org.apache.drill.yarn.appMaster

Class ClusterControllerImpl

java.lang.Object

org.apache.drill.yarn.appMaster.ClusterControllerImpl

All Implemented Interfaces:

ClusterController, RegistryHandler

public class ClusterControllerImpl
extends Object
implements ClusterController

Controls the Drill cluster by representing the current cluster state with a desired state, taking corrective action to keep the cluster in the desired state. The cluster as a whole has a state, as do each task (node) within the cluster.

This class is designed to allow unit tests. In general, testing the controller on a live cluster is tedious. This class encapsulates the controller algorithm so it can be driven by a simulated cluster.

This object is shared between threads, thus synchronized.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ClusterControllerImpl.State Controller lifecycle state.

Field Summary

Fields

Modifier and Type	Field and Description
protected int	maxRetries Maximum number of retries for each task launch.

Constructor Summary

Constructors

Constructor and Description
ClusterControllerImpl(AMYarnFacade yarn)

Method Summary

Modifier and Type	Method and Description
boolean	cancelTask(int id) Cancels the given task, reducing the target task count.
void	completionAck(Task task, String propertyKey)
void	containerAllocated(Task task)
void	containerReleased(Task task)
void	containersAllocated(List<org.apache.hadoop.yarn.api.records.Container> containers) The RM has allocated one or more containers in response to container requests submitted to the RM.
void	containersCompleted(List<org.apache.hadoop.yarn.api.records.ContainerStatus> statuses) The Resource Manager reports that containers have completed with the given statuses.
void	containerStarted(org.apache.hadoop.yarn.api.records.ContainerId containerId) The NM reports that a container has successfully started.
void	containerStopped(org.apache.hadoop.yarn.api.records.ContainerId containerId) The Node Manager reports that a container has stopped.
void	enableFailureCheck(boolean flag)
void	fireLifecycleChange(TaskLifecycleListener.Event event, EventContext context)
int	getFreeNodeCount() Get the approximate number of free YARN nodes (those that can accept a task request.) Starts with the number of nodes from the node inventory, then subtracts any in-flight requests (which do not, by definition, have node allocated.)
List<Task>	getHistory()
int	getMaxRetries()
NodeInventory	getNodeInventory()
List<SchedulerStateActions>	getPools()
float	getProgress()
Object	getProperty(String key)
ClusterControllerImpl.State	getState()
int	getStopTimeoutMs()
int	getTargetCount() Return the target number of tasks that the controller seeks to maintain.
AMYarnFacade	getYarn()
boolean	isLive()
boolean	isTaskLive(int id)
void	registerLifecycleListener(TaskLifecycleListener listener)
void	registerScheduler(Scheduler scheduler) Define a task type.
void	registryDown()
void	releaseHost(String hostName)
void	reserveHost(String hostName)
void	resizeDelta(int delta) Request to resize the Drill cluster by a relative amount.
int	resizeTo(int n) Request to resize the Drill cluster to the given size.
void	setMaxRetries(int value)
void	setProperty(String key, Object value)
void	shutDown() Indicates a request to gracefully shut down the cluster.
void	startAck(Task task, String propertyKey, Object value)
void	started() Called when the caller has completed start-up and the controller should become live.
void	stopTaskFailed(org.apache.hadoop.yarn.api.records.ContainerId containerId, Throwable t) The Node Manager API reports that a request sent to the NM to stop a task has failed.
boolean	succeeded()
boolean	supportsDiskResource() Whether this distribution of YARN supports disk resources.
void	taskEnded(Task task)
void	taskGroupCompleted(SchedulerStateActions taskGroup)
void	taskRetried(Task task)
void	taskStartFailed(org.apache.hadoop.yarn.api.records.ContainerId containerId, Throwable t) The RM API reports that an attempt to start a container has failed locally.
void	tick(long curTime) Called by the timer ("pulse") thread to trigger time-based events.
void	updateRMStatus() Get an update from YARN on available resources.
void	visit(ControllerVisitor visitor) Allow an observer to see a consistent view of the controller's state by performing the visit in a synchronized block.
void	visitTasks(TaskVisitor visitor) Allow an observer to see a consistent view of the controller's task state by performing the visit in a synchronized block.
boolean	waitForCompletion() Called by the main thread to wait for the normal shutdown of the controller.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

maxRetries

protected int maxRetries

Maximum number of retries for each task launch.

Constructor Detail

ClusterControllerImpl

public ClusterControllerImpl(AMYarnFacade yarn)

Method Detail

enableFailureCheck

public void enableFailureCheck(boolean flag)

Specified by:

enableFailureCheck in interface ClusterController

registerScheduler

public void registerScheduler(Scheduler scheduler)

Define a task type. Registration order is important: the controller starts task in the order that they are registered. Must happen before the YARN callbacks start.

Specified by:

registerScheduler in interface ClusterController

Parameters:

scheduler -

started

public void started()
             throws YarnFacadeException,
                    AMException

Called when the caller has completed start-up and the controller should become live.

Specified by:

started in interface ClusterController

Throws:

YarnFacadeException

AMException

tick

public void tick(long curTime)

Description copied from interface: ClusterController

Called by the timer ("pulse") thread to trigger time-based events.

Specified by:

tick in interface ClusterController

getFreeNodeCount

public int getFreeNodeCount()

Get the approximate number of free YARN nodes (those that can accept a task request.) Starts with the number of nodes from the node inventory, then subtracts any in-flight requests (which do not, by definition, have node allocated.)

This approximation does not consider whether the node has sufficient resources to run a task; only whether the node itself exists.

Specified by:

getFreeNodeCount in interface ClusterController

Returns:

The approximate number of free YARN nodes.

updateRMStatus

public void updateRMStatus()

Get an update from YARN on available resources.

Specified by:

updateRMStatus in interface ClusterController

containersAllocated

public void containersAllocated(List<org.apache.hadoop.yarn.api.records.Container> containers)

Description copied from interface: ClusterController

The RM has allocated one or more containers in response to container requests submitted to the RM.

Specified by:

containersAllocated in interface ClusterController

Parameters:

containers - the set of containers provided by YARN

containerStarted

public void containerStarted(org.apache.hadoop.yarn.api.records.ContainerId containerId)

Description copied from interface: ClusterController

The NM reports that a container has successfully started.

Specified by:

containerStarted in interface ClusterController

Parameters:

containerId - the container which started

taskStartFailed

public void taskStartFailed(org.apache.hadoop.yarn.api.records.ContainerId containerId,
                            Throwable t)

Description copied from interface: ClusterController

The RM API reports that an attempt to start a container has failed locally.

Specified by:

taskStartFailed in interface ClusterController

Parameters:

containerId - the container that failed to launch

t - the error that occurred

containerStopped

public void containerStopped(org.apache.hadoop.yarn.api.records.ContainerId containerId)

Description copied from interface: ClusterController

The Node Manager reports that a container has stopped.

Specified by:

containerStopped in interface ClusterController

containersCompleted

public void containersCompleted(List<org.apache.hadoop.yarn.api.records.ContainerStatus> statuses)

Description copied from interface: ClusterController

The Resource Manager reports that containers have completed with the given statuses. Find the task for each container and mark them as completed.

Specified by:

containersCompleted in interface ClusterController

getProgress

public float getProgress()

Specified by:

getProgress in interface ClusterController

stopTaskFailed

public void stopTaskFailed(org.apache.hadoop.yarn.api.records.ContainerId containerId,
                           Throwable t)

Description copied from interface: ClusterController

The Node Manager API reports that a request sent to the NM to stop a task has failed.

Specified by:

stopTaskFailed in interface ClusterController

Parameters:

containerId - the container that failed to stop

t - the reason that the stop request failed

resizeDelta

public void resizeDelta(int delta)

Description copied from interface: ClusterController

Request to resize the Drill cluster by a relative amount.

Specified by:

resizeDelta in interface ClusterController

Parameters:

delta - the amount of change. Can be positive (to grow) or negative (to shrink the cluster)

resizeTo

public int resizeTo(int n)

Description copied from interface: ClusterController

Request to resize the Drill cluster to the given size.

Specified by:

resizeTo in interface ClusterController

Parameters:

n - the desired cluster size

shutDown

public void shutDown()

Description copied from interface: ClusterController

Indicates a request to gracefully shut down the cluster.

Specified by:

shutDown in interface ClusterController

waitForCompletion

public boolean waitForCompletion()

Description copied from interface: ClusterController

Called by the main thread to wait for the normal shutdown of the controller. Such shutdown occurs when the admin sends a sutdown command from the UI or REST API.

Specified by:

waitForCompletion in interface ClusterController

isLive

public boolean isLive()

succeeded

public boolean succeeded()

containerAllocated

public void containerAllocated(Task task)

getYarn

public AMYarnFacade getYarn()

containerReleased

public void containerReleased(Task task)

taskEnded

public void taskEnded(Task task)

taskRetried

public void taskRetried(Task task)

taskGroupCompleted

public void taskGroupCompleted(SchedulerStateActions taskGroup)

getMaxRetries

public int getMaxRetries()

getStopTimeoutMs

public int getStopTimeoutMs()

reserveHost

public void reserveHost(String hostName)

Specified by:

reserveHost in interface RegistryHandler

releaseHost

public void releaseHost(String hostName)

Specified by:

releaseHost in interface RegistryHandler

getNodeInventory

public NodeInventory getNodeInventory()

setProperty

public void setProperty(String key,
                        Object value)

Specified by:

setProperty in interface ClusterController

getProperty

public Object getProperty(String key)

Specified by:

getProperty in interface ClusterController

registerLifecycleListener

public void registerLifecycleListener(TaskLifecycleListener listener)

Specified by:

registerLifecycleListener in interface ClusterController

fireLifecycleChange

public void fireLifecycleChange(TaskLifecycleListener.Event event,
                                EventContext context)

setMaxRetries

public void setMaxRetries(int value)

Specified by:

setMaxRetries in interface ClusterController

getTargetCount

public int getTargetCount()

Description copied from interface: ClusterController

Return the target number of tasks that the controller seeks to maintain. This is the sum across all pools.

Specified by:

getTargetCount in interface ClusterController

getState

public ClusterControllerImpl.State getState()

visit

public void visit(ControllerVisitor visitor)

Description copied from interface: ClusterController

Allow an observer to see a consistent view of the controller's state by performing the visit in a synchronized block.

Specified by:

visit in interface ClusterController

getPools

public List<SchedulerStateActions> getPools()

visitTasks

public void visitTasks(TaskVisitor visitor)

Description copied from interface: ClusterController

Allow an observer to see a consistent view of the controller's task state by performing the visit in a synchronized block.

Specified by:

visitTasks in interface ClusterController

getHistory

public List<Task> getHistory()

isTaskLive

public boolean isTaskLive(int id)

Specified by:

isTaskLive in interface ClusterController

cancelTask

public boolean cancelTask(int id)

Description copied from interface: ClusterController

Cancels the given task, reducing the target task count. Called from the UI to allow the user to select the specific task to end when reducing cluster size.

Specified by:

cancelTask in interface ClusterController

completionAck

public void completionAck(Task task,
                          String propertyKey)

Specified by:

completionAck in interface RegistryHandler

startAck

public void startAck(Task task,
                     String propertyKey,
                     Object value)

Specified by:

startAck in interface RegistryHandler

supportsDiskResource

public boolean supportsDiskResource()

Description copied from interface: ClusterController

Whether this distribution of YARN supports disk resources.

Specified by:

supportsDiskResource in interface ClusterController

Returns:

True if this distribution of YARN supports disk resources. False otherwise.

registryDown

public void registryDown()

Specified by:

registryDown in interface RegistryHandler