|
docs contribution · merged ✓
the optional label that bricks clusters
github.com/openshift/openshift-docs · PR #112643
|
|
🐣 first — what is limited live migration?
OpenShift lets you migrate your CNI from OpenShift SDN → OVN-Kubernetes without taking the cluster down.
This is called limited live migration — nodes migrate one by one while the cluster stays up.
After migration: you run a cleanup step to remove the migration annotation.
Run cleanup too early → cluster stuck in a hybrid SDN+OVN zombie state. No easy way out — requires a ticket and engineering involvement to recover.
|
|
|
🐛 one word caused it all
|
1. verification was marked "Optional"
The verification step in initiating-limited-live-migration.adoc
had [Optional] in the title.
users read "optional" → skip it → run cleanup immediately
|
|
2. gate language was lost in a prior refactor
PR #78178 split the original monolithic module. During the split,
the gate language — "Complete the following steps only if the migration succeeds" — was dropped.
the procedure became a linear flow with no safety barrier before cleanup
|
|
3. 🔴 only 2 checks — not nearly enough
original verification checked only: network type + status condition.
MCP rollout, node machineconfig convergence, and Operator health were never checked.
migration could look "done" while nodes were still mid-rollout
|
|
|
|
💭 what changed?
|
before
[Optional] Verify migration
✓ check network type
✓ check status condition
↓
oc annotate ... - ← cleanup
↓
🔥 cluster half-SDN half-OVN
|
|
after
Verify migration (mandatory)
✓ network type = OVNKubernetes
✓ NetworkTypeMigrationCompleted
✓ all MCP pools updated
✓ all nodes: currentConfig=done
✓ all ClusterOperators healthy
⚠ WARNING: do not proceed
until all checks pass
↓
oc annotate ... - ← safe ✓
|
|
|
|
💡 the rule
never mark a safety gate as "Optional".
Users follow the happy path. If a step looks skippable, they skip it.
If skipping it can brick the cluster — it is not optional.
When docs split a procedure across modules, re-read the gate language.
Refactors that move content without moving its warnings ship silent regressions.
|
|
|
|