基于Kafka+Flink+Redis的电商大屏实时计算案例

前言

一年一度的双11又要到了,阿里的双11销量大屏可以说是一道特殊的风景线。实时大屏(real-time dashboard)正在被越来越多的企业采用,用来及时呈现关键的数据指标。并且在实际操作中,肯定也不会仅仅计算一两个维度。由于Flink的“真·流式计算”这一特点,它比Spark Streaming要更适合大屏应用。本文从笔者的实际工作经验抽象出简单的模型,并简要叙述计算流程(当然大部分都是源码)。

数据格式与接入

简化的子订单消息体如下。

{
    "userId": 234567,
    "orderId": 2902306918400,
    "subOrderId": 2902306918401,
    "siteId": 10219,
    "siteName": "site_blabla",
    "cityId": 101,
    "cityName": "北京市",
    "warehouseId": 636,
    "merchandiseId": 187699,
    "price": 299,
    "quantity": 2,
    "orderStatus": 1,
    "isNewOrder": 0,
    "timestamp": 1572963672217
}

由于订单可能会包含多种商品,故会被拆分成子订单来表示,每条JSON消息表示一个子订单。现在要按照自然日来统计以下指标,并以1秒的刷新频率呈现在大屏上:

  • 每个站点(站点ID即siteId)的总订单数、子订单数、销量与GMV;
  • 当前销量排名前N的商品(商品ID即merchandiseId)与它们的销量。

由于大屏的最大诉求是实时性,等待迟到数据显然不太现实,因此我们采用处理时间作为时间特征,并以1分钟的频率做checkpointing。

    StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
    env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime);
    env.enableCheckpointing(60 * 1000, CheckpointingMode.EXACTLY_ONCE);
    env.getCheckpointConfig().setCheckpointTimeout(30 * 1000);

然后订阅Kafka的订单消息作为数据源。

    Properties consumerProps = ParameterUtil.getFromResourceFile("kafka.properties");
    DataStream<String> sourceStream = env
      .addSource(new FlinkKafkaConsumer011<>(
        ORDER_EXT_TOPIC_NAME,                        // topic
        new SimpleStringSchema(),                    // deserializer
        consumerProps                                // consumer properties
      ))
      .setParallelism(PARTITION_COUNT)
      .name("source_kafka_" + ORDER_EXT_TOPIC_NAME)
      .uid("source_kafka_" + ORDER_EXT_TOPIC_NAME);

给带状态的算子设定算子ID(通过调用uid()方法)是个好习惯,能够保证Flink应用从保存点重启时能够正确恢复状态现场。为了尽量稳妥,Flink官方也建议为每个算子都显式地设定ID,参考官方文档

接下来将JSON数据转化为POJO,JSON框架采用FastJSON。

    DataStream<SubOrderDetail> orderStream = sourceStream
      .map(message -> JSON.parseObject(message, SubOrderDetail.class))
      .name("map_sub_order_detail").uid("map_sub_order_detail");

JSON已经是预先处理好的标准化格式,所以POJO类SubOrderDetail的写法可以通过Lombok极大地简化。如果JSON的字段有不规范的,那么就需要手写Getter和Setter,并用@JSONField注解来指明。

@Getter
@Setter
@NoArgsConstructor
@AllArgsConstructor
@ToString
public class SubOrderDetail implements Serializable {
  private static final long serialVersionUID = 1L;
  
  private long userId;
  private long orderId;
  private long subOrderId;
  private long siteId;
  private String siteName;
  private long cityId;
  private String cityName;
  private long warehouseId;
  private long merchandiseId;
  private long price;
  private long quantity;
  private int orderStatus;
  private int isNewOrder;
  private long timestamp;
}

统计站点指标

将子订单流按站点ID分组,开1天的滚动窗口,并同时设定ContinuousProcessingTimeTrigger触发器,以1秒周期触发计算。注意处理时间的时区问题,这是老生常谈了。

    WindowedStream<SubOrderDetail, Tuple, TimeWindow> siteDayWindowStream = orderStream
      .keyBy("siteId")
      .window(TumblingProcessingTimeWindows.of(Time.days(1), Time.hours(-8)))
      .trigger(ContinuousProcessingTimeTrigger.of(Time.seconds(1)));

接下来写个聚合函数。

    DataStream<OrderAccumulator> siteAggStream = siteDayWindowStream
      .aggregate(new OrderAndGmvAggregateFunc())
      .name("aggregate_site_order_gmv").uid("aggregate_site_order_gmv");
  public static final class OrderAndGmvAggregateFunc
    implements AggregateFunction<SubOrderDetail, OrderAccumulator, OrderAccumulator> {
    private static final long serialVersionUID = 1L;

    @Override
    public OrderAccumulator createAccumulator() {
      return new OrderAccumulator();
    }

    @Override
    public OrderAccumulator add(SubOrderDetail record, OrderAccumulator acc) {
      if (acc.getSiteId() == 0) {
        acc.setSiteId(record.getSiteId());
        acc.setSiteName(record.getSiteName());
      }
      acc.addOrderId(record.getOrderId());
      acc.addSubOrderSum(1);
      acc.addQuantitySum(record.getQuantity());
      acc.addGmv(record.getPrice() * record.getQuantity());
      return acc;
    }

    @Override
    public OrderAccumulator getResult(OrderAccumulator acc) {
      return acc;
    }

    @Override
    public OrderAccumulator merge(OrderAccumulator acc1, OrderAccumulator acc2) {
      if (acc1.getSiteId() == 0) {
        acc1.setSiteId(acc2.getSiteId());
        acc1.setSiteName(acc2.getSiteName());
      }
      acc1.addOrderIds(acc2.getOrderIds());
      acc1.addSubOrderSum(acc2.getSubOrderSum());
      acc1.addQuantitySum(acc2.getQuantitySum());
      acc1.addGmv(acc2.getGmv());
      return acc1;
    }
  }

累加器类OrderAccumulator的实现很简单,看源码就大概知道它的结构了,因此不再多废话。唯一需要注意的是订单ID可能重复,所以需要用名为orderIds的HashSet来保存它。HashSet应付我们目前的数据规模还是没太大问题的,如果是海量数据,就考虑换用HyperLogLog吧。

接下来就该输出到Redis供呈现端查询了。这里有个问题:一秒内有数据变化的站点并不多,而ContinuousProcessingTimeTrigger每次触发都会输出窗口里全部的聚合数据,这样做了很多无用功,并且还会增大Redis的压力。所以,我们在聚合结果后再接一个ProcessFunction,代码如下。

    DataStream<Tuple2<Long, String>> siteResultStream = siteAggStream
      .keyBy(0)
      .process(new OutputOrderGmvProcessFunc(), TypeInformation.of(new TypeHint<Tuple2<Long, String>>() {}))
      .name("process_site_gmv_changed").uid("process_site_gmv_changed");
  public static final class OutputOrderGmvProcessFunc
    extends KeyedProcessFunction<Tuple, OrderAccumulator, Tuple2<Long, String>> {
    private static final long serialVersionUID = 1L;

    private MapState<Long, OrderAccumulator> state;

    @Override
    public void open(Configuration parameters) throws Exception {
      super.open(parameters);
      state = this.getRuntimeContext().getMapState(new MapStateDescriptor<>(
        "state_site_order_gmv",
        Long.class,
        OrderAccumulator.class)
      );
    }

    @Override
    public void processElement(OrderAccumulator value, Context ctx, Collector<Tuple2<Long, String>> out) throws Exception {
      long key = value.getSiteId();
      OrderAccumulator cachedValue = state.get(key);

      if (cachedValue == null || value.getSubOrderSum() != cachedValue.getSubOrderSum()) {
        JSONObject result = new JSONObject();
        result.put("site_id", value.getSiteId());
        result.put("site_name", value.getSiteName());
        result.put("quantity", value.getQuantitySum());
        result.put("orderCount", value.getOrderIds().size());
        result.put("subOrderCount", value.getSubOrderSum());
        result.put("gmv", value.getGmv());
        out.collect(new Tuple2<>(key, result.toJSONString());
        state.put(key, value);
      }
    }

    @Override
    public void close() throws Exception {
      state.clear();
      super.close();
    }
  }

说来也简单,就是用一个MapState状态缓存当前所有站点的聚合数据。由于数据源是以子订单为单位的,因此如果站点ID在MapState中没有缓存,或者缓存的子订单数与当前子订单数不一致,表示结果有更新,这样的数据才允许输出。

最后就可以安心地接上Redis Sink了,结果会被存进一个Hash结构里。

    // 看官请自己构造合适的FlinkJedisPoolConfig
    FlinkJedisPoolConfig jedisPoolConfig = ParameterUtil.getFlinkJedisPoolConfig(false, true);
    siteResultStream
      .addSink(new RedisSink<>(jedisPoolConfig, new GmvRedisMapper()))
      .name("sink_redis_site_gmv").uid("sink_redis_site_gmv")
      .setParallelism(1);
  public static final class GmvRedisMapper implements RedisMapper<Tuple2<Long, String>> {
    private static final long serialVersionUID = 1L;
    private static final String HASH_NAME_PREFIX = "RT:DASHBOARD:GMV:";

    @Override
    public RedisCommandDescription getCommandDescription() {
      return new RedisCommandDescription(RedisCommand.HSET, HASH_NAME_PREFIX);
    }

    @Override
    public String getKeyFromData(Tuple2<Long, String> data) {
      return String.valueOf(data.f0);
    }

    @Override
    public String getValueFromData(Tuple2<Long, String> data) {
      return data.f1;
    }

    @Override
    public Optional<String> getAdditionalKey(Tuple2<Long, String> data) {
      return Optional.of(
        HASH_NAME_PREFIX +
        new LocalDateTime(System.currentTimeMillis()).toString(Consts.TIME_DAY_FORMAT) +
        "SITES"
      );
    }
  }

商品Top N

我们可以直接复用前面产生的orderStream,玩法与上面的GMV统计大同小异。这里用1秒滚动窗口就可以了。

    WindowedStream<SubOrderDetail, Tuple, TimeWindow> merchandiseWindowStream = orderStream
      .keyBy("merchandiseId")
      .window(TumblingProcessingTimeWindows.of(Time.seconds(1)));

    DataStream<Tuple2<Long, Long>> merchandiseRankStream = merchandiseWindowStream
      .aggregate(new MerchandiseSalesAggregateFunc(), new MerchandiseSalesWindowFunc())
      .name("aggregate_merch_sales").uid("aggregate_merch_sales")
      .returns(TypeInformation.of(new TypeHint<Tuple2<Long, Long>>() { }));

聚合函数与窗口函数的实现更加简单了,最终返回的是商品ID与商品销量的二元组。

  public static final class MerchandiseSalesAggregateFunc
    implements AggregateFunction<SubOrderDetail, Long, Long> {
    private static final long serialVersionUID = 1L;

    @Override
    public Long createAccumulator() {
      return 0L;
    }

    @Override
    public Long add(SubOrderDetail value, Long acc) {
      return acc + value.getQuantity();
    }

    @Override
    public Long getResult(Long acc) {
      return acc;
    }

    @Override
    public Long merge(Long acc1, Long acc2) {
      return acc1 + acc2;
    }
  }


  public static final class MerchandiseSalesWindowFunc
    implements WindowFunction<Long, Tuple2<Long, Long>, Tuple, TimeWindow> {
    private static final long serialVersionUID = 1L;

    @Override
    public void apply(
      Tuple key,
      TimeWindow window,
      Iterable<Long> accs,
      Collector<Tuple2<Long, Long>> out) throws Exception {
      long merchId = ((Tuple1<Long>) key).f0;
      long acc = accs.iterator().next();
      out.collect(new Tuple2<>(merchId, acc));
    }
  }

既然数据最终都要落到Redis,那么我们完全没必要在Flink端做Top N的统计,直接利用Redis的有序集合(zset)就行了,商品ID作为field,销量作为分数值,简单方便。不过flink-redis-connector项目中默认没有提供ZINCRBY命令的实现(必须再吐槽一次),我们可以自己加,步骤参照之前写过的那篇加SETEX的命令的文章,不再赘述。RedisMapper的写法如下。

  public static final class RankingRedisMapper implements RedisMapper<Tuple2<Long, Long>> {
    private static final long serialVersionUID = 1L;
    private static final String ZSET_NAME_PREFIX = "RT:DASHBOARD:RANKING:";

    @Override
    public RedisCommandDescription getCommandDescription() {
      return new RedisCommandDescription(RedisCommand.ZINCRBY, ZSET_NAME_PREFIX);
    }

    @Override
    public String getKeyFromData(Tuple2<Long, Long> data) {
      return String.valueOf(data.f0);
    }

    @Override
    public String getValueFromData(Tuple2<Long, Long> data) {
      return String.valueOf(data.f1);
    }

    @Override
    public Optional<String> getAdditionalKey(Tuple2<Long, Long> data) {
      return Optional.of(
        ZSET_NAME_PREFIX +
        new LocalDateTime(System.currentTimeMillis()).toString(Consts.TIME_DAY_FORMAT) + ":" +
        "MERCHANDISE"
      );
    }
  }

后端取数时,用ZREVRANGE命令即可取出指定排名的数据了。只要数据规模不是大到难以接受,并且有现成的Redis,这个方案完全可以作为各类Top N需求的通用实现。

The End

大屏的实际呈现需要保密,截图自然是没有的。以下是提交执行时Flink Web UI给出的执行计划(实际有更多的统计任务,不止3个Sink)。通过复用源数据,可以在同一个Flink job内实现更多统计需求。

晚安晚安。