Sentry 后端云原生中间件实践 ClickHouse PaaS ，为 Snuba 事件分析引擎提供动力
2023-02-13 12:07:38来源：黑客下午茶
ClickHouse PaaS 云原生多租户平台(Altinity.Cloud)
官网：https://altinity.cloud
PaaS 架构概览
(资料图)
设计一个拥有云原生编排能力、支持多云环境部署、自动化运维、弹性扩缩容、故障自愈等特性，同时提供租户隔离、权限管理、操作审计等企业级能力的高性能、低成本的分布式中间件服务是真挺难的。
SaaS 模式交付给用户
Sentry Snuba 事件大数据分析引擎架构概览
Snuba 是一个在 Clickhouse 基础上提供丰富数据模型、快速摄取消费者和查询优化器的服务。以搜索和提供关于 Sentry 事件数据的聚合引擎。
数据完全存储在 Clickhouse 表和物化视图中，它通过输入流(目前只有 Kafka 主题)摄入，可以通过时间点查询或流查询(订阅)进行查询。
文档：https://getsentry.github.io/snuba/architecture/overview.html
Kubernetes ClickHouse Operator什么是 Kubernetes Operator？Kubernetes Operator 是一种封装、部署和管理 Kubernetes 应用的方法。我们使用 Kubernetes API（应用编程接口）和 kubectl 工具在 Kubernetes 上部署并管理 Kubernetes 应用。
https://kubernetes.io/zh-cn/docs/concepts/extend-kubernetes/operator/Altinity Operator for ClickHouseAltinity：ClickHouse Operator 业界领先开源提供商。
Altinity：https://altinity.com/GitHub：https://github.com/Altinity/clickhouse-operatorYoutube：https://www.youtube.com/@Altinity当然这种多租户隔离的 ClickHouse 中间件 PaaS 云平台，公司或云厂商几乎是不开源的。
RadonDB ClickHousehttps://github.com/radondb/radondb-clickhouse-operatorhttps://github.com/radondb/radondb-clickhouse-kubernetes云厂商(青云)基于 altinity-clickhouse-operator 定制的。对于快速部署生产集群做了些优化。
Helm + Operator 快速上云 ClickHouse 集群云原生实验环境VKE K8S Cluster，Vultr托管集群(v1.23.14)Kubesphere v3.3.1 集群可视化管理，全栈的 Kubernetes 容器云 PaaS 解决方案。Longhorn 1.14，Kubernetes 的云原生分布式块存储。部署 clickhouse-operator这里我们使用 RadonDB 定制的 Operator。
values.operator.yaml定制如下两个参数：# operator 监控集群所有 namespace 的 clickhouse 部署watchAllNamespaces: true# 启用 operator 指标监控enablePrometheusMonitor: truehelm 部署 operator：
cd vip-k8s-paas/10-cloud-native-clickhouse# 部署在 kube-systemhelm install clickhouse-operator ./clickhouse-operator -f values.operator.yaml -n kube-systemkubectl -n kube-system get po | grep clickhouse-operator# clickhouse-operator-6457c6dcdd-szgpd 1/1 Running 0 3m33skubectl -n kube-system get svc | grep clickhouse-operator# clickhouse-operator-metrics ClusterIP 10.110.129.244部署 clickhouse-cluster8888/TCP 4m18skubectl api-resources | grep clickhouse# clickhouseinstallations chi clickhouse.radondb.com/v1 true ClickHouseInstallation# clickhouseinstallationtemplates chit clickhouse.radondb.com/v1 true ClickHouseInstallationTemplate# clickhouseoperatorconfigurations chopconf clickhouse.radondb.com/v1 true ClickHouseOperatorConfiguration
这里我们使用 RadonDB 定制的 clickhouse-cluster helm charts。快速部署 2 shards + 2 replicas + 3 zk nodes 的集群。
values.cluster.yaml定制：clickhouse: clusterName: snuba-clickhouse-nodes shardscount: 2 replicascount: 2...zookeeper: install: true replicas: 3helm 部署 clickhouse-cluster：
kubectl create ns cloud-clickhousehelm install clickhouse ./clickhouse-cluster -f values.cluster.yaml -n cloud-clickhousekubectl get po -n cloud-clickhouse# chi-clickhouse-snuba-ck-nodes-0-0-0 3/3 Running 5 (6m13s ago) 16m# chi-clickhouse-snuba-ck-nodes-0-1-0 3/3 Running 1 (5m33s ago) 6m23s# chi-clickhouse-snuba-ck-nodes-1-0-0 3/3 Running 1 (4m58s ago) 5m44s# chi-clickhouse-snuba-ck-nodes-1-1-0 3/3 Running 1 (4m28s ago) 5m10s# zk-clickhouse-0 1/1 Running 0 17m# zk-clickhouse-1 1/1 Running 0 17m# zk-clickhouse-2 1/1 Running 0 17m借助 Operator 快速扩展 clickhouse 分片集群使用如下命令，将shardsCount改为3：
kubectl edit chi/clickhouse -n cloud-clickhouse查看 pods：
kubectl get po -n cloud-clickhouse# NAME READY STATUS RESTARTS AGE# chi-clickhouse-snuba-ck-nodes-0-0-0 3/3 Running 5 (24m ago) 34m# chi-clickhouse-snuba-ck-nodes-0-1-0 3/3 Running 1 (23m ago) 24m# chi-clickhouse-snuba-ck-nodes-1-0-0 3/3 Running 1 (22m ago) 23m# chi-clickhouse-snuba-ck-nodes-1-1-0 3/3 Running 1 (22m ago) 23m# chi-clickhouse-snuba-ck-nodes-2-0-0 3/3 Running 1 (108s ago) 2m33s# chi-clickhouse-snuba-ck-nodes-2-1-0 3/3 Running 1 (72s ago) 119s# zk-clickhouse-0 1/1 Running 0 35m# zk-clickhouse-1 1/1 Running 0 35m# zk-clickhouse-2 1/1 Running 0 35m
发现多出chi-clickhouse-snuba-ck-nodes-2-0-0与chi-clickhouse-snuba-ck-nodes-2-1-0。分片与副本已自动由Operator新建。
小试牛刀ReplicatedMergeTree+Distributed+Zookeeper 构建多分片多副本集群连接 clickhouse我们进入 Pod, 使用原生命令行客户端clickhouse-client连接。
kubectl exec -it chi-clickhouse-snuba-ck-nodes-0-0-0 -n cloud-clickhouse -- bashkubectl exec -it chi-clickhouse-snuba-ck-nodes-0-1-0 -n cloud-clickhouse -- bashkubectl exec -it chi-clickhouse-snuba-ck-nodes-1-0-0 -n cloud-clickhouse -- bashkubectl exec -it chi-clickhouse-snuba-ck-nodes-1-1-0 -n cloud-clickhouse -- bashkubectl exec -it chi-clickhouse-snuba-ck-nodes-2-0-0 -n cloud-clickhouse -- bashkubectl exec -it chi-clickhouse-snuba-ck-nodes-2-1-0 -n cloud-clickhouse -- bash
我们直接通过终端分别进入这 6 个 pod。然后进行测试：
clickhouse-client --multiline -u username -h ip --password passowrd# clickhouse-client -m创建分布式数据库查看system.clusters
select * from system.clusters;
2.创建名为test的数据库
create database test on cluster "snuba-ck-nodes";# 删除：drop database test on cluster "snuba-ck-nodes";在各个节点查看，都已存在test数据库。
show databases;创建本地表(ReplicatedMergeTree)建表语句如下：
在集群中各个节点test数据库中创建t_local本地表，采用ReplicatedMergeTree表引擎，接受两个参数：
zoo_path— zookeeper 中表的路径，针对表同一个分片的不同副本，定义相同路径。"/clickhouse/tables/{shard}/test/t_local"
replica_name— zookeeper 中表的副本名称CREATE TABLE test.t_local on cluster "snuba-ck-nodes"( EventDate DateTime, CounterID UInt32, UserID UInt32)ENGINE = ReplicatedMergeTree("/clickhouse/tables/{shard}/test/t_local", "{replica}")PARTITION BY toYYYYMM(EventDate)ORDER BY (CounterID, EventDate, intHash32(UserID))SAMPLE BY intHash32(UserID);宏(macros)占位符：建表语句参数包含的宏替换占位符(如：{replica})。会被替换为配置文件里 macros 部分的值。
查看集群中 clickhouse 分片&副本节点configmap：
kubectl get configmap -n cloud-clickhouse | grep clickhouseNAME DATA AGEchi-clickhouse-common-configd 6 20hchi-clickhouse-common-usersd 6 20hchi-clickhouse-deploy-confd-snuba-ck-nodes-0-0 2 20hchi-clickhouse-deploy-confd-snuba-ck-nodes-0-1 2 20hchi-clickhouse-deploy-confd-snuba-ck-nodes-1-0 2 20hchi-clickhouse-deploy-confd-snuba-ck-nodes-1-1 2 20hchi-clickhouse-deploy-confd-snuba-ck-nodes-2-0 2 19hchi-clickhouse-deploy-confd-snuba-ck-nodes-2-1 2 19h
查看节点配置值：
kubectl describe configmap chi-clickhouse-deploy-confd-snuba-ck-nodes-0-0 -n cloud-clickhouse创建对应的分布式表(Distributed)
CREATE TABLE test.t_dist on cluster "snuba-ck-nodes"( EventDate DateTime, CounterID UInt32, UserID UInt32)ENGINE = Distributed("snuba-ck-nodes", test, t_local, rand());# drop table test.t_dist on cluster "snuba-ck-nodes";这里，Distributed 引擎的所用的四个参数：
cluster - 服务为配置中的集群名(snuba-ck-nodes)database - 远程数据库名(test)table - 远程数据表名(t_local)sharding_key - (可选) 分片key(CounterID/rand())查看相关表，如：
use test;show tables;# t_dist# t_local
通过分布式表插入几条数据：
# 插入INSERT INTO test.t_dist VALUES ("2022-12-16 00:00:00", 1, 1),("2023-01-01 00:00:00",2, 2),("2023-02-01 00:00:00",3, 3);任一节点查询数据：
select * from test.t_dist;实战，为 Snuba 引擎提供 ClickHouse PaaS拆解与分析 Sentry Helm Charts
在我们迁移到 Kubernetes Operator 之前，我们先拆解与分析下 sentry-charts 中自带的 clickhouse & zookeeper charts。
非官方 Sentry Helm Charts：
https://github.com/sentry-kubernetes/charts他的Chart.yaml如下：
apiVersion: v2appVersion: 22.11.0dependencies:- condition: sourcemaps.enabled name: memcached repository: https://charts.bitnami.com/bitnami version: 6.1.5- condition: redis.enabled name: redis repository: https://charts.bitnami.com/bitnami version: 16.12.1- condition: kafka.enabled name: kafka repository: https://charts.bitnami.com/bitnami version: 16.3.2- condition: clickhouse.enabled name: clickhouse repository: https://sentry-kubernetes.github.io/charts version: 3.2.0- condition: zookeeper.enabled name: zookeeper repository: https://charts.bitnami.com/bitnami version: 9.0.0- alias: rabbitmq condition: rabbitmq.enabled name: rabbitmq repository: https://charts.bitnami.com/bitnami version: 8.32.2- condition: postgresql.enabled name: postgresql repository: https://charts.bitnami.com/bitnami version: 10.16.2- condition: nginx.enabled name: nginx repository: https://charts.bitnami.com/bitnami version: 12.0.4description: A Helm chart for Kubernetesmaintainers:- name: sentry-kubernetesname: sentrytype: applicationversion: 17.9.0
这个 sentry-charts 将所有中间件 helm charts 耦合依赖在一起部署，不适合 sentry 微服务 & 中间件集群扩展。更高级的做法是每个中间件拥有定制的 Kubernetes Operator(如：clickhouse-operator) & 独立的 K8S 集群，形成中间件 PaaS 平台对外提供服务。
这里我们拆分中间件 charts 到独立的 namespace 或单独的集群运维。设计为：
ZooKeeper 命名空间：cloud-zookeeper-paasClickHouse 命名空间：cloud-clickhouse-paas独立部署 ZooKeeper Helm Chart这里 zookeeper chart 采用的是 bitnami/zookeeper，他的仓库地址如下：
https://github.com/bitnami/charts/tree/master/bitnami/zookeeperhttps://github.com/bitnami/containers/tree/main/bitnami/zookeeperZooKeeper Operator 会在后续文章专项讨论。创建命名空间：kubectl create ns cloud-zookeeper-paas简单定制下values.yaml：
# 暴露下 prometheus 监控所需的服务metrics: containerPort: 9141 enabled: true........service: annotations: {} clusterIP: "" disableBaseClientPort: false externalTrafficPolicy: Cluster extraPorts: [] headless: annotations: {} publishNotReadyAddresses: true loadBalancerIP: "" loadBalancerSourceRanges: [] nodePorts: client: "" tls: "" ports: client: 2181 election: 3888 follower: 2888 tls: 3181 sessionAffinity: None type: ClusterIP注意：在使用支持外部负载均衡器的云提供商的服务时，需设置 Sevice 的 type 的值为 "LoadBalancer"， 将为 Service 提供负载均衡器。来自外部负载均衡器的流量将直接重定向到后端 Pod 上，不过实际它们是如何工作的，这要依赖于云提供商。
helm 部署：helm install zookeeper ./zookeeper -f values.yaml -n cloud-zookeeper-paas
集群内，可使用zookeeper.cloud-zookeeper-paas.svc.cluster.local:2181对外提供服务。
zkCli 连接 ZooKeeper：export POD_NAME=$(kubectl get pods --namespace cloud-zookeeper-paas -l "app.kubernetes.io/name=zookeeper,app.kubernetes.io/instance=zookeeper,app.kubernetes.io/compnotallow=zookeeper" -o jsnotallow="{.items[0].metadata.name}")kubectl -n cloud-zookeeper-paas exec -it $POD_NAME -- zkCli.sh# test[zk: localhost:2181(CONNECTED) 0] ls /[zookeeper][zk: localhost:2181(CONNECTED) 1] ls /zookeeper[config, quota][zk: localhost:2181(CONNECTED) 2] quit# 外部访问# kubectl port-forward --namespace cloud-zookeeper-paas svc/zookeeper 2181: & zkCli.sh 127.0.0.1:2181查看zoo.cfgkubectl -n cloud-zookeeper-paas exec -it $POD_NAME -- cat /opt/bitnami/zookeeper/conf/zoo.cfg
# The number of milliseconds of each ticktickTime=2000# The number of ticks that the initial# synchronization phase can takeinitLimit=10# The number of ticks that can pass between# sending a request and getting an acknowledgementsyncLimit=5# the directory where the snapshot is stored.# do not use /tmp for storage, /tmp here is just# example sakes.dataDir=/bitnami/zookeeper/data# the port at which the clients will connectclientPort=2181# the maximum number of client connections.# increase this if you need to handle more clientsmaxClientCnxns=60## Be sure to read the maintenance section of the# administrator guide before turning on autopurge.## https://zookeeper.apache.org/doc/current/zookeeperAdmin.html#sc_maintenance## The number of snapshots to retain in dataDirautopurge.snapRetainCount=3# Purge task interval in hours# Set to "0" to disable auto purge featureautopurge.purgeInterval=0## Metrics Providers## https://prometheus.io Metrics ExportermetricsProvider.className=org.apache.zookeeper.metrics.prometheus.PrometheusMetricsProvider#metricsProvider.httpHost=0.0.0.0metricsProvider.httpPort=9141metricsProvider.exportJvmInfo=truepreAllocSize=65536snapCount=100000maxCnxns=0recnotallow=falsequorumListenOnAllIPs=false4lw.commands.whitelist=srvr, mntr, ruokmaxSessinotallow=40000admin.serverPort=8080admin.enableServer=trueserver.1=zookeeper-0.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181server.2=zookeeper-1.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181server.3=zookeeper-2.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181独立部署 ClickHouse Helm Chart
这里 clickhouse chart 采用的是 sentry-kubernetes/charts 自己维护的一个版本：
sentry snuba 目前对于 clickhouse 21.x 等以上版本支持的并不友好，这里的镜像版本是yandex/clickhouse-server:20.8.19.4。https://github.com/sentry-kubernetes/charts/tree/develop/clickhouseClickHouse Operator + ClickHouse Keeper 会在后续文章专项讨论。这个自带的 clickhouse-charts 存在些问题，Service 部分需简单修改下允许配置 "type:LoadBalancer" or "type:NodePort"。
注意：在使用支持外部负载均衡器的云提供商的服务时，需设置 Sevice 的 type 的值为 "LoadBalancer"， 将为 Service 提供负载均衡器。来自外部负载均衡器的流量将直接重定向到后端 Pod 上，不过实际它们是如何工作的，这要依赖于云提供商。
创建命名空间：kubectl create ns cloud-clickhouse-paas简单定制下values.yaml：
注意上面zoo.cfg的 3 个 zookeeper 实例的地址：
server.1=zookeeper-0.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181server.2=zookeeper-1.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181server.3=zookeeper-2.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local:2888:3888;2181
# 修改 zookeeper_serversclickhouse: configmap: zookeeper_servers: config: - hostTemplate: "zookeeper-0.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local" index: clickhouse port: "2181" - hostTemplate: "zookeeper-1.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local" index: clickhouse port: "2181" - hostTemplate: "zookeeper-2.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local" index: clickhouse port: "2181" enabled: true operation_timeout_ms: "10000" session_timeout_ms: "30000"# 暴露下 prometheus 监控所需的服务metrics: enabled: true
当然这里也可以不用 Headless Service，因为是同一个集群的不同 namespace 的内部访问，所以也可简单填入 ClusterIP 类型 Sevice：
# 修改 zookeeper_serversclickhouse: configmap: zookeeper_servers: config: - hostTemplate: "zookeeper.cloud-zookeeper-paas.svc.cluster.local" index: clickhouse port: "2181" enabled: true operation_timeout_ms: "10000" session_timeout_ms: "30000"# 暴露下 prometheus 监控所需的服务metrics: enabled: truehelm 部署：
helm install clickhouse ./clickhouse -f values.yaml -n cloud-clickhouse-paas连接 clickhouse
kubectl -n cloud-clickhouse-paas exec -it clickhouse-0 -- clickhouse-client --multiline --host="clickhouse-1.clickhouse-headless.cloud-clickhouse-paas"验证集群
show databases;select * from system.clusters;select * from system.zookeeper where path = "/clickhouse";当前 ClickHouse 集群的 ConfigMap
kubectl get configmap -n cloud-clickhouse-paas | grep clickhouse
clickhouse-config 1 28hclickhouse-metrica 1 28hclickhouse-users 1 28hclickhouse-config(config.xml)
clickhouse-metrica(metrica.xml)/var/lib/clickhouse/ /var/lib/clickhouse/tmp/ /var/lib/clickhouse/user_files/ /var/lib/clickhouse/format_schemas/ /etc/clickhouse-server/metrica.d/metrica.xml users.xml clickhouse 0.0.0.0 8123 9000 9009 4096 3 100 8589934592 5368709120 UTC 022 false 3600 3600 60 1 system query_log
toYYYYMM(event_date) 7500 system query_thread_log
toYYYYMM(event_date) 7500 /clickhouse/task_queue/ddl trace /var/log/clickhouse-server/clickhouse-server.log /var/log/clickhouse-server/clickhouse-server.err.log 1000M 10
clickhouse-users(users.xml)zookeeper-0.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local 2181 zookeeper-1.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local 2181 zookeeper-2.zookeeper-headless.cloud-zookeeper-paas.svc.cluster.local 2181 30000 10000 true clickhouse-0.clickhouse-headless.cloud-clickhouse-paas.svc.cluster.local 9000 default true true clickhouse-1.clickhouse-headless.cloud-clickhouse-paas.svc.cluster.local 9000 default true true clickhouse-2.clickhouse-headless.cloud-clickhouse-paas.svc.cluster.local 9000 default true
Sentry Helm Charts 定制接入 ClickHouse PaaS, 单集群多节点
我们简单修改values.yml
禁用 sentry-charts 中的 clickHouse & zookeeperclickhouse: enabled: falsezookeeper: enabled: false修改externalClickhouse
externalClickhouse: database: default host: "clickhouse.cloud-clickhouse-paas.svc.cluster.local" httpPort: 8123 password: "" singleNode: false clusterName: "clickhouse" tcpPort: 9000 username: default
注意：
这里只是简单的集群内部接入1个多节点分片集群，而 Snuba 系统的设计是允许你接入多个 ClickHouse多节点多分片多副本集群，将多个 Schema 分散到不同的集群，从而实现超大规模吞吐。因为是同一个集群的不同 namespace 的内部访问，所以这里简单填入类型为 ClusterIP Sevice 即可。注意这里singleNode要设置成false。因为我们是多节点，同时我们需要提供clusterName：源码分析：这将用于确定：以及确定来使用不同的 ClickHouse Table Engines 等。当然，ClickHouse 本身是一个单独的技术方向，这里就不展开讨论了。将运行哪些迁移（仅本地或本地和分布式表）
查询中的差异 - 例如是否选择了 _local 或 _dist 表
部署helm install sentry ./sentry -f values.yaml -n sentry验证 _local 与 _dist 表以及 system.zookeeper
kubectl -n cloud-clickhouse-paas exec -it clickhouse-0 -- clickhouse-client --multiline --host="clickhouse-1.clickhouse-headless.cloud-clickhouse-paas"show databases;show tables;select * from system.zookeeper where path = "/clickhouse";高级部分 & 超大规模吞吐接入 ClickHouse多集群/多节点/多分片/多副本的中间件 PaaS
独立部署多套 VKE LoadBlancer+ VKE K8S Cluster + ZooKeeper-Operator + ClickHouse-Operator，分散 Schema 到不同的集群以及多节点分片。
分析 Snuba 系统设计查看测试用例源码，了解系统设计与高阶配置关于针对 ClickHouse 集群各个分片、副本之间的读写负载均衡、连接池等问题。Snuba 在系统设计、代码层面部分就已经做了充分的考虑以及优化。
关于 ClickHouse Operator 独立的多个云原生编排集群以及 Snuba 系统设计等高级部分会在 VIP 专栏直播课单独讲解。
