Spark Metrics: how to access executor and worker data?

6
votes

Note: I am using Spark on YARN

I have been trying out the Metric System implemented in Spark. I enabled the ConsoleSink and the CsvSink, and enabled JvmSource for all four instances (driver, master, executor, worker). However I only have driver outputs, and no worker/executor/master data in the console and csv target directory.

After having read this question, I wonder if I do have to ship something to executors when submitting a job.

My submit command: ./bin/spark-submit --class org.apache.spark.examples.SparkPi lib/spark-examples-1.5.0-hadoop2.6.0.jar 10

Bellow is my metric.properties file:

# Enable JmxSink for all instances by class name
*.sink.jmx.class=org.apache.spark.metrics.sink.JmxSink

# Enable ConsoleSink for all instances by class name
*.sink.console.class=org.apache.spark.metrics.sink.ConsoleSink

# Polling period for ConsoleSink
*.sink.console.period=10

*.sink.console.unit=seconds

#######################################
# worker instance overlap polling period
worker.sink.console.period=5

worker.sink.console.unit=seconds
#######################################

# Master instance overlap polling period
master.sink.console.period=15

master.sink.console.unit=seconds

# Enable CsvSink for all instances
*.sink.csv.class=org.apache.spark.metrics.sink.CsvSink
#driver.sink.csv.class=org.apache.spark.metrics.sink.CsvSink

# Polling period for CsvSink
*.sink.csv.period=10

*.sink.csv.unit=seconds

# Polling directory for CsvSink
*.sink.csv.directory=/opt/spark-1.5.0-bin-hadoop2.6/csvSink/

# Worker instance overlap polling period
worker.sink.csv.period=10

worker.sink.csv.unit=second

# Enable Slf4jSink for all instances by class name
#*.sink.slf4j.class=org.apache.spark.metrics.sink.Slf4jSink

# Polling period for Slf4JSink
#*.sink.slf4j.period=1

#*.sink.slf4j.unit=minutes


# Enable jvm source for instance master, worker, driver and executor
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.JvmSource

executor.source.jvm.class=org.apache.spark.metrics.source.JvmSource

And here is a listing of the csv files created by Spark. I am looking forward to access the same data for Spark executors (which are also JVMs).

app-20160812135008-0013.driver.BlockManager.disk.diskSpaceUsed_MB.csv
app-20160812135008-0013.driver.BlockManager.memory.maxMem_MB.csv
app-20160812135008-0013.driver.BlockManager.memory.memUsed_MB.csv
app-20160812135008-0013.driver.BlockManager.memory.remainingMem_MB.csv
app-20160812135008-0013.driver.jvm.heap.committed.csv
app-20160812135008-0013.driver.jvm.heap.init.csv
app-20160812135008-0013.driver.jvm.heap.max.csv
app-20160812135008-0013.driver.jvm.heap.usage.csv
app-20160812135008-0013.driver.jvm.heap.used.csv
app-20160812135008-0013.driver.jvm.non-heap.committed.csv
app-20160812135008-0013.driver.jvm.non-heap.init.csv
app-20160812135008-0013.driver.jvm.non-heap.max.csv
app-20160812135008-0013.driver.jvm.non-heap.usage.csv
app-20160812135008-0013.driver.jvm.non-heap.used.csv
app-20160812135008-0013.driver.jvm.pools.Code-Cache.committed.csv
app-20160812135008-0013.driver.jvm.pools.Code-Cache.init.csv
app-20160812135008-0013.driver.jvm.pools.Code-Cache.max.csv
app-20160812135008-0013.driver.jvm.pools.Code-Cache.usage.csv
app-20160812135008-0013.driver.jvm.pools.Code-Cache.used.csv
app-20160812135008-0013.driver.jvm.pools.Compressed-Class-Space.committed.csv
app-20160812135008-0013.driver.jvm.pools.Compressed-Class-Space.init.csv
app-20160812135008-0013.driver.jvm.pools.Compressed-Class-Space.max.csv
app-20160812135008-0013.driver.jvm.pools.Compressed-Class-Space.usage.csv
app-20160812135008-0013.driver.jvm.pools.Compressed-Class-Space.used.csv
app-20160812135008-0013.driver.jvm.pools.Metaspace.committed.csv
app-20160812135008-0013.driver.jvm.pools.Metaspace.init.csv
app-20160812135008-0013.driver.jvm.pools.Metaspace.max.csv
app-20160812135008-0013.driver.jvm.pools.Metaspace.usage.csv
app-20160812135008-0013.driver.jvm.pools.Metaspace.used.csv
app-20160812135008-0013.driver.jvm.pools.PS-Eden-Space.committed.csv
app-20160812135008-0013.driver.jvm.pools.PS-Eden-Space.init.csv
app-20160812135008-0013.driver.jvm.pools.PS-Eden-Space.max.csv
app-20160812135008-0013.driver.jvm.pools.PS-Eden-Space.usage.csv
app-20160812135008-0013.driver.jvm.pools.PS-Eden-Space.used.csv
app-20160812135008-0013.driver.jvm.pools.PS-Old-Gen.committed.csv
app-20160812135008-0013.driver.jvm.pools.PS-Old-Gen.init.csv
app-20160812135008-0013.driver.jvm.pools.PS-Old-Gen.max.csv
app-20160812135008-0013.driver.jvm.pools.PS-Old-Gen.usage.csv
app-20160812135008-0013.driver.jvm.pools.PS-Old-Gen.used.csv
app-20160812135008-0013.driver.jvm.pools.PS-Survivor-Space.committed.csv
app-20160812135008-0013.driver.jvm.pools.PS-Survivor-Space.init.csv
app-20160812135008-0013.driver.jvm.pools.PS-Survivor-Space.max.csv
app-20160812135008-0013.driver.jvm.pools.PS-Survivor-Space.usage.csv
app-20160812135008-0013.driver.jvm.pools.PS-Survivor-Space.used.csv
app-20160812135008-0013.driver.jvm.PS-MarkSweep.count.csv
app-20160812135008-0013.driver.jvm.PS-MarkSweep.time.csv
app-20160812135008-0013.driver.jvm.PS-Scavenge.count.csv
app-20160812135008-0013.driver.jvm.PS-Scavenge.time.csv
app-20160812135008-0013.driver.jvm.total.committed.csv
app-20160812135008-0013.driver.jvm.total.init.csv
app-20160812135008-0013.driver.jvm.total.max.csv
app-20160812135008-0013.driver.jvm.total.used.csv
DAGScheduler.job.activeJobs.csv
DAGScheduler.job.allJobs.csv
DAGScheduler.messageProcessingTime.csv
DAGScheduler.stage.failedStages.csv
DAGScheduler.stage.runningStages.csv
DAGScheduler.stage.waitingStages.csv
1
apache-sparkmonitoringhadoop-yarnmetrics
I am having this issue, did you ever find a solution? – Babbleshack

1 Answers

4
votes

Since you have not given the command you tried, I am assuming that you are not passing metrics.properties. To pass the metrics.propertis the command should be

spark-submit <other parameters> --files metrics.properties 
--conf spark.metrics.conf=metrics.properties

Note metrics.properties has to be specified in --files & --conf, the --files will transfer the metrics.properties file to the executors. Since you can see the output on driver and not on executors I think you are missing the --files option.