docs: A-4 — a dead pod reports healthier than a live one

Kubernetes keeps calling the node Ready for forty seconds while users already see failures, and the pod on the powered-off machine stays ready=true because its kubelet can no longer contradict itself. Eviction waits another five minutes, then the StatefulSet refuses to recreate its pod and the replacement Deployment pod cannot schedule because the local-path volume is pinned to the dead node.

Killing the server node instead shows the opposite shape: containerd keeps the workload running while the API server, Traefik and the observability stack disappear, so the outage is the missing path rather than the missing application. Traefik at one replica is the ingress single point of failure.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
DongHyeonka
2026-09-04 12:25:30 +09:00
co-authored by Claude Opus 5
parent 99b689e715
commit d0666c5ba0
15 changed files with 866 additions and 0 deletions
@@ -0,0 +1,20 @@
=== A-4 기준선 ===
kc-lab-1 Ready true
kc-lab-2 Ready <none>
a2-probe true kc-lab-2
keycloak-0 true kc-lab-2
keycloak-1 true kc-lab-1
postgres-7b474b88c8-2gf27 true kc-lab-2
=== PVC 가 어느 노드에 묶여 있는가 (재배치 가능성) ===
persistentvolumeclaim/postgres-data → kc-lab-2
=== 서비스 정상 확인 ===
https://auth.hyeonworks.com/realms/master HTTP 200
=== VM 상태 ===
--------------------------
1 kc-lab-1 running
2 kc-lab-2 running
@@ -0,0 +1,17 @@
=== 워커 노드(kc-lab-2) 전원 차단 — virsh destroy 는 종료 신호가 없다 ===
차단 시각: 12:07:43
Domain 'kc-lab-2' destroyed
+15초 node=Ready | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 000
+30초 node=Ready | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 000
+45초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+60초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+75초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+90초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+105초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+120초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+135초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+150초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+165초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
+180초 node=NotReady | keycloak-0=Running postgres-7b474b88c8-2gf27=Running | 외부 HTTP 503
@@ -0,0 +1,28 @@
=== 파드 상태의 진실 — Running 인데 노드가 없다 ===
a2-probe Running true kc-lab-2 <none>
keycloak-0 Running true kc-lab-2 <none>
keycloak-1 Running false kc-lab-1 <none>
postgres-7b474b88c8-2gf27 Running true kc-lab-2 <none>
=== 재배치가 시도되었는가 ===
10m Warning Unhealthy pod/keycloak-0 Readiness probe failed: Get "http://10.42.1.67:9000/health/ready": context deadline exceeded (Client.Timeout exceeded while awaiting headers)
3m15s Warning NodeNotReady pod/postgres-7b474b88c8-2gf27 Node is not ready
3m15s Warning NodeNotReady pod/keycloak-0 Node is not ready
3m15s Warning NodeNotReady pod/a2-probe Node is not ready
2m27s Warning Unhealthy pod/keycloak-1 Readiness probe failed: Get "http://10.42.0.35:9000/health/ready": context deadline exceeded (Client.Timeout exceeded while awaiting headers)
2s Warning Unhealthy pod/keycloak-1 Readiness probe failed: HTTP probe failed with statuscode: 503
=== 노드 taint — 쿠버네티스가 붙인 것 ===
node.kubernetes.io/unreachable=:NoSchedule
node.kubernetes.io/unreachable=:NoExecute
=== Prometheus 가 본 것 (kc-lab-1 에 있어 살아남았다) ===
up{job=keycloak pod=keycloak-1 } = 1
up{job=keycloak pod=keycloak-0 } = 0
up{job=kubelet pod=- } = 1
up{job=kubelet pod=- } = 0
up{job=node-exporter pod=kc-lab-1 } = 1
up{job=node-exporter pod=kc-lab-2 } = 0
up{job=prometheus pod=- } = 1
=== 진입점이 처음 40초간 000 이었던 이유 — nginx upstream ===
@@ -0,0 +1,15 @@
=== nginx 설정 위치 찾기 ===
=== NoExecute taint 의 tolerationSeconds — 언제 축출되는가 ===
node.kubernetes.io/not-ready NoExecute tolerationSeconds=300
node.kubernetes.io/unreachable NoExecute tolerationSeconds=300
=== 5분 축출 시점까지 관찰 ===
+210초 a2-probe:Running keycloak-0:Running keycloak-1:Running postgres-7b474b88c8-2gf27:Running
+240초 a2-probe:Running keycloak-0:Running keycloak-1:Running postgres-7b474b88c8-2gf27:Running
+270초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
+300초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
+330초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
+360초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
+390초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
+420초 a2-probe:Terminating keycloak-0:Terminating keycloak-1:Running postgres-7b474b88c8-2gf27:Terminating postgres-7b474b88c8-9cmsv:Pending
@@ -0,0 +1,18 @@
=== 새 postgres 가 Pending 인 이유 ===
Events:
Type Reason Age From Message
---- ------ ---- ---- -------
Warning FailedScheduling 4m45s default-scheduler 0/2 nodes are available: 1 node(s) didn't match PersistentVolume's node affinity, 1 node(s) had untolerated taint(s). no new claims to deallocate, preemption: 0/2 nodes are available: 2 Preemption is not helpful for scheduling.
=== keycloak-0 대체 파드가 안 생기는 이유 (StatefulSet) ===
keycloak 2 <none> 1
keycloak-0 1/1 Terminating 0 30m
keycloak-1 0/1 Running 0 143m
=== 복구 — 노드 재기동 ===
재기동 시각: 12:16:31
Domain 'kc-lab-2' started
+30초 node=Ready | Running 파드 3 개 | 외부 HTTP 503
+60초 node=Ready | Running 파드 3 개 | 외부 HTTP 200
→ 서비스 복귀
@@ -0,0 +1,19 @@
=== 복구 확인 ===
keycloak-0 1/1 Running 0 68s
keycloak-1 1/1 Running 0 144m
postgres-7b474b88c8-9cmsv 1/1 Running 0 4m20s
=== kc-lab-1(k3s server)에 무엇이 있는가 — 이게 곧 영향 범위다 ===
keycloak-lab keycloak-1
kube-system coredns-54996dc9b4-8k8fj
kube-system helm-install-traefik-crd-q29b5
kube-system local-path-provisioner-77b9867795-g27z8
kube-system metrics-server-6dc596dfb8-7xxq4
kube-system svclb-traefik-5eb6a9a1-qwwk5
kube-system traefik-5d6fcf895-wpfhr
observability grafana-845b5678cf-b6gvc
observability node-exporter-9qk9w
observability prometheus-6774f94f7c-pzr2t
=== Traefik replica 수 (진입점의 단일 장애점인가) ===
traefik 1 1
@@ -0,0 +1,19 @@
=== 컨트롤 플레인 노드(kc-lab-1) 전원 차단 ===
차단 시각: 12:18:08
Domain 'kc-lab-1' destroyed
+20초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+40초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+60초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+80초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+100초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+120초 외부 auth=000 grafana=502 | kubectl: Unable to connect to the server: dial tcp
+140초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
+160초 외부 auth=000 grafana=000 | kubectl: Unable to connect to the server: dial tcp
=== 살아 있는 노드에서 직접 확인 — 워크로드는 도는가 ===
CONTAINER IMAGE CREATED STATE NAME ATTEMPT POD ID POD NAMESPACE
e5f777900b762 60e153026e8f5 4 minutes ago Running keycloak 0 640d4dafaefb3 keycloak-0 keycloak-lab
6
@@ -0,0 +1,13 @@
=== 컨트롤 플레인 노드 복구 ===
재기동: 12:23:39
Domain 'kc-lab-1' started
+30초 외부=502 | kc-lab-1=Ready kc-lab-2=Ready
+60초 외부=200 | kc-lab-1=Ready kc-lab-2=Ready
→ 서비스 복귀 (총 60초)
=== 최종 상태 ===
keycloak-0 1/1 Running 0 7m57s
keycloak-1 1/1 Running 1 (<invalid> ago) 151m
postgres-7b474b88c8-9cmsv 1/1 Running 0 11m
(prometheus port-forward 재연결 필요)
+23
View File
@@ -0,0 +1,23 @@
# A-4 — 노드 전원 차단 증거
2026-09-04 12:0712:24 KST
해설: [`docs/experiment-a4-node-loss.md`](../../experiment-a4-node-loss.md)
| 파일 | 무엇을 보여주는가 |
|---|---|
| `01-baseline.txt` | 차단 전 — 양쪽 Ready, **PVC 가 kc-lab-2 에 못박혀 있음**(재배치 불가의 원인), 외부 200 |
| `02-worker-node-killed.txt` | `virsh destroy kc-lab-2`**40초간 노드가 Ready 로 남아 있고** 외부는 이미 `000`. 이후 `503` |
| `03-state-during-loss.txt` | **죽은 파드가 `ready=true`, 산 파드가 `ready=false`.** `up` 은 정확히 0. `unreachable` taint |
| `04-eviction-timing.txt` | `tolerationSeconds=300`**5분 뒤** 축출, 새 postgres 는 `Pending` |
| `05-recovery.txt` | `FailedScheduling: didn't match PersistentVolume's node affinity`, StatefulSet `DESIRED=2 CURRENT=1`. 노드 복귀 후 **60초** |
| `06-control-plane-inventory.txt` | kc-lab-1 에 있는 것 목록 — **Traefik `replicas=1`** |
| `07-control-plane-loss.txt` | `virsh destroy kc-lab-1` — 외부 `000`, `kubectl` 불통. **그런데 `crictl ps` 로 보면 keycloak-0 은 Running** |
| `08-control-plane-recovery.txt` | 60초 만에 복귀 |
| `a4-up-dropped-per-node.png` | Grafana — `up` 이 노드별로 0 으로 떨어지는 구간. 12:1812:23 은 **0 이 아니라 데이터 없음**(관측자가 같이 죽음) |
## 핵심 네 줄
1. **쿠버네티스는 40초 동안 노드가 살아 있다고 말한다.** 사용자는 이미 장애를 겪는 중이다.
2. **죽은 파드의 상태는 화석이다.** `ready=true` 인 파드가 꺼진 기계 위에 있다.
3. **StatefulSet 은 대체 파드를 만들지 않고, PVC 는 재배치를 막는다.** 사람이 개입해야 한다.
4. **컨트롤 플레인 상실 ≠ 워크로드 상실.** 컨테이너는 계속 돌고, 들어갈 문만 사라진다.
Binary file not shown.

After

Width:  |  Height:  |  Size: 79 KiB