Flink real-time JOIN: tables 4+2 TB

# Implementing Real-Time JOIN for Large Tables on Apache Flink: Basic Approach

We need to enable real-time updates for the INNER JOIN result between the users table (4 TB) and domains (2 TB). The query returns user_id, firstname, lastname, and domain_name. Traditional views overload the data source (OLTP), while simple materialization via CDC into DWH doesn't provide the required speed.

Example data:

users:

| id | firstname | lastname |

|----|-----------|----------|

| 1 | Egor | Myasnik |

| 2 | Pavel | Hvastun |

| 3 | Mitya | Volk |

domains:

| id | user_id | domain_name |

|----|---------|-------------|

| 1 | 1 | Approval |

| 2 | 1 | Rejection |

| 3 | 1 | Stoppage |

| 4 | 3 | Cancellation|

Analysis of Alternatives

• Direct view on OLTP — unacceptable due to load on the OLTP system.
• CDC → Kafka → DWH — unloads OLTP, but read speed from DWH is similar to the original. Adds two layers: Kafka and the view.

Both approaches fail to provide real-time updates for the JOIN result.

Flink Architecture

Solution: CDC streams into Kafka → Flink Table API → stateful JOIN → sink (console for MVP).

Maven Dependencies

<dependency>
    <groupId>org.apache.flink</groupId>
    <artifactId>flink-connector-kafka</artifactId>
    <version>3.2.0-1.19</version>
</dependency>
<dependency>
    <groupId>org.apache.flink</groupId>
    <artifactId>flink-sql-client</artifactId>
    <version>1.19</version>
</dependency>
<dependency>
    <groupId>org.apache.flink</groupId>
    <artifactId>flink-java</artifactId>
    <version>${flink.version}</version>
</dependency>
<dependency>
    <groupId>org.apache.flink</groupId>
    <artifactId>flink-table-planner-loader</artifactId>
    <version>${flink.version}</version>
</dependency>

Data Models

public class User implements Serializable {
    public Integer id;
    public String firstname;
    public String lastname;
    // getters/setters
    public static User fromRow(Row row) { /* mapper */ }
}

public class Domain implements Serializable {
    public Integer id;
    public Integer user_id;
    public String domain_name;
    // getters/setters
    public static Domain fromRow(Row row) { /* mapper */ }
}

Creating Kafka Tables

tableEnv.executeSql("CREATE TABLE users (" +
    "`before` ROW<id: INT, firstname: STRING, lastname: STRING>," +
    "`op` STRING," +
    "`after` ROW<id: INT, firstname: STRING, lastname: STRING>," +
    ") WITH (" +
    "'connector' = 'kafka'," +
    "'topic' = 'users_topic'," +
    "'properties.bootstrap.servers' = 'kafka-brokers'," +
    "'properties.group.id' = 'users_consumer_group'," +
    "'scan.startup.mode' = 'earliest'");

DataStream<User> users = tableEnv.toDataStream(tableEnv.from("users")).map(User::fromRow);

// Similarly for domains
tableEnv.executeSql("CREATE TABLE domains (...)");
DataStream<Domain> domains = tableEnv.toDataStream(tableEnv.from("domains")).map(Domain::fromRow);

Stateful JOIN Implementation

Joining streams by key user_id == domain.user_id using KeyedCoProcessFunction:

users
    .connect(domains)
    .keyBy(
        user -> user.id,
        domain -> domain.user_id
    )
    .process(new Join1())
    .print();

KeyedCoProcessFunction

public class Join1 extends KeyedCoProcessFunction<Integer, User, Domain, Output> {
    private MapState<Integer, Domain> domainsState;
    private ValueState<User> usersState;

    @Override
    public void processElement1(User user, Context ctx, Collector<Output> out) throws Exception {
        usersState.update(user);
        for (Domain domain : domainsState.values()) {
            out.collect(new Output(user.id, user.firstname, user.lastname, domain.domain_name));
        }
    }

    @Override
    public void processElement2(Domain domain, Context ctx, Collector<Output> out) throws Exception {
        domainsState.put(domain.id, domain);
        User user = usersState.value();
        if (user != null) {
            out.collect(new Output(user.id, user.firstname, user.lastname, domain.domain_name));
        }
    }

    @Override
    public void open(OpenContext openContext) throws Exception {
        usersState = getRuntimeContext().getState(new ValueStateDescriptor<>("users", User.class));
        domainsState = getRuntimeContext().getMapState(
            new MapStateDescriptor<>("domains", Integer.class, Domain.class));
    }

    public static class Output implements Serializable {
        public Integer user_id;
        public String firstname, lastname, domain_name;
        // constructors/getters/setters
    }
}

MapState is used for domains due to the one-to-many relationship. State operations are O(1).

Key Points

• Real-time updates: JOIN result is up-to-date upon CDC event arrival.
• Stateful processing: ValueState for users (1:1), MapState for domains (1:N).
• Scalability: proven on terabyte-scale volumes with OLAP sink.
• Performance: O(1) state access, low latency.
• Next steps: handle DELETE/UPDATE, RocksDB, skew balancing.

Production Recommendations

• Use OLAP sink (ClickHouse, Pinot) instead of console.
• Process op flags from CDC for correct UPDATE/DELETE handling.
• Use RocksDB for state backend with large volumes.
• Monitor key skew and distribute partitions.
• Test scaling on a cluster.

— Editorial Team