Ysia Tamas

1598393640

Spring Cloud Stream with Apache Kafka & RabbitMQ Producer Consumer example

In this video, You will learn how to create a Kafka & RabbitMQ Producer and Consumer by using spring cloud stream

#spring #rabbitmq #kafka

What is GEEK

Buddha Community

Spring Cloud Stream with Apache Kafka & RabbitMQ Producer Consumer example

Enhance Amazon Aurora Read/Write Capability with ShardingSphere-JDBC

1. Introduction

Amazon Aurora is a relational database management system (RDBMS) developed by AWS(Amazon Web Services). Aurora gives you the performance and availability of commercial-grade databases with full MySQL and PostgreSQL compatibility. In terms of high performance, Aurora MySQL and Aurora PostgreSQL have shown an increase in throughput of up to 5X over stock MySQL and 3X over stock PostgreSQL respectively on similar hardware. In terms of scalability, Aurora achieves enhancements and innovations in storage and computing, horizontal and vertical functions.

Aurora supports up to 128TB of storage capacity and supports dynamic scaling of storage layer in units of 10GB. In terms of computing, Aurora supports scalable configurations for multiple read replicas. Each region can have an additional 15 Aurora replicas. In addition, Aurora provides multi-primary architecture to support four read/write nodes. Its Serverless architecture allows vertical scaling and reduces typical latency to under a second, while the Global Database enables a single database cluster to span multiple AWS Regions in low latency.

Aurora already provides great scalability with the growth of user data volume. Can it handle more data and support more concurrent access? You may consider using sharding to support the configuration of multiple underlying Aurora clusters. To this end, a series of blogs, including this one, provides you with a reference in choosing between Proxy and JDBC for sharding.

1.1 Why sharding is needed

AWS Aurora offers a single relational database. Primary-secondary, multi-primary, and global database, and other forms of hosting architecture can satisfy various architectural scenarios above. However, Aurora doesn’t provide direct support for sharding scenarios, and sharding has a variety of forms, such as vertical and horizontal forms. If we want to further increase data capacity, some problems have to be solved, such as cross-node database Join, associated query, distributed transactions, SQL sorting, page turning, function calculation, database global primary key, capacity planning, and secondary capacity expansion after sharding.

1.2 Sharding methods

It is generally accepted that when the capacity of a MySQL table is less than 10 million, the time spent on queries is optimal because at this time the height of its BTREE index is between 3 and 5. Data sharding can reduce the amount of data in a single table and distribute the read and write loads to different data nodes at the same time. Data sharding can be divided into vertical sharding and horizontal sharding.

1. Advantages of vertical sharding

  • Address the coupling of business system and make clearer.
  • Implement hierarchical management, maintenance, monitoring, and expansion to data of different businesses, like micro-service governance.
  • In high concurrency scenarios, vertical sharding removes the bottleneck of IO, database connections, and hardware resources on a single machine to some extent.

2. Disadvantages of vertical sharding

  • After splitting the library, Join can only be implemented by interface aggregation, which will increase the complexity of development.
  • After splitting the library, it is complex to process distributed transactions.
  • There is a large amount of data on a single table and horizontal sharding is required.

3. Advantages of horizontal sharding

  • There is no such performance bottleneck as a large amount of data on a single database and high concurrency, and it increases system stability and load capacity.
  • The business modules do not need to be split due to minor modification on the application client.

4. Disadvantages of horizontal sharding

  • Transaction consistency across shards is hard to be guaranteed;
  • The performance of associated query in cross-library Join is poor.
  • It’s difficult to scale the data many times and maintenance is a big workload.

Based on the analysis above, and the available studis on popular sharding middleware, we selected ShardingSphere, an open source product, combined with Amazon Aurora to introduce how the combination of these two products meets various forms of sharding and how to solve the problems brought by sharding.

ShardingSphere is an open source ecosystem including a set of distributed database middleware solutions, including 3 independent products, Sharding-JDBC, Sharding-Proxy & Sharding-Sidecar.

2. ShardingSphere introduction:

The characteristics of Sharding-JDBC are:

  1. With the client end connecting directly to the database, it provides service in the form of jar and requires no extra deployment and dependence.
  2. It can be considered as an enhanced JDBC driver, which is fully compatible with JDBC and all kinds of ORM frameworks.
  3. Applicable in any ORM framework based on JDBC, such as JPA, Hibernate, Mybatis, Spring JDBC Template or direct use of JDBC.
  4. Support any third-party database connection pool, such as DBCP, C3P0, BoneCP, Druid, HikariCP;
  5. Support any kind of JDBC standard database: MySQL, Oracle, SQLServer, PostgreSQL and any databases accessible to JDBC.
  6. Sharding-JDBC adopts decentralized architecture, applicable to high-performance light-weight OLTP application developed with Java

Hybrid Structure Integrating Sharding-JDBC and Applications

Sharding-JDBC’s core concepts

Data node: The smallest unit of a data slice, consisting of a data source name and a data table, such as ds_0.product_order_0.

Actual table: The physical table that really exists in the horizontal sharding database, such as product order tables: product_order_0, product_order_1, and product_order_2.

Logic table: The logical name of the horizontal sharding databases (tables) with the same schema. For instance, the logic table of the order product_order_0, product_order_1, and product_order_2 is product_order.

Binding table: It refers to the primary table and the joiner table with the same sharding rules. For example, product_order table and product_order_item are sharded by order_id, so they are binding tables with each other. Cartesian product correlation will not appear in the multi-tables correlating query, so the query efficiency will increase greatly.

Broadcast table: It refers to tables that exist in all sharding database sources. The schema and data must consist in each database. It can be applied to the small data volume that needs to correlate with big data tables to query, dictionary table and configuration table for example.

3. Testing ShardingSphere-JDBC

3.1 Example project

Download the example project code locally. In order to ensure the stability of the test code, we choose shardingsphere-example-4.0.0 version.

git clone https://github.com/apache/shardingsphere-example.git

Project description:

shardingsphere-example
  ├── example-core
  │   ├── config-utility
  │   ├── example-api
  │   ├── example-raw-jdbc
  │   ├── example-spring-jpa #spring+jpa integration-based entity,repository
  │   └── example-spring-mybatis
  ├── sharding-jdbc-example
  │   ├── sharding-example
  │   │   ├── sharding-raw-jdbc-example
  │   │   ├── sharding-spring-boot-jpa-example #integration-based sharding-jdbc functions
  │   │   ├── sharding-spring-boot-mybatis-example
  │   │   ├── sharding-spring-namespace-jpa-example
  │   │   └── sharding-spring-namespace-mybatis-example
  │   ├── orchestration-example
  │   │   ├── orchestration-raw-jdbc-example
  │   │   ├── orchestration-spring-boot-example #integration-based sharding-jdbc governance function
  │   │   └── orchestration-spring-namespace-example
  │   ├── transaction-example
  │   │   ├── transaction-2pc-xa-example #sharding-jdbc sample of two-phase commit for a distributed transaction
  │   │   └──transaction-base-seata-example #sharding-jdbc distributed transaction seata sample
  │   ├── other-feature-example
  │   │   ├── hint-example
  │   │   └── encrypt-example
  ├── sharding-proxy-example
  │   └── sharding-proxy-boot-mybatis-example
  └── src/resources
        └── manual_schema.sql  

Configuration file description:

application-master-slave.properties #read/write splitting profile
application-sharding-databases-tables.properties #sharding profile
application-sharding-databases.properties       #library split profile only
application-sharding-master-slave.properties    #sharding and read/write splitting profile
application-sharding-tables.properties          #table split profile
application.properties                         #spring boot profile

Code logic description:

The following is the entry class of the Spring Boot application below. Execute it to run the project.

The execution logic of demo is as follows:

3.2 Verifying read/write splitting

As business grows, the write and read requests can be split to different database nodes to effectively promote the processing capability of the entire database cluster. Aurora uses a reader/writer endpoint to meet users' requirements to write and read with strong consistency, and a read-only endpoint to meet the requirements to read without strong consistency. Aurora's read and write latency is within single-digit milliseconds, much lower than MySQL's binlog-based logical replication, so there's a lot of loads that can be directed to a read-only endpoint.

Through the one primary and multiple secondary configuration, query requests can be evenly distributed to multiple data replicas, which further improves the processing capability of the system. Read/write splitting can improve the throughput and availability of system, but it can also lead to data inconsistency. Aurora provides a primary/secondary architecture in a fully managed form, but applications on the upper-layer still need to manage multiple data sources when interacting with Aurora, routing SQL requests to different nodes based on the read/write type of SQL statements and certain routing policies.

ShardingSphere-JDBC provides read/write splitting features and it is integrated with application programs so that the complex configuration between application programs and database clusters can be separated from application programs. Developers can manage the Shard through configuration files and combine it with ORM frameworks such as Spring JPA and Mybatis to completely separate the duplicated logic from the code, which greatly improves the ability to maintain code and reduces the coupling between code and database.

3.2.1 Setting up the database environment

Create a set of Aurora MySQL read/write splitting clusters. The model is db.r5.2xlarge. Each set of clusters has one write node and two read nodes.

3.2.2 Configuring Sharding-JDBC

application.properties spring boot Master profile description:

You need to replace the green ones with your own environment configuration.

# Jpa automatically creates and drops data tables based on entities
spring.jpa.properties.hibernate.hbm2ddl.auto=create-drop
spring.jpa.properties.hibernate.dialect=org.hibernate.dialect.MySQL5Dialect
spring.jpa.properties.hibernate.show_sql=true

#spring.profiles.active=sharding-databases
#spring.profiles.active=sharding-tables
#spring.profiles.active=sharding-databases-tables
#Activate master-slave configuration item so that sharding-jdbc can use master-slave profile
spring.profiles.active=master-slave
#spring.profiles.active=sharding-master-slave

application-master-slave.properties sharding-jdbc profile description:

spring.shardingsphere.datasource.names=ds_master,ds_slave_0,ds_slave_1
# data souce-master
spring.shardingsphere.datasource.ds_master.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master.password=Your master DB password
spring.shardingsphere.datasource.ds_master.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master.jdbc-url=Your primary DB data sourceurl spring.shardingsphere.datasource.ds_master.username=Your primary DB username
# data source-slave
spring.shardingsphere.datasource.ds_slave_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_slave_0.password= Your slave DB password
spring.shardingsphere.datasource.ds_slave_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_slave_0.jdbc-url=Your slave DB data source url
spring.shardingsphere.datasource.ds_slave_0.username= Your slave DB username
# data source-slave
spring.shardingsphere.datasource.ds_slave_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_slave_1.password= Your slave DB password
spring.shardingsphere.datasource.ds_slave_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_slave_1.jdbc-url= Your slave DB data source url
spring.shardingsphere.datasource.ds_slave_1.username= Your slave DB username
# Routing Policy Configuration
spring.shardingsphere.masterslave.load-balance-algorithm-type=round_robin
spring.shardingsphere.masterslave.name=ds_ms
spring.shardingsphere.masterslave.master-data-source-name=ds_master
spring.shardingsphere.masterslave.slave-data-source-names=ds_slave_0,ds_slave_1
# sharding-jdbc configures the information storage mode
spring.shardingsphere.mode.type=Memory
# start shardingsphere log,and you can see the conversion from logical SQL to actual SQL from the print
spring.shardingsphere.props.sql.show=true

 

3.2.3 Test and verification process description

  • Test environment data initialization: Spring JPA initialization automatically creates tables for testing.

  • Write data to the master instance

As shown in the ShardingSphere-SQL log figure below, the write SQL is executed on the ds_master data source.

  • Data query operations are performed on the slave library.

As shown in the ShardingSphere-SQL log figure below, the read SQL is executed on the ds_slave data source in the form of polling.

[INFO ] 2022-04-02 19:43:39,376 --main-- [ShardingSphere-SQL] Rule Type: master-slave 
[INFO ] 2022-04-02 19:43:39,376 --main-- [ShardingSphere-SQL] SQL: select orderentit0_.order_id as order_id1_1_, orderentit0_.address_id as address_2_1_, 
orderentit0_.status as status3_1_, orderentit0_.user_id as user_id4_1_ from t_order orderentit0_ ::: DataSources: ds_slave_0 
---------------------------- Print OrderItem Data -------------------
Hibernate: select orderiteme1_.order_item_id as order_it1_2_, orderiteme1_.order_id as order_id2_2_, orderiteme1_.status as status3_2_, orderiteme1_.user_id 
as user_id4_2_ from t_order orderentit0_ cross join t_order_item orderiteme1_ where orderentit0_.order_id=orderiteme1_.order_id
[INFO ] 2022-04-02 19:43:40,898 --main-- [ShardingSphere-SQL] Rule Type: master-slave 
[INFO ] 2022-04-02 19:43:40,898 --main-- [ShardingSphere-SQL] SQL: select orderiteme1_.order_item_id as order_it1_2_, orderiteme1_.order_id as order_id2_2_, orderiteme1_.status as status3_2_, 
orderiteme1_.user_id as user_id4_2_ from t_order orderentit0_ cross join t_order_item orderiteme1_ where orderentit0_.order_id=orderiteme1_.order_id ::: DataSources: ds_slave_1 

Note: As shown in the figure below, if there are both reads and writes in a transaction, Sharding-JDBC routes both read and write operations to the master library. If the read/write requests are not in the same transaction, the corresponding read requests are distributed to different read nodes according to the routing policy.

@Override
@Transactional // When a transaction is started, both read and write in the transaction go through the master library. When closed, read goes through the slave library and write goes through the master library
public void processSuccess() throws SQLException {
    System.out.println("-------------- Process Success Begin ---------------");
    List<Long> orderIds = insertData();
    printData();
    deleteData(orderIds);
    printData();
    System.out.println("-------------- Process Success Finish --------------");
}

3.2.4 Verifying Aurora failover scenario

The Aurora database environment adopts the configuration described in Section 2.2.1.

3.2.4.1 Verification process description

  1. Start the Spring-Boot project

2. Perform a failover on Aurora’s console

3. Execute the Rest API request

4. Repeatedly execute POST (http://localhost:8088/save-user) until the call to the API failed to write to Aurora and eventually recovered successfully.

5. The following figure shows the process of executing code failover. It takes about 37 seconds from the time when the latest SQL write is successfully performed to the time when the next SQL write is successfully performed. That is, the application can be automatically recovered from Aurora failover, and the recovery time is about 37 seconds.

3.3 Testing table sharding-only function

3.3.1 Configuring Sharding-JDBC

application.properties spring boot master profile description

# Jpa automatically creates and drops data tables based on entities
spring.jpa.properties.hibernate.hbm2ddl.auto=create-drop
spring.jpa.properties.hibernate.dialect=org.hibernate.dialect.MySQL5Dialect
spring.jpa.properties.hibernate.show_sql=true
#spring.profiles.active=sharding-databases
#Activate sharding-tables configuration items
#spring.profiles.active=sharding-tables
#spring.profiles.active=sharding-databases-tables
# spring.profiles.active=master-slave
#spring.profiles.active=sharding-master-slave

application-sharding-tables.properties sharding-jdbc profile description

## configure primary-key policy
spring.shardingsphere.sharding.tables.t_order.key-generator.column=order_id
spring.shardingsphere.sharding.tables.t_order.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order.key-generator.props.worker.id=123
spring.shardingsphere.sharding.tables.t_order_item.actual-data-nodes=ds.t_order_item_$->{0..1}
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.sharding-column=order_id
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.algorithm-expression=t_order_item_$->{order_id % 2}
spring.shardingsphere.sharding.tables.t_order_item.key-generator.column=order_item_id
spring.shardingsphere.sharding.tables.t_order_item.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order_item.key-generator.props.worker.id=123
# configure the binding relation of t_order and t_order_item
spring.shardingsphere.sharding.binding-tables[0]=t_order,t_order_item
# configure broadcast tables
spring.shardingsphere.sharding.broadcast-tables=t_address
# sharding-jdbc mode
spring.shardingsphere.mode.type=Memory
# start shardingsphere log
spring.shardingsphere.props.sql.show=true

 

3.3.2 Test and verification process description

1. DDL operation

JPA automatically creates tables for testing. When Sharding-JDBC routing rules are configured, the client executes DDL, and Sharding-JDBC automatically creates corresponding tables according to the table splitting rules. If t_address is a broadcast table, create a t_address because there is only one master instance. Two physical tables t_order_0 and t_order_1 will be created when creating t_order.

2. Write operation

As shown in the figure below, Logic SQL inserts a record into t_order. When Sharding-JDBC is executed, data will be distributed to t_order_0 and t_order_1 according to the table splitting rules.

When t_order and t_order_item are bound, the records associated with order_item and order are placed on the same physical table.

3. Read operation

As shown in the figure below, perform the join query operations to order and order_item under the binding table, and the physical shard is precisely located based on the binding relationship.

The join query operations on order and order_item under the unbound table will traverse all shards.

3.4 Testing database sharding-only function

3.4.1 Setting up the database environment

Create two instances on Aurora: ds_0 and ds_1

When the sharding-spring-boot-jpa-example project is started, tables t_order, t_order_itemt_address will be created on two Aurora instances.

3.4.2 Configuring Sharding-JDBC

application.properties springboot master profile description

# Jpa automatically creates and drops data tables based on entities
spring.jpa.properties.hibernate.hbm2ddl.auto=create
spring.jpa.properties.hibernate.dialect=org.hibernate.dialect.MySQL5Dialect
spring.jpa.properties.hibernate.show_sql=true

# Activate sharding-databases configuration items
spring.profiles.active=sharding-databases
#spring.profiles.active=sharding-tables
#spring.profiles.active=sharding-databases-tables
#spring.profiles.active=master-slave
#spring.profiles.active=sharding-master-slave

application-sharding-databases.properties sharding-jdbc profile description

spring.shardingsphere.datasource.names=ds_0,ds_1
# ds_0
spring.shardingsphere.datasource.ds_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_0.jdbc-url= spring.shardingsphere.datasource.ds_0.username= 
spring.shardingsphere.datasource.ds_0.password=
# ds_1
spring.shardingsphere.datasource.ds_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_1.jdbc-url= 
spring.shardingsphere.datasource.ds_1.username= 
spring.shardingsphere.datasource.ds_1.password=
spring.shardingsphere.sharding.default-database-strategy.inline.sharding-column=user_id
spring.shardingsphere.sharding.default-database-strategy.inline.algorithm-expression=ds_$->{user_id % 2}
spring.shardingsphere.sharding.binding-tables=t_order,t_order_item
spring.shardingsphere.sharding.broadcast-tables=t_address
spring.shardingsphere.sharding.default-data-source-name=ds_0

spring.shardingsphere.sharding.tables.t_order.actual-data-nodes=ds_$->{0..1}.t_order
spring.shardingsphere.sharding.tables.t_order.key-generator.column=order_id
spring.shardingsphere.sharding.tables.t_order.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order.key-generator.props.worker.id=123
spring.shardingsphere.sharding.tables.t_order_item.actual-data-nodes=ds_$->{0..1}.t_order_item
spring.shardingsphere.sharding.tables.t_order_item.key-generator.column=order_item_id
spring.shardingsphere.sharding.tables.t_order_item.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order_item.key-generator.props.worker.id=123
# sharding-jdbc mode
spring.shardingsphere.mode.type=Memory
# start shardingsphere log
spring.shardingsphere.props.sql.show=true

 

3.4.3 Test and verification process description

1. DDL operation

JPA automatically creates tables for testing. When Sharding-JDBC’s library splitting and routing rules are configured, the client executes DDL, and Sharding-JDBC will automatically create corresponding tables according to table splitting rules. If t_address is a broadcast table, physical tables will be created on ds_0 and ds_1. The three tables, t_address, t_order and t_order_item will be created on ds_0 and ds_1 respectively.

2. Write operation

For the broadcast table t_address, each record written will also be written to the t_address tables of ds_0 and ds_1.

The tables t_order and t_order_item of the slave library are written on the table in the corresponding instance according to the slave library field and routing policy.

3. Read operation

Query order is routed to the corresponding Aurora instance according to the routing rules of the slave library .

Query Address. Since address is a broadcast table, an instance of address will be randomly selected and queried from the nodes used.

As shown in the figure below, perform the join query operations to order and order_item under the binding table, and the physical shard is precisely located based on the binding relationship.

3.5 Verifying sharding function

3.5.1 Setting up the database environment

As shown in the figure below, create two instances on Aurora: ds_0 and ds_1

When the sharding-spring-boot-jpa-example project is started, physical tables t_order_01, t_order_02, t_order_item_01,and t_order_item_02 and global table t_address will be created on two Aurora instances.

3.5.2 Configuring Sharding-JDBC

application.properties springboot master profile description

# Jpa automatically creates and drops data tables based on entities
spring.jpa.properties.hibernate.hbm2ddl.auto=create
spring.jpa.properties.hibernate.dialect=org.hibernate.dialect.MySQL5Dialect
spring.jpa.properties.hibernate.show_sql=true
# Activate sharding-databases-tables configuration items
#spring.profiles.active=sharding-databases
#spring.profiles.active=sharding-tables
spring.profiles.active=sharding-databases-tables
#spring.profiles.active=master-slave
#spring.profiles.active=sharding-master-slave

application-sharding-databases.properties sharding-jdbc profile description

spring.shardingsphere.datasource.names=ds_0,ds_1
# ds_0
spring.shardingsphere.datasource.ds_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_0.jdbc-url= 306/dev?useSSL=false&characterEncoding=utf-8
spring.shardingsphere.datasource.ds_0.username= 
spring.shardingsphere.datasource.ds_0.password=
spring.shardingsphere.datasource.ds_0.max-active=16
# ds_1
spring.shardingsphere.datasource.ds_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_1.jdbc-url= 
spring.shardingsphere.datasource.ds_1.username= 
spring.shardingsphere.datasource.ds_1.password=
spring.shardingsphere.datasource.ds_1.max-active=16
# default library splitting policy
spring.shardingsphere.sharding.default-database-strategy.inline.sharding-column=user_id
spring.shardingsphere.sharding.default-database-strategy.inline.algorithm-expression=ds_$->{user_id % 2}
spring.shardingsphere.sharding.binding-tables=t_order,t_order_item
spring.shardingsphere.sharding.broadcast-tables=t_address
# Tables that do not meet the library splitting policy are placed on ds_0
spring.shardingsphere.sharding.default-data-source-name=ds_0
# t_order table splitting policy
spring.shardingsphere.sharding.tables.t_order.actual-data-nodes=ds_$->{0..1}.t_order_$->{0..1}
spring.shardingsphere.sharding.tables.t_order.table-strategy.inline.sharding-column=order_id
spring.shardingsphere.sharding.tables.t_order.table-strategy.inline.algorithm-expression=t_order_$->{order_id % 2}
spring.shardingsphere.sharding.tables.t_order.key-generator.column=order_id
spring.shardingsphere.sharding.tables.t_order.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order.key-generator.props.worker.id=123
# t_order_item table splitting policy
spring.shardingsphere.sharding.tables.t_order_item.actual-data-nodes=ds_$->{0..1}.t_order_item_$->{0..1}
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.sharding-column=order_id
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.algorithm-expression=t_order_item_$->{order_id % 2}
spring.shardingsphere.sharding.tables.t_order_item.key-generator.column=order_item_id
spring.shardingsphere.sharding.tables.t_order_item.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order_item.key-generator.props.worker.id=123
# sharding-jdbc mdoe
spring.shardingsphere.mode.type=Memory
# start shardingsphere log
spring.shardingsphere.props.sql.show=true

 

3.5.3 Test and verification process description

1. DDL operation

JPA automatically creates tables for testing. When Sharding-JDBC’s sharding and routing rules are configured, the client executes DDL, and Sharding-JDBC will automatically create corresponding tables according to table splitting rules. If t_address is a broadcast table, t_address will be created on both ds_0 and ds_1. The three tables, t_address, t_order and t_order_item will be created on ds_0 and ds_1 respectively.

2. Write operation

For the broadcast table t_address, each record written will also be written to the t_address tables of ds_0 and ds_1.

The tables t_order and t_order_item of the sub-library are written to the table on the corresponding instance according to the slave library field and routing policy.

3. Read operation

The read operation is similar to the library split function verification described in section2.4.3.

3.6 Testing database sharding, table sharding and read/write splitting function

3.6.1 Setting up the database environment

The following figure shows the physical table of the created database instance.

3.6.2 Configuring Sharding-JDBC

application.properties spring boot master profile description

# Jpa automatically creates and drops data tables based on entities
spring.jpa.properties.hibernate.hbm2ddl.auto=create
spring.jpa.properties.hibernate.dialect=org.hibernate.dialect.MySQL5Dialect
spring.jpa.properties.hibernate.show_sql=true

# activate sharding-databases-tables configuration items
#spring.profiles.active=sharding-databases
#spring.profiles.active=sharding-tables
#spring.profiles.active=sharding-databases-tables
#spring.profiles.active=master-slave
spring.profiles.active=sharding-master-slave

application-sharding-master-slave.properties sharding-jdbc profile description

The url, name and password of the database need to be changed to your own database parameters.

spring.shardingsphere.datasource.names=ds_master_0,ds_master_1,ds_master_0_slave_0,ds_master_0_slave_1,ds_master_1_slave_0,ds_master_1_slave_1
spring.shardingsphere.datasource.ds_master_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_0.jdbc-url= spring.shardingsphere.datasource.ds_master_0.username= 
spring.shardingsphere.datasource.ds_master_0.password=
spring.shardingsphere.datasource.ds_master_0.max-active=16
spring.shardingsphere.datasource.ds_master_0_slave_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_0_slave_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_0_slave_0.jdbc-url= spring.shardingsphere.datasource.ds_master_0_slave_0.username= 
spring.shardingsphere.datasource.ds_master_0_slave_0.password=
spring.shardingsphere.datasource.ds_master_0_slave_0.max-active=16
spring.shardingsphere.datasource.ds_master_0_slave_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_0_slave_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_0_slave_1.jdbc-url= spring.shardingsphere.datasource.ds_master_0_slave_1.username= 
spring.shardingsphere.datasource.ds_master_0_slave_1.password=
spring.shardingsphere.datasource.ds_master_0_slave_1.max-active=16
spring.shardingsphere.datasource.ds_master_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_1.jdbc-url= 
spring.shardingsphere.datasource.ds_master_1.username= 
spring.shardingsphere.datasource.ds_master_1.password=
spring.shardingsphere.datasource.ds_master_1.max-active=16
spring.shardingsphere.datasource.ds_master_1_slave_0.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_1_slave_0.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_1_slave_0.jdbc-url=
spring.shardingsphere.datasource.ds_master_1_slave_0.username=
spring.shardingsphere.datasource.ds_master_1_slave_0.password=
spring.shardingsphere.datasource.ds_master_1_slave_0.max-active=16
spring.shardingsphere.datasource.ds_master_1_slave_1.type=com.zaxxer.hikari.HikariDataSource
spring.shardingsphere.datasource.ds_master_1_slave_1.driver-class-name=com.mysql.jdbc.Driver
spring.shardingsphere.datasource.ds_master_1_slave_1.jdbc-url= spring.shardingsphere.datasource.ds_master_1_slave_1.username=admin
spring.shardingsphere.datasource.ds_master_1_slave_1.password=
spring.shardingsphere.datasource.ds_master_1_slave_1.max-active=16
spring.shardingsphere.sharding.default-database-strategy.inline.sharding-column=user_id
spring.shardingsphere.sharding.default-database-strategy.inline.algorithm-expression=ds_$->{user_id % 2}
spring.shardingsphere.sharding.binding-tables=t_order,t_order_item
spring.shardingsphere.sharding.broadcast-tables=t_address
spring.shardingsphere.sharding.default-data-source-name=ds_master_0
spring.shardingsphere.sharding.tables.t_order.actual-data-nodes=ds_$->{0..1}.t_order_$->{0..1}
spring.shardingsphere.sharding.tables.t_order.table-strategy.inline.sharding-column=order_id
spring.shardingsphere.sharding.tables.t_order.table-strategy.inline.algorithm-expression=t_order_$->{order_id % 2}
spring.shardingsphere.sharding.tables.t_order.key-generator.column=order_id
spring.shardingsphere.sharding.tables.t_order.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order.key-generator.props.worker.id=123
spring.shardingsphere.sharding.tables.t_order_item.actual-data-nodes=ds_$->{0..1}.t_order_item_$->{0..1}
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.sharding-column=order_id
spring.shardingsphere.sharding.tables.t_order_item.table-strategy.inline.algorithm-expression=t_order_item_$->{order_id % 2}
spring.shardingsphere.sharding.tables.t_order_item.key-generator.column=order_item_id
spring.shardingsphere.sharding.tables.t_order_item.key-generator.type=SNOWFLAKE
spring.shardingsphere.sharding.tables.t_order_item.key-generator.props.worker.id=123
# master/slave data source and slave data source configuration
spring.shardingsphere.sharding.master-slave-rules.ds_0.master-data-source-name=ds_master_0
spring.shardingsphere.sharding.master-slave-rules.ds_0.slave-data-source-names=ds_master_0_slave_0, ds_master_0_slave_1
spring.shardingsphere.sharding.master-slave-rules.ds_1.master-data-source-name=ds_master_1
spring.shardingsphere.sharding.master-slave-rules.ds_1.slave-data-source-names=ds_master_1_slave_0, ds_master_1_slave_1
# sharding-jdbc mode
spring.shardingsphere.mode.type=Memory
# start shardingsphere log
spring.shardingsphere.props.sql.show=true

 

3.6.3 Test and verification process description

1. DDL operation

JPA automatically creates tables for testing. When Sharding-JDBC’s library splitting and routing rules are configured, the client executes DDL, and Sharding-JDBC will automatically create corresponding tables according to table splitting rules. If t_address is a broadcast table, t_address will be created on both ds_0 and ds_1. The three tables, t_address, t_order and t_order_item will be created on ds_0 and ds_1 respectively.

2. Write operation

For the broadcast table t_address, each record written will also be written to the t_address tables of ds_0 and ds_1.

The tables t_order and t_order_item of the slave library are written to the table on the corresponding instance according to the slave library field and routing policy.

3. Read operation

The join query operations on order and order_item under the binding table are shown below.

3. Conclusion

As an open source product focusing on database enhancement, ShardingSphere is pretty good in terms of its community activitiy, product maturity and documentation richness.

Among its products, ShardingSphere-JDBC is a sharding solution based on the client-side, which supports all sharding scenarios. And there’s no need to introduce an intermediate layer like Proxy, so the complexity of operation and maintenance is reduced. Its latency is theoretically lower than Proxy due to the lack of intermediate layer. In addition, ShardingSphere-JDBC can support a variety of relational databases based on SQL standards such as MySQL/PostgreSQL/Oracle/SQL Server, etc.

However, due to the integration of Sharding-JDBC with the application program, it only supports Java language for now, and is strongly dependent on the application programs. Nevertheless, Sharding-JDBC separates all sharding configuration from the application program, which brings relatively small changes when switching to other middleware.

In conclusion, Sharding-JDBC is a good choice if you use a Java-based system and have to to interconnect with different relational databases — and don’t want to bother with introducing an intermediate layer.

Author

Sun Jinhua

A senior solution architect at AWS, Sun is responsible for the design and consult on cloud architecture. for providing customers with cloud-related design and consulting services. Before joining AWS, he ran his own business, specializing in building e-commerce platforms and designing the overall architecture for e-commerce platforms of automotive companies. He worked in a global leading communication equipment company as a senior engineer, responsible for the development and architecture design of multiple subsystems of LTE equipment system. He has rich experience in architecture design with high concurrency and high availability system, microservice architecture design, database, middleware, IOT etc.

Royce  Reinger

Royce Reinger

1659714540

Ruby-kafka: A Ruby Client Library for Apache Kafka

ruby-kafka

A Ruby client library for Apache Kafka, a distributed log and message bus. The focus of this library will be operational simplicity, with good logging and metrics that can make debugging issues easier.

Installation

Add this line to your application's Gemfile:

gem 'ruby-kafka'

And then execute:

$ bundle

Or install it yourself as:

$ gem install ruby-kafka

Compatibility

 Producer APIConsumer API
Kafka 0.8Full support in v0.4.xUnsupported
Kafka 0.9Full support in v0.4.xFull support in v0.4.x
Kafka 0.10Full support in v0.5.xFull support in v0.5.x
Kafka 0.11Full support in v0.7.xLimited support
Kafka 1.0Limited supportLimited support
Kafka 2.0Limited supportLimited support
Kafka 2.1Limited supportLimited support
Kafka 2.2Limited supportLimited support
Kafka 2.3Limited supportLimited support
Kafka 2.4Limited supportLimited support
Kafka 2.5Limited supportLimited support
Kafka 2.6Limited supportLimited support
Kafka 2.7Limited supportLimited support

This library is targeting Kafka 0.9 with the v0.4.x series and Kafka 0.10 with the v0.5.x series. There's limited support for Kafka 0.8, and things should work with Kafka 0.11, although there may be performance issues due to changes in the protocol.

  • Kafka 0.8: Full support for the Producer API in ruby-kafka v0.4.x, but no support for consumer groups. Simple message fetching works.
  • Kafka 0.9: Full support for the Producer and Consumer API in ruby-kafka v0.4.x.
  • Kafka 0.10: Full support for the Producer and Consumer API in ruby-kafka v0.5.x. Note that you must run version 0.10.1 or higher of Kafka due to limitations in 0.10.0.
  • Kafka 0.11: Full support for Producer API, limited support for Consumer API in ruby-kafka v0.7.x. New features in 0.11.x includes new Record Batch format, idempotent and transactional production. The missing feature is dirty reading of Consumer API.
  • Kafka 1.0: Everything that works with Kafka 0.11 should still work, but so far no features specific to Kafka 1.0 have been added.
  • Kafka 2.0: Everything that works with Kafka 1.0 should still work, but so far no features specific to Kafka 2.0 have been added.
  • Kafka 2.1: Everything that works with Kafka 2.0 should still work, but so far no features specific to Kafka 2.1 have been added.
  • Kafka 2.2: Everything that works with Kafka 2.1 should still work, but so far no features specific to Kafka 2.2 have been added.
  • Kafka 2.3: Everything that works with Kafka 2.2 should still work, but so far no features specific to Kafka 2.3 have been added.
  • Kafka 2.4: Everything that works with Kafka 2.3 should still work, but so far no features specific to Kafka 2.4 have been added.
  • Kafka 2.5: Everything that works with Kafka 2.4 should still work, but so far no features specific to Kafka 2.5 have been added.
  • Kafka 2.6: Everything that works with Kafka 2.5 should still work, but so far no features specific to Kafka 2.6 have been added.
  • Kafka 2.7: Everything that works with Kafka 2.6 should still work, but so far no features specific to Kafka 2.7 have been added.

This library requires Ruby 2.1 or higher.

Usage

Please see the documentation site for detailed documentation on the latest release. Note that the documentation on GitHub may not match the version of the library you're using – there are still being made many changes to the API.

Setting up the Kafka Client

A client must be initialized with at least one Kafka broker, from which the entire Kafka cluster will be discovered. Each client keeps a separate pool of broker connections. Don't use the same client from more than one thread.

require "kafka"

# The first argument is a list of "seed brokers" that will be queried for the full
# cluster topology. At least one of these *must* be available. `client_id` is
# used to identify this client in logs and metrics. It's optional but recommended.
kafka = Kafka.new(["kafka1:9092", "kafka2:9092"], client_id: "my-application")

You can also use a hostname with seed brokers' IP addresses:

kafka = Kafka.new("seed-brokers:9092", client_id: "my-application", resolve_seed_brokers: true)

Producing Messages to Kafka

The simplest way to write a message to a Kafka topic is to call #deliver_message:

kafka = Kafka.new(...)
kafka.deliver_message("Hello, World!", topic: "greetings")

This will write the message to a random partition in the greetings topic. If you want to write to a specific partition, pass the partition parameter:

# Will write to partition 42.
kafka.deliver_message("Hello, World!", topic: "greetings", partition: 42)

If you don't know exactly how many partitions are in the topic, or if you'd rather have some level of indirection, you can pass in partition_key instead. Two messages with the same partition key will always be assigned to the same partition. This is useful if you want to make sure all messages with a given attribute are always written to the same partition, e.g. all purchase events for a given customer id.

# Partition keys assign a partition deterministically.
kafka.deliver_message("Hello, World!", topic: "greetings", partition_key: "hello")

Kafka also supports message keys. When passed, a message key can be used instead of a partition key. The message key is written alongside the message value and can be read by consumers. Message keys in Kafka can be used for interesting things such as Log Compaction. See Partitioning for more information.

# Set a message key; the key will be used for partitioning since no explicit
# `partition_key` is set.
kafka.deliver_message("Hello, World!", key: "hello", topic: "greetings")

Efficiently Producing Messages

While #deliver_message works fine for infrequent writes, there are a number of downsides:

  • Kafka is optimized for transmitting messages in batches rather than individually, so there's a significant overhead and performance penalty in using the single-message API.
  • The message delivery can fail in a number of different ways, but this simplistic API does not provide automatic retries.
  • The message is not buffered, so if there is an error, it is lost.

The Producer API solves all these problems and more:

# Instantiate a new producer.
producer = kafka.producer

# Add a message to the producer buffer.
producer.produce("hello1", topic: "test-messages")

# Deliver the messages to Kafka.
producer.deliver_messages

#produce will buffer the message in the producer but will not actually send it to the Kafka cluster. Buffered messages are only delivered to the Kafka cluster once #deliver_messages is called. Since messages may be destined for different partitions, this could involve writing to more than one Kafka broker. Note that a failure to send all buffered messages after the configured number of retries will result in Kafka::DeliveryFailed being raised. This can be rescued and ignored; the messages will be kept in the buffer until the next attempt.

Read the docs for Kafka::Producer for more details.

Asynchronously Producing Messages

A normal producer will block while #deliver_messages is sending messages to Kafka, possibly for tens of seconds or even minutes at a time, depending on your timeout and retry settings. Furthermore, you have to call #deliver_messages manually, with a frequency that balances batch size with message delay.

In order to avoid blocking during message deliveries you can use the asynchronous producer API. It is mostly similar to the synchronous API, with calls to #produce and #deliver_messages. The main difference is that rather than blocking, these calls will return immediately. The actual work will be done in a background thread, with the messages and operations being sent from the caller over a thread safe queue.

# `#async_producer` will create a new asynchronous producer.
producer = kafka.async_producer

# The `#produce` API works as normal.
producer.produce("hello", topic: "greetings")

# `#deliver_messages` will return immediately.
producer.deliver_messages

# Make sure to call `#shutdown` on the producer in order to avoid leaking
# resources. `#shutdown` will wait for any pending messages to be delivered
# before returning.
producer.shutdown

By default, the delivery policy will be the same as for a synchronous producer: only when #deliver_messages is called will the messages be delivered. However, the asynchronous producer offers two complementary policies for automatic delivery:

  1. Trigger a delivery once the producer's message buffer reaches a specified threshold. This can be used to improve efficiency by increasing the batch size when sending messages to the Kafka cluster.
  2. Trigger a delivery at a fixed time interval. This puts an upper bound on message delays.

These policies can be used alone or in combination.

# `async_producer` will create a new asynchronous producer.
producer = kafka.async_producer(
  # Trigger a delivery once 100 messages have been buffered.
  delivery_threshold: 100,

  # Trigger a delivery every 30 seconds.
  delivery_interval: 30,
)

producer.produce("hello", topic: "greetings")

# ...

When calling #shutdown, the producer will attempt to deliver the messages and the method call will block until that has happened. Note that there's no guarantee that the messages will be delivered.

Note: if the calling thread produces messages faster than the producer can write them to Kafka, you'll eventually run into problems. The internal queue used for sending messages from the calling thread to the background worker has a size limit; once this limit is reached, a call to #produce will raise Kafka::BufferOverflow.

Serialization

This library is agnostic to which serialization format you prefer. Both the value and key of a message is treated as a binary string of data. This makes it easier to use whatever serialization format you want, since you don't have to do anything special to make it work with ruby-kafka. Here's an example of encoding data with JSON:

require "json"

# ...

event = {
  "name" => "pageview",
  "url" => "https://example.com/posts/123",
  # ...
}

data = JSON.dump(event)

producer.produce(data, topic: "events")

There's also an example of encoding messages with Apache Avro.

Partitioning

Kafka topics are partitioned, with messages being assigned to a partition by the client. This allows a great deal of flexibility for the users. This section describes several strategies for partitioning and how they impact performance, data locality, etc.

Load Balanced Partitioning

When optimizing for efficiency, we either distribute messages as evenly as possible to all partitions, or make sure each producer always writes to a single partition. The former ensures an even load for downstream consumers; the latter ensures the highest producer performance, since message batching is done per partition.

If no explicit partition is specified, the producer will look to the partition key or the message key for a value that can be used to deterministically assign the message to a partition. If there is a big number of different keys, the resulting distribution will be pretty even. If no keys are passed, the producer will randomly assign a partition. Random partitioning can be achieved even if you use message keys by passing a random partition key, e.g. partition_key: rand(100).

If you wish to have the producer write all messages to a single partition, simply generate a random value and re-use that as the partition key:

partition_key = rand(100)

producer.produce(msg1, topic: "messages", partition_key: partition_key)
producer.produce(msg2, topic: "messages", partition_key: partition_key)

# ...

You can also base the partition key on some property of the producer, for example the host name.

Semantic Partitioning

By assigning messages to a partition based on some property of the message, e.g. making sure all events tracked in a user session are assigned to the same partition, downstream consumers can make simplifying assumptions about data locality. In this example, a consumer can keep process local state pertaining to a user session knowing that all events for the session will be read from a single partition. This is also called semantic partitioning, since the partition assignment is part of the application behavior.

Typically it's sufficient to simply pass a partition key in order to guarantee that a set of messages will be assigned to the same partition, e.g.

# All messages with the same `session_id` will be assigned to the same partition.
producer.produce(event, topic: "user-events", partition_key: session_id)

However, sometimes it's necessary to select a specific partition. When doing this, make sure that you don't pick a partition number outside the range of partitions for the topic:

partitions = kafka.partitions_for("events")

# Make sure that we don't exceed the partition count!
partition = some_number % partitions

producer.produce(event, topic: "events", partition: partition)

Compatibility with Other Clients

There's no standardized way to assign messages to partitions across different Kafka client implementations. If you have a heterogeneous set of clients producing messages to the same topics it may be important to ensure a consistent partitioning scheme. This library doesn't try to implement all schemes, so you'll have to figure out which scheme the other client is using and replicate it. An example:

partitions = kafka.partitions_for("events")

# Insert your custom partitioning scheme here:
partition = PartitioningScheme.assign(partitions, event)

producer.produce(event, topic: "events", partition: partition)

Another option is to configure a custom client partitioner that implements call(partition_count, message) and uses the same schema as the other client. For example:

class CustomPartitioner
  def call(partition_count, message)
    ...
  end
end
  
partitioner = CustomPartitioner.new
Kafka.new(partitioner: partitioner, ...)

Or, simply create a Proc handling the partitioning logic instead of having to add a new class. For example:

partitioner = -> (partition_count, message) { ... }
Kafka.new(partitioner: partitioner, ...)

Supported partitioning schemes

In order for semantic partitioning to work a partition_key must map to the same partition number every time. The general approach, and the one used by this library, is to hash the key and mod it by the number of partitions. There are many different algorithms that can be used to calculate a hash. By default crc32 is used. murmur2 is also supported for compatibility with Java based Kafka producers.

To use murmur2 hashing pass it as an argument to Partitioner. For example:

Kafka.new(partitioner: Kafka::Partitioner.new(hash_function: :murmur2))

Buffering and Error Handling

The producer is designed for resilience in the face of temporary network errors, Kafka broker failovers, and other issues that prevent the client from writing messages to the destination topics. It does this by employing local, in-memory buffers. Only when messages are acknowledged by a Kafka broker will they be removed from the buffer.

Typically, you'd configure the producer to retry failed attempts at sending messages, but sometimes all retries are exhausted. In that case, Kafka::DeliveryFailed is raised from Kafka::Producer#deliver_messages. If you wish to have your application be resilient to this happening (e.g. if you're logging to Kafka from a web application) you can rescue this exception. The failed messages are still retained in the buffer, so a subsequent call to #deliver_messages will still attempt to send them.

Note that there's a maximum buffer size; by default, it's set to 1,000 messages and 10MB. It's possible to configure both these numbers:

producer = kafka.producer(
  max_buffer_size: 5_000,           # Allow at most 5K messages to be buffered.
  max_buffer_bytesize: 100_000_000, # Allow at most 100MB to be buffered.
  ...
)

A final note on buffers: local buffers give resilience against broker and network failures, and allow higher throughput due to message batching, but they also trade off consistency guarantees for higher availability and resilience. If your local process dies while messages are buffered, those messages will be lost. If you require high levels of consistency, you should call #deliver_messages immediately after #produce.

Message Durability

Once the client has delivered a set of messages to a Kafka broker the broker will forward them to its replicas, thus ensuring that a single broker failure will not result in message loss. However, the client can choose when the leader acknowledges the write. At one extreme, the client can choose fire-and-forget delivery, not even bothering to check whether the messages have been acknowledged. At the other end, the client can ask the broker to wait until all its replicas have acknowledged the write before returning. This is the safest option, and the default. It's also possible to have the broker return as soon as it has written the messages to its own log but before the replicas have done so. This leaves a window of time where a failure of the leader will result in the messages being lost, although this should not be a common occurrence.

Write latency and throughput are negatively impacted by having more replicas acknowledge a write, so if you require low-latency, high throughput writes you may want to accept lower durability.

This behavior is controlled by the required_acks option to #producer and #async_producer:

# This is the default: all replicas must acknowledge.
producer = kafka.producer(required_acks: :all)

# This is fire-and-forget: messages can easily be lost.
producer = kafka.producer(required_acks: 0)

# This only waits for the leader to acknowledge.
producer = kafka.producer(required_acks: 1)

Unless you absolutely need lower latency it's highly recommended to use the default setting (:all).

Message Delivery Guarantees

There are basically two different and incompatible guarantees that can be made in a message delivery system such as Kafka:

  1. at-most-once delivery guarantees that a message is at most delivered to the recipient once. This is useful only if delivering the message twice carries some risk and should be avoided. Implicit is the fact that there's no guarantee that the message will be delivered at all.
  2. at-least-once delivery guarantees that a message is delivered, but it may be delivered more than once. If the final recipient de-duplicates messages, e.g. by checking a unique message id, then it's even possible to implement exactly-once delivery.

Of these two options, ruby-kafka implements the second one: when in doubt about whether a message has been delivered, a producer will try to deliver it again.

The guarantee is made only for the synchronous producer and boils down to this:

producer = kafka.producer

producer.produce("hello", topic: "greetings")

# If this line fails with Kafka::DeliveryFailed we *may* have succeeded in delivering
# the message to Kafka but won't know for sure.
producer.deliver_messages

# If we get to this line we can be sure that the message has been delivered to Kafka!

That is, once #deliver_messages returns we can be sure that Kafka has received the message. Note that there are some big caveats here:

  • Depending on how your cluster and topic is configured the message could still be lost by Kafka.
  • If you configure the producer to not require acknowledgements from the Kafka brokers by setting required_acks to zero there is no guarantee that the message will ever make it to a Kafka broker.
  • If you use the asynchronous producer there's no guarantee that messages will have been delivered after #deliver_messages returns. A way of blocking until a message has been delivered with the asynchronous producer may be implemented in the future.

It's possible to improve your chances of success when calling #deliver_messages, at the price of a longer max latency:

producer = kafka.producer(
  # The number of retries when attempting to deliver messages. The default is
  # 2, so 3 attempts in total, but you can configure a higher or lower number:
  max_retries: 5,

  # The number of seconds to wait between retries. In order to handle longer
  # periods of Kafka being unavailable, increase this number. The default is
  # 1 second.
  retry_backoff: 5,
)

Note that these values affect the max latency of the operation; see Understanding Timeouts for an explanation of the various timeouts and latencies.

If you use the asynchronous producer you typically don't have to worry too much about this, as retries will be done in the background.

Compression

Depending on what kind of data you produce, enabling compression may yield improved bandwidth and space usage. Compression in Kafka is done on entire messages sets rather than on individual messages. This improves the compression rate and generally means that compressions works better the larger your buffers get, since the message sets will be larger by the time they're compressed.

Since many workloads have variations in throughput and distribution across partitions, it's possible to configure a threshold for when to enable compression by setting compression_threshold. Only if the defined number of messages are buffered for a partition will the messages be compressed.

Compression is enabled by passing the compression_codec parameter to #producer with the name of one of the algorithms allowed by Kafka:

  • :snappy for Snappy compression.
  • :gzip for gzip compression.
  • :lz4 for LZ4 compression.
  • :zstd for zstd compression.

By default, all message sets will be compressed if you specify a compression codec. To increase the compression threshold, set compression_threshold to an integer value higher than one.

producer = kafka.producer(
  compression_codec: :snappy,
  compression_threshold: 10,
)

Producing Messages from a Rails Application

A typical use case for Kafka is tracking events that occur in web applications. Oftentimes it's advisable to avoid having a hard dependency on Kafka being available, allowing your application to survive a Kafka outage. By using an asynchronous producer, you can avoid doing IO within the individual request/response cycles, instead pushing that to the producer's internal background thread.

In this example, a producer is configured in a Rails initializer:

# config/initializers/kafka_producer.rb
require "kafka"

# Configure the Kafka client with the broker hosts and the Rails
# logger.
$kafka = Kafka.new(["kafka1:9092", "kafka2:9092"], logger: Rails.logger)

# Set up an asynchronous producer that delivers its buffered messages
# every ten seconds:
$kafka_producer = $kafka.async_producer(
  delivery_interval: 10,
)

# Make sure to shut down the producer when exiting.
at_exit { $kafka_producer.shutdown }

In your controllers, simply call the producer directly:

# app/controllers/orders_controller.rb
class OrdersController
  def create
    @order = Order.create!(params[:order])

    event = {
      order_id: @order.id,
      amount: @order.amount,
      timestamp: Time.now,
    }

    $kafka_producer.produce(event.to_json, topic: "order_events")
  end
end

Consuming Messages from Kafka

Note: If you're just looking to get started with Kafka consumers, you might be interested in visiting the Higher level libraries section that lists ruby-kafka based frameworks. Read on, if you're interested in either rolling your own executable consumers or if you want to learn more about how consumers work in Kafka.

Consuming messages from a Kafka topic with ruby-kafka is simple:

require "kafka"

kafka = Kafka.new(["kafka1:9092", "kafka2:9092"])

kafka.each_message(topic: "greetings") do |message|
  puts message.offset, message.key, message.value
end

While this is great for extremely simple use cases, there are a number of downsides:

  • You can only fetch from a single topic at a time.
  • If you want to have multiple processes consume from the same topic, there's no way of coordinating which processes should fetch from which partitions.
  • If the process dies, there's no way to have another process resume fetching from the point in the partition that the original process had reached.

Consumer Groups

The Consumer API solves all of the above issues, and more. It uses the Consumer Groups feature released in Kafka 0.9 to allow multiple consumer processes to coordinate access to a topic, assigning each partition to a single consumer. When a consumer fails, the partitions that were assigned to it are re-assigned to other members of the group.

Using the API is simple:

require "kafka"

kafka = Kafka.new(["kafka1:9092", "kafka2:9092"])

# Consumers with the same group id will form a Consumer Group together.
consumer = kafka.consumer(group_id: "my-consumer")

# It's possible to subscribe to multiple topics by calling `subscribe`
# repeatedly.
consumer.subscribe("greetings")

# Stop the consumer when the SIGTERM signal is sent to the process.
# It's better to shut down gracefully than to kill the process.
trap("TERM") { consumer.stop }

# This will loop indefinitely, yielding each message in turn.
consumer.each_message do |message|
  puts message.topic, message.partition
  puts message.offset, message.key, message.value
end

Each consumer process will be assigned one or more partitions from each topic that the group subscribes to. In order to handle more messages, simply start more processes.

Consumer Checkpointing

In order to be able to resume processing after a consumer crashes, each consumer will periodically checkpoint its position within each partition it reads from. Since each partition has a monotonically increasing sequence of message offsets, this works by committing the offset of the last message that was processed in a given partition. Kafka handles these commits and allows another consumer in a group to resume from the last commit when a member crashes or becomes unresponsive.

By default, offsets are committed every 10 seconds. You can increase the frequency, known as the offset commit interval, to limit the duration of double-processing scenarios, at the cost of a lower throughput due to the added coordination. If you want to improve throughput, and double-processing is of less concern to you, then you can decrease the frequency. Set the commit interval to zero in order to disable the timer-based commit trigger entirely.

In addition to the time based trigger it's possible to trigger checkpointing in response to n messages having been processed, known as the offset commit threshold. This puts a bound on the number of messages that can be double-processed before the problem is detected. Setting this to 1 will cause an offset commit to take place every time a message has been processed. By default this trigger is disabled (set to zero).

It is possible to trigger an immediate offset commit by calling Consumer#commit_offsets. This blocks the caller until the Kafka cluster has acknowledged the commit.

Stale offsets are periodically purged by the broker. The broker setting offsets.retention.minutes controls the retention window for committed offsets, and defaults to 1 day. The length of the retention window, known as offset retention time, can be changed for the consumer.

Previously committed offsets are re-committed, to reset the retention window, at the first commit and periodically at an interval of half the offset retention time.

consumer = kafka.consumer(
  group_id: "some-group",

  # Increase offset commit frequency to once every 5 seconds.
  offset_commit_interval: 5,

  # Commit offsets when 100 messages have been processed.
  offset_commit_threshold: 100,

  # Increase the length of time that committed offsets are kept.
  offset_retention_time: 7 * 60 * 60
)

For some use cases it may be necessary to control when messages are marked as processed. Note that since only the consumer position within each partition can be saved, marking a message as processed implies that all messages in the partition with a lower offset should also be considered as having been processed.

The method Consumer#mark_message_as_processed marks a message (and all those that precede it in a partition) as having been processed. This is an advanced API that you should only use if you know what you're doing.

# Manually controlling checkpointing:

# Typically you want to use this API in order to buffer messages until some
# special "commit" message is received, e.g. in order to group together
# transactions consisting of several items.
buffer = []

# Messages will not be marked as processed automatically. If you shut down the
# consumer without calling `#mark_message_as_processed` first, the consumer will
# not resume where you left off!
consumer.each_message(automatically_mark_as_processed: false) do |message|
  # Our messages are JSON with a `type` field and other stuff.
  event = JSON.parse(message.value)

  case event.fetch("type")
  when "add_to_cart"
    buffer << event
  when "complete_purchase"
    # We've received all the messages we need, time to save the transaction.
    save_transaction(buffer)

    # Now we can set the checkpoint by marking the last message as processed.
    consumer.mark_message_as_processed(message)

    # We can optionally trigger an immediate, blocking offset commit in order
    # to minimize the risk of crashing before the automatic triggers have
    # kicked in.
    consumer.commit_offsets

    # Make the buffer ready for the next transaction.
    buffer.clear
  end
end

Topic Subscriptions

For each topic subscription it's possible to decide whether to consume messages starting at the beginning of the topic or to just consume new messages that are produced to the topic. This policy is configured by setting the start_from_beginning argument when calling #subscribe:

# Consume messages from the very beginning of the topic. This is the default.
consumer.subscribe("users", start_from_beginning: true)

# Only consume new messages.
consumer.subscribe("notifications", start_from_beginning: false)

Once the consumer group has checkpointed its progress in the topic's partitions, the consumers will always start from the checkpointed offsets, regardless of start_from_beginning. As such, this setting only applies when the consumer initially starts consuming from a topic.

Shutting Down a Consumer

In order to shut down a running consumer process cleanly, call #stop on it. A common pattern is to trap a process signal and initiate the shutdown from there:

consumer = kafka.consumer(...)

# The consumer can be stopped from the command line by executing
# `kill -s TERM <process-id>`.
trap("TERM") { consumer.stop }

consumer.each_message do |message|
  ...
end

Consuming Messages in Batches

Sometimes it is easier to deal with messages in batches rather than individually. A batch is a sequence of one or more Kafka messages that all belong to the same topic and partition. One common reason to want to use batches is when some external system has a batch or transactional API.

# A mock search index that we'll be keeping up to date with new Kafka messages.
index = SearchIndex.new

consumer.subscribe("posts")

consumer.each_batch do |batch|
  puts "Received batch: #{batch.topic}/#{batch.partition}"

  transaction = index.transaction

  batch.messages.each do |message|
    # Let's assume that adding a document is idempotent.
    transaction.add(id: message.key, body: message.value)
  end

  # Once this method returns, the messages have been successfully written to the
  # search index. The consumer will only checkpoint a batch *after* the block
  # has completed without an exception.
  transaction.commit!
end

One important thing to note is that the client commits the offset of the batch's messages only after the entire batch has been processed.

Balancing Throughput and Latency

There are two performance properties that can at times be at odds: throughput and latency. Throughput is the number of messages that can be processed in a given timespan; latency is the time it takes from a message is written to a topic until it has been processed.

In order to optimize for throughput, you want to make sure to fetch as many messages as possible every time you do a round trip to the Kafka cluster. This minimizes network overhead and allows processing data in big chunks.

In order to optimize for low latency, you want to process a message as soon as possible, even if that means fetching a smaller batch of messages.

There are three values that can be tuned in order to balance these two concerns.

  • min_bytes is the minimum number of bytes to return from a single message fetch. By setting this to a high value you can increase the processing throughput. The default value is one byte.
  • max_wait_time is the maximum number of seconds to wait before returning data from a single message fetch. By setting this high you also increase the processing throughput – and by setting it low you set a bound on latency. This configuration overrides min_bytes, so you'll always get data back within the time specified. The default value is one second. If you want to have at most five seconds of latency, set max_wait_time to 5. You should make sure max_wait_time * num brokers + heartbeat_interval is less than session_timeout.
  • max_bytes_per_partition is the maximum amount of data a broker will return for a single partition when fetching new messages. The default is 1MB, but increasing this number may lead to better throughtput since you'll need to fetch less frequently. Setting it to a lower value is not recommended unless you have so many partitions that it's causing network and latency issues to transfer a fetch response from a broker to a client. Setting the number too high may result in instability, so be careful.

The first two settings can be passed to either #each_message or #each_batch, e.g.

# Waits for data for up to 5 seconds on each broker, preferring to fetch at least 5KB at a time.
# This can wait up to num brokers * 5 seconds.
consumer.each_message(min_bytes: 1024 * 5, max_wait_time: 5) do |message|
  # ...
end

The last setting is configured when subscribing to a topic, and can vary between topics:

# Fetches up to 5MB per partition at a time for better throughput.
consumer.subscribe("greetings", max_bytes_per_partition: 5 * 1024 * 1024)

consumer.each_message do |message|
  # ...
end

Customizing Partition Assignment Strategy

In some cases, you might want to assign more partitions to some consumers. For example, in applications inserting some records to a database, the consumers running on hosts nearby the database can process more messages than the consumers running on other hosts. You can use a custom assignment strategy by passing an object that implements #call as the argument assignment_strategy like below:

class CustomAssignmentStrategy
  def initialize(user_data)
    @user_data = user_data
  end

  # Assign the topic partitions to the group members.
  #
  # @param cluster [Kafka::Cluster]
  # @param members [Hash<String, Kafka::Protocol::JoinGroupResponse::Metadata>] a hash
  #   mapping member ids to metadata
  # @param partitions [Array<Kafka::ConsumerGroup::Assignor::Partition>] a list of
  #   partitions the consumer group processes
  # @return [Hash<String, Array<Kafka::ConsumerGroup::Assignor::Partition>] a hash
  #   mapping member ids to partitions.
  def call(cluster:, members:, partitions:)
    ...
  end
end

strategy = CustomAssignmentStrategy.new("some-host-information")
consumer = kafka.consumer(group_id: "some-group", assignment_strategy: strategy)

members is a hash mapping member IDs to metadata, and partitions is a list of partitions the consumer group processes. The method call must return a hash mapping member IDs to partitions. For example, the following strategy assigns partitions randomly:

class RandomAssignmentStrategy
  def call(cluster:, members:, partitions:)
    member_ids = members.keys
    partitions.each_with_object(Hash.new {|h, k| h[k] = [] }) do |partition, partitions_per_member|
      partitions_per_member[member_ids[rand(member_ids.count)]] << partition
    end
  end
end

If the strategy needs user data, you should define the method user_data that returns user data on each consumer. For example, the following strategy uses the consumers' IP addresses as user data:

class NetworkTopologyAssignmentStrategy
  def user_data
    Socket.ip_address_list.find(&:ipv4_private?).ip_address
  end

  def call(cluster:, members:, partitions:)
    # Display the pair of the member ID and IP address
    members.each do |id, metadata|
      puts "#{id}: #{metadata.user_data}"
    end

    # Assign partitions considering the network topology
    ...
  end
end

Note that the strategy uses the class name as the default protocol name. You can change it by defining the method protocol_name:

class NetworkTopologyAssignmentStrategy
  def protocol_name
    "networktopology"
  end

  def user_data
    Socket.ip_address_list.find(&:ipv4_private?).ip_address
  end

  def call(cluster:, members:, partitions:)
    ...
  end
end

As the method call might receive different user data from what it expects, you should avoid using the same protocol name as another strategy that uses different user data.

Thread Safety

You typically don't want to share a Kafka client object between threads, since the network communication is not synchronized. Furthermore, you should avoid using threads in a consumer unless you're very careful about waiting for all work to complete before returning from the #each_message or #each_batch block. This is because checkpointing assumes that returning from the block means that the messages that have been yielded have been successfully processed.

You should also avoid sharing a synchronous producer between threads, as the internal buffers are not thread safe. However, the asynchronous producer should be safe to use in a multi-threaded environment. This is because producers, when instantiated, get their own copy of any non-thread-safe data such as network sockets. Furthermore, the asynchronous producer has been designed in such a way to only a single background thread operates on this data while any foreground thread with a reference to the producer object can only send messages to that background thread over a safe queue. Therefore it is safe to share an async producer object between many threads.

Logging

It's a very good idea to configure the Kafka client with a logger. All important operations and errors are logged. When instantiating your client, simply pass in a valid logger:

logger = Logger.new("log/kafka.log")
kafka = Kafka.new(logger: logger, ...)

By default, nothing is logged.

Instrumentation

Most operations are instrumented using Active Support Notifications. In order to subscribe to notifications, make sure to require the notifications library:

require "active_support/notifications"
require "kafka"

The notifications are namespaced based on their origin, with separate namespaces for the producer and the consumer.

In order to receive notifications you can either subscribe to individual notification names or use regular expressions to subscribe to entire namespaces. This example will subscribe to all notifications sent by ruby-kafka:

ActiveSupport::Notifications.subscribe(/.*\.kafka$/) do |*args|
  event = ActiveSupport::Notifications::Event.new(*args)
  puts "Received notification `#{event.name}` with payload: #{event.payload.inspect}"
end

All notification events have the client_id key in the payload, referring to the Kafka client id.

Producer Notifications

produce_message.producer.kafka is sent whenever a message is produced to a buffer. It includes the following payload:

  • value is the message value.
  • key is the message key.
  • topic is the topic that the message was produced to.
  • buffer_size is the size of the producer buffer after adding the message.
  • max_buffer_size is the maximum size of the producer buffer.

deliver_messages.producer.kafka is sent whenever a producer attempts to deliver its buffered messages to the Kafka brokers. It includes the following payload:

  • attempts is the number of times delivery was attempted.
  • message_count is the number of messages for which delivery was attempted.
  • delivered_message_count is the number of messages that were acknowledged by the brokers - if this number is smaller than message_count not all messages were successfully delivered.

Consumer Notifications

All notifications have group_id in the payload, referring to the Kafka consumer group id.

process_message.consumer.kafka is sent whenever a message is processed by a consumer. It includes the following payload:

  • value is the message value.
  • key is the message key.
  • topic is the topic that the message was consumed from.
  • partition is the topic partition that the message was consumed from.
  • offset is the message's offset within the topic partition.
  • offset_lag is the number of messages within the topic partition that have not yet been consumed.

start_process_message.consumer.kafka is sent before process_message.consumer.kafka, and contains the same payload. It is delivered before the message is processed, rather than after.

process_batch.consumer.kafka is sent whenever a message batch is processed by a consumer. It includes the following payload:

  • message_count is the number of messages in the batch.
  • topic is the topic that the message batch was consumed from.
  • partition is the topic partition that the message batch was consumed from.
  • highwater_mark_offset is the message batch's highest offset within the topic partition.
  • offset_lag is the number of messages within the topic partition that have not yet been consumed.

start_process_batch.consumer.kafka is sent before process_batch.consumer.kafka, and contains the same payload. It is delivered before the batch is processed, rather than after.

join_group.consumer.kafka is sent whenever a consumer joins a consumer group. It includes the following payload:

  • group_id is the consumer group id.

sync_group.consumer.kafka is sent whenever a consumer is assigned topic partitions within a consumer group. It includes the following payload:

  • group_id is the consumer group id.

leave_group.consumer.kafka is sent whenever a consumer leaves a consumer group. It includes the following payload:

  • group_id is the consumer group id.

seek.consumer.kafka is sent when a consumer first seeks to an offset. It includes the following payload:

  • group_id is the consumer group id.
  • topic is the topic we are seeking in.
  • partition is the partition we are seeking in.
  • offset is the offset we have seeked to.

heartbeat.consumer.kafka is sent when a consumer group completes a heartbeat. It includes the following payload:

  • group_id is the consumer group id.
  • topic_partitions is a hash of { topic_name => array of assigned partition IDs }

Connection Notifications

  • request.connection.kafka is sent whenever a network request is sent to a Kafka broker. It includes the following payload:
    • api is the name of the API that was called, e.g. produce or fetch.
    • request_size is the number of bytes in the request.
    • response_size is the number of bytes in the response.

Monitoring

It is highly recommended that you monitor your Kafka client applications in production. Typical problems you'll see are:

  • high network error rates, which may impact performance and time-to-delivery;
  • producer buffer growth, which may indicate that producers are unable to deliver messages at the rate they're being produced;
  • consumer processing errors, indicating exceptions are being raised in the processing code;
  • frequent consumer rebalances, which may indicate unstable network conditions or consumer configurations.

You can quite easily build monitoring on top of the provided instrumentation hooks. In order to further help with monitoring, a prebuilt Statsd and Datadog reporter is included with ruby-kafka.

What to Monitor

We recommend monitoring the following:

  • Low-level Kafka API calls:
    • The rate of API call errors to the total number of calls by both API and broker.
    • The API call throughput by both API and broker.
    • The API call latency by both API and broker.
  • Producer-level metrics:
    • Delivery throughput by topic.
    • The latency of deliveries.
    • The producer buffer fill ratios.
    • The async producer queue sizes.
    • Message delivery delays.
    • Failed delivery attempts.
  • Consumer-level metrics:
    • Message processing throughput by topic.
    • Processing latency by topic.
    • Processing errors by topic.
    • Consumer lag (how many messages are yet to be processed) by topic/partition.
    • Group join/sync/leave by client host.

Reporting Metrics to Statsd

The Statsd reporter is automatically enabled when the kafka/statsd library is required. You can optionally change the configuration.

require "kafka/statsd"

# Default is "ruby_kafka".
Kafka::Statsd.namespace = "custom-namespace"

# Default is "127.0.0.1".
Kafka::Statsd.host = "statsd.something.com"

# Default is 8125.
Kafka::Statsd.port = 1234

Reporting Metrics to Datadog

The Datadog reporter is automatically enabled when the kafka/datadog library is required. You can optionally change the configuration.

# This enables the reporter:
require "kafka/datadog"

# Default is "ruby_kafka".
Kafka::Datadog.namespace = "custom-namespace"

# Default is "127.0.0.1".
Kafka::Datadog.host = "statsd.something.com"

# Default is 8125.
Kafka::Datadog.port = 1234

Understanding Timeouts

It's important to understand how timeouts work if you have a latency sensitive application. This library allows configuring timeouts on different levels:

Network timeouts apply to network connections to individual Kafka brokers. There are two config keys here, each passed to Kafka.new:

  • connect_timeout sets the number of seconds to wait while connecting to a broker for the first time. When ruby-kafka initializes, it needs to connect to at least one host in seed_brokers in order to discover the Kafka cluster. Each host is tried until there's one that works. Usually that means the first one, but if your entire cluster is down, or there's a network partition, you could wait up to n * connect_timeout seconds, where n is the number of seed brokers.
  • socket_timeout sets the number of seconds to wait when reading from or writing to a socket connection to a broker. After this timeout expires the connection will be killed. Note that some Kafka operations are by definition long-running, such as waiting for new messages to arrive in a partition, so don't set this value too low. When configuring timeouts relating to specific Kafka operations, make sure to make them shorter than this one.

Producer timeouts can be configured when calling #producer on a client instance:

  • ack_timeout is a timeout executed by a broker when the client is sending messages to it. It defines the number of seconds the broker should wait for replicas to acknowledge the write before responding to the client with an error. As such, it relates to the required_acks setting. It should be set lower than socket_timeout.
  • retry_backoff configures the number of seconds to wait after a failed attempt to send messages to a Kafka broker before retrying. The max_retries setting defines the maximum number of retries to attempt, and so the total duration could be up to max_retries * retry_backoff seconds. The timeout can be arbitrarily long, and shouldn't be too short: if a broker goes down its partitions will be handed off to another broker, and that can take tens of seconds.

When sending many messages, it's likely that the client needs to send some messages to each broker in the cluster. Given n brokers in the cluster, the total wait time when calling Kafka::Producer#deliver_messages can be up to

n * (connect_timeout + socket_timeout + retry_backoff) * max_retries

Make sure your application can survive being blocked for so long.

Security

Encryption and Authentication using SSL

By default, communication between Kafka clients and brokers is unencrypted and unauthenticated. Kafka 0.9 added optional support for encryption and client authentication and authorization. There are two layers of security made possible by this:

Encryption of Communication

By enabling SSL encryption you can have some confidence that messages can be sent to Kafka over an untrusted network without being intercepted.

In this case you just need to pass a valid CA certificate as a string when configuring your Kafka client:

kafka = Kafka.new(["kafka1:9092"], ssl_ca_cert: File.read('my_ca_cert.pem'))

Without passing the CA certificate to the client it would be impossible to protect against man-in-the-middle attacks.

Using your system's CA cert store

If you want to use the CA certs from your system's default certificate store, you can use:

kafka = Kafka.new(["kafka1:9092"], ssl_ca_certs_from_system: true)

This configures the store to look up CA certificates from the system default certificate store on an as needed basis. The location of the store can usually be determined by: OpenSSL::X509::DEFAULT_CERT_FILE

Client Authentication

In order to authenticate the client to the cluster, you need to pass in a certificate and key created for the client and trusted by the brokers.

NOTE: You can disable hostname validation by passing ssl_verify_hostname: false.

kafka = Kafka.new(
  ["kafka1:9092"],
  ssl_ca_cert: File.read('my_ca_cert.pem'),
  ssl_client_cert: File.read('my_client_cert.pem'),
  ssl_client_cert_key: File.read('my_client_cert_key.pem'),
  ssl_client_cert_key_password: 'my_client_cert_key_password',
  ssl_verify_hostname: false,
  # ...
)

Once client authentication is set up, it is possible to configure the Kafka cluster to authorize client requests.

Using JKS Certificates

Typically, Kafka certificates come in the JKS format, which isn't supported by ruby-kafka. There's a wiki page that describes how to generate valid X509 certificates from JKS certificates.

Authentication using SASL

Kafka has support for using SASL to authenticate clients. Currently GSSAPI, SCRAM and PLAIN mechanisms are supported by ruby-kafka.

NOTE: With SASL for authentication, it is highly recommended to use SSL encryption. The default behavior of ruby-kafka enforces you to use SSL and you need to configure SSL encryption by passing ssl_ca_cert or enabling ssl_ca_certs_from_system. However, this strict SSL mode check can be disabled by setting sasl_over_ssl to false while initializing the client.

GSSAPI

In order to authenticate using GSSAPI, set your principal and optionally your keytab when initializing the Kafka client:

kafka = Kafka.new(
  ["kafka1:9092"],
  sasl_gssapi_principal: 'kafka/kafka.example.com@EXAMPLE.COM',
  sasl_gssapi_keytab: '/etc/keytabs/kafka.keytab',
  # ...
)

AWS MSK (IAM)

In order to authenticate using IAM w/ an AWS MSK cluster, set your access key, secret key, and region when initializing the Kafka client:

k = Kafka.new(
  ["kafka1:9092"],
  sasl_aws_msk_iam_access_key_id: 'iam_access_key',
  sasl_aws_msk_iam_secret_key_id: 'iam_secret_key',
  sasl_aws_msk_iam_aws_region: 'us-west-2',
  ssl_ca_certs_from_system: true,
  # ...
)

PLAIN

In order to authenticate using PLAIN, you must set your username and password when initializing the Kafka client:

kafka = Kafka.new(
  ["kafka1:9092"],
  ssl_ca_cert: File.read('/etc/openssl/cert.pem'),
  sasl_plain_username: 'username',
  sasl_plain_password: 'password'
  # ...
)

SCRAM

Since 0.11 kafka supports SCRAM.

kafka = Kafka.new(
  ["kafka1:9092"],
  sasl_scram_username: 'username',
  sasl_scram_password: 'password',
  sasl_scram_mechanism: 'sha256',
  # ...
)

OAUTHBEARER

This mechanism is supported in kafka >= 2.0.0 as of KIP-255

In order to authenticate using OAUTHBEARER, you must set the client with an instance of a class that implements a token method (the interface is described in Kafka::Sasl::OAuth) which returns an ID/Access token.

Optionally, the client may implement an extensions method that returns a map of key-value pairs. These can be sent with the SASL/OAUTHBEARER initial client response. This is only supported in kafka >= 2.1.0.

class TokenProvider
  def token
    "some_id_token"
  end
end
# ...
client = Kafka.new(
  ["kafka1:9092"],
  sasl_oauth_token_provider: TokenProvider.new
)

Topic management

In addition to producing and consuming messages, ruby-kafka supports managing Kafka topics and their configurations. See the Kafka documentation for a full list of topic configuration keys.

List all topics

Return an array of topic names.

kafka = Kafka.new(["kafka:9092"])
kafka.topics
# => ["topic1", "topic2", "topic3"]

Create a topic

kafka = Kafka.new(["kafka:9092"])
kafka.create_topic("topic")

By default, the new topic has 1 partition, replication factor 1 and default configs from the brokers. Those configurations are customizable:

kafka = Kafka.new(["kafka:9092"])
kafka.create_topic("topic",
  num_partitions: 3,
  replication_factor: 2,
  config: {
    "max.message.bytes" => 100000
  }
)

Create more partitions for a topic

After a topic is created, you can increase the number of partitions for the topic. The new number of partitions must be greater than the current one.

kafka = Kafka.new(["kafka:9092"])
kafka.create_partitions_for("topic", num_partitions: 10)

Fetch configuration for a topic (alpha feature)

kafka = Kafka.new(["kafka:9092"])
kafka.describe_topic("topic", ["max.message.bytes", "retention.ms"])
# => {"max.message.bytes"=>"100000", "retention.ms"=>"604800000"}

Alter a topic configuration (alpha feature)

Update the topic configurations.

NOTE: This feature is for advanced usage. Only use this if you know what you're doing.

kafka = Kafka.new(["kafka:9092"])
kafka.alter_topic("topic", "max.message.bytes" => 100000, "retention.ms" => 604800000)

Delete a topic

kafka = Kafka.new(["kafka:9092"])
kafka.delete_topic("topic")

After a topic is marked as deleted, Kafka only hides it from clients. It would take a while before a topic is completely deleted.

Design

The library has been designed as a layered system, with each layer having a clear responsibility:

  • The network layer handles low-level connection tasks, such as keeping open connections to each Kafka broker, reconnecting when there's an error, etc. See Kafka::Connection for more details.
  • The protocol layer is responsible for encoding and decoding the Kafka protocol's various structures. See Kafka::Protocol for more details.
  • The operational layer provides high-level operations, such as fetching messages from a topic, that may involve more than one API request to the Kafka cluster. Some complex operations are made available through Kafka::Cluster, which represents an entire cluster, while simpler ones are only available through Kafka::Broker, which represents a single Kafka broker. In general, Kafka::Cluster is the high-level API, with more polish.
  • The API layer provides APIs to users of the libraries. The Consumer API is implemented in Kafka::Consumer while the Producer API is implemented in Kafka::Producer and Kafka::AsyncProducer.
  • The configuration layer provides a way to set up and configure the client, as well as easy entrypoints to the various APIs. Kafka::Client implements the public APIs. For convenience, the method Kafka.new can instantiate the class for you.

Note that only the API and configuration layers have any backwards compatibility guarantees – the other layers are considered internal and may change without warning. Don't use them directly.

Producer Design

The producer is designed with resilience and operational ease of use in mind, sometimes at the cost of raw performance. For instance, the operation is heavily instrumented, allowing operators to monitor the producer at a very granular level.

The producer has two main internal data structures: a list of pending messages and a message buffer. When the user calls Kafka::Producer#produce, a message is appended to the pending message list, but no network communication takes place. This means that the call site does not have to handle the broad range of errors that can happen at the network or protocol level. Instead, those errors will only happen once Kafka::Producer#deliver_messages is called. This method will go through the pending messages one by one, making sure they're assigned a partition. This may fail for some messages, as it could require knowing the current configuration for the message's topic, necessitating API calls to Kafka. Messages that cannot be assigned a partition are kept in the list, while the others are written into the message buffer. The producer then figures out which topic partitions are led by which Kafka brokers so that messages can be sent to the right place – in Kafka, it is the responsibility of the client to do this routing. A separate produce API request will be sent to each broker; the response will be inspected; and messages that were acknowledged by the broker will be removed from the message buffer. Any messages that were not acknowledged will be kept in the buffer.

If there are any messages left in either the pending message list or the message buffer after this operation, Kafka::DeliveryFailed will be raised. This exception must be rescued and handled by the user, possibly by calling #deliver_messages at a later time.

Asynchronous Producer Design

The synchronous producer allows the user fine-grained control over when network activity and the possible errors arising from that will take place, but it requires the user to handle the errors nonetheless. The async producer provides a more hands-off approach that trades off control for ease of use and resilience.

Instead of writing directly into the pending message list, Kafka::AsyncProducer writes the message to an internal thread-safe queue, returning immediately. A background thread reads messages off the queue and passes them to a synchronous producer.

Rather than triggering message deliveries directly, users of the async producer will typically set up automatic triggers, such as a timer.

Consumer Design

The Consumer API is designed for flexibility and stability. The first is accomplished by not dictating any high-level object model, instead opting for a simple loop-based approach. The second is accomplished by handling group membership, heartbeats, and checkpointing automatically. Messages are marked as processed as soon as they've been successfully yielded to the user-supplied processing block, minimizing the cost of processing errors.

Development

After checking out the repo, run bin/setup to install dependencies. Then, run rake spec to run the tests. You can also run bin/console for an interactive prompt that will allow you to experiment.

Note: the specs require a working Docker instance, but should work out of the box if you have Docker installed. Please create an issue if that's not the case.

If you would like to contribute to ruby-kafka, please join our Slack team and ask how best to do it.

Circle CI

Support and Discussion

If you've discovered a bug, please file a Github issue, and make sure to include all the relevant information, including the version of ruby-kafka and Kafka that you're using.

If you have other questions, or would like to discuss best practises, how to contribute to the project, or any other ruby-kafka related topic, join our Slack team!

Roadmap

Version 0.4 will be the last minor release with support for the Kafka 0.9 protocol. It is recommended that you pin your dependency on ruby-kafka to ~> 0.4.0 in order to receive bugfixes and security updates. New features will only target version 0.5 and up, which will be incompatible with the Kafka 0.9 protocol.

v0.4

Last stable release with support for the Kafka 0.9 protocol. Bug and security fixes will be released in patch updates.

v0.5

Latest stable release, with native support for the Kafka 0.10 protocol and eventually newer protocol versions. Kafka 0.9 is no longer supported by this release series.

Higher level libraries

Currently, there are three actively developed frameworks based on ruby-kafka, that provide higher level API that can be used to work with Kafka messages and two libraries for publishing messages.

Message processing frameworks

Racecar - A simple framework that integrates with Ruby on Rails to provide a seamless way to write, test, configure, and run Kafka consumers. It comes with sensible defaults and conventions.

Karafka - Framework used to simplify Apache Kafka based Ruby and Rails applications development. Karafka provides higher abstraction layers, including Capistrano, Docker and Heroku support.

Phobos - Micro framework and library for applications dealing with Apache Kafka. It wraps common behaviors needed by consumers and producers in an easy and convenient API.

Message publishing libraries

DeliveryBoy – A library that integrates with Ruby on Rails, making it easy to publish Kafka messages from any Rails application.

WaterDrop – A library for Ruby and Ruby on Rails applications, to easy publish Kafka messages in both sync and async way.

Why Create A New Library?

There are a few existing Kafka clients in Ruby:

  • Poseidon seems to work for Kafka 0.8, but the project is unmaintained and has known issues.
  • Hermann wraps the C library librdkafka and seems to be very efficient, but its API and mode of operation is too intrusive for our needs.
  • jruby-kafka is a great option if you're running on JRuby.

We needed a robust client that could be used from our existing Ruby apps, allowed our Ops to monitor operation, and provided flexible error handling. There didn't exist such a client, hence this project.

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/zendesk/ruby-kafka.

Author: Zendesk
Source Code: https://github.com/zendesk/ruby-kafka 
License: Apache-2.0 license

#ruby #kafka 

Adaline  Kulas

Adaline Kulas

1594162500

Multi-cloud Spending: 8 Tips To Lower Cost

A multi-cloud approach is nothing but leveraging two or more cloud platforms for meeting the various business requirements of an enterprise. The multi-cloud IT environment incorporates different clouds from multiple vendors and negates the dependence on a single public cloud service provider. Thus enterprises can choose specific services from multiple public clouds and reap the benefits of each.

Given its affordability and agility, most enterprises opt for a multi-cloud approach in cloud computing now. A 2018 survey on the public cloud services market points out that 81% of the respondents use services from two or more providers. Subsequently, the cloud computing services market has reported incredible growth in recent times. The worldwide public cloud services market is all set to reach $500 billion in the next four years, according to IDC.

By choosing multi-cloud solutions strategically, enterprises can optimize the benefits of cloud computing and aim for some key competitive advantages. They can avoid the lengthy and cumbersome processes involved in buying, installing and testing high-priced systems. The IaaS and PaaS solutions have become a windfall for the enterprise’s budget as it does not incur huge up-front capital expenditure.

However, cost optimization is still a challenge while facilitating a multi-cloud environment and a large number of enterprises end up overpaying with or without realizing it. The below-mentioned tips would help you ensure the money is spent wisely on cloud computing services.

  • Deactivate underused or unattached resources

Most organizations tend to get wrong with simple things which turn out to be the root cause for needless spending and resource wastage. The first step to cost optimization in your cloud strategy is to identify underutilized resources that you have been paying for.

Enterprises often continue to pay for resources that have been purchased earlier but are no longer useful. Identifying such unused and unattached resources and deactivating it on a regular basis brings you one step closer to cost optimization. If needed, you can deploy automated cloud management tools that are largely helpful in providing the analytics needed to optimize the cloud spending and cut costs on an ongoing basis.

  • Figure out idle instances

Another key cost optimization strategy is to identify the idle computing instances and consolidate them into fewer instances. An idle computing instance may require a CPU utilization level of 1-5%, but you may be billed by the service provider for 100% for the same instance.

Every enterprise will have such non-production instances that constitute unnecessary storage space and lead to overpaying. Re-evaluating your resource allocations regularly and removing unnecessary storage may help you save money significantly. Resource allocation is not only a matter of CPU and memory but also it is linked to the storage, network, and various other factors.

  • Deploy monitoring mechanisms

The key to efficient cost reduction in cloud computing technology lies in proactive monitoring. A comprehensive view of the cloud usage helps enterprises to monitor and minimize unnecessary spending. You can make use of various mechanisms for monitoring computing demand.

For instance, you can use a heatmap to understand the highs and lows in computing visually. This heat map indicates the start and stop times which in turn lead to reduced costs. You can also deploy automated tools that help organizations to schedule instances to start and stop. By following a heatmap, you can understand whether it is safe to shut down servers on holidays or weekends.

#cloud computing services #all #hybrid cloud #cloud #multi-cloud strategy #cloud spend #multi-cloud spending #multi cloud adoption #why multi cloud #multi cloud trends #multi cloud companies #multi cloud research #multi cloud market

Mireille  Von

Mireille Von

1625334540

Kafka Streams using Spring Cloud Stream | Microservices Example | Tech Primers

This video covers how to leverage Kafka Streams using Spring Cloud stream by creating multiple spring boot microservices

📌 Related Links

🔗 Kafka setup: https://docs.confluent.io/platform/current/quickstart/cos-docker-quickstart.html
🔗 Public Domain API: https://domainsdb.info/

📌 Related Videos

🔗 Spring Boot with Spring Kafka Producer example - https://youtu.be/NjHYWEV_E_o
🔗 Spring Boot with Spring Kafka Consumer example - https://youtu.be/IncG0_XSSBg

📌 Related Playlist

🔗Spring Boot Primer - https://www.youtube.com/playlist?list=PLTyWtrsGknYegrUmDZB6rcqMotOFZKvbn
🔗Spring Cloud Primer - https://www.youtube.com/playlist?list=PLTyWtrsGknYeOJHtd3Ll93GRf28hrjlHV
🔗Spring Microservices Primer - https://www.youtube.com/playlist?list=PLTyWtrsGknYdZlO7LAZFEElWkEk59Y2ak
🔗Spring JPA Primer - https://www.youtube.com/playlist?list=PLTyWtrsGknYdt079e1pyvpgLrJ48RQ1LK
🔗Java 8 Streams - https://www.youtube.com/playlist?list=PLTyWtrsGknYdqY_7lwcbJ1z4bvc5yEEZl
🔗Spring Security Primer - https://www.youtube.com/playlist?list=PLTyWtrsGknYe0Sba9o-JRtnRlkl4gXMQl

💥 Join TechPrimers Slack Community: https://bit.ly/JoinTechPrimers
💥 Telegram: https://t.me/TechPrimers
💥 TechPrimer HindSight (Blog): https://medium.com/TechPrimers
💥 Website: http://techprimers.com
💥 Slack Community: https://techprimers.slack.com
💥 Twitter: https://twitter.com/TechPrimers
💥 Facebook: http://fb.me/TechPrimers
💥 GitHub: https://github.com/TechPrimers or https://techprimers.github.io/

🎬Video Editing: FCP


🔥 Disclaimer/Policy:
The content/views/opinions posted here are solely mine and the code samples created by me are open sourced.
You are free to use the code samples in Github after forking and you can modify it for your own use.
All the videos posted here are copyrighted. You cannot re-distribute videos on this channel in other channels or platforms.
#KafkaStreams #SpringCloudStream #TechPrimers

#kafka streams #kafka #spring cloud stream #spring cloud

Ysia Tamas

1598393640

Spring Cloud Stream with Apache Kafka & RabbitMQ Producer Consumer example

In this video, You will learn how to create a Kafka & RabbitMQ Producer and Consumer by using spring cloud stream

#spring #rabbitmq #kafka