在 Spark Executors 上向 Kafka 提交偏移量 [英] Commit Offsets to Kafka on Spark Executors
问题描述
我从 Kafka 获取事件，在 Spark 上丰富/过滤/转换它们，然后将它们存储在 ES 中.我正在向 Kafka 提交偏移量
I am getting events from Kafka, enriching/filtering/transforming them on Spark and then storing them in ES. I am committing back the offsets to Kafka
我有两个问题:
(1) 我目前的 Spark 工作非常慢
我有 50 个分区用于一个主题和 20 个执行程序.每个 executor 有 2 个内核和 4g 内存.我的驱动程序有 8g 内存.我消耗了 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
I have 50 partitions for a topic and 20 executors. Each executor has 2 cores and 4g of memory each. My driver has 8g of memory. I am consuming 1000 events/partition/second and my batch interval is 10 seconds. This means, I am consuming 500000 events in 10 seconds
我的ES集群如下:
20 个分片/索引
3 个主实例 c5.xlarge.elasticsearch
3 master instances c5.xlarge.elasticsearch
12 个实例 m4.xlarge.elasticsearch
12 instances m4.xlarge.elasticsearch
磁盘/节点 = 1024 GB 所以总共 12 TB
disk / node = 1024 GB so 12 TB in total
而且我遇到了巨大的调度和处理延迟
And I am getting huge scheduling and processing delays
(2) 如何在执行器上提交偏移量?
目前，我在执行器上丰富/转换/过滤我的事件，然后使用 BulkRequest 将所有内容发送到 ES.这是一个同步过程.如果我得到正面反馈，我会将偏移量列表发送给驱动程序.如果没有，我会发回一个空列表.在驱动程序上，我向 Kafka 提交了偏移量.我相信，应该有一种方法，我可以在执行程序上提交偏移量，但我不知道如何将 kafka Stream 传递给执行程序:
Currently, I enrich/transform/filter my events on executors and then send everything to ES using BulkRequest. It's a synchronous process. If I get positive feedback, I send the offset list to driver. If not, I send back an empty list. On the driver, I commit offsets to Kafka. I believe, there should be a way, where I can commit offsets on executors but I don't know how to pass kafka Stream to executors:
((CanCommitOffsets) kafkaStream.inputDStream()).commitAsync(offsetRanges, this::onComplete);
这是向需要Kafka Stream的Kafka提交偏移量的代码
This is the code for committing offsets to Kafka which requires Kafka Stream
这是我的整体代码:
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();
});
推荐答案
您可以将偏移逻辑移到 rdd 循环上方...我正在使用以下模板以获得更好的偏移处理和性能
You can move the offset logic above the rdd loop ... I am using below template for better offset handling and performance
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);
});
这篇关于在 Spark Executors 上向 Kafka 提交偏移量的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
