一、概述
正值 TiDB v6.0 (DMR) 发布，本地虚拟机体验一把 TiSpark ，感受一下 TiDB 的强大！
二、TiSpark 简述
2.1 TiSpark 是什么？
TiSpark 是 PingCAP 为解决用户复杂 OLAP 需求而推出的产品。TiSpark 本身是 Spark 的一个扩展，利用了 Spark 提供的 Extensions 机制。
2.2 依赖其他组件吗？
TiSpark 依赖于 TiKV 集群和 Placement Driver (PD)，也需要你搭建一个 Spark 集群（已有或全新搭建）。
2.3 架构在 TiDB 中的位置？
TiSpark 是将 Spark SQL 直接运行在分布式存储引擎 TiKV 上的 OLAP 解决方案。
2.4 TiSpark 能做什么？
- 使用 TiSpark 进行数据分析和 ETL (Extraction-Transformation-Loading 的缩写,中文名为数据抽取、转换和加载）。
2.5 TiSpark 的优势是什么？
-
简化了系统架构和运维
- 从数据集群的角度看，TiSpark + TiDB 可以让用户无需进行脆弱和难以维护的 ETL，直接在同一个平台进行事务和分析两种工作，简化了系统架构和运维。
- 分布式写入 TiKV
- 相比使用 Spark 结合 JDBC 的方式写入 TiDB，分布式写入 TiKV 可以实现事务（要么全部数据写入成功，要么全部都写入失败）。
- 使用 Spark 生态圈提供的多种工具进行数据处理
- 用户借助 TiSpark 项目可以在 TiDB 上使用 Spark 生态圈提供的多种工具进行数据处理。例如，使用 TiSpark 进行数据分析和 ETL；使用 TiKV 作为机器学习的数据源；借助调度系统产生定时报表等等。
- 支持鉴权和授权（TiSpark 2.5.0 版本以上）
- 提高集群安全性、更好的优化 Tispark 的读写请求逻辑，拆分业务模块提高性能。
三、安装 TiDB 集群和 TiSpark
环境： Centos 7.3 虚拟机 2C 4G TiDB V6.0
版本说明
TiDB V6.0.0
TiSpark V2.4.1 （tispark-v2.4.1-any-any.tar.gz）
Spark V2.4.3 (spark-v2.4.3-any-any.tar.gz)
准备工作
- Centos 7.3
使用 Virtualbox 导入介质配置虚拟机 Centos 7.3 大约1分钟配置完毕！（这里不再详细说明）
- ssh 设置
调大 sshd 服务的连接数限制
修改 /etc/ssh/sshd_config 将 MaxSessions 调至 100
sed -i 's/#MaxSessions.*/MaxSessions 100/g' /etc/ssh/sshd_config
# 重启 sshd 服务
systemctl restart sshd
- 安装tiup ,cluster组件
# 安装tiup ,cluster组件
curl --proto '=https' --tlsv1.2 -sSf https://tiup-mirrors.pingcap.com/install.sh | sh
source .bash_profile
tiup cluster
# 升级
tiup update --self && tiup update cluster
3.1 安装 TiDB 集群 Cluster111 和监控
3.1.1 Cluster111 拓扑
# 参考 https://github.com/pingcap/docs-cn/blob/master/config-templates/complex-mini.yaml
# cluster111.yaml
# # Global variables are applied to all deployments and used as the default value of
# # the deployments if a specific deployment value is missing.
global:
user: "tidb"
ssh_port: 22
deploy_dir: "/tidb-deploy"
data_dir: "/tidb-data"
# # Monitored variables are applied to all the machines.
monitored:
node_exporter_port: 9100
blackbox_exporter_port: 9115
# deploy_dir: "/tidb-deploy/monitored-9100"
# data_dir: "/tidb-data/monitored-9100"
# log_dir: "/tidb-deploy/monitored-9100/log"
# # Server configs are used to specify the runtime configuration of TiDB components.
# # All configuration items can be found in TiDB docs:
# # - TiDB: https://pingcap.com/docs/stable/reference/configuration/tidb-server/configuration-file/
# # - TiKV: https://pingcap.com/docs/stable/reference/configuration/tikv-server/configuration-file/
# # - PD: https://pingcap.com/docs/stable/reference/configuration/pd-server/configuration-file/
# # All configuration items use points to represent the hierarchy, e.g:
# # readpool.storage.use-unified-pool
# #
# # You can overwrite this configuration via the instance-level `config` field.
server_configs:
tidb:
log.slow-threshold: 300
binlog.enable: false
binlog.ignore-error: false
tikv:
# server.grpc-concurrency: 4
# raftstore.apply-pool-size: 2
# raftstore.store-pool-size: 2
# rocksdb.max-sub-compactions: 1
# storage.block-cache.capacity: "16GB"
# readpool.unified.max-thread-count: 12
readpool.storage.use-unified-pool: false
readpool.coprocessor.use-unified-pool: true
pd:
schedule.leader-schedule-limit: 4
schedule.region-schedule-limit: 2048
schedule.replica-schedule-limit: 64
pd_servers:
- host: 10.0.2.15
# ssh_port: 22
# name: "pd-1"
# client_port: 2379
# peer_port: 2380
# deploy_dir: "/tidb-deploy/pd-2379"
# data_dir: "/tidb-data/pd-2379"
# log_dir: "/tidb-deploy/pd-2379/log"
# numa_node: "0,1"
# # The following configs are used to overwrite the `server_configs.pd` values.
# config:
# schedule.max-merge-region-size: 20
# schedule.max-merge-region-keys: 200000
tidb_servers:
- host: 10.0.2.15
# ssh_port: 22
# port: 4000
# status_port: 10080
# deploy_dir: "/tidb-deploy/tidb-4000"
# log_dir: "/tidb-deploy/tidb-4000/log"
# numa_node: "0,1"
# # The following configs are used to overwrite the `server_configs.tidb` values.
# config:
# log.slow-query-file: tidb-slow-overwrited.log
tikv_servers:
- host: 10.0.2.15
# ssh_port: 22
# port: 20160
# status_port: 20180
# deploy_dir: "/tidb-deploy/tikv-20160"
# data_dir: "/tidb-data/tikv-20160"
# log_dir: "/tidb-deploy/tikv-20160/log"
# numa_node: "0,1"
# # The following configs are used to overwrite the `server_configs.tikv` values.
# config:
# server.grpc-concurrency: 4
# server.labels: { zone: "zone1", dc: "dc1", host: "host1" }
monitoring_servers:
- host: 10.0.2.15
# ssh_port: 22
# port: 9090
# deploy_dir: "/tidb-deploy/prometheus-8249"
# data_dir: "/tidb-data/prometheus-8249"
# log_dir: "/tidb-deploy/prometheus-8249/log"
grafana_servers:
- host: 10.0.2.15
# port: 3000
# deploy_dir: /tidb-deploy/grafana-3000
alertmanager_servers:
- host: 10.0.2.15
# ssh_port: 22
# web_port: 9093
# cluster_port: 9094
# deploy_dir: "/tidb-deploy/alertmanager-9093"
# data_dir: "/tidb-data/alertmanager-9093"
# log_dir: "/tidb-deploy/alertmanager-9093/log"
3.1.2 安装 Cluster1111
- 查看 TiUP 支持的最新可用版本 （选择 v6.0.0）
tiup list tidb
- 安装cluster111
# tiup cluster deploy <cluster-name> <tidb-version> ./topo.yaml --user root -p
tiup cluster check ./cluster111.yml --user root -p
tiup cluster deploy cluster111 v6.0.0 ./cluster111.yml --user root -p
# 会提示输入密码
# 提示输入y/n
# 提示下面信息表示成功
# “Cluster `cluster111` deployed successfully, you can start it with command: `tiup cluster start cluster111 --init`”
- 查看集群
```shell
# 查看集群
tiup cluster list
# 初始化集群
tiup cluster start cluster111 --init
# 查看集群
tiup cluster display cluster111
注意：
- 使用
tiup cluster start cluster111 --init将给root用户生成随机密码- 如果不加--init 将不生成随机密码
- 演示期间把密码修改为123456。ALTER USER 'root' IDENTIFIED BY '123456';
3.1.3 查看 Dashboard
- 虚拟机设置端口转发 4000->4000,2379,2379
- http://127.0.0.1:2379/dashboard root=123456
3.1.4 Mysql 客户端连接 TiDB
mysql -h127.0.0.1 -uroot -P4000 -p
3.2 没有 Spark 集群的环境下安装 TiSpark
推荐使用 Spark Standalone 方式部署即扩容的方式安装 TiSpark。在安装 TiDB 集群的时候，一同安装 TiSpark 也是可以的（通过配置集群拓扑文件）。
3.2.1 TiSpark 最小拓扑
参考 https://github.com/pingcap/docs-cn/blob/master/config-templates/complex-tispark.yaml
仅保留 tispark的部分
# cluster111-v6.0.0-tispark.yaml
tispark_masters:
- host: 10.0.2.15
# ssh_port: 22
# port: 7077
# web_port: 8080
# deploy_dir: "/tidb-deploy/tispark-master-7077"
# java_home: "/usr/local/bin/java-1.8.0"
# spark_config:
# spark.driver.memory: "2g"
# spark.eventLog.enabled: "False"
# spark.tispark.grpc.framesize: 268435456
# spark.tispark.grpc.timeout_in_sec: 100
# spark.tispark.meta.reload_period_in_sec: 60
# spark.tispark.request.command.priority: "Low"
# spark.tispark.table.scan_concurrency: 256
# spark_env:
# SPARK_EXECUTOR_CORES: 5
# SPARK_EXECUTOR_MEMORY: "10g"
# SPARK_WORKER_CORES: 5
# SPARK_WORKER_MEMORY: "10g"
# NOTE: multiple worker nodes on the same host is not supported by Spark
tispark_workers:
- host: 10.0.2.15
# ssh_port: 22
# port: 7078
# web_port: 8081
# deploy_dir: "/tidb-deploy/tispark-worker-7078"
# java_home: "/usr/local/bin/java-1.8.0"
3.2.2 安装 TiSpark
- 安装 openjdk8
yum -y install java-1.8.0-openjdk java-1.8.0-openjdk-devel
编辑 ~/.bashrc 文件 结尾追加如下内容：
export JAVA_HOME=$(dirname $(dirname $(readlink $(readlink $(which javac)))))
export PATH=\$PATH:\$JAVA_HOME/bin
export CLASSPATH=.:\$JAVA_HOME/jre/lib:\$JAVA_HOME/lib:\$JAVA_HOME/lib/tools.jar
- 验证 jdk
[root@tispark vagrant]# java -version
openjdk version "1.8.0_332"
OpenJDK Runtime Environment (build 1.8.0_332-b09)
OpenJDK 64-Bit Server VM (build 25.332-b09, mixed mode)
- 扩容的方式安装 TiSpark
tiup cluster scale-out cluster111 ./cluster111-v6.0.0-tispark.yaml -uroot -p
输入 y 继续安装
上图中有 2 个信息：
- 扩容安装 1 个Tispark master 和1个Tispark work 成功
- 扩容的 2 个节点启动失败（因为需要配置 TiSpark 和启动 Spark 服务端）
下图证实了这点。
3.3 已有 Spark 集群的环境下安装 TiSpark
如果在已有 Spark 集群上运行 TiSpark，无需重启集群。可以使用 Spark 的
--jars参数将 TiSpark 作为依赖引入
# spark-shell --jars $TISPARK_FOLDER/tispark-${name_with_version}.jar
/tidb-deploy/tispark-master-7077/bin/spark-shell --jars /tidb-deploy/tispark-master-7077/jars/tispark-assembly-2.4.1.jar
四、启动Spark和TiSpark
4.1 配置 TiSpark
tispark-master中的 /tidb-deploy/tispark-master-7077/conf/spark-defaults.conf 中增加如下配置：
# sql扩展类 spark.sql.extensions org.apache.spark.sql.TiExtensions # master节点 spark.master spark://10.0.2.15:7077 # pd节点 多个pd用逗号隔开 如：10.16.20.1:2379,10.16.20.2:2379,10.16.20.3:2379 spark.tispark.pd.addresses 10.0.2.15:2379
4.2 启动 Spark Standalone 方式部署的本地 Spark 集群
已有 Spark 集群可略过此部分！
/tidb-deploy/tispark-master-7077/sbin/start-all.sh
也可以分开启动：先启动 Master，再启动 Slave
启动之后，过几秒钟 验证 TiSpark 是否启动
tiup cluster display cluster111
如果没有启动，可以手动启动 TiSpark 节点
# 手动启动 tispark节点
tiup cluster start cluster111 -N 10.0.2.15:707,10.0.2.15:7078
五、测试
5.1 导入样例数据
curl -L http://download.pingcap.org/tispark-sample-data.tar.gz -o tispark-sample-data.tar.gz
tar -zxvf tispark-sample-data.tar.gz
cd tispark-sample-data
# 导入示例数据
mysql --local-infile=1 -h 127.0.0.1 -P 4000 -u root < dss.ddl
登录 TiDB 并验证数据
包含 TPCH_001 库及以下表：
验证数据量
select c.*,1 as ordernum from (
select concat(
'select \'',
TABLE_name,
'\' tablename , count(*) ct from ',
TABLE_name,
' union all'
) as sqlstr from information_schema.tables
where TABLE_SCHEMA='TPCH_001') c union all
select 'select 0,0 from dual order by ct desc',0 from dual order by ordernum desc ;
-- 执行上面sql查出来的sql语句
select 'CUSTOMER' tablename , count(*) ct from CUSTOMER union all
select 'NATION' tablename , count(*) ct from NATION union all
select 'REGION' tablename , count(*) ct from REGION union all
select 'PART' tablename , count(*) ct from PART union all
select 'SUPPLIER' tablename , count(*) ct from SUPPLIER union all
select 'PARTSUPP' tablename , count(*) ct from PARTSUPP union all
select 'ORDERS' tablename , count(*) ct from ORDERS union all
select 'LINEITEM' tablename , count(*) ct from LINEITEM union all
select 0,0 from dual order by ct desc
5.2 spark-shell 像使用原生 Spark 一样
# 启动spark-shell
/tidb-deploy/tispark-master-7077/bin/spark-shell
scala> spark.sql("use tpch_001")
res3: org.apache.spark.sql.DataFrame = []
scala> spark.sql("select count(*) from lineitem").show
+--------+
|count(1)|
+--------+
| 60175|
+--------+
scala> spark.sql(
| """select
| | l_returnflag,
| | l_linestatus,
| | sum(l_quantity) as sum_qty,
| | sum(l_extendedprice) as sum_base_price,
| | sum(l_extendedprice * (1 - l_discount)) as sum_disc_price,
| | sum(l_extendedprice * (1 - l_discount) * (1 + l_tax)) as sum_charge,
| | avg(l_quantity) as avg_qty,
| | avg(l_extendedprice) as avg_price,
| | avg(l_discount) as avg_disc,
| | count(*) as count_order
| |from
| | lineitem
| |where
| | l_shipdate <= date '1998-12-01' - interval '90' day
| |group by
| | l_returnflag,
| | l_linestatus
| |order by
| | l_returnflag,
| | l_linestatus
| """.stripMargin).show
# 如下图结果
更多样例请参考 pingcap/tispark-test
5.3 spark-sql 像使用 sql 一样
# 启动spark-sql
cd /tidb-deploy/tispark-master-7077/bin/
./spark-sql
# web ui 需要宿主机端口转发到虚拟机4041->4041
# http://127.0.0.1:4041/#spark
# http://127.0.0.1:4040/#spark-sql
# http://127.0.0.1:8080/#Spark Master
示例如下：
# show databases;
spark-sql> show databases;
22/04/09 16:36:02 INFO PDClient: Switched to new leader: [leaderInfo: 10.0.2.15:2379]
22/04/09 16:36:08 INFO ReflectionUtil$: tispark class url: file:/tidb-deploy/tispark-master-7077/jars/tispark-assembly-2.4.1.jar
22/04/09 16:36:08 INFO ReflectionUtil$: spark wrapper class url: jar:file:/tidb-deploy/tispark-master-7077/jars/tispark-assembly-2.4.1.jar!/resources/spark-wrapper-spark-2_4/
22/04/09 16:36:08 INFO HiveMetaStore: 0: get_databases: *
22/04/09 16:36:08 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_databases: *
22/04/09 16:36:09 INFO CodeGenerator: Code generated in 864.536604 ms
default
test
tpch_001
mysql
Time taken: 20.766 seconds, Fetched 4 row(s)
22/04/09 16:36:09 INFO SparkSQLCLIDriver: Time taken: 20.766 seconds, Fetched 4 row(s)
# use tpch_001;
spark-sql> use tpch_001;
22/04/09 16:38:18 INFO HiveMetaStore: 0: get_database: tpch_001
22/04/09 16:38:18 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_database: tpch_001
Time taken: 0.353 seconds
22/04/09 16:38:18 INFO SparkSQLCLIDriver: Time taken: 0.353 seconds
# show tables;
spark-sql> show tables;
22/04/09 16:38:21 INFO HiveMetaStore: 0: get_database: tpch_001
22/04/09 16:38:21 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_database: tpch_001
22/04/09 16:38:22 INFO HiveMetaStore: 0: get_database: global_temp
22/04/09 16:38:22 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_database: global_temp
22/04/09 16:38:22 INFO HiveMetaStore: 0: get_database: default
22/04/09 16:38:22 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_database: default
22/04/09 16:38:22 INFO HiveMetaStore: 0: get_database: default
22/04/09 16:38:22 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_database: default
22/04/09 16:38:22 INFO HiveMetaStore: 0: get_tables: db=default pat=*
22/04/09 16:38:22 INFO audit: ugi=root ip=unknown-ip-addr cmd=get_tables: db=default pat=*
22/04/09 16:38:22 INFO CodeGenerator: Code generated in 47.307373 ms
tpch_001 customer false
tpch_001 lineitem false
tpch_001 nation false
tpch_001 orders false
tpch_001 part false
tpch_001 partsupp false
tpch_001 region false
tpch_001 supplier false
Time taken: 1.038 seconds, Fetched 8 row(s)
22/04/09 16:38:22 INFO SparkSQLCLIDriver: Time taken: 1.038 seconds, Fetched 8 row(s)
# select count(*) from lineitem
spark-sql> select count(*) from lineitem;
# 长sql语句
spark-sql> select
l_returnflag,
l_linestatus,
sum(l_quantity) as sum_qty,
sum(l_extendedprice) as sum_base_price,
sum(l_extendedprice * (1 - l_discount)) as sum_disc_price,
sum(l_extendedprice * (1 - l_discount) * (1 + l_tax)) as sum_charge,
avg(l_quantity) as avg_qty,
avg(l_extendedprice) as avg_price,
avg(l_discount) as avg_disc,
count(*) as count_order
from
lineitem
where
l_shipdate <= '1998-09-02'
group by
l_returnflag,
l_linestatus
order by
l_returnflag,
l_linestatus;
5.4 TiSpark 写数据
5.4.1 配置 allow_spark_sql
在通过spark-sql写数据时提示 “SparkSQL entry for tispark write is disabled. Set spark.tispark.write.allow_spark_sql to enable.”。
解决方式： 需要在 conf/spark-defaults.conf 里面配置一下下面的参数
vim /tidb-deploy/tispark-master-7077/conf/spark-defaults.conf #增加 配置 spark.tispark.write.allow_spark_sql true
5.4.2 写数据测试
前提准备
- 必须先创建数据表结构
- 目标 TiDB 表必须有主键
-- 源表（没有明示主键）
CREATE TABLE `CUSTOMER` (
`C_CUSTKEY` int(11) NOT NULL,
`C_NAME` varchar(25) NOT NULL,
`C_ADDRESS` varchar(40) NOT NULL,
`C_NATIONKEY` int(11) NOT NULL,
`C_PHONE` char(15) NOT NULL,
`C_ACCTBAL` decimal(15,2) NOT NULL,
`C_MKTSEGMENT` char(10) NOT NULL,
`C_COMMENT` varchar(117) NOT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin;
-- 目标表 CUSTOMER_2
CREATE TABLE `CUSTOMER_2` (
`C_CUSTKEY` int(11) NOT NULL,
`C_NAME` varchar(25) NOT NULL,
`C_ADDRESS` varchar(40) NOT NULL,
`C_NATIONKEY` int(11) NOT NULL,
`C_PHONE` char(15) NOT NULL,
`C_ACCTBAL` decimal(15,2) NOT NULL,
`C_MKTSEGMENT` char(10) NOT NULL,
`C_COMMENT` varchar(117) NOT NULL,
PRIMARY key (`C_CUSTKEY`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin
TiSpark 写 TiDB
完整代码如下
package org.example.demo.spark;
import org.apache.spark.SparkConf;
import org.apache.spark.sql.SaveMode;
import org.apache.spark.sql.SparkSession;
import java.util.HashMap;
import java.util.Map;
public class SparkDemo {
public static void main(String[] args) {
String pd_addr = "10.0.2.15:2379";
String tidb_addr = "10.0.2.15";
SparkConf conf = new SparkConf()
.set("spark.sql.extensions", "org.apache.spark.sql.TiExtensions")
.set("spark.tispark.pd.addresses", pd_addr);
SparkSession spark = SparkSession
.builder()
.appName("SparkDemo")
.config(conf)
.getOrCreate();
try{
String source_db_name = "TPCH_001";
String source_table_name = "CUSTOMER";
String target_db_name = "TPCH_001";
String target_table_name = "CUSTOMER_2";
String username = "root";
String password = "";
//通过 TiSpark 将 DataFrame 批量写入 TiDB
Map<String, String> tiOptionMap = new HashMap<String, String>();
tiOptionMap.put("tidb.addr", tidb_addr);
tiOptionMap.put("tidb.port", "4000");
tiOptionMap.put("tidb.user", username);
tiOptionMap.put("tidb.password", password);
tiOptionMap.put("replace", "true");
tiOptionMap.put("spark.tispark.pd.addresses", pd_addr);
spark.sql("use "+source_db_name);
String source_sql = "select * from "+source_table_name;
spark.sql(source_sql)
.write()
.format("tidb")
.options(tiOptionMap)
.option("database", target_db_name)
.option("table", target_table_name)
.mode(SaveMode.Append)
.save();
}catch (Exception e){
e.printStackTrace();
}
spark.stop();
}
}
5.4.3 执行 Spark-submit
MASTER=spark://10.0.2.15:7077 ./bin/spark-submit --class org.example.demo.spark.SparkDemo /usr/local0/webserver/tispark/sparkdemo1-1.0.jar
5.4.4 数据验证
select 'CUSTOMER (源表）' as '表名',count(*) as '记录数' from CUSTOMER
union all
select 'CUSTOMER_2 （目标表）',count(*) from CUSTOMER_2;
数据复合预期！
注意：
1、如果出现错误提示 “Initial job has not accepted any resources; check your cluster UI to ensure that workers are registered and have sufficient resources” 解决：
1）确保 tispark-worker 已经激活；
2）再检查资源：打开 http://127.0.0.1:8080/ 发现 running 发现进程过多，可能导致资源不足，随即关闭进程，重新执行。
2、更多示例 spark-sql 的示例
https://github.com/pingcap/tispark-test/tree/master/tpch/mysql
六、总结
1、本次实验使用版本为 TiSpark 2.4.1 和 Spark 2.4.3 ，如果对新版的 TiSpark 比较感兴趣可以关注后面的文章。
2、TiSpark 在写 TiDB 的时候注意下面几点：
1）目标表必须存在；
2）目标表有明示主键（不算 _tidb_rowid）。
3、一路过来，体验使用 TiSpark 还算顺利，给 PingCAP 的同学们点个赞！
4、更多 TiSpark 的特性还需要继续探索！
5、如果文章中的表述有不当的地方、请私信留言！
谢谢！
参考
https://docs.pingcap.com/zh/tidb/v6.0/tispark-overview#tispark-用户指南 https://docs.pingcap.com/zh/tidb/v6.0/get-started-with-tispark#tispark-快速上手 https://docs.pingcap.com/zh/tidb/v6.0/tispark-deployment-topology/#tispark-部署拓扑 https://zhuanlan.zhihu.com/p/270265931#TiSpark 批量写入 TiDB 原理与实现