org.apache.hadoop.hbase.master

Class ClusterStatusPublisher

java.lang.Object

org.apache.hadoop.hbase.ScheduledChore

org.apache.hadoop.hbase.master.ClusterStatusPublisher

All Implemented Interfaces:

Runnable

@InterfaceAudience.Private
public class ClusterStatusPublisher
extends ScheduledChore

Class to publish the cluster status to the client. This allows them to know immediately the dead region servers, hence to cut the connection they have with them, eventually stop waiting on the socket. This improves the mean time to recover, and as well allows to increase on the client the different timeouts, as the dead servers will be detected separately.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ClusterStatusPublisher.MulticastPublisher
static interface	ClusterStatusPublisher.Publisher

Field Summary

Fields

Modifier and Type	Field and Description
private boolean	connected
static int	DEFAULT_STATUS_PUBLISH_PERIOD
static Class<? extends ClusterStatusPublisher.Publisher>	DEFAULT_STATUS_PUBLISHER_CLASS
private long	lastMessageTime
private ConcurrentMap<ServerName,Integer>	lastSent
private HMaster	master
static int	MAX_SERVER_PER_MESSAGE We want to limit the size of the protobuf message sent, do fit into a single packet.
private int	messagePeriod
static int	NB_SEND If a server dies, we're sending the information multiple times in case a receiver misses the message.
private ClusterStatusPublisher.Publisher	publisher
static String	STATUS_PUBLISH_PERIOD The minimum time between two status messages, in milliseconds.
static String	STATUS_PUBLISHER_CLASS The implementation class used to publish the status.

Constructor Summary

Constructors

Modifier	Constructor and Description
protected	ClusterStatusPublisher()
	ClusterStatusPublisher(HMaster master, org.apache.hadoop.conf.Configuration conf, Class<? extends ClusterStatusPublisher.Publisher> publisherClass)

Method Summary

Modifier and Type	Method and Description
protected void	chore() The task to execute on each scheduled execution of the Chore
protected void	cleanup() Override to run cleanup tasks when the Chore encounters an error and must stop running
protected List<ServerName>	generateDeadServersListToSend() Create the dead server to send.
protected List<Pair<ServerName,Long>>	getDeadServers(long since) Get the servers which died since a given timestamp.
private boolean	isConnected()

Methods inherited from class org.apache.hadoop.hbase.ScheduledChore

cancel, cancel, choreForTesting, getInitialDelay, getName, getPeriod, getStopper, getTimeUnit, initialChore, isInitialChoreComplete, isScheduled, run, toString, triggerNow

Methods inherited from class java.lang.Object

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

Field Detail

STATUS_PUBLISHER_CLASS

public static final String STATUS_PUBLISHER_CLASS

The implementation class used to publish the status. Default is null (no publish). Use org.apache.hadoop.hbase.master.ClusterStatusPublisher.MulticastPublisher to multicast the status.

See Also:

Constant Field Values

DEFAULT_STATUS_PUBLISHER_CLASS

public static final Class<? extends ClusterStatusPublisher.Publisher> DEFAULT_STATUS_PUBLISHER_CLASS

STATUS_PUBLISH_PERIOD

public static final String STATUS_PUBLISH_PERIOD

The minimum time between two status messages, in milliseconds.

See Also:

Constant Field Values

DEFAULT_STATUS_PUBLISH_PERIOD

public static final int DEFAULT_STATUS_PUBLISH_PERIOD

See Also:

Constant Field Values

lastMessageTime

private long lastMessageTime

master

private final HMaster master

messagePeriod

private final int messagePeriod

lastSent

private final ConcurrentMap<ServerName,Integer> lastSent

publisher

private ClusterStatusPublisher.Publisher publisher

connected

private boolean connected

MAX_SERVER_PER_MESSAGE

public static final int MAX_SERVER_PER_MESSAGE

We want to limit the size of the protobuf message sent, do fit into a single packet. a reasonable size for ip / ethernet is less than 1Kb.

See Also:

Constant Field Values

NB_SEND

public static final int NB_SEND

If a server dies, we're sending the information multiple times in case a receiver misses the message.

See Also:

Constant Field Values

Constructor Detail

ClusterStatusPublisher

public ClusterStatusPublisher(HMaster master,
                              org.apache.hadoop.conf.Configuration conf,
                              Class<? extends ClusterStatusPublisher.Publisher> publisherClass)
                       throws IOException

Throws:

IOException

ClusterStatusPublisher

protected ClusterStatusPublisher()

Method Detail

chore

protected void chore()

Description copied from class: ScheduledChore

The task to execute on each scheduled execution of the Chore

Specified by:

chore in class ScheduledChore

cleanup

protected void cleanup()

Description copied from class: ScheduledChore

Override to run cleanup tasks when the Chore encounters an error and must stop running

Overrides:

cleanup in class ScheduledChore

isConnected

private boolean isConnected()

generateDeadServersListToSend

protected List<ServerName> generateDeadServersListToSend()

Create the dead server to send. A dead server is sent NB_SEND times. We send at max MAX_SERVER_PER_MESSAGE at a time. if there are too many dead servers, we send the newly dead first.

getDeadServers

protected List<Pair<ServerName,Long>> getDeadServers(long since)

Get the servers which died since a given timestamp. protected because it can be subclassed by the tests.