Spark應用監控解決方案--使用Prometheus和Grafana監控Spark應用

Spark任務啓動後，咱們一般都是經過跳板機去Spark UI界面查看對應任務的信息，一旦任務多了以後，這將會是讓人頭疼的問題。若是能將全部任務信息集中起來監控，那將會是很完美的事情。

經過Spark官網指導文檔，發現Spark只支持如下sink

Each instance can report to zero or more sinks. Sinks are contained in the org.apache.spark.metrics.sink package:

• ConsoleSink: Logs metrics information to the console.
• CSVSink: Exports metrics data to CSV files at regular intervals.
• JmxSink: Registers metrics for viewing in a JMX console.
• MetricsServlet: Adds a servlet within the existing Spark UI to serve metrics data as JSON data.
• GraphiteSink: Sends metrics to a Graphite node.
• Slf4jSink: Sends metrics to slf4j as log entries.
• StatsdSink: Sends metrics to a StatsD node.

 

沒有比較經常使用的Influxdb和Prometheus ~~~

谷歌一把發現要支持influxdb須要使用第三方包，比較有參考意義的是這篇，Monitoring Spark Streaming with InfluxDB and Grafana ,在提交任務的時候增長file和配置文件，但成功永遠不會這麼輕鬆。。。

寫入influxdb的數據都是以application_id命名的，相似這種application_1533838659288_1030_1_jvm_heap_usage，也就是說每一個任務的指標都是在單獨的表，最終咱們展現在grafana不還得一個一個配置麼？

顯然這個不是我想要的結果，最終目的就是：一次配置後每提交一個任務自動會在監控上看到。

谷歌是治癒一切的良藥，終究找到一個比較完美的解決方案，就是經過graphite_exporter中轉數據後接入Prometheus，再經過grafana展現出來。

因此，目前已經實踐可行的方案有兩個

方案一：

監控數據直接寫入influxdb，再經過grafana讀取數據作展現，步驟以下：

1.在spark下 conf/metrics.properties 加入如下配置

master.source.jvm.class=org.apache.spark.metrics.source.JvmSource worker.source.jvm.class=org.apache.spark.metrics.source.JvmSource driver.source.jvm.class=org.apache.spark.metrics.source.JvmSourc executor.source.jvm.class=org.apache.spark.metrics.source.JvmSource

*.sink.influx.class=org.apache.spark.metrics.sink.InfluxDbSink
*.sink.influx.protocol=http
*.sink.influx.host=xx.xx.xx.xx
*.sink.influx.port=8086
*.sink.influx.database=sparkonyarn
*.sink.influx.auth=admin:admin

2.在提交任務的時候增長如下配置，並確保如下jar存在

--files /spark/conf/metrics.properties \

--conf spark.metrics.conf=metrics.properties \
--jars /spark/jars/metrics-influxdb-1.1.8.jar,/spark/jars/spark-influx-sink-0.4.0.jar \
--conf spark.driver.extraClassPath=metrics-influxdb-1.1.8.jar:spark-influx-sink-0.4.0.jar \
--conf spark.executor.extraClassPath=metrics-influxdb-1.1.8.jar:spark-influx-sink-0.4.0.jar

缺點：application_id發生變化須要從新配置grafana

 

方案二（目前在用的）：

經過graphite_exporter將原生數據經過映射文件轉化爲有 label 維度的 Prometheus 數據

1.下載graphite_exporter，解壓後執行如下命令，其中graphite_exporter_mapping須要咱們本身建立，內容爲數據映射文件

 nohup ./graphite_exporter --graphite.mapping-config=graphite_exporter_mapping &  

例如

 

mappings: - match: '*.*.jvm.*.*' name: jvm_memory_usage labels: application: $1 executor_id: $2 mem_type: $3 qty: $4

會將數據轉化成 metric name 爲 jvm_memory_usage，label 爲 application，executor_id，mem_type，qty 的格式。

application_1533838659288_1030_1_jvm_heap_usage -> jvm_memory_usage{application="application_1533838659288_1030",executor_id="driver",mem_type="heap",qty="usage"}

2.配置 Prometheus 從 graphite_exporter 獲取數據，重啓prometheus

/path/to/prometheus/prometheus.yml

scrape_configs: - job_name: 'spark' static_configs: - targets: ['localhost:9108']

3.在spark下 conf/metrics.properties 加入如下配置

master.source.jvm.class=org.apache.spark.metrics.source.JvmSource worker.source.jvm.class=org.apache.spark.metrics.source.JvmSource driver.source.jvm.class=org.apache.spark.metrics.source.JvmSourc executor.source.jvm.class=org.apache.spark.metrics.source.JvmSource

*.sink.graphite.class=org.apache.spark.metrics.sink.GraphiteSink
*.sink.graphite.protocol=tcp
*.sink.graphite.host=xx.xx.xx.xx
*.sink.graphite.port=9109
*.sink.graphite.period=5
*.sink.graphite.unit=seconds

4.提交spark任務的時候增長  --files /spark/conf/metrics.properties

5.最後在grafana建立prometheus數據源，建立須要的指標，最終效果以下，有新提交的任務不須要再配置監控，直接選擇application_id就能夠看對應的信息

 

須要用到的jar包

https://repo1.maven.org/maven2/com/izettle/metrics-influxdb/1.1.8/metrics-influxdb-1.1.8.jar

https://mvnrepository.com/artifact/com.palantir.spark.influx/spark-influx-sink

 

模板 

mappings:
- match: '*.*.executor.filesystem.*.*'
  name: filesystem_usage
  labels:
    application: $1
    executor_id: $2
    fs_type: $3
    qty: $4

- match: '*.*.executor.threadpool.*'
  name: executor_tasks
  labels:
    application: $1
    executor_id: $2
    qty: $3

- match: '*.*.executor.jvmGCTime.count'
  name: jvm_gcTime_count
  labels:
    application: $1
    executor_id: $2

- match: '*.*.executor.*.*'
  name: executor_info
  labels:
    application: $1
    executor_id: $2
    type: $3
    qty: $4

- match: '*.*.jvm.*.*'
  name: jvm_memory_usage
  labels:
    application: $1
    executor_id: $2
    mem_type: $3
    qty: $4

- match: '*.*.jvm.pools.*.*'
  name: jvm_memory_pools
  labels:
    application: $1
    executor_id: $2
    mem_type: $3
    qty: $4

- match: '*.*.BlockManager.*.*'
  name: block_manager
  labels:
    application: $1
    executor_id: $2
    type: $3
    qty: $4

- match: '*.driver.DAGScheduler.*.*'
  name: DAG_scheduler
  labels:
    application: $1
    type: $2
    qty: $3

- match: '*.driver.*.*.*.*'
  name: task_info
  labels:
    application: $1
    task: $2
    type1: $3
    type2: $4
    qty: $5

graphite_exporter_mapping

 

參考資料

https://github.com/palantir/spark-influx-sink

https://spark.apache.org/docs/latest/monitoring.html

https://www.linkedin.com/pulse/monitoring-spark-streaming-influxdb-grafana-christian-g%C3%BCgi

https://github.com/prometheus/prometheus/wiki/Default-port-allocations

https://github.com/prometheus/graphite_exporter

https://prometheus.io/download/

https://rokroskar.github.io/monitoring-spark-on-hadoop-with-prometheus-and-grafana.html

https://blog.csdn.net/lsshlsw/article/details/82670508

https://www.jianshu.com/p/274380bb0974