Question

在flink中实现过程功能时面对竞争条件连接的流。我有两个正在共享的Cache Map 被并行调用的函数processElement1和processElement2 通过2个不同的线程。

Streams1 --->（发送报价数据）

Streams2 --->（发送lms（忠诚度管理系统数据））

connect=Streams1.connect(Streams2);

connect.process(new TriggerStream);

在TriggerStream Class中，我使用唯一的ID：MemberId作为unique Key存储数据，以将＆lookup data存储在缓存中。当数据流入时，我没有得到一致的结果

class LRUConcurrentCache<K,V>{
    private final Map<K,V> cache;
    private final int maxEntries;
    public LRUConcurrentCache(final int maxEntries) {
        this.cache = new LinkedHashMap<K,V>(maxEntries, 0.75F, true) {
            private static final long serialVersionUID = -1236481390177598762L;
            @Override
            protected boolean removeEldestEntry(Map.Entry<K,V> eldest){
                return size() > maxEntries;
            }
        };
    }
    //Why we cant lock on the key
    public void put(K key, V value) {
        synchronized(key) {
            cache.put(key, value);
        }
    }

    //get methode
    public V get(K key) {
        synchronized(key) {
            return cache.get(key);
        }
    }



public class TriggerStream extends CoProcessFunction<IOffer, LMSData, String> {


    private static final long serialVersionUID = 1L;
    LRUCache cache; 
    private String offerNode;
    String updatedValue, retrivedValue;
    Subscriber subscriber;

    TriggerStream(){
        this.cache== new LRUCache(10);
    }



@Override
    public void processElement1(IOffer offer) throws Exception {
        try {
            ObjectMapper mapper = new ObjectMapper();
            mapper.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
            mapper.enableDefaultTyping();
            // TODO Auto-generated method stub
            IOffer latestOffer = offer;

            //Check the subscriber is there or not

            retrivedValue = cache.get(latestOffer.getMemberId().toString());
            if ((retrivedValue == null)) {
                //Subscriber is the class that is used and converted as Json String & then store into map
                Subscriber subscriber = new Subscriber();
                subscriber.setMemberId(latestOffer.getMemberId());
                ArrayList<IOffer> offerList = new ArrayList<IOffer>();
                offerList.add(latestOffer);
                subscriber.setOffers(offerList);
                updatedValue = mapper.writeValueAsString(subscriber);
                cache.set(subscriber.getMemberId().toString(), updatedValue);
            } else {
                Subscriber subscriber = mapper.readValue(retrivedValue, Subscriber.class);
                List<IOffer> offers = subscriber.getOffers();
                offers.add(latestOffer);
                updatedValue= mapper.writeValueAsString(subscriber);
                cache.set(subscriber.getMemberId().toString(), subscriberUpdatedValue);
            }
        } catch (Exception pb) {
            applicationlogger.error("Exception in Offer Loading:"+pb);
            applicationlogger.debug("*************************FINISHED OFFER LOADING*******************************");
        }
        applicationlogger.debug("*************************FINISHED OFFER LOADING*******************************");

    }

@Override
    public void processElement2(LMSData lms) throws Exception {
        try {
            ObjectMapper mapper = new ObjectMapper();
            mapper.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
            mapper.enableDefaultTyping();
            // TODO Auto-generated method stub

            //Check the subscriber is there or not

            retrivedValue = cache.get(lms.getMemberId().toString());
            if(retrivedValue !=null){
                Subscriber subscriber = mapper.readValue(retrivedValue, Subscriber.class);
                //do some calculations 
                String updatedValue = mapper.writeValueAsString(subscriber);
                //Update value
                cache.set(subscriber.getMemberId().toString(), updatedValue);
            }

        } catch (Exception pb) {
            applicationlogger.error("Exception in Offer Loading:"+pb);
            applicationlogger.debug("*************************FINISHED OFFER LOADING*******************************");
        }
        applicationlogger.debug("*************************FINISHED OFFER LOADING*******************************");

    }

}

Answer 1

Flink不保证CoProcessFunction（或任何其他Co * Function）以何种顺序提取数据。跨分布式并行任务维护某种确定性顺序会太昂贵。

相反，您必须在功能中使用状态和可能的计时器来解决该问题。函数中的LRUCache应该保持状态（可能是keyed state）。否则，如果发生故障，它将丢失。您可以为第一个流添加另一个状态并缓冲记录，直到第二个流的查找值到达为止。

在Flink中面对竞态条件在Apache Flink中连接流

1 个答案: