[zzzz465]

Background

• Azure AKS cloud-node-manager can drop the ExternalIP from a Node's status.addresses for one of its 5-minute reconcile cycles
  • Hostname and InternalIP stay, and the same ExternalIP comes back 5 minutes later
  • The public IP keeps working during that time
• The migration controller treats this as an address mismatch and marks every post-Ready GameServer on that Node as Unhealthy
  • As a result, every match running on the Node is terminated
• A GameServer scheduled onto the Node while the ExternalIP is missing gets the InternalIP as its address
  • It keeps the InternalIP after the ExternalIP comes back, because the InternalIP is still one of the Node's addresses

Changes

• Defer the migration check while the ExternalIP is transiently missing
  • Condition: the pod's nodeName is unchanged, the GameServer's address type is ExternalIP (based on status.addresses), and the Node has no ExternalIP
  • Record the first time this is seen per GameServer in memory, and re-check with EnqueueAfter for 15 minutes
  • If the ExternalIP is still missing after 15 minutes, fall back to the existing behavior
  • Clear the record when the pod is no longer a migration candidate or is deleted
• Fix GameServers that were given an InternalIP while the ExternalIP was missing
  • Condition: the GameServer's address type is InternalIP or InternalDNS, and the Node has an ExternalIP or ExternalDNS
  • Before Ready: re-apply the address
  • Ready or Reserved: move to Unhealthy so it is recreated
  • Allocated: keep it, to preserve the running session

[Copilot]

🟡 Changes recommended

Controller image builds fail with the existing build context, the image omits dependency distribution artifacts, and migration handling can misclassify servers or mishandle grace periods.

7 open issues

• Fix controller image build contexts and forward version/build tags · New
• Include dependency notices and source archive in the final image · New
• Reset migration timers when Node ExternalIP returns · New
• Preserve migration deadline until the API update succeeds · New
• Prevent stale migration timers from carrying over to replacement Pods · New
• Match address types using address() precedence and validity rules · New
• Validate ExternalIP entries before treating them as available · New

What changed in this PR

This PR aims to prevent transient Node ExternalIP omissions from terminating running GameServers, while correcting internal addresses when external connectivity returns.

Changes:

• Adds a 15-minute migration grace period with per-Pod tracking and delayed retries.
• Handles restored external addresses according to GameServer state, with tests for deferral, expiry, and recovery.
• Switches the controller image to a multi-stage source build and expands the image workflow to include it.

File	Description
pkg/​gameservers/​migration.go	Adds grace-period tracking and external-address recovery handling.
pkg/​gameservers/​migration_test.go	Tests missing and restored addresses, state transitions, and grace-period timing.
cmd/​controller/​Dockerfile	Builds the controller from source in a multi-stage image.
.github/​workflows/​build-images.yml	Builds and publishes allocator and controller images through a matrix.

_{Review effort: Balanced}

[mnthe]

엣지케이스나 상태관리 찌꺼기 피드백들 좀 있네요, 체크해보면 좋을 것 같아요

그리고 언헬시 노드에 Pod을 받는게 아니라 ExtenralIP 없으면 안받고있다가 풀리면 받는게 나을 것 같긴 한데, 주소가 있는걸로 판단해서 어려운거죠?

[zzzz465]

엣지케이스나 상태관리 찌꺼기 피드백들 좀 있네요, 체크해보면 좋을 것 같아요

Copilot 리뷰 기반으로 수정했습니다.

그리고 언헬시 노드에 Pod을 받는게 아니라 ExtenralIP 없으면 안받고있다가 풀리면 받는게 나을 것 같긴 한데, 주소가 있는걸로 판단해서 어려운거죠?

문제점은 node.status.address 만 읽는 것이라 public IP 가 없는 노드, 일시적으로 IP 가 빠진 노드를 구분하지 못해서 그렇습니다. GameServer CR 에 custom status 를 추가해서 기존에 IP 가 있었는지 여부를 기록하면 되긴 해요.

버그 수정 이외에 추가적인 분기를 만드는건 원본 소스에서 기능이 달라지는 것이라 검토를 좀 더 하면 좋겠습니다