实战自定义Flink SQL Connector（w/ Flink 1.11 & Redis）

Foreword

Flink SQL之所以简洁易用而功能强大，其中一个重要因素就是其拥有丰富的Connector(连接器)组件。Connector是Flink与外部系统交互的载体，并分为负责读取的Source和负责写入的Sink两大类。不过，Flink SQL内置的Connector有可能无法cover实际业务中的种种需求，需要我们自行定制。好在社区已经提供了一套标准化、易于扩展的体系，用户只要按照规范面向接口编程，就能轻松打造自己的Connector。本文就在现有Bahir Flink项目的基础上逐步实现一个SQL化的Redis Connector。

Introducing DynamicTableSource/Sink

当前(Flink 1.11+)Flink SQL Connector的架构简图如下所示，设计文档可参见FLIP-95。

动态表(dynamic table)一直都是Flink SQL流批一体化的重要概念，也是上述架构中Planning阶段的核心。而自定义Connector的主要工作就是实现基于动态表的Source/Sink，还包括上游产生它的工厂，以及下游在Runtime阶段实际执行Source/Sink逻辑的RuntimeProvider。Metadata阶段的表元数据则由Catalog维护。

前方海量代码预警。

Implementing RedisDynamicTableFactory

DynamicTableFactory需要具备以下功能：

• 定义与校验建表时传入的各项参数；
• 获取表的元数据；
• 定义读写数据时的编码/解码格式(非必需)；
• 创建可用的DynamicTable[Source/Sink]实例。

实现了DynamicTable[Source/Sink]Factory接口的工厂类骨架如下所示。

public class RedisDynamicTableFactory implements DynamicTableSourceFactory, DynamicTableSinkFactory {
  @Override
  public DynamicTableSource createDynamicTableSource(Context context) { }
 
  @Override
  public DynamicTableSink createDynamicTableSink(Context context) { }
 
  @Override
  public String factoryIdentifier() { }
 
  @Override
  public Set<ConfigOption<?>> requiredOptions() { }
 
  @Override
  public Set<ConfigOption<?>> optionalOptions() { }
}

首先来定义Redis Connector需要的各项参数，利用内置的ConfigOption/ConfigOptions类即可。它们的含义都很简单，不再赘述。

public static final ConfigOption<String> MODE = ConfigOptions
    .key("mode")
    .stringType()
    .defaultValue("single");
  public static final ConfigOption<String> SINGLE_HOST = ConfigOptions
    .key("single.host")
    .stringType()
    .defaultValue(Protocol.DEFAULT_HOST);
  public static final ConfigOption<Integer> SINGLE_PORT = ConfigOptions
    .key("single.port")
    .intType()
    .defaultValue(Protocol.DEFAULT_PORT);
  public static final ConfigOption<String> CLUSTER_NODES = ConfigOptions
    .key("cluster.nodes")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<String> SENTINEL_NODES = ConfigOptions
    .key("sentinel.nodes")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<String> SENTINEL_MASTER = ConfigOptions
    .key("sentinel.master")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<String> PASSWORD = ConfigOptions
    .key("password")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<String> COMMAND = ConfigOptions
    .key("command")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<Integer> DB_NUM = ConfigOptions
    .key("db-num")
    .intType()
    .defaultValue(Protocol.DEFAULT_DATABASE);
  public static final ConfigOption<Integer> TTL_SEC = ConfigOptions
    .key("ttl-sec")
    .intType()
    .noDefaultValue();
  public static final ConfigOption<Integer> CONNECTION_TIMEOUT_MS = ConfigOptions
    .key("connection.timeout-ms")
    .intType()
    .defaultValue(Protocol.DEFAULT_TIMEOUT);
  public static final ConfigOption<Integer> CONNECTION_MAX_TOTAL = ConfigOptions
    .key("connection.max-total")
    .intType()
    .defaultValue(GenericObjectPoolConfig.DEFAULT_MAX_TOTAL);
  public static final ConfigOption<Integer> CONNECTION_MAX_IDLE = ConfigOptions
    .key("connection.max-idle")
    .intType()
    .defaultValue(GenericObjectPoolConfig.DEFAULT_MAX_IDLE);
  public static final ConfigOption<Boolean> CONNECTION_TEST_ON_BORROW = ConfigOptions
    .key("connection.test-on-borrow")
    .booleanType()
    .defaultValue(GenericObjectPoolConfig.DEFAULT_TEST_ON_BORROW);
  public static final ConfigOption<Boolean> CONNECTION_TEST_ON_RETURN = ConfigOptions
    .key("connection.test-on-return")
    .booleanType()
    .defaultValue(GenericObjectPoolConfig.DEFAULT_TEST_ON_RETURN);
  public static final ConfigOption<Boolean> CONNECTION_TEST_WHILE_IDLE = ConfigOptions
    .key("connection.test-while-idle")
    .booleanType()
    .defaultValue(GenericObjectPoolConfig.DEFAULT_TEST_WHILE_IDLE);
  public static final ConfigOption<String> LOOKUP_ADDITIONAL_KEY = ConfigOptions
    .key("lookup.additional-key")
    .stringType()
    .noDefaultValue();
  public static final ConfigOption<Integer> LOOKUP_CACHE_MAX_ROWS = ConfigOptions
    .key("lookup.cache.max-rows")
    .intType()
    .defaultValue(-1);
  public static final ConfigOption<Integer> LOOKUP_CACHE_TTL_SEC = ConfigOptions
    .key("lookup.cache.ttl-sec")
    .intType()
    .defaultValue(-1);

接下来分别覆写requiredOptions()和optionalOptions()方法，它们分别返回Connector的必需参数集合和可选参数集合。

@Override
  public Set<ConfigOption<?>> requiredOptions() {
    Set<ConfigOption<?>> requiredOptions = new HashSet<>();
    requiredOptions.add(MODE);
    requiredOptions.add(COMMAND);
    return requiredOptions;
  }
 
  @Override
  public Set<ConfigOption<?>> optionalOptions() {
    Set<ConfigOption<?>> optionalOptions = new HashSet<>();
    optionalOptions.add(SINGLE_HOST);
    optionalOptions.add(SINGLE_PORT);
    // 其他14个参数略去......
    optionalOptions.add(LOOKUP_CACHE_TTL_SEC);
    return optionalOptions;
  }

然后分别覆写createDynamicTableSource()与createDynamicTableSink()方法，创建DynamicTableSource和DynamicTableSink实例。在创建之前，我们可以利用内置的TableFactoryHelper工具类来校验传入的参数，当然也可以自己编写校验逻辑。另外，通过关联的上下文对象还能获取到表的元数据。代码如下，稍后会编写具体的Source/Sink类。

@Override
  public DynamicTableSource createDynamicTableSource(Context context) {
    FactoryUtil.TableFactoryHelper helper = createTableFactoryHelper(this, context);
    helper.validate();
 
    ReadableConfig options = helper.getOptions();
    validateOptions(options);
 
    TableSchema schema = context.getCatalogTable().getSchema();
    return new RedisDynamicTableSource(options, schema);
  }
 
  @Override
  public DynamicTableSink createDynamicTableSink(Context context) {
    FactoryUtil.TableFactoryHelper helper = createTableFactoryHelper(this, context);
    helper.validate();
 
    ReadableConfig options = helper.getOptions();
    validateOptions(options);
 
    TableSchema schema = context.getCatalogTable().getSchema();
    return new RedisDynamicTableSink(options, schema);
  }
 
  private void validateOptions(ReadableConfig options) {
    switch (options.get(MODE)) {
      case "single":
        if (StringUtils.isEmpty(options.get(SINGLE_HOST))) {
          throw new IllegalArgumentException("Parameter single.host must be provided in single mode");
        }
        break;
      case "cluster":
        if (StringUtils.isEmpty(options.get(CLUSTER_NODES))) {
          throw new IllegalArgumentException("Parameter cluster.nodes must be provided in cluster mode");
        }
        break;
      case "sentinel":
        if (StringUtils.isEmpty(options.get(SENTINEL_NODES)) || StringUtils.isEmpty(options.get(SENTINEL_MASTER))) {
          throw new IllegalArgumentException("Parameters sentinel.nodes and sentinel.master must be provided in sentinel mode");
        }
        break;
      default:
        throw new IllegalArgumentException("Invalid Redis mode. Must be single/cluster/sentinel");
    }
  }

在factoryIdentifier()方法内指定工厂类的标识符，该标识符就是建表时必须填写的connector参数的值。

@Override
  public String factoryIdentifier() {
    return "redis";
  }

笔者在之前的文章中介绍过，Flink SQL采用Java SPI机制来发现与加载表工厂类。所以最后不要忘了classpath的META-INF/services目录下创建一个名为org.apache.flink.table.factories.Factory的文件，并写入我们自定义的工厂类的全限定名，如：org.apache.flink.streaming.connectors.redis.dynamic.RedisDynamicTableFactory。

Implementing RedisDynamicTableSink

Bahir Flink项目已经提供了基于DataStream API的RedisSink，我们可以利用它来直接构建RedisDynamicTableSink，减少重复工作。实现了DynamicTableSink接口的类骨架如下。

public class RedisDynamicTableSink implements DynamicTableSink {
  private final ReadableConfig options;
  private final TableSchema schema;
 
  public RedisDynamicTableSink(ReadableConfig options, TableSchema schema) {
    this.options = options;
    this.schema = schema;
  }
 
  @Override
  public ChangelogMode getChangelogMode(ChangelogMode changelogMode) { }
 
  @Override
  public SinkRuntimeProvider getSinkRuntimeProvider(Context context) { }
 
  @Override
  public DynamicTableSink copy() { }
 
  @Override
  public String asSummaryString() { }
}

getChangelogMode()方法需要返回该Sink可以接受的change log行的类别。由于向Redis写入的数据可以是只追加的，也可以是带有回撤语义的(如各种聚合数据)，因此支持INSERT、UPDATE_BEFORE和UPDATE_AFTER类别。

@Override
  public ChangelogMode getChangelogMode(ChangelogMode changelogMode) {
    return ChangelogMode.newBuilder()
      .addContainedKind(RowKind.INSERT)
      .addContainedKind(RowKind.UPDATE_BEFORE)
      .addContainedKind(RowKind.UPDATE_AFTER)
      .build();
  }

接下来需要实现SinkRuntimeProvider，即编写SinkFunction供底层运行时调用。由于RedisSink已经是现成的SinkFunction了，我们只需要写好通用的RedisMapper，顺便做一些前置的校验工作(如检查表的列数以及数据类型)即可。getSinkRuntimeProvider()方法与RedisMapper的代码如下，很容易理解。

@Override
  public SinkRuntimeProvider getSinkRuntimeProvider(Context context) {
    Preconditions.checkNotNull(options, "No options supplied");
 
    FlinkJedisConfigBase jedisConfig = Util.getFlinkJedisConfig(options);
    Preconditions.checkNotNull(jedisConfig, "No Jedis config supplied");
 
    RedisCommand command = RedisCommand.valueOf(options.get(COMMAND).toUpperCase());
 
    int fieldCount = schema.getFieldCount();
    if (fieldCount != (needAdditionalKey(command) ? 3 : 2)) {
      throw new ValidationException("Redis sink only supports 2 or 3 columns");
    }
 
    DataType[] dataTypes = schema.getFieldDataTypes();
    for (int i = 0; i < fieldCount; i++) {
      if (!dataTypes[i].getLogicalType().getTypeRoot().equals(LogicalTypeRoot.VARCHAR)) {
        throw new ValidationException("Redis connector only supports STRING type");
      }
    }
 
    RedisMapper<RowData> mapper = new RedisRowDataMapper(options, command);
    RedisSink<RowData> redisSink = new RedisSink<>(jedisConfig, mapper);
    return SinkFunctionProvider.of(redisSink);
  }
 
  private static boolean needAdditionalKey(RedisCommand command) {
    return command.getRedisDataType() == RedisDataType.HASH || command.getRedisDataType() == RedisDataType.SORTED_SET;
  }
 
  public static final class RedisRowDataMapper implements RedisMapper<RowData> {
    private static final long serialVersionUID = 1L;
 
    private final ReadableConfig options;
    private final RedisCommand command;
 
    public RedisRowDataMapper(ReadableConfig options, RedisCommand command) {
      this.options = options;
      this.command = command;
    }
 
    @Override
    public RedisCommandDescription getCommandDescription() {
      return new RedisCommandDescription(command, "default-additional-key");
    }
 
    @Override
    public String getKeyFromData(RowData data) {
      return data.getString(needAdditionalKey(command) ? 1 : 0).toString();
    }
 
    @Override
    public String getValueFromData(RowData data) {
      return data.getString(needAdditionalKey(command) ? 2 : 1).toString();
    }
 
    @Override
    public Optional<String> getAdditionalKey(RowData data) {
      return needAdditionalKey(command) ? Optional.of(data.getString(0).toString()) : Optional.empty();
    }
 
    @Override
    public Optional<Integer> getAdditionalTTL(RowData data) {
      return options.getOptional(TTL_SEC);
    }
  }

剩下的copy()和asSummaryString()方法就很简单了。

@Override
  public DynamicTableSink copy() {
    return new RedisDynamicTableSink(options, schema);
  }
 
  @Override
  public String asSummaryString() {
    return "Redis Dynamic Table Sink";
  }

Implementing RedisDynamicTableSource

与DynamicTableSink不同，DynamicTableSource又根据其特性分为两类，即ScanTableSource和LookupTableSource。顾名思义，前者能够扫描外部系统中的所有或部分数据，并且支持谓词下推、分区下推之类的特性；而后者不会感知到外部系统中数据的全貌，而是根据一个或者多个key去执行点查询并返回结果。

考虑到在数仓体系中Redis一般作为维度库使用，因此我们需要实现的是LookupTableSource接口。实现该接口的RedisDynamicTableSource类如下所示，大体结构与Sink类似。

public class RedisDynamicTableSource implements LookupTableSource {
  private final ReadableConfig options;
  private final TableSchema schema;
 
  public RedisDynamicTableSource(ReadableConfig options, TableSchema schema) {
    this.options = options;
    this.schema = schema;
  }
 
  @Override
  public LookupRuntimeProvider getLookupRuntimeProvider(LookupContext context) {
    Preconditions.checkArgument(context.getKeys().length == 1 && context.getKeys()[0].length == 1, "Redis source only supports lookup by single key");
    
    int fieldCount = schema.getFieldCount();
    if (fieldCount != 2) {
      throw new ValidationException("Redis source only supports 2 columns");
    }
 
    DataType[] dataTypes = schema.getFieldDataTypes();
    for (int i = 0; i < fieldCount; i++) {
      if (!dataTypes[i].getLogicalType().getTypeRoot().equals(LogicalTypeRoot.VARCHAR)) {
        throw new ValidationException("Redis connector only supports STRING type");
      }
    }
 
    return TableFunctionProvider.of(new RedisRowDataLookupFunction(options));
  }
 
  @Override
  public DynamicTableSource copy() {
    return new RedisDynamicTableSource(options, schema);
  }
 
  @Override
  public String asSummaryString() {
    return "Redis Dynamic Table Source";
  }
}

根据Flink框架本身的要求，用于执行点查询的LookupRuntimeProvider必须是TableFunction(同步)或者AsyncTableFunction(异步)。由于Bahir Flink项目采用的Jedis是同步客户端，故本文只给出同步版本的实现，异步版本可以换用其他客户端(如Redisson或Vert.x Redis Client)。RedisRowDataLookupFunction的代码如下。

public static class RedisRowDataLookupFunction extends TableFunction<RowData> {
    private static final long serialVersionUID = 1L;
 
    private final ReadableConfig options;
    private final String command;
    private final String additionalKey;
    private final int cacheMaxRows;
    private final int cacheTtlSec;
 
    private RedisCommandsContainer commandsContainer;
    private transient Cache<RowData, RowData> cache;
 
    public RedisRowDataLookupFunction(ReadableConfig options) {
      Preconditions.checkNotNull(options, "No options supplied");
      this.options = options;
 
      command = options.get(COMMAND).toUpperCase();
      Preconditions.checkArgument(command.equals("GET") || command.equals("HGET"), "Redis table source only supports GET and HGET commands");
 
      additionalKey = options.get(LOOKUP_ADDITIONAL_KEY);
      cacheMaxRows = options.get(LOOKUP_CACHE_MAX_ROWS);
      cacheTtlSec = options.get(LOOKUP_CACHE_TTL_SEC);
    }
 
    @Override
    public void open(FunctionContext context) throws Exception {
      super.open(context);
 
      FlinkJedisConfigBase jedisConfig = Util.getFlinkJedisConfig(options);
      commandsContainer = RedisCommandsContainerBuilder.build(jedisConfig);
      commandsContainer.open();
 
      if (cacheMaxRows > 0 && cacheTtlSec > 0) {
        cache = CacheBuilder.newBuilder()
          .expireAfterWrite(cacheTtlSec, TimeUnit.SECONDS)
          .maximumSize(cacheMaxRows)
          .build();
      }
    }
 
    @Override
    public void close() throws Exception {
      if (cache != null) {
        cache.invalidateAll();
      }
      if (commandsContainer != null) {
        commandsContainer.close();
      }
      super.close();
    }
 
    public void eval(Object obj) {
      RowData lookupKey = GenericRowData.of(obj);
      if (cache != null) {
        RowData cachedRow = cache.getIfPresent(lookupKey);
        if (cachedRow != null) {
          collect(cachedRow);
          return;
        }
      }
 
      StringData key = lookupKey.getString(0);
      String value = command.equals("GET") ? commandsContainer.get(key.toString()) : commandsContainer.hget(additionalKey, key.toString());
      RowData result = GenericRowData.of(key, StringData.fromString(value));
 
      cache.put(lookupKey, result);
      collect(result);
    }
  }

有三点需要注意：

• Redis维度数据一般用String或Hash类型存储，因此命令支持GET与HGET。如果使用Hash类型，需要在参数中额外传入它的key，不能像Sink一样动态指定；
• 为了避免每来一条数据都请求Redis，需要设计缓存，上面利用的是Guava Cache。在Redis中查不到的数据也要缓存，防止穿透；
• TableFunction必须有一个签名为eval(Object)或eval(Object...)的方法。在本例中实际输出的数据类型为ROW<STRING, STRING>，在Flink Table的类型体系中应表示为RowData(StringData, StringData)。

Using Redis SQL Connector

来实际应用一下吧。先创建一张表示Hash结构的Redis Sink表。

CREATE TABLE rtdw_dws.redis_test_order_stat_dashboard (
  hashKey STRING,
  cityId STRING,
  data STRING,
  PRIMARY KEY (hashKey) NOT ENFORCED
) WITH (
  'connector' = 'redis',
  'mode' = 'single',
  'single.host' = '172.16.200.124',
  'single.port' = '6379',
  'db-num' = '10',
  'command' = 'HSET',
  'ttl-sec' = '86400',
  'connection.max-total' = '5',
  'connection.timeout-ms' = '5000',
  'connection.test-while-idle' = 'true'
)

然后读取Kafka中的订单流，统计一些简单的数据，并写入Redis。

/*
tableEnvConfig.setBoolean("table.dynamic-table-options.enabled", true)
tableEnvConfig.setBoolean("table.exec.emit.early-fire.enabled", true)
tableEnvConfig.setString("table.exec.emit.early-fire.delay", "5s")
tableEnv.createTemporarySystemFunction("MapToJsonString", classOf[MapToJsonString])
*/
INSERT INTO rtdw_dws.redis_test_order_stat_dashboard
SELECT
  CONCAT('dashboard:city_stat:', p.orderDay) AS hashKey,
  CAST(p.cityId AS STRING) AS cityId,
  MapToJsonString(MAP[
    'subOrderNum', CAST(p.subOrderNum AS STRING),
    'buyerNum', CAST(p.buyerNum AS STRING),
    'gmv', CAST(p.gmv AS STRING)
  ]) AS data
FROM (
  SELECT
    cityId,
    SUBSTR(tss, 0, 10) AS orderDay,
    COUNT(1) AS subOrderNum,
    COUNT(DISTINCT userId) AS buyerNum,
    SUM(quantity * merchandisePrice) AS gmv
  FROM rtdw_dwd.kafka_order_done_log /*+ OPTIONS('scan.startup.mode'='latest-offset','properties.group.id'='fsql_redis_test_order_stat_dashboard') */
  GROUP BY TUMBLE(procTime, INTERVAL '1' DAY), cityId, SUBSTR(tss, 0, 10)
) p

观察结果~

再看一下Redis作为维度表的使用，仍然以Hash结构为例。

CREATE TABLE rtdw_dim.redis_test_city_info (
  cityId STRING,
  cityName STRING
) WITH (
  'connector' = 'redis',
  'mode' = 'single',
  'single.host' = '172.16.200.124',
  'single.port' = '6379',
  'db-num' = '9',
  'command' = 'HGET',
  'connection.timeout-ms' = '5000',
  'connection.test-while-idle' = 'true',
  'lookup.additional-key' = 'rtdw_dim:test_city_info',
  'lookup.cache.max-rows' = '1000',
  'lookup.cache.ttl-sec' = '600'
)

为了方便观察结果，创建一张Print Sink表输出数据，然后将Kafka流表与Redis维表做Temporal Join，SQL语句如下。

CREATE TABLE test.print_redis_test_dim_join (
  tss STRING,
  cityId BIGINT,
  cityName STRING
) WITH (
  'connector' = 'print'
)
 
INSERT INTO test.print_redis_test_dim_join
SELECT a.tss, a.cityId, b.cityName
FROM rtdw_dwd.kafka_order_done_log /*+ OPTIONS('scan.startup.mode'='latest-offset','properties.group.id'='fsql_redis_source_test') */ AS a
LEFT JOIN rtdw_dim.redis_test_city_info FOR SYSTEM_TIME AS OF a.procTime AS b ON CAST(a.cityId AS STRING) = b.cityId
WHERE a.orderType = 12

查看输出~

4> +I(2021-03-04 20:44:48,10264,漳州市)
3> +I(2021-03-04 20:45:26,10030,常德市)
4> +I(2021-03-04 20:45:23,10332,桂林市)
7> +I(2021-03-04 20:45:26,10031,九江市)
9> +I(2021-03-04 20:45:23,10387,惠州市)
4> +I(2021-03-04 20:45:19,10607,芜湖市)
3> +I(2021-03-04 20:45:25,10364,无锡市)

The End

通过上面的示例，相信看官已经能够根据自己的需求灵活地定制Flink SQL Connector了。本文未详述的ScanTableSource、异步LookupTableSource和Encoding/Decoding Format也会在之后的文章中择机讲解。

最近春寒料峭，民那注意增减衣物。

晚安晚安。