The behavior of MR3 is specified by the configuration file mr3-site.xml in the classpath. Below we describe all configuration keys for MR3 which are divided into 12 sections:

  • MR3Runtime: configuration keys relevant to all components (MR3Client, DAGAppMaster, ContainerWorker)
  • MR3Client: configuration keys that are consumed or set by MR3Client
  • DAGAppMaster: configuration keys that are consumed or set by DAGAppMaster
  • ContainerGroup: configuration keys that specify properties of ContainerGroups
  • DAG: configuration keys that specify properties of DAGs
  • ContainerWorker: configuration keys for ContainerWorkers
  • Memory usage and autoscaling: configuration keys for autoscaling
  • TokenRenewer: configuration keys related to Kerberos and token renewal
  • HistoryLogger: configuration keys for history logging
  • tez.common.counters.Limits: configuration keys for Tez counters (which MR3 borrows from Tez)
  • Kubernetes: configuration keys for running MR3 on Kubernetes
  • Spark on MR3: configuration keys for Spark on MR3

MR3Runtime

Name Default value Description
mr3.runtime tez tez: use Tez 0.9.1 runtime. asm: use ASM runtime.
mr3.master.mode yarn local-thread: DAGAppMaster starts as a new thread inside MR3Client. local-process: DAGAppMaster starts as a new process on the same machine where MR3Client is running. yarn: DAGAppMaster starts as a new container in the Hadoop cluster. kubernetes: DAGAppMaster starts as a pod in the Kubernetes cluster. For more details, see DAGAppMaster and ContainerWorker Modes.
mr3.am.acls.enabled true true: enable ACLs for DAGAppMaster and DAGs. false: disable ACLS for DAGAppMaster and DAGs.
mr3.cluster.additional.classpath Additional classpath for DAGAppMaster and ContainerWorkers
mr3.cluster.use.hadoop-libs false true: include the classpath defined in YarnConfiguration.YARN_APPLICATION_CLASSPATH. false: do not include the classpath defined in YarnConfiguration.YARN_APPLICATION_CLASSPATH.
mr3.am.max.java.heap.fraction 0.8 Fraction of memory to be allocated for Java heap in DAGAppMaster
mr3.container.max.java.heap.fraction 0.8 Fraction of memory to be allocated for Java heap in ContainerWorkers
mr3.async.logging true true: use asynchronous logging. false: use synchronous logging.

MR3Client

Name Default value Description
mr3.lib.uris ${liburis} URIs for the MR3 library jar files
mr3.aux.uris ${auxuris} URIs for the MR3 auxiliary jar files
mr3.queue.name Name of the Yarn queue to which the MR3 job is submitted. On Yarn, the user can exploit it to set a limit on the total resources consumed by MR3. Not used on Kubernetes.
mr3.application.tags Comma-separated list of application tags for the MR3 job
mr3.application.scheduling.properties.map Comma-separated list of scheduling properties for the MR3 job (e.g., foo1=bar1,foo2=bar2)
mr3.application.am.node.label Node label expression for DAGAppMaster on Yarn
mr3.application.worker.node.label Node label expression for ContainerWorkers on Yarn
mr3.credentials.path Path to the credentials for MR3
mr3.am.staging-dir /tmp/${user.name}/staging Staging directory for DAGAppMaster
mr3.am.staging.dir.check.ownership.permission true true: check the ownership and directory permission of the staging directory. false: do not check. Set to false if the staging directory reside on S3 which has notion of ownership and directory permission.
mr3.am.resource.memory.mb 4096 Size of memory in MB for DAGAppMaster
mr3.am.resource.cpu.cores 1 Number of cores for DAGAppMaster
mr3.am.max.app.attempts 2 Max number of Yarn ApplicationAttempts for the MR3 job
mr3.am.log.level INFO Logging level for DAGAppMaster
mr3.am.local.working-dir ”/tmp/${user.name}/mr3/working-dir” Local working directory for DAGAppMaster running in LocalThread or LocalProcess mode
mr3.am.local.log-dir ”/tmp/${user.name}/mr3/log-dir” Logging directory for DAGAppMaster running in LocalThread or LocalProcess mode
mr3.cancel.delegation.tokens.on.completion true true: cancel delegation tokens when the MR3 job completes. false: do not cancel delegation tokens.
mr3.dag.status.pollinterval.ms 1000 Time interval in milliseconds for retrieving the status of running DAGs
mr3.am.session.mode false true: create MR3 SessionClient. false: create MR3 JobClient. For more details, see MR3Client.
mr3.am.session.share.dag.client.rpc true For MR3 SessionClient only. true: all DAGClients share a commmon DAGClientRPC object. false: each DAGClient creates its own DAGClientRPC object.
mr3.session.client.timeout.secs 120 Time in seconds for terminating MR3 SessionClient with a timeout
Name Default value Description
mr3.am.launch.cmd-opts -server -Djava.net.preferIPv4Stack=true -Dhadoop.metrics.log.level=WARN -XX:+UseNUMA -XX:+UseParallelGC Command-line options for launching DAGAppMaster
mr3.am.launch.env LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HADOOP_HOME/lib/native/ Environment variables for launching DAGAppMaster. Each entry takes either “VAR=VALUE” separated by “=” or “VAR”. In the latter case, the value in the system environment is used.

DAGAppMaster

Name Default value Description
mr3.yarn.priority 0 Priority of the MR3 Yarn application
mr3.am.worker.mode local Type of the resource scheduler created in DAGAppMaster. The actual type of ContainerWorkers is specified by mr3.container.resourcescheduler.type for each ContainerGroup. local: ContainerWorkers start as threads inside DAGAppMaster. yarn: ContainerWorkers start as containers in a Hadoop cluster. kubernetes: ContainerWorkers start as Pods in a Kubernetes cluster. process: ContainerWorkers executed by users contact DAGAppMaster. For more details, see DAGAppMaster and ContainerWorker Modes.
mr3.am.max.num.concurrent.dags 128 Max number of DAGs that can run concurrently in DAGAppMaster
mr3.am.shutdown.rightaway true true: DAGAppMaster does not wait until MR3Client retrieves the final states of all DAGs. false: DAGAppMaster waits until MR3Client retrieves the final states of all DAGs.
mr3.am.shutdown.sleep.max.ms 5000 Time in milliseconds to wait until MR3Client retrieves the final states of all DAGs
mr3.am.delete.local.working-dir true true: DAGAppMaster running in LocalThread or LocalProcess mode deletes its local working directory upon termination. false: DAGAppMaster running in LocalThread or LocalProcess mode does not delete its local working directory upon termination. Set to true to ensure that DAGAppMaster in LocalProcess mode terminates properly.
mr3.am.taskcommunicator.type protobuf protobuf: use Protobuf for communication between TaskCommunicator and ContainerWorkers. protowritable: use Protobuf + Writable for communication between TaskCommunicator and ContainerWorkers. writable: use Writable for communication between TaskCommunicator and ContainerWorkers. direct: use direct communication between TaskCommunicator and local ContainerWorkers.
mr3.am.taskcommunicator.thread.count 30 Number of threads in TaskCommunicator for serving requests from ContainerWorkers
mr3.am.rm.heartbeat.interval.ms 1000 Time interval in milliseconds of heartbeats in AMRMClientAsync
mr3.dag.priority.scheme fifo fifo: assign DAG priorities on FIFO basis. concurrent: assign the same priority to all DAGs. Not set for individual DAGs. For more details, see Task Scheduling.
mr3.vertex.priority.scheme intact Scheme for assigning priorities to Vertexes. Available options: intact, roots, leaves, postorder, normalize. Not set for individual DAGs. For more details, see Task Scheduling.
mr3.am.client.thread-count 32 Number of threads in DAGClientServer for serving requests from MR3Clients
mr3.heartbeat.task.timeout.ms 120000 Time in milliseconds for triggering heartbeat timeout for TaskAttempts (counted after being fetched by ContainerWorkers)
mr3.heartbeat.container.timeout.ms 600000 Time in milliseconds for triggering heartbeat timeout for ContainerWorkers. Should be (much) larger than the total time for sleeping due to mr3.container.task.failure.num.sleeps.
mr3.task.heartbeat.timeout.check.ms 30000 Time interval in milliseconds for checking heartbeat timeout for TaskAttempts
mr3.container.heartbeat.timeout.check.ms 15000 Time interval in milliseconds for checking heartbeat timeout for ContainerWorkers
mr3.dag.timeout.kill.check.ms 15000 Time interval in milliseconds for checking DAG timeout
mr3.container.idle.timeout.ms 300000 Time in milliseconds for triggering timeout for idle ContainerWorkers
mr3.am.node-blacklisting.enabled false true: enable node blacklisting. false: disable node blacklisting. For more details, see Node Blacklisting.
mr3.am.maxtaskfailure.percent 5 Percentage of TaskAttempt failures that triggers node blacklisting
mr3.am.max.safe.resource.percent.blacklisted 50 Max percentage of resource to be allocated to a node that is blacklisted
mr3.am.min.safe.resource.percent.blacklisted 10 Min percentage of resource to be allocated to a node that is blacklisted
mr3.dag.delete.local.dir true true: ask ContainerWorkeres to delete DAG-local directories. false: do not ask (as in Spark on MR3).
mr3.dag.recovery.enabled true true: enable DAG recovery when DAGAppMaster restarts. false: disable DAG recovery when DAGAppMaster restarts.
mr3.am.max.finished.reported.dags 10 Max number of DAGs whose final states are kept in DAGAppMaster after being reported to MR3Client
mr3.am.generate.dag.graph.viz false true: create DOT graph files showing the structure of DAGs on the working directory of DAGAppMaster. false: do not create DOT graph files.
Name Default value Description
mr3.am.local.resourcescheduler.min.memory.mb 256 Min size of memory in MB to reserve for all local ContainerWorkers running in DAGAppMaster
mr3.am.local.resourcescheduler.max.memory.mb 4096 Max size of memory in MB to reserve for all local ContainerWorkers running in DAGAppMaster
mr3.am.local.resourcescheduler.max.cpu.cores 16 Max number of cores for all local ContainerWorkers running in DAGAppMaster
mr3.am.local.resourcescheduler.native.fraction 0.0 Fraction of memory to be allocated for native memory for all local ContainerWorkers running in DAGAppMaster
mr3.am.resourcescheduler.max.requests.per.taskscheduler 10 Max number of containers that TaskScheduler can request to Yarn ResourceScheduler at once

ContainerGroup

Name Default value Description
mr3.container.scheduler.scheme none none: do not recycle ContainerWorkers. fifo: use FIFO scheduling for recycling ContainerWorkers. fair: use fair scheduling for recycling ContainerWorkers. For more details, see Recycling ContainerWorkers.
mr3.container.scheduler.remove.empty.kind false true: remove ContainerKind with no ContainerGroups and reserve ContainrWorkers for recycling only if their ContainerKind has multiple ContainerGroups. false: never remove ContainerKind and reserve all ContainrWorkers for recycling.
mr3.dag.queue.scheme common Scheme for mapping DAGs to queues of TaskAttempts in TaskScheduler. Available options: common, individual, capacity. Can be set for individual ContainerGroups. For more details, see Task Scheduling.
mr3.dag.queue.capacity.specs default:0 Comma-separated list of specifications for capacity scheduling. Each entry consists of a queue name and the minimum capacity in percentage. Queues are specified in the order of priority. E.g., high=50,medium=30,default=20,background=0. Can be set for individual ContainerGroups.
mr3.taskattempt.queue.scheme simple Scheme for managing the queue of TaskAttempts in TaskScheduler. Available options: basic, simple, opt, indexed, spark. Can be set for individual ContainerGroups. For common workloads, simple should be used. Scheduling in opt and indexed can be slightly faster than in simple. basic does not use the optimization for increasing the temporal locality of intermediate data and should be used only for performance comparison. spark uses a Spark-style TaskScheduler which schedules consumer Tasks after all producer Tasks are finished.
mr3.taskattempt.queue.scheme.use.delay false true: use delay scheduling. false: do not use delay scheduling. Can be set for individual ContainerGroups. Set to true when using LLAP IO cache in Hive on MR3.
mr3.container.stop.cross.dag.reuse true true: stop cross-DAG container reuse for the current ContainerGroup. false: do not update the current ContainerGroup with regard to cross-DAG container reuse. Can be set for individual ContainerGroups.
mr3.container.reuse false true: reuse ContainerWorkers in the current ContainerGroup. false: use each ContainerWorker only for a single TaskAttempt. Can be set for individual ContainerGroups.
mr3.container.resourcescheduler.type local Type of ContainerWorkers. local: create local ContainerWorkers in DAGAppMaster for the current ContainerGroup. yarn: create Yarn ContainerWorkers for the current ContainerGroup. kubernetes: create Kubernetes ContainerWorkers for the current ContainerGroup. process: use ContainerWorkers executed by users. Can be set for individual ContainerGroups.
mr3.container.combine.taskattempts false true: allow multiple TaskAttempts to run concurrently in a ContainerWorker. false: allow only one TaskAttempt to run at a time in a ContainerWorker. Can be set for individual ContainerGroups.
mr3.container.mix.taskattempts true true: allow TaskAttempts from different DAGs to run concurrently in a ContainerWorker. false: use each ContainerWorker for a single DAG exclusively. Can be set for individual ContainerGroups.
mr3.container.max.num.workers Int.MaxValue Max number of ContainerWorkers that can be created by a ContainerGroup. Can be set for individual ContainerGroups.
mr3.container.log.level INFO Logging level for ContainerWorkers
mr3.use.daemon.shufflehandler 0 Number of shuffle handlers in each ContainerWorker. On Kubernetes, a value of 0 causes the creation of processes for shuffle handlers.
mr3.daemon.shuffle.service-id Service identifier for the shuffle handler
mr3.daemon.shuffle.port Port number for the shuffle handler
mr3.daemon.task.message.buffer.size 16 Size of the message queue for each DaemonTask
Name Default value Description
mr3.container.launch.cmd-opts -server -Djava.net.preferIPv4Stack=true -Dhadoop.metrics.log.level=WARN -XX:+UseNUMA -XX:+UseParallelGC Command-line options for launching ContainerWorkers
mr3.container.launch.env LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HADOOP_HOME/lib/native/ Environment variables for launching ContainerWorkers. Each entry takes either “VAR=VALUE” separated by “=” or “VAR”. In the latter case, the value in the system environment is used.
mr3.container.kill.policy container.kill.wait.workervertex container.kill.wait.workervertex: stop a ContainerWorker only if no more TaskAttempts are to arrive. container.kill.nowait: stop a ContainerWorker right away if it is not serving any TaskAttempt.

DAG

Name Default value Description
mr3.am.min.cluster.resource.memory.mb 102400 Min size of memory in MB that DAGAppMaster assumes as the cluster resource when initializing Map Tasks. Can be set for individual DAGs.
mr3.am.min.cluster.resource.cpu.cores 100 Min number of cores that DAGAppMaster assumes as the cluster resource when initializing Map Tasks. Can be set for individual DAGs.
mr3.am.task.max.failed.attempts 3 Max number of TaskAttempts to create for Task. Must be greater than zero. Can be set for individual DAGs. For more details, see Fault Tolerance.
mr3.am.task.no.retry.errors Comma-separated list of names of Exceptions and Errors that prevent the re-execution of Tasks. Can be set for individual DAGs. For more details, see Fault Tolerance.
mr3.am.task.retry.on.fatal.error false true: retry even if TaskAttempts fail with fatal errors. false: do not retry if TaskAttempts fail with fatal errors. Can be set for individual DAGs. For more details, see Fault Tolerance.
mr3.am.notify.destination.vertex.complete false true: notify ContainerWorker of the completion of all destination Vertexes so that it can delete the directory for intermediate data of the source Vertex. false: do not notify. Can be set for individual DAGs.
mr3.am.commit-all-outputs-on-dag-success true true: commit the output of all Vertexes when DAG completes successfully. false: commit the output when Vertex completes successfully. Can be set for individual DAGs.
mr3.am.permit.custom.user.class false true: allow custom classes for VertexManager, InputInitializer, OutputCommitter. false: do not allow custom classes. Can be set for individual DAGs. Set to true in order to use JDBC storage handlers when running Hive on MR3 on Hadoop.
mr3.am.task.concurrent.run.threshold.percent 100 Percentage of Tasks that complete before starting speculative execution. Can be set to an integer between 1 and 100. If set to 100, speculative execution of TaskAttempts is disabled. Can be set for individual DAGs. For more details, see Speculative Execution.
mr3.am.task.concurrent.run.min.threshold.ms 10000 Minimum of the maximum execution time (in milliseconds) of Tasks that complete before starting speculative execution. For example, a value of 10000 means that if all Tasks complete within 10 seconds before starting speculative execution, we use 10 seconds as their maximum execution time. Can be set for individual DAGs.
mr3.am.task.concurrent.run.multiplier 2.0d Multiplier of the maximum execution time of Tasks that complete before starting speculative execution. Can be set for individual DAGs.
mr3.am.task.concurrent.run.enable.root.vertex false true: Speculative execution is effective on root Vertexes with no ancestors. false: Speculative execution is not effective on root Vertexes.
mr3.dag.queue.name default Name of the Task queue to which the current DAG belongs. Used with capacity scheduling. If an invalid name is given, the default value of default is used. Can be set for individual DAGs.
mr3.dag.vertex.schedule.by.stage false true: A Vertex creates Tasks only after all source Vertexes are completed. false: A Vertex can creates Tasks while source Vertexes are running.
mr3.dag.route.event.after.source.vertex false true: A Vertex receives events only after all source Vertexes are completed. false: A Vertex can receive events while source Vertexes are running.
mr3.dag.include.indeterminate.vertex false true: The DAG contains indeterminate Vertexes whose output can vary at each execution. Fault tolerance is not supported when fetch failures occur. false: The DAG contains no indeterminate Vertexes.
mr3.dag.create.daemon.vertex.always false true: create DaemonVertex in every DAG. false: do not create DaemonVertex except in the creator DAG (which creates ContainerGroups)
mr3.dag.timeout.kill.threshold.secs 0 Maximum execution time (in seconds) of DAGs. Set to 0 in order to disable timeout check. Can be set for individual DAGs.

ContainerWorker

name default value description
mr3.container.get.command.interval.ms 2000 Time interval in milliseconds for retrieving commands in ContainerWorkers that are currently serving TaskAttempts
mr3.container.busy.wait.interval.ms 100 Time interval in milliseconds for retrieving commands in idle ContainerWorkers
mr3.task.am.heartbeat.interval.ms 250 Time interval in milliseconds for sending heartbeats from TaskAttempts
mr3.task.am.heartbeat.duration.interval.ms 15000 Time interval in milliseconds for sending durations in heartbeats from TaskAttempts. It also determines the granularity of updating the duration of TaskAttempts in speculative execution. For more details, see Speculative Execution.
mr3.task.am.heartbeat.counter.interval.ms 60000 Time interval in milliseconds for sending counters in heartbeats from TaskAttempts
mr3.task.max.events.per.heartbeat 500 Max number of task events to include in a heartbeat reply
mr3.container.thread.keep.alive.time.ms 4000 Time in milliseconds for keeping threads serving TaskAttempts in ContainerWorkers
mr3.container.command.num.waits.in.reserved 180 Number of times that reserved ContainerWorker attempts to contact DAGAppMaster at an interval of 1 second. Ensure mr3.container.command.num.waits.in.reserved * 1 second > mr3.k8s.pod.creation.timeout.ms on Kubernetes with autoscaling so that new requests for ContainerWorkers can be made while reserved ContainerWorkers are still alive.
mr3.container.command.num.waits.to.kill 6 Number of times that ContainerWorker attempts to contact DAGAppMaster at an interval of 1 second to re-establish the connection. A failed attempt takes about 10 seconds.
mr3.container.use.termination.checker true true: check whether TaskAttempts terminate successfully or not after termination requests. If a TaskAttempt fails to terminate, terminate the ContainerWorker. false: do not check. Do not set to false in production environments.
mr3.container.terminate.on.fatal.error false true: always terminate ContainerWorkers that throw fatal errors such as OutOfMemoryError. false: do not terminate ContainerWorkers that manage to recover from fatal errors.
mr3.container.termination.checker.timeout.ms 300000 Time in milliseconds before checking the termination of a TaskAttempt after a termination request. With the default value, the ContainerWorker checks whether a TaskAttempt has properly terminated within 300 seconds after the termination request. If the TaskAttempt has not terminated, the whole ContainerWorker is shut down. The user should not use too small a value (e.g., 30000 for 30 seconds) because closing HTTP connections to shuffle handlers can take long.
mr3.container.task.failure.num.sleeps 0 Number of times to sleep (15 seconds each by default) in a ContainerWorker thread after a TaskAttempt fails. Before and after each sleep, the thread tries to allocate a memory block of 1GB to trigger garbage collection. For example, if set to 2, the sequence is: allocate 1GB, sleep 15 seconds, allocate 1GB, sleep 15 seconds, allocate 1GB. If set to 0, do not sleep and do not try to allocate a memory block. For more details, see Basic Performance Tuning. For Hive on MR3, do not set to a non-zero value when executing interactive queries with limit operators because all active Tasks are killed when the number of records reaches the limit.
mr3.container.task.failure.sleep.period.secs 15 Time in seconds to sleep after a TaskAttempt fails
mr3.container.runtime.auto.start.input false true: automatically start LogicalInputs in RuntimeTasks. false: do not automatically start LogicalInputs. Setting it to true can have negative effects on the performance because a ContainerWorker may fetch the same input data multiple times via broadcast edges.
mr3.container.close.filesystem.ugi true true: call FileSystem.closeAllForUGI() after finishing each DAG in ContainerWorkers. false: do not call (for Spark on MR3).
mr3.container.use.framework.counters false true: collect framework counters (on garbage collection and process statistics) in ContainerWorkers. false: do not collect framework counters.
mr3.container.localize.python.working.dir.unsafe false true: localize Python resources (*.py or *.PY) in the working directory of ContainerWorkers. false: do not localize Python resources in the working directory. For more details, see Using User Defined Functions.
mr3.container.use.am.credentials.for.daemon true true: use the credentials of DAGAppMaster for all DaemonTaskAttempts. false: use the credentials of the DAG for all its DaemonTaskAttempts.
name default value description
mr3.container.elastic.execution.memory.commit.ratio 1.0 Multiplier for memory to be allocated to each TaskAttempt. For exmple, a value of 1.5 means that a taksattempt created with memory resource of 4GB is actually allocated 6GB of memory in a ContainerWorker.

Memory usage and autoscaling (see Autoscaling)

Name Default value Description
mr3.memory.usage.check.scheme none average: calculate the average memory usage of the current window. maximum: calculate the maximum memory usage of the current window. none: do not calculate memory usage.
mr3.memory.usage.check.window.length.secs 600 Window length in seconds for calculating memory usage
mr3.check.memory.usage.event.interval.secs 10 Time interval in seconds for 1) generating events for calculating memory usage for autoscaling and 2) updating the number of ContainerWorkers belonging to each ContainerGroup under fair scheduling when recycling ContainerWorkers
mr3.enable.auto.scaling false true: enable autoscaling. false: disable autoscaling.
mr3.auto.scale.out.threshold.percent 80 Minimum percentage of memory usage to trigger scale-out
mr3.auto.scale.in.threshold.percent 50 Maximum percentage of memory usage to trigger scale-in
mr3.auto.scale.in.min.hosts 1 Minimum number of nodes that should remain when performing scale-in
mr3.auto.scale.out.grace.period.secs 300 Cooldown period in seconds after triggering scale-out
mr3.auto.scale.in.delay.after.scale.out.secs 60 Minimum time in seconds to wait after leaving scale-out before triggering scale-in
mr3.auto.scale.in.grace.period.secs 300 Cooldown period in seconds after triggering scale-in
mr3.auto.scale.in.wait.dag.finished true true: wait until all running DAGs complete before terminating containers in the event of scale-in. false: do not wait and terminate containers immediately.
mr3.auto.scale.out.num.initial.containers 0 If greater than zero: number of containers to add in the case of scale-out when no containers are running. If zero or less: not used.
mr3.auto.scale.out.num.increment.containers 0 If greater than zero: number of containers to add in the case of scale-out. If zero or less: use mr3.auto.scale.out.threshold.percent to calculate the number of containers to add.
mr3.auto.scale.in.num.decrement.hosts 0 If greater than zero: number of hosts to remove in the case of scale-in. If zero or less: use mr3.auto.scale.in.threshold.percent to calculate the number of hosts to remove.

Prometheus

Name Default value Description
mr3.prometheus.enable.metrics false true: DAGAppMaster runs a Prometheus client to export metrics. false: DAGAppMaster does not run a Prometheus client.
mr3.prometheus.enable.jvm.metrics false true: export Java VM metrics from DAGAppMaster (using io.prometheus.client.hotspot.DefaultExports). false: do not export Java VM metrics.
mr3.prometheus.httpserver.port 9890 Port number for the Prometheus client
mr3.prometheus.worker.enable.metrics false true: Every ContainerWorker runs a Prometheus client to export metrics. false: ContainerWorkers do not run Prometheus clients.
mr3.prometheus.worker.enable.jvm.metrics false true: export Java VM metrics from ContainerWorkers (using io.prometheus.client.hotspot.DefaultExports). false: do not export Java VM metrics from ContainerWorkers.
mr3.prometheus.worker.httpserver.port 0 Port number for the Prometheus clients for ContainerWorkers. Use 0 if multiple ContainerWorkers run on the same node.

TokenRenewer

Name Default value Description
mr3.principal Kerberos principal
mr3.keytab Location of the Kerberos keytab file
mr3.token.renewal.fraction 0.75 Fraction of the token renewal interval for renewing tokens conservatively
mr3.token.renewal.retry.interval.ms 3600000 Time interval in milliseconds for retrying token renewal
mr3.token.renewal.num.credentials.files 5 Max number of credential files to keep for token renewal
mr3.token.renewal.hdfs.enabled false true: automatically renew HDFS tokens. false: do not renew HDFS tokens.
mr3.token.renewal.hive.enabled false true: automatically renew Hive tokens. false: do not renew Hive tokens.
mr3.am.token.renewal.paths Path that specifies FileSystem for token renewal. If empty, DAGAppMaster uses the staging directory.
mr3.token.renewal.pass.credentials.via.memory true true: DAGAppMaster passes credentials to ContainerWorkers directly via messages. false: DAGAppMaster stores credentials on HDFS.

HistoryLogger (see Running MR3-UI)

Name Default value Description
mr3.app.history.logging.enabled false true: enable history logging for Yarn applications and ContainerWorkers. false: disable history logging for Yarn applications and ContainerWorkers.
mr3.dag.history.logging.enabled false true: enable history logging for DAGs. false: disable history logging for DAGs.
mr3.task.history.logging.enabled false true: enable history logging for Tasks. false: disable history logging for Tasks.

tez.common.counters.Limits

Name Default value Description
tez.counters.max 1200 Max number of Tez counters
tez.counters.max.groups 500 Max number of Tez counters groups
tez.counters.counter-name.max-length 64 Max length of Tez counter names
tez.counters.group-name.max-length 256 Max length of Tez counters group names

Kubernetes

Name Default value Description
mr3.k8s.api.server.url https://kubernetes.default.svc URL for the Kubernetes API server
mr3.k8s.client.config.file Configuration file for creating a Kubernetes client (e.g., ~/.kube/config)
mr3.k8s.service.account.use.token.ca.cert.path true true: use mr3.k8s.service.account.token.path and mr3.k8s.service.account.token.ca.cert.path. false: do not use.
mr3.k8s.service.account.token.path /var/run/secrets/kubernetes.io/serviceaccount/token Token path for ServiceAccount for the Kubernetes client. Used only when mr3.k8s.service.account.use.token.ca.cert.path is set to true.
mr3.k8s.service.account.token.ca.cert.path /var/run/secrets/kubernetes.io/serviceaccount/ca.crt Certificate path for ServiceAccount for the Kubernetes client. Used only when mr3.k8s.service.account.use.token.ca.cert.path is set to true.
Name Default value Description
mr3.k8s.namespace mr3 Kubernetes namespace to use when creating a Kubernetes client
mr3.k8s.am.service.host Host associated with the Service for DAGAppMaster Pod. Set only when MR3Client runs outside the Kubernetes cluster and the user creates a Service manually. If not set, MR3 obtains the host from the Service created by MR3Client, e.g., service-master-4110-0.hivemr3.svc.cluster.local.
mr3.k8s.am.service.port 80 Port associated with the Service for DAGAppMaster Pod
mr3.k8s.nodes.polling.interval.ms 60000 Time interval in milliseconds for querying states of Kubernetes Nodes
mr3.k8s.pods.polling.interval.ms 15000 Time interval in milliseconds for querying Pod states
mr3.k8s.pod.creation.timeout.ms 30000 Time in milliseconds to wait until a new Pod is created
mr3.k8s.pod.image.pull.policy IfNotPresent Image pull policy for Pods
mr3.k8s.pod.image.pull.secrets Image pull secrets for Pods
mr3.k8s.pod.master.serviceaccount ServiceAccount for DAGAppMaster Pod
mr3.k8s.pod.worker.serviceaccount ServiceAccount for ContainerWorker Pods
mr3.k8s.pod.master.image Docker image for DAGAppMaster containers
mr3.k8s.pod.master.user User for DAGAppMaster Pod
mr3.k8s.pod.master.emptydirs Comma-separated list of directories where emptyDir volumes are mounted for DAGAppMaster
mr3.k8s.pod.master.hostpaths Comma-separated list of directories (on each node) to which hostPath volumes point for DAGAppMaster. For example, /data1/k8s,/data2/k8s,/data3/k8s mounts three hostPath volumes created from three local directories of the node where DAGAppMaster Pod is to run.
mr3.k8s.pod.master.node.selector Comma-separated list of node selectors for DAGAppMaster Pod (e.g., masternode=true,hivemr3=true)
mr3.k8s.pod.master.toleration.specs Comma-separated list of toleration specifications for DAGAppMaster Pod. The format of each entry is [key]:[operator]:[value]:[effect]:[toleration in seconds] where [value] and :[toleration in seconds] are optional.
mr3.k8s.master.working.dir Working directory for DAGAppMaster container
mr3.k8s.master.command /usr/bin/java Command for launching Java VM for DAGAppMaster container
mr3.k8s.master.pod.affinity.match.label Label for specifying Pod affinity for DAGAppMaster Pod (e.g., hivemr3_app=hiveserver2). Internally MR3 uses preferredDuringSchedulingIgnoredDuringExecution with a weight of 100.
mr3.k8s.master.pod.additional.labels Comma-separated list of additional labels for DAGAppMaster Pod (e.g., foo=bar,hivemr3_aux=prometheus)
mr3.k8s.master.pod.cpu.limit.multiplier 1.0 Multiplier for the CPU resource limit for DAGAppMaster Pod
mr3.k8s.pod.worker.image Docker image for ContainerWorker containers
mr3.k8s.pod.worker.user User for ContainerWorker Pods
mr3.k8s.pod.worker.emptydirs Comma-separated list of directories where emptyDir volumes are mounted for ContainerWorkers. These volumes become local directories where intermediate data is written.
mr3.k8s.pod.worker.hostpaths Comma-separated list of directories (on each node) to which hostPath volumes point for ContainerWorkers. For example, /data1/k8s,/data2/k8s,/data3/k8s mounts three hostPath volumes created from three local directories of the node where ContainerWorker Pods are to run. These volumes become local directories where intermediate data is written.
mr3.k8s.pod.worker.additional.hostpaths Comma-separated list of additional directories (on each node) to which hostPath volumes point for ContainerWorkers
mr3.k8s.pod.worker.node.selector Comma-separated list of node selectors for ContainerWorker Pods (e.g., workernode=true,hivemr3=true)
mr3.k8s.pod.worker.toleration.specs Comma-separated list of toleration specifications for ContainerWorker Pods. The format of each entry is [key]:[operator]:[value]:[effect]:[toleration in seconds] where [value] and :[toleration in seconds] are optional.
mr3.k8s.pod.worker.node.affinity.specs Comma-separated list of node affinity specifications for ContainerWorker Pods. The format of each entry is [key]:[operator]:[value#1]:...:[value#N] (e.g., key1:In:value1:value2:value3). Internally MR3 uses requiredDuringSchedulingIgnoredDuringExecution.
mr3.k8s.worker.working.dir Working directory for ContainerWorker containers
mr3.k8s.java.io.tmpdir /tmp Temporary directory for Java in ContainerWorker containers
mr3.k8s.worker.command /usr/bin/java Command for launching Java VM for ContainerWorker containers
mr3.k8s.worker.total.max.memory.gb 1048576 Max memory in GB for all ContainerWorker Pods
mr3.k8s.worker.total.max.cpu.cores 1048576 Max number of cores for all ContainerWorker Pods
mr3.k8s.pod.cpu.cores.max.multiplier 1.0 Multiplier for the limit of CPU cores for each ContainerWorker Pod. For example, a value of 2.0 means that the CPU limit of a ContainerWorker Pod is twice its CPU request.
mr3.k8s.pod.memory.max.multiplier 1.0 Multiplier for the limit of memory for each ContainerWorker Pod. For example, a value of 2.0 means that the memory limit of a ContainerWorker Pod is twice its memory request. We do not recommend a value of larger than 1.0 unless every node has more memory than needed (because ContainerWorker Pods may be killed).
mr3.k8s.conf.dir.configmap Name of the ConfigMap carrying all configuration files (such as mr3-site.xml)
mr3.k8s.conf.dir.mount.dir Mount path for the ConfigMap carrying all configuration files
mr3.k8s.keytab.secret Name of the Secret (containing the Keytab file) to be used by DAGAppMaster
mr3.k8s.worker.secret Name of the Secret to be used by ContainerWorkers
mr3.k8s.keytab.mount.dir Mount path for the Secret in DAGAppMaster (containing the keytab file) and in ContainerWorkers
mr3.k8s.keytab.mount.file File name for the Secret containing the keytab file. mr3.k8s.keytab.mount.dir and mr3.k8s.keytab.mount.file specify the full path for the keytab file mounted inside ContainerWorker Pods.
mr3.k8s.mount.keytab.secret false true: mount mr3.k8s.keytab.secret to mr3.k8s.keytab.mount.dir. false: do not mount. Set to true when: 1) mr3.token.renewal.hdfs.enabled is set to true; 2) mr3.token.renewal.hive.enabled is set to true; 3) secure shuffle is used (tez.runtime.shuffle.ssl.enable is set to true in tez-site.xml) so as to pass keystore and truststore files.
mr3.k8s.mount.worker.secret false true: mount mr3.k8s.worker.secret to mr3.k8s.keytab.mount.dir. false: do not mount. Set to true when secure shuffle is used (tez.runtime.shuffle.ssl.enable is set to true in tez-site.xml) so as to pass keystore and truststore files.
mr3.k8s.host.aliases Comma-separated list of pairs of a hostname and an IP address. For example, foo=1.1.1.1,bar=2.2.2.2 registers host foo as IP address 1.1.1.1 in DAGAppMaster and ContainerWorker Pods, and so on. For more details, see Performance Tuning for Kubernetes.
mr3.k8s.shuffle.process.ports Comma-separated list of port numbers for shuffle handlers. For example, 15500,15510,15520,15530,15540 creates 5 shuffle handlers with port number 15500 to 15540. For more details, see Using the MR3 Shuffle Handler.
mr3.k8s.shufflehandler.process.memory.mb 1024 Size of memory in MB for the container for shuffle handlers
mr3.k8s.readiness.probe.initial.delay.secs 10 Time in seconds before performing the first readiness probe
mr3.k8s.readiness.probe.period.secs 20 Time interval in seconds for performing the readiness probe
mr3.k8s.liveness.probe.initial.delay.secs 20 Time in seconds before performing the first liveness probe
mr3.k8s.liveness.probe.period.secs 40 Time interval in seconds for performing the liveness probe
Name Default value Description
mr3.k8s.master.persistentvolumeclaim.mounts Comma-separated list of pairs of a PersistentVolumeClaim and its mount point for DAGAppMaster Pod
mr3.k8s.worker.persistentvolumeclaim.mounts Comma-separated list of pairs of a PersistentVolumeClaim and its mount point. For example, foo1=bar1,foo2=bar2,foo3=bar3 mounts PersistentVolumeClaim foo1 on directory bar1 in ContainerWorker Pods, and so on.
mr3.k8s.pod.worker.security.context.sysctls Comma-separated list of sysctl properties to be set by an init container in a ContainerWorker Pod. Example: net.core.somaxconn=16384,net.ipv4.ip_local_port_range='1024 65535'. For more details, see Configuring Kernel Parameters.
mr3.k8s.pod.worker.init.container.command Shell command to be executed by an init container in a ContainerWorker Pod. Before executing the shell command, the init container mounts hostPath volumes specified by mr3.k8s.pod.worker.hostpaths. Example: chown 1000:1000 /data1/k8s/; ls -alt /data1/k8s.
mr3.k8s.pod.worker.init.container.image Docker image for init containers when mr3.k8s.pod.worker.security.context.sysctls is set. busybox works okay.

Spark on MR3

Name Default value Description
mr3.spark.delay.scheduling.interval.ms 1000 Time interval in milliseconds of checking delay scheduling
mr3.dag.queue.scheme common Scheme for mapping DAGs to queues of TaskAttempts in TaskScheduler. Available options: common (corresponding to Spark FIFO scheme), individual (corresponding to Spark FAIR scheme). Can be set for individual Spark applications.
mr3.dag.priority.scheme fifo fifo: use Spark job priorities for DAG priorities. concurrent: assign the same priority to all DAGs. Not set for individual DAGs.
mr3.taskattempt.queue.scheme opt Scheme for managing the queue of TaskAttempts in TaskScheduler. Available option: opt (implementing delay scheduling).
mr3.vertex.priority.scheme normalize intact: use Spark stage IDs for Vertex priorities. normalize: set Vertex priorties to 0. Not set for individual DAGs.
Calendar Oct 2, 2024