Master and Worker configuration

Master and Worker daemons will only read configuration from conf/gear.conf.

Master reads configuration from section master and gearpump:

master {
}
gearpump{
}

Worker reads configuration from section worker and gearpump:

worker {
}
gearpump{
}

Configuration for user submitted application job

For user application job, it will read configuration file gear.conf and application.conf from classpath, while application.conf has higher priority. The default classpath contains:

  1. conf/
  2. current working directory.

For example, you can put a application.conf on your working directory, and then it will be effective when you submit a new job application.

Logging

To change the log level, you need to change both gear.conf, and log4j.properties.

To change the log level for master and worker daemon

Please change log4j.rootLevel in log4j.properties, gearpump-master.akka.loglevel and gearpump-worker.akka.loglevel in gear.conf.

To change the log level for application job

Please change log4j.rootLevel in log4j.properties, and akka.loglevel in gear.conf or application.conf.

Gearpump Default Configuration

This is the default configuration for gear.conf.

config item default value description
gearpump.hostname "127.0.0.1" hostname of current machine. If you are using local mode, then set this to 127.0.0.1. If you are using cluster mode, make sure this hostname can be accessed by other machines.
gearpump.cluster.masters ["127.0.0.1:3000"] Config to set the master nodes of the cluster. If there are multiple master in the list, then the master nodes runs in HA mode. For example, you may start three master, on node1: bin/master -ip node1 -port 3000, on node2: bin/master -ip node2 -port 3000, on node3: bin/master -ip node3 -port 3000, then you need to set gearpump.cluster.masters = ["node1:3000","node2:3000","node3:3000"]
gearpump.task-dispatcher "gearpump.shared-thread-pool-dispatcher" default dispatcher for task actor
gearpump.metrics.enabled true flag to enable the metrics system
gearpump.metrics.sample-rate 1 We will take one sample every gearpump.metrics.sample-rate data points. Note it may have impact that the statistics on UI portal is not accurate. Change it to 1 if you want accurate metrics in UI
gearpump.metrics.report-interval-ms 15000 we will report once every 15 seconds
gearpump.metrics.reporter "akka" available value: "graphite", "akka", "logfile" which write the metrics data to different places.
gearpump.retainHistoryData.hours 72 max hours of history data to retain, Note: Due to implementation limitation(we store all history in memory), please don't set this to too big which may exhaust memory.
gearpump.retainHistoryData.intervalMs 3600000 time interval between two data points for history data (unit: ms). Usually this is set to a big value so that we only store coarse-grain data
gearpump.retainRecentData.seconds 300 max seconds of recent data to retain. This is for the fine-grain data
gearpump.retainRecentData.intervalMs 15000 time interval between two data points for recent data (unit: ms)
gearpump.log.daemon.dir "logs" The log directory for daemon processes(relative to current working directory)
gearpump.log.application.dir "logs" The log directory for applications(relative to current working directory)
gearpump.serializers a map custom serializer for streaming application, e.g. "scala.Array" = ""
gearpump.worker.slots 1000 How many slots each worker contains
gearpump.appmaster.vmargs "-server -Xss1M -XX:+HeapDumpOnOutOfMemoryError -XX:+UseConcMarkSweepGC -XX:CMSInitiatingOccupancyFraction=80 -XX:+UseParNewGC -XX:NewRatio=3 -Djava.rmi.server.hostname=localhost" JVM arguments for AppMaster
gearpump.appmaster.extraClasspath "" JVM default class path for AppMaster
gearpump.executor.vmargs "-server -Xss1M -XX:+HeapDumpOnOutOfMemoryError -XX:+UseConcMarkSweepGC -XX:CMSInitiatingOccupancyFraction=80 -XX:+UseParNewGC -XX:NewRatio=3 -Djava.rmi.server.hostname=localhost" JVM arguments for executor
gearpump.executor.extraClasspath "" JVM default class path for executor
gearpump.jarstore.rootpath "jarstore/" Define where the submitted jar file will be stored. This path follows the hadoop path schema. For HDFS, use hdfs://host:port/path/, and HDFS HA, hdfs://namespace/path/; if you want to store on master nodes, then use local directory. jarstore.rootpath = "jarstore/" will point to relative directory where master is started. jarstore.rootpath = "/jarstore/" will point to absolute directory on master server
gearpump.scheduling.scheduler-class "io.gearpump.cluster.scheduler.PriorityScheduler" Class to schedule the applications.
gearpump.services.host "127.0.0.1" dashboard UI host address
gearpump.services.port 8090 dashboard UI host port
gearpump.netty.buffer-size 5242880 netty connection buffer size
gearpump.netty.max-retries 30 maximum number of retries for a netty client to connect to remote host
gearpump.netty.base-sleep-ms 100 base sleep time for a netty client to retry a connection. Actual sleep time is a multiple of this value
gearpump.netty.max-sleep-ms 1000 maximum sleep time for a netty client to retry a connection
gearpump.netty.message-batch-size 262144 netty max batch size
gearpump.netty.flush-check-interval 10 max flush interval for the netty layer, in milliseconds
gearpump.netty.dispatcher "gearpump.shared-thread-pool-dispatcher" default dispatcher for netty client and server
gearpump.shared-thread-pool-dispatcher default Dispatcher with "fork-join-executor" default shared thread pool dispatcher
gearpump.single-thread-dispatcher PinnedDispatcher default single thread dispatcher
gearpump.serialization-framework "io.gearpump.serializer.FastKryoSerializationFramework" Gearpump has built-in serialization framework using Kryo. Users are allowed to use a different serialization framework, like Protobuf. See io.gearpump.serializer.FastKryoSerializationFramework to find how a custom serialization framework can be defined
worker.executor-share-same-jvm-as-worker false whether the executor actor is started in the same jvm(process) from which running the worker actor, the intention of this setting is for the convenience of single machine debugging, however, the app jar need to be added to the worker's classpath when you set it true and have a 'real' worker in the cluster