说明
在k8s环境部署Apache Zookeeper 3.9.5，高可用，多pod，适配Hadoop。
准备配置文件
注意，配置文件里使用了hadoop作为namespace，需修改为自己期望的namespace
v1.yaml内容如下
bash
# ========== ZooKeeper 3节点集群 ==========
# 用于 HDFS NameNode HA 自动故障转移（ZKFC）。若用于其他环境，应当修改启动命令中的hadoop字段为真实namespace
# 使用官方 zookeeper:3.9.5 镜像
# 3.9.x 变化：基于 JDK 17+，配置项 ZOO_4LINEAR_ENABLE 已移除（默认启用）
---
apiVersion: v1
kind: Service
metadata:
name: zookeeper
labels:
app: zookeeper
spec:
clusterIP: None # Headless Service，用于 Pod DNS 发现
publishNotReadyAddresses: true # 让等待方能在 Pod 未 Ready 时也解析到 IP
ports:
- port: 2181
name: client
- port: 2888
name: peer
- port: 3888
name: election
selector:
app: zookeeper
---
apiVersion: v1
kind: Service
metadata:
name: zookeeper-external
labels:
app: zookeeper
spec:
type: NodePort
ports:
- port: 2181
name: client
targetPort: 2181
# nodePort 留空，系统自动分配
selector:
app: zookeeper
---
apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
name: zookeeper-pdb
spec:
minAvailable: 2
selector:
matchLabels:
app: zookeeper
---
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: zookeeper
spec:
serviceName: zookeeper
replicas: 3
# OrderedReady（默认）：Pod 按序启动，确保前一个就绪后才启动下一个
# 必须用 OrderedReady 而非 Parallel：Parallel 会导致 3 个 Pod 同时启动，
# 此时 Headless Service DNS 尚未完全传播，ZK 节点找不到 Peer 无法形成
# Quorum，进程崩溃退出 (exit code 14)，产生多次无效重启。
# OrderedReady 保证 zookeeper-1 启动时 zookeeper-0 的 DNS 已就绪。
podManagementPolicy: OrderedReady
selector:
matchLabels:
app: zookeeper
template:
metadata:
labels:
app: zookeeper
spec:
subdomain: zookeeper
priorityClassName: hadoop-namenode-high
# ZK 节点分散到不同 Node（保证仲裁安全）
affinity:
podAntiAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
- labelSelector:
matchLabels:
app: zookeeper
topologyKey: kubernetes.io/hostname
terminationGracePeriodSeconds: 30
containers:
- name: zookeeper
# ZK 3.9.5 官方镜像（华为云镜像源）
# 重要：3.9.5 Docker entrypoint 在使用 ZOO_SERVERS 时不写 clientPort 到 zoo.cfg，
# 导致 plain 客户端端口被禁用 (plain=disabled)，端口 2181 不监听。
# 解决方案：覆盖 command，手动写完整 zoo.cfg（含 clientPort=2181），然后前台启动 ZK。
image: swr.cn-north-4.myhuaweicloud.com/ddn-k8s/docker.io/library/zookeeper:3.9.5
imagePullPolicy: IfNotPresent
command:
- /bin/bash
- -ec
- |
# ===== 阶段1：准备环境 =====
set -e
mkdir -p /data /datalog /conf /logs
# 写入完整的 zoo.cfg（显式包含 clientPort=2181）
cat > /conf/zoo.cfg << 'ZKCFG'
tickTime=2000
dataDir=/data
dataLogDir=/datalog
clientPort=2181
initLimit=10
syncLimit=2
autopurge.snapRetainCount=3
autopurge.purgeInterval=1
server.1=zookeeper-0.zookeeper.hadoop:2888:3888
server.2=zookeeper-1.zookeeper.hadoop:2888:3888
server.3=zookeeper-2.zookeeper.hadoop:2888:3888
ZKCFG
echo "=== /conf/zoo.cfg ==="
cat /conf/zoo.cfg
echo "=== /data/myid ==="
cat /data/myid
# JDK 17 反射权限（Hadoop ZKFC 依赖）
export SERVER_JVMFLAGS="--add-opens java.base/java.io=ALL-UNNAMED --add-opens java.base/java.lang=ALL-UNNAMED --add-opens java.base/java.util=ALL-UNNAMED --add-opens java.base/java.net=ALL-UNNAMED"
# ===== 阶段2：SIGTERM 陷阱（转发给 ZK 进程，确保优雅关闭） =====
ZK_PID=""
_term() {
echo "$(date): Received SIGTERM, stopping ZooKeeper..."
if [ -n "$ZK_PID" ] && kill -0 "$ZK_PID" 2>/dev/null; then
kill -TERM "$ZK_PID"
wait "$ZK_PID" 2>/dev/null
fi
exit 0
}
trap _term TERM INT
# ===== 阶段3：启动 ZK（带重试，吸收 Quorum 组建期间的偶发崩溃） =====
# OrderedReady 保证前置 Pod 已就绪，但 ZK 处于 LOOKING 状态时
# readiness probe 已通过，新 Pod 加入 Quorum 仍可能有瞬态时序竞争，
# 导致 ZK 偶发退出 (exit code 14)。通过进程内重试吸收，避免 Pod 级重启。
set +e
MAX_RETRY=3
RETRY=0
while [ $RETRY -lt $MAX_RETRY ]; do
echo "$(date): Starting ZooKeeper (attempt $((RETRY+1))/$MAX_RETRY)..."
zkServer.sh start-foreground &
ZK_PID=$!
wait $ZK_PID
EXIT=$?
ZK_PID=""
echo "$(date): ZooKeeper exited with code $EXIT"
if [ $EXIT -eq 0 ]; then
echo "$(date): ZooKeeper exited normally"
exit 0
fi
RETRY=$((RETRY+1))
if [ $RETRY -lt $MAX_RETRY ]; then
echo "$(date): Will retry in 5 seconds..."
sleep 5
fi
done
echo "$(date): Max retries ($MAX_RETRY) exhausted, giving up"
exit $EXIT
ports:
- containerPort: 2181
name: client
- containerPort: 2888
name: peer
- containerPort: 3888
name: election
resources:
requests:
memory: "1Gi"
cpu: "500m"
limits:
memory: "2Gi"
cpu: "1000m"
# startupProbe：ZK 3.9.5 + JDK 17 启动较慢，需要充足启动时间
# startupProbe 通过后，liveness/readiness 才开始工作
startupProbe:
tcpSocket:
port: 2181
initialDelaySeconds: 10
periodSeconds: 10
timeoutSeconds: 3
failureThreshold: 30 # 10 + 10×30 = 310s 最大启动容忍
livenessProbe:
tcpSocket:
port: 2181
periodSeconds: 15
timeoutSeconds: 3
failureThreshold: 5 # 75s 容忍（startup 通过后 ZK 已稳定）
readinessProbe:
tcpSocket:
port: 2181
periodSeconds: 10
timeoutSeconds: 3
failureThreshold: 3
volumeMounts:
- name: zookeeper-data
mountPath: /data
# initContainer：从 Pod 名提取序号写入 /data/myid
# zoo.cfg 由主容器的 command 直接写入 /conf/zoo.cfg（绕过有 bug 的 Docker entrypoint）
initContainers:
- name: fix-myid
image: swr.cn-north-4.myhuaweicloud.com/ddn-k8s/docker.io/library/zookeeper:3.9.5
imagePullPolicy: IfNotPresent
command: ["/bin/bash", "-c"]
args:
- |
# 从 Pod 名提取序号：zookeeper-0 → 1，zookeeper-1 → 2，zookeeper-2 → 3
POD_NAME=$(hostname)
ORD=${POD_NAME##*-}
MY_ID=$((ORD + 1))
echo "Pod: $POD_NAME -> ZOO_MY_ID=$MY_ID"
echo $MY_ID > /data/myid
echo "Written myid=$MY_ID to /data/myid"
cat /data/myid
volumeMounts:
- name: zookeeper-data
mountPath: /data
volumeClaimTemplates:
- metadata:
name: zookeeper-data
spec:
accessModes: [ "ReadWriteOnce" ]
# 不指定 storageClassName，自动使用集群默认 StorageClass
resources:
requests:
storage: 10Gi部署
再次提醒，我部署到了namespace hadoop里，需改为你期望的namespace，包括文件中配置的hadoop字段都替换。
bash
kubectl apply -f v1.yaml -n hadoop