Spark metrics实现KafkaSink

背景

监控是Spark非常重要的一部分。Spark的运行情况是由ListenerBus以及MetricsSystem 来完成的。通过Spark的Metrics系统，我们可以把Spark Metrics的收集到的信息发送到各种各样的Sink，比如HTTP、JMX以及CSV文件。
目前支持的Sink包括：

• ConsoleSink

• CSVSink

• JmxSink

• MetricsServlet

• GraphiteSink

• GangliaSink

有时我们需要实时获取metrics数据通过spark分析展示等需求，这个时候若有个KafkaSink将metrics指标数据实时往kafka发送那就太方便了，故有了这篇博文。

实践

所有的Sink都需要继承Sink这个特质：

private[spark] trait Sink {  def start(): Unit  def stop(): Unit  def report(): Unit
}

当该Sink注册到metrics系统中时，会调用start方法进行一些初始化操作，再通过report方式进行真正的输出操作，stop方法可以进行一些连接关闭等操作。直接上代码：

package org.apache.spark.metrics.sinkimport java.util.concurrent.TimeUnitimport java.util.{Locale, Properties}import com.codahale.metrics.MetricRegistryimport org.apache.kafka.clients.producer.KafkaProducerimport org.apache.spark.SecurityManagerimport org.apache.spark.internal.Loggingprivate[spark] class KafkaSink(val property: Properties, val registry: MetricRegistry,
                               securityMgr: SecurityManager) extends Sink with Logging{

    val KAFKA_KEY_PERIOD = "period"
    val KAFKA_DEFAULT_PERIOD = 10

    val KAFKA_KEY_UNIT = "unit"
    val KAFKA_DEFAULT_UNIT = "SECONDS"

    val KAFKA_TOPIC = "topic"
    val KAFKA_DEFAULT_TOPIC = "kafka-sink-topic"

    val KAFAK_BROKERS = "kafka-brokers"
    val KAFAK_DEFAULT_BROKERS = "XXX:9092"

    val TOPIC = Option(property.getProperty(KAFKA_TOPIC)).getOrElse(KAFKA_DEFAULT_TOPIC)
    val BROKERS = Option(property.getProperty(KAFAK_BROKERS)).getOrElse(throw new IllegalStateException("kafka-brokers is null!"))    private val kafkaProducerConfig = new Properties()
    kafkaProducerConfig.put("bootstrap.servers",BROKERS)
    kafkaProducerConfig.put("key.serializer", "org.apache.kafka.common.serialization.StringSerializer")
    kafkaProducerConfig.put("value.serializer", "org.apache.kafka.common.serialization.StringSerializer")    private val producer = new KafkaProducer[String, String](kafkaProducerConfig)    private val reporter: KafkaReporter = KafkaReporter.forRegistry(registry)
        .topic(TOPIC)
        .build(producer)


    val pollPeriod = Option(property.getProperty(KAFKA_KEY_PERIOD)) match {        case Some(s) => s.toInt        case None => KAFKA_DEFAULT_PERIOD
    }

    val pollUnit: TimeUnit = Option(property.getProperty(KAFKA_KEY_UNIT)) match {        case Some(s) => TimeUnit.valueOf(s.toUpperCase(Locale.ROOT))        case None => TimeUnit.valueOf(KAFKA_DEFAULT_UNIT)
    }

    override def start(): Unit = {        log.info("I4 Metrics System KafkaSink Start ......")
        reporter.start(pollPeriod, pollUnit)
    }

    override def stop(): Unit = {        log.info("I4 Metrics System KafkaSink Stop ......")
        reporter.stop()
        producer.close()
    }

    override def report(): Unit = {        log.info("I4 Metrics System KafkaSink Report ......")
        reporter.report()
    }
}

KafkaReporter类：

package org.apache.spark.metrics.sink;import com.alibaba.fastjson.JSONObject;import com.codahale.metrics.*;import com.twitter.bijection.Injection;import com.twitter.bijection.avro.GenericAvroCodecs;import org.apache.avro.Schema;import org.apache.avro.generic.GenericData;import org.apache.avro.generic.GenericRecord;import org.apache.kafka.clients.producer.KafkaProducer;import org.apache.kafka.clients.producer.ProducerRecord;import org.slf4j.Logger;import org.slf4j.LoggerFactory;import java.util.Map;import java.util.SortedMap;import java.util.concurrent.TimeUnit;public class KafkaReporter  extends ScheduledReporter  {    private static final Logger LOGGER = LoggerFactory.getLogger(KafkaReporter.class);    public static KafkaReporter.Builder forRegistry(MetricRegistry registry) {        return new KafkaReporter.Builder(registry);
    }    private KafkaProducer producer;    private Clock clock;    private String topic;    private KafkaReporter(MetricRegistry registry,
                        TimeUnit rateUnit,
                        TimeUnit durationUnit,
                        MetricFilter filter,
                        Clock clock,
                        String topic,
                        KafkaProducer producer) {        super(registry, "kafka-reporter", filter, rateUnit, durationUnit);        this.producer = producer;        this.topic = topic;        this.clock = clock;
    }    @Override
    public void report(SortedMap<String, Gauge> gauges, SortedMap<String, Counter> counters, SortedMap<String, Histogram> histograms, SortedMap<String, Meter> meters, SortedMap<String, Timer> timers) {        final long timestamp = TimeUnit.MILLISECONDS.toSeconds(clock.getTime());        // Gauge
        for (Map.Entry<String, Gauge> entry : gauges.entrySet()) {
            reportGauge(timestamp,entry.getKey(), entry.getValue());
        }        // Histogram//        for (Map.Entry<String, Histogram> entry : histograms.entrySet()) {//            reportHistogram(timestamp, entry.getKey(), entry.getValue());//        }
    }    private void reportGauge(long timestamp, String name, Gauge gauge) {
        report(timestamp, name, gauge.getValue());
    }    private void reportHistogram(long timestamp, String name, Histogram histogram) {        final Snapshot snapshot = histogram.getSnapshot();
        report(timestamp, name, snapshot.getMax());
    }    private void report(long timestamp, String name,  Object values) {
        JSONObject jsonObject = new JSONObject();
        jsonObject.put("name",name);
        jsonObject.put("timestamp",timestamp);
        jsonObject.put("value",values);
        producer.send(new ProducerRecord(topic,name, jsonObject.toJSONString()));
    }    public static class Builder {        private final MetricRegistry registry;        private TimeUnit rateUnit;        private TimeUnit durationUnit;        private MetricFilter filter;        private Clock clock;        private String topic;        private Builder(MetricRegistry registry) {            this.registry = registry;            this.rateUnit = TimeUnit.SECONDS;            this.durationUnit = TimeUnit.MILLISECONDS;            this.filter = MetricFilter.ALL;            this.clock = Clock.defaultClock();
        }        /**
         * Convert rates to the given time unit.
         *
         * @param rateUnit a unit of time
         * @return {@code this}
         */
        public KafkaReporter.Builder convertRatesTo(TimeUnit rateUnit) {            this.rateUnit = rateUnit;            return this;
        }        /**
         * Convert durations to the given time unit.
         *
         * @param durationUnit a unit of time
         * @return {@code this}
         */
        public KafkaReporter.Builder convertDurationsTo(TimeUnit durationUnit) {            this.durationUnit = durationUnit;            return this;
        }        /**
         * Use the given {@link Clock} instance for the time.
         *
         * @param clock a {@link Clock} instance
         * @return {@code this}
         */
        public Builder withClock(Clock clock) {            this.clock = clock;            return this;
        }        /**
         * Only report metrics which match the given filter.
         *
         * @param filter a {@link MetricFilter}
         * @return {@code this}
         */
        public KafkaReporter.Builder filter(MetricFilter filter) {            this.filter = filter;            return this;
        }        /**
         * Only report metrics which match the given filter.
         *
         * @param topic a
         * @return {@code this}
         */
        public KafkaReporter.Builder topic(String topic) {            this.topic = topic;            return this;
        }        /**
         * Builds a {@link KafkaReporter} with the given properties, writing {@code .csv} files to the
         * given directory.
         *
         * @return a {@link KafkaReporter}
         */
        public KafkaReporter build(KafkaProducer producer) {            return new KafkaReporter(registry,
                    rateUnit,
                    durationUnit,
                    filter,
                    clock,
                    topic,
                    producer);
        }
    }
}

其中的report方法就是获取各种类型指标，并进行对应的输出操作的时机。

如何使用

可在配置文件或者程序中设定需要注册的sink，并带上对应的参数即可：

spark.metrics.conf.*.sink.kafka.class=org.apache.spark.metrics.sink.KafkaSink
spark.metrics.conf.*.sink.kafka.kafka-brokers=XXX:9092

作者：BIGUFO
链接：https://www.jianshu.com/p/cee005368b61