Skip to main content
Version: next

Day-2 operations

Preview. These pages are for operators of P6 on Kubernetes (Kubernetes and OpenShift), the first path planned for Stable. Facts come from the manifests, setup.sh and the code as of 2026-10-13. Where something is not documented anywhere in the repository, the page says so rather than guessing. Single-VM and Compose notes appear only where they differ.

PageCovers
UpgradesRe-running setup.sh, image pinning, rollout behaviour, version skew, rollback
Redis backup and recoveryWhat lives in Redis, what a loss costs, which secrets to back up separately, key rotation
ObservabilityHealth endpoints, the JSON metrics, probes, log lines, the audit stream
Sizing and scalingResource defaults, workers and turns per worker, replicas, sandbox capacity
HardeningA checklist per deployment path, from the security model
TroubleshootingSymptom, cause, check and fix

Other deployment paths​

The pages above are written for Kubernetes. For the single VM and the sandbox service, the same day-2 tasks are on their own pages:

TaskSingle VMSandbox service
UpgradeUpgrading an existing VMUpgrading
Check health and statusChecking itStatus and the audit lines
Redis and its dataRedis persistenceUninstalling (what is kept, what is deleted)
Reboots and host outagesRebootsA KVM-host outage
TroubleshootingThe units' start failuresReproducing a member's issue
LimitsWhat round one does not claimLimits operators must know

Gaps you should know about​

These are properties of the current deployment, not documentation omissions:

  • Single points of failure. Redis, the relay, the control plane and (by default) the supervisor each run one replica, and there is no PodDisruptionBudget or anti-affinity, so a node drain can take them all at once.
  • No tested backup or rollback procedure exists in the repository.
  • Sessions have no TTL and per-session Redis memory has not been measured, while Redis is limited to 256Mi.
  • /metrics is JSON, not Prometheus format, and the relay and sandboxes have no health endpoints.
  • A drain protects only short turns (about 15 to 20 seconds), although the pod's grace period is 120 seconds.
  • No capacity numbers are published. The worker and turn defaults are trial values, not measurements.