive lines:
    • ephemeral: Robin's appendonly no / save "" appear exactly once, before the user config;
    • persistent: Robin adds none.
  • TestBuildConfigMap_Ephemeral / TestBuildConfigMap_NonEphemeral updated.
• CHANGELOG.md: breaking-change entry with migration notes.

Breaking change

Persistent clusters that do not set appendonly in spec.config will no longer have AOF enabled. They fall back to the Redis defaults: RDB snapshots with the default save points, AOF off.

• Upgrading Robin does not change already deployed clusters. The ConfigMap is only regenerated when:
  • the Redkey changes (config, image, resources, labels, etc.), which goes through the upgrade flow and recycles the pods;
  • the cluster scales up from zero;
  • the password is rotated.
• Cluster mode: the upgrade migrates the slots before recycling each node, so no data is lost.
• Standalone mode: the restarted pod reloads the dump.rdb written at shutdown. If spec.config sets save "", it starts empty.
• Password rotation: the ConfigMap is rewritten without recycling the pods, so the change applies on the next pod restart. After a crash, writes since the last RDB snapshot are lost.
• Migration: add appendonly yes (and save ... if needed) to spec.config of persistent clusters to keep AOF.
• Not affected: deployments using the Helm chart, because the chart already renders explicit save / appendonly.

Additional context

• go test ./internal/...: 617 passed. make test-integration: 58/58. make lint: 0 issues.
• Validated end-to-end together with redkey-operator#113: the new E2E spec persistence passed 5/5 on kind with this Robin build.
• Redis 8.10 behaviour checked in Docker: without a save line RDB stays on with the default save points; without an appendonly line AOF stays off.
• Operator docs and E2E alignment: redkey-operator#113.