Question

我从Kafka获取事件，在Spark上丰富/过滤/转换事件，然后将其存储在ES中。我将抵消额返还给Kafka

我有两个问题/问题：

（1）我当前的Spark作业非常慢

我有一个主题的50个分区和20个执行程序。每个执行器都有2个核心和4g的内存。我的驱动程序有8克内存。我正在消耗1000个事件/分区/秒，我的批处理间隔是10秒。这意味着我在10秒内消耗了500000个事件

我的ES群集如下：

20个碎片/索引

3个主实例c5.xlarge.elasticsearch

12个实例m4.xlarge.elasticsearch

磁盘/节点= 1024 GB，所以总共12 TB

我的计划和处理延迟越来越大

（2）如何在执行器上提交偏移量？

当前，我在执行程序上丰富/转换/过滤事件，然后使用 BulkRequest 将所有内容发送到ES。这是一个同步过程。如果得到积极的反馈，我会将偏移量列表发送给驱动程序。如果没有，我发回一个空清单。在驱动程序上，我将偏移量提交给Kafka。我相信，应该有一种方法，我可以在执行器上提交偏移量，但是我不知道如何将kafka Stream传递给执行器：

((CanCommitOffsets) kafkaStream.inputDStream()).commitAsync(offsetRanges, this::onComplete);

这是向Kafka提交偏移量的代码，需要Kafka Stream

这是我的总体代码：

 kafkaStream.foreachRDD( // kafka topic
                rdd -> { // runs on driver
                    rdd.cache();
                    String batchIdentifier =
                            Long.toHexString(Double.doubleToLongBits(Math.random()));

                    LOGGER.info("@@ [" + batchIdentifier + "] Starting batch ...");

                    Instant batchStart = Instant.now();

                    List<OffsetRange> offsetsToCommit =
                            rdd.mapPartitionsWithIndex( // kafka partition
                                    (index, eventsIterator) -> { // runs on worker

                                        OffsetRange[] offsetRanges = ((HasOffsetRanges) rdd.rdd()).offsetRanges();

                                        LOGGER.info(
                                                "@@ Consuming " + offsetRanges[index].count() + " events" + " partition: " + index
                                        );

                                        if (!eventsIterator.hasNext()) {
                                            return Collections.emptyIterator();
                                        }

                                        // get single ES documents
                                        List<SingleEventBaseDocument> eventList = getSingleEventBaseDocuments(eventsIterator);

                                        // build request wrappers
                                        List<InsertRequestWrapper> requestWrapperList = getRequestsToInsert(eventList, offsetRanges[index]);

                                        LOGGER.info(
                                                "@@ Processed " + offsetRanges[index].count() + " events" + " partition: " + index + " list size: " + eventList.size()
                                        );

                                        BulkResponse bulkItemResponses = elasticSearchRepository.addElasticSearchDocumentsSync(requestWrapperList);

                                        if (!bulkItemResponses.hasFailures()) {
                                            return Arrays.asList(offsetRanges).iterator();
                                        }

                                        elasticSearchRepository.close();
                                        return Collections.emptyIterator();
                                    },
                                    true
                            ).collect();

                    LOGGER.info(
                            "@@ [" + batchIdentifier + "] Collected all offsets in " + (Instant.now().toEpochMilli() - batchStart.toEpochMilli()) + "ms"
                    );

                    OffsetRange[] offsets = new OffsetRange[offsetsToCommit.size()];

                    for (int i = 0; i < offsets.length ; i++) {
                        offsets[i] = offsetsToCommit.get(i);
                    }

                    try {
                        offsetManagementMapper.commit(offsets);
                    } catch (Exception e) {
                        // ignore
                    }

                    LOGGER.info(
                            "@@ [" + batchIdentifier + "] Finished batch of " + offsetsToCommit.size() + " messages " +
                                    "in " + (Instant.now().toEpochMilli() - batchStart.toEpochMilli()) + "ms"
                    );
                    rdd.unpersist();
                });

Answer 1

您可以将偏移逻辑移到rdd循环上方...我正在使用下面的模板以获得更好的偏移处理和性能

JavaInputDStream<ConsumerRecord<String, String>> kafkaStream = KafkaUtils.createDirectStream(jssc,
                LocationStrategies.PreferConsistent(),
                ConsumerStrategies.<String, String>Subscribe(topics, kafkaParams));



        kafkaStream.foreachRDD( kafkaStreamRDD -> {
            //fetch kafka offsets for manually commiting it later
            OffsetRange[] offsetRanges = ((HasOffsetRanges) kafkaStreamRDD.rdd()).offsetRanges();

            //filter unwanted data
            kafkaStreamRDD.filter(
                    new Function<ConsumerRecord<String, String>, Boolean>() {
                @Override
                public Boolean call(ConsumerRecord<String, String> kafkaRecord) throws Exception {
                    if(kafkaRecord!=null) {
                        if(!StringUtils.isAnyBlank(kafkaRecord.key() , kafkaRecord.value())) {
                            return Boolean.TRUE;
                        }
                    }
                    return Boolean.FALSE;
                }
            }).foreachPartition( kafkaRecords -> {

                // init connections here

                while(kafkaRecords.hasNext()) {
                    ConsumerRecord<String, String> kafkaConsumerRecord = kafkaRecords.next();
                    // work here
                }

            });
            //commit offsets
            ((CanCommitOffsets) kafkaStream.inputDStream()).commitAsync(offsetRanges);
        });

在Spark执行器上将偏移提交给Kafka

1 个答案: