org.apache.slider.server.appmaster.state

Class RoleHistory

java.lang.Object

org.apache.slider.server.appmaster.state.RoleHistory

public class RoleHistory
extends Object

The Role History.

Synchronization policy: all public operations are synchronized. Protected methods are in place for testing -no guarantees are made.

Inner classes have no synchronization guarantees; they should be manipulated in these classes and not externally.

Note that as well as some methods marked visible for testing, there is the option for the time generator method, now() to be overridden so that a repeatable time series can be used.

Field Summary

Fields

Modifier and Type	Field and Description
protected static org.slf4j.Logger	log

Constructor Summary

Constructors

Constructor and Description
RoleHistory(Collection<RoleStatus> roles, AbstractClusterServices recordFactory) Instantiate

Method Summary

Methods

Modifier and Type	Method and Description
void	addNewRole(RoleStatus roleStatus) Add a new role
Map<CharSequence,Integer>	buildMappingForHistoryFile() Build the mapping entry for persisting to the role history
void	buildRecentNodeLists() (After the start), rebuild the availability data structures
List<AbstractRMOperation>	cancelOutstandingAARequests() Escalate operation as triggered by external timer.
List<AbstractRMOperation>	cancelRequestsForRole(RoleStatus role, int toCancel) Cancel a number of outstanding requests for a role -that is, not actual containers, just requests for new ones.
boolean	canPlaceAANodes() Does the RoleHistory have enough information about the YARN cluster to start placing AA requests? That is: has it the node map and any label information needed?
NodeMap	cloneNodemap() Get a clone of the nodemap.
List<NodeInstance>	cloneRecentNodeList(int role) Get a clone of the available list
void	dump() Print the history to the log.
List<AbstractRMOperation>	escalateOutstandingRequests() Escalate operation as triggered by external timer.
List<NodeInstance>	findNodeForNewAAInstance(RoleStatus role) Find a node for use
NodeInstance	findRecentNodeForNewInstance(RoleStatus role) Find a node for use
int	getClusterSize() Get the total size of the cluster -the number of NodeInstances
NodeInstance	getExistingNodeInstance(org.apache.hadoop.yarn.api.records.Container container) Get the node instance of a container if there's an entry in the history
NodeInstance	getExistingNodeInstance(String hostname) Get the node instance of a host if defined
org.apache.hadoop.fs.Path	getHistoryPath() Get the path used for history files
NodeInformation	getNodeInformation(String hostname, Map<Integer,String> naming) Get the information on a node
Map<String,NodeInformation>	getNodeInformationSnapshot(Map<Integer,String> naming) Get snapshot of the node map
long	getNodesUpdatedTime() Get the last time the nodes were updated from YARN
NodeEntry	getOrCreateNodeEntry(org.apache.hadoop.yarn.api.records.Container container) Get the node entry of a container
NodeInstance	getOrCreateNodeInstance(org.apache.hadoop.yarn.api.records.Container container) Get the node instance of a container -always returns something
NodeInstance	getOrCreateNodeInstance(String hostname) Get the node instance for the specific node -creating it if needed
List<NodeInstance>	getRecentNodesForRoleId(int id) Get the nodes for an ID -may be null
int	getRoleSize()
long	getSaveTime()
long	getStartTime()
long	getThawedDataTime()
void	insert(Collection<NodeInstance> nodes) Insert a list of nodes into the map; overwrite any with that name.
boolean	isDirty()
List<NodeInstance>	listActiveNodes(int role) Get the list of active nodes ...
List<OutstandingRequest>	listOpenRequests() Get a snapshot of the outstanding placed request list
List<OutstandingRequest>	listPlacedRequests() Get a snapshot of the outstanding placed request list
ProviderRole	lookupRole(int roleId) Lookup a role by ID
protected boolean	markContainerFinished(org.apache.hadoop.yarn.api.records.Container container, boolean wasReleased, boolean shortLived, ContainerOutcome outcome) Mark a container finished; if it was released then that is treated differently.
protected long	now() Get current time.
void	onBootstrap() Handler for bootstrap event: there was no history to thaw
ContainerAllocationResults	onContainerAllocated(org.apache.hadoop.yarn.api.records.Container container, long desiredCount, long actualCount) A container has been allocated on a node -update the data structures
void	onContainerAssigned(org.apache.hadoop.yarn.api.records.Container container) A container has been assigned to a role instance on a node -update the data structures
void	onContainerReleaseSubmitted(org.apache.hadoop.yarn.api.records.Container container) A container release request was issued
void	onContainerStarted(org.apache.hadoop.yarn.api.records.Container container) Container start event
void	onContainerStartSubmitted(org.apache.hadoop.yarn.api.records.Container container, RoleInstance instance) Event: a container start has been submitted
boolean	onFailedContainer(org.apache.hadoop.yarn.api.records.Container container, boolean shortLived, ContainerOutcome outcome) App state notified of a container completed -but as it wasn't being released it is marked as failed
boolean	onNodeManagerContainerStartFailed(org.apache.hadoop.yarn.api.records.Container container) A container failed to start: update the node entry state and return the container to the queue
boolean	onNodesUpdated(List<org.apache.hadoop.yarn.api.records.NodeReport> updatedNodes) Update failedNodes and nodemap based on the node state
boolean	onReleaseCompleted(org.apache.hadoop.yarn.api.records.Container container) App state notified of a container completed
boolean	onStart(org.apache.hadoop.fs.FileSystem fs, org.apache.hadoop.fs.Path historyDir) Start up
boolean	onThaw() Handle the start process after the history has been rebuilt, and after any gc/purge
List<org.apache.hadoop.yarn.api.records.Container>	prepareAllocationList(List<org.apache.hadoop.yarn.api.records.Container> allocatedContainers) Perform any pre-allocation operations on the list of allocated containers based on knowledge of system state.
protected void	putRole(RoleStatus roleStatus) safety check: make sure the role is unique amongst the role stats...which is extended with the new role
int	rebuild(LoadedRoleHistory loadedRoleHistory) rebuild the placement history from the loaded role history
void	register(MetricsAndMonitoring metrics) Register all metrics with the metrics infra
OutstandingRequest	requestContainerForAARole(RoleStatus role) Find a node for an AA role and request an instance on that (or a location-less instance)
OutstandingRequest	requestContainerForRole(RoleStatus role) Find a node for a role and request an instance on that (or a location-less instance)
OutstandingRequest	requestInstanceOnNode(NodeInstance node, RoleStatus role, org.apache.hadoop.yarn.api.records.Resource resource) Request an instance on a given node.
protected void	reset() Reset the variables -this does not adjust the fixed attributes of the history, but the nodemap and failed node map are cleared.
void	resetFailedRecently() reset the failed recently counters
void	saved(long timestamp) Tell the history that it has been saved; marks itself as clean
org.apache.hadoop.fs.Path	saveHistory(long time) Save the history to its location using the timestamp as part of the filename.
org.apache.hadoop.fs.Path	saveHistoryIfDirty() Save the history with the current timestamp if it is dirty; return the path saved to if this is the case
void	setDirty(boolean dirty)
void	setThawedDataTime(long thawedDataTime)
void	touch() Mark ourselves as dirty

Methods inherited from class java.lang.Object

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

Field Detail

log

protected static final org.slf4j.Logger log

Constructor Detail

RoleHistory

public RoleHistory(Collection<RoleStatus> roles,
           AbstractClusterServices recordFactory)
            throws BadConfigException

Instantiate

Parameters:

roles - initial role list

recordFactory - yarn record factory

Throws:

BadConfigException

Method Detail

reset

protected void reset()
              throws BadConfigException

Reset the variables -this does not adjust the fixed attributes of the history, but the nodemap and failed node map are cleared.

Throws:

BadConfigException

register

public void register(MetricsAndMonitoring metrics)

Register all metrics with the metrics infra

Parameters:

metrics - metrics

putRole

protected void putRole(RoleStatus roleStatus)
                throws BadConfigException

safety check: make sure the role is unique amongst the role stats...which is extended with the new role

Parameters:

roleStatus - role

Throws:

ArrayIndexOutOfBoundsException

BadConfigException

addNewRole

public void addNewRole(RoleStatus roleStatus)
                throws BadConfigException

Add a new role

Parameters:

roleStatus - new role

Throws:

BadConfigException

lookupRole

public ProviderRole lookupRole(int roleId)

Lookup a role by ID

Parameters:

roleId - role Id

Returns:

role or null if not found

rebuild

public int rebuild(LoadedRoleHistory loadedRoleHistory)
            throws BadConfigException

rebuild the placement history from the loaded role history

Parameters:

loadedRoleHistory - loaded history

Returns:

the number of entries discarded

Throws:

BadConfigException - if there is a problem rebuilding the state

getStartTime

public long getStartTime()

getSaveTime

public long getSaveTime()

getThawedDataTime

public long getThawedDataTime()

setThawedDataTime

public void setThawedDataTime(long thawedDataTime)

getRoleSize

public int getRoleSize()

getClusterSize

public int getClusterSize()

Get the total size of the cluster -the number of NodeInstances

Returns:

a count

isDirty

public boolean isDirty()

setDirty

public void setDirty(boolean dirty)

saved

public void saved(long timestamp)

Tell the history that it has been saved; marks itself as clean

Parameters:

timestamp - timestamp -updates the savetime field

cloneNodemap

public NodeMap cloneNodemap()

Get a clone of the nodemap. The instances inside are not cloned

Returns:

the map

getNodeInformationSnapshot

public Map<String,NodeInformation> getNodeInformationSnapshot(Map<Integer,String> naming)

Get snapshot of the node map

Parameters:

naming - naming map of priority to enty name; entries must be unique. It's OK to be incomplete, for those the list falls back to numbers.

Returns:

a snapshot of the current node state

getNodeInformation

public NodeInformation getNodeInformation(String hostname,
                                 Map<Integer,String> naming)

Get the information on a node

Parameters:

hostname - hostname

naming - naming map of priority to enty name; entries must be unique. It's OK to be incomplete, for those the list falls back to numbers.

Returns:

the information about that host, or null if there is none

getOrCreateNodeInstance

public NodeInstance getOrCreateNodeInstance(String hostname)

Get the node instance for the specific node -creating it if needed

Parameters:

hostname - node address

Returns:

the instance

insert

public void insert(Collection<NodeInstance> nodes)

Insert a list of nodes into the map; overwrite any with that name. This is a bulk operation for testing. Important: this does not update the available node lists, these must be rebuilt afterwards.

Parameters:

nodes - collection of nodes.

now

protected long now()

Get current time. overrideable for test subclasses

Returns:

current time in millis

touch

public void touch()

Mark ourselves as dirty

resetFailedRecently

public void resetFailedRecently()

reset the failed recently counters

getHistoryPath

public org.apache.hadoop.fs.Path getHistoryPath()

Get the path used for history files

Returns:

the directory used for history files

saveHistory

public org.apache.hadoop.fs.Path saveHistory(long time)
                                      throws IOException

Save the history to its location using the timestamp as part of the filename. The saveTime and dirty fields are updated

Parameters:

time - timestamp timestamp to use as the save time

Returns:

the path saved to

Throws:

IOException - IO problems

saveHistoryIfDirty

public org.apache.hadoop.fs.Path saveHistoryIfDirty()
                                             throws IOException

Save the history with the current timestamp if it is dirty; return the path saved to if this is the case

Returns:

the path or null if the history was not saved

Throws:

IOException - failed to save for some reason

onStart

public boolean onStart(org.apache.hadoop.fs.FileSystem fs,
              org.apache.hadoop.fs.Path historyDir)
                throws BadConfigException

Start up

Parameters:

fs - filesystem

historyDir - path in FS for history

Returns:

true if the history was thawed

Throws:

BadConfigException

onBootstrap

public void onBootstrap()

Handler for bootstrap event: there was no history to thaw

onThaw

public boolean onThaw()
               throws BadConfigException

Handle the start process after the history has been rebuilt, and after any gc/purge

Throws:

BadConfigException

buildRecentNodeLists

public void buildRecentNodeLists()

(After the start), rebuild the availability data structures

getRecentNodesForRoleId

public List<NodeInstance> getRecentNodesForRoleId(int id)

Get the nodes for an ID -may be null

Parameters:

id - role ID

Returns:

potentially null list

findRecentNodeForNewInstance

public NodeInstance findRecentNodeForNewInstance(RoleStatus role)

Find a node for use

Parameters:

role - role

Returns:

the instance, or null for none

findNodeForNewAAInstance

public List<NodeInstance> findNodeForNewAAInstance(RoleStatus role)

Find a node for use

Parameters:

role - role

Returns:

the instance, or null for none

requestInstanceOnNode

public OutstandingRequest requestInstanceOnNode(NodeInstance node,
                                       RoleStatus role,
                                       org.apache.hadoop.yarn.api.records.Resource resource)

Request an instance on a given node. An outstanding request is created & tracked, with the relevant node entry for that role updated.

The role status entries will also be tracked

Returns the request that is now being tracked. If the node instance is not null, it's details about the role is incremented

Parameters:

node - node to target or null for "any"

role - role to request

Returns:

the request

requestContainerForRole

public OutstandingRequest requestContainerForRole(RoleStatus role)

Find a node for a role and request an instance on that (or a location-less instance)

Parameters:

role - role status

Returns:

a request ready to go, or null if this is an AA request and no location can be found.

requestContainerForAARole

public OutstandingRequest requestContainerForAARole(RoleStatus role)

Find a node for an AA role and request an instance on that (or a location-less instance)

Parameters:

role - role status

Returns:

a request ready to go, or null if no location can be found.

listActiveNodes

public List<NodeInstance> listActiveNodes(int role)

Get the list of active nodes ... walks the node map so is O(nodes)

Parameters:

role - role index

Returns:

a possibly empty list of nodes with an instance of that node

getOrCreateNodeEntry

public NodeEntry getOrCreateNodeEntry(org.apache.hadoop.yarn.api.records.Container container)

Get the node entry of a container

Parameters:

container - container to look up

Returns:

the entry

Throws:

RuntimeException - if the container has no hostname

getOrCreateNodeInstance

public NodeInstance getOrCreateNodeInstance(org.apache.hadoop.yarn.api.records.Container container)

Get the node instance of a container -always returns something

Parameters:

container - container to look up

Returns:

a (possibly new) node instance

Throws:

RuntimeException - if the container has no hostname

getExistingNodeInstance

public NodeInstance getExistingNodeInstance(String hostname)

Get the node instance of a host if defined

Parameters:

hostname - hostname to look up

Returns:

a node instance or null

Throws:

RuntimeException - if the container has no hostname

getExistingNodeInstance

public NodeInstance getExistingNodeInstance(org.apache.hadoop.yarn.api.records.Container container)

Get the node instance of a container if there's an entry in the history

Parameters:

container - container to look up

Returns:

a node instance or null

Throws:

RuntimeException - if the container has no hostname

prepareAllocationList

public List<org.apache.hadoop.yarn.api.records.Container> prepareAllocationList(List<org.apache.hadoop.yarn.api.records.Container> allocatedContainers)

Perform any pre-allocation operations on the list of allocated containers based on knowledge of system state. Currently this places requested hosts ahead of unrequested ones.

Parameters:

allocatedContainers - list of allocated containers

Returns:

list of containers potentially reordered

onContainerAllocated

public ContainerAllocationResults onContainerAllocated(org.apache.hadoop.yarn.api.records.Container container,
                                              long desiredCount,
                                              long actualCount)

A container has been allocated on a node -update the data structures

Parameters:

container - container

desiredCount - desired #of instances

actualCount - current count of instances

Returns:

The allocation outcome

onContainerAssigned

public void onContainerAssigned(org.apache.hadoop.yarn.api.records.Container container)

A container has been assigned to a role instance on a node -update the data structures

Parameters:

container - container

onContainerStartSubmitted

public void onContainerStartSubmitted(org.apache.hadoop.yarn.api.records.Container container,
                             RoleInstance instance)

Event: a container start has been submitted

Parameters:

container - container being started

instance - instance bound to the container

onContainerStarted

public void onContainerStarted(org.apache.hadoop.yarn.api.records.Container container)

Container start event

Parameters:

container - container that just started

onNodeManagerContainerStartFailed

public boolean onNodeManagerContainerStartFailed(org.apache.hadoop.yarn.api.records.Container container)

A container failed to start: update the node entry state and return the container to the queue

Parameters:

container - container that failed

Returns:

true if the node was queued

canPlaceAANodes

public boolean canPlaceAANodes()

Does the RoleHistory have enough information about the YARN cluster to start placing AA requests? That is: has it the node map and any label information needed?

Returns:

true if the caller can start requesting AA nodes

getNodesUpdatedTime

public long getNodesUpdatedTime()

Get the last time the nodes were updated from YARN

Returns:

the update time or zero if never updated.

onNodesUpdated

public boolean onNodesUpdated(List<org.apache.hadoop.yarn.api.records.NodeReport> updatedNodes)

Update failedNodes and nodemap based on the node state

Parameters:

updatedNodes - list of updated nodes

Returns:

true if a review should be triggered.

onContainerReleaseSubmitted

public void onContainerReleaseSubmitted(org.apache.hadoop.yarn.api.records.Container container)

A container release request was issued

Parameters:

container - container submitted

onReleaseCompleted

public boolean onReleaseCompleted(org.apache.hadoop.yarn.api.records.Container container)

App state notified of a container completed

Parameters:

container - completed container

Returns:

true if the node was queued

onFailedContainer

public boolean onFailedContainer(org.apache.hadoop.yarn.api.records.Container container,
                        boolean shortLived,
                        ContainerOutcome outcome)

App state notified of a container completed -but as it wasn't being released it is marked as failed

Parameters:

container - completed container

shortLived - was the container short lived?

outcome -

Returns:

true if the node is considered available for work

markContainerFinished

protected boolean markContainerFinished(org.apache.hadoop.yarn.api.records.Container container,
                            boolean wasReleased,
                            boolean shortLived,
                            ContainerOutcome outcome)

Mark a container finished; if it was released then that is treated differently. history is touch()-ed

Parameters:

container - completed container

wasReleased - was the container released?

shortLived - was the container short lived?

outcome -

Returns:

true if the node was queued

dump

public void dump()

Print the history to the log. This is for testing and diagnostics

buildMappingForHistoryFile

public Map<CharSequence,Integer> buildMappingForHistoryFile()

Build the mapping entry for persisting to the role history

Returns:

a mapping object

cloneRecentNodeList

public List<NodeInstance> cloneRecentNodeList(int role)

Get a clone of the available list

Parameters:

role - role index

Returns:

a clone of the list

listPlacedRequests

public List<OutstandingRequest> listPlacedRequests()

Get a snapshot of the outstanding placed request list

Returns:

a list of the requests outstanding at the time of requesting

listOpenRequests

public List<OutstandingRequest> listOpenRequests()

Get a snapshot of the outstanding placed request list

Returns:

a list of the requests outstanding at the time of requesting

escalateOutstandingRequests

public List<AbstractRMOperation> escalateOutstandingRequests()

Escalate operation as triggered by external timer.

Returns:

a (usually empty) list of cancel/request operations.

cancelOutstandingAARequests

public List<AbstractRMOperation> cancelOutstandingAARequests()

Escalate operation as triggered by external timer.

Returns:

a (usually empty) list of cancel/request operations.

cancelRequestsForRole

public List<AbstractRMOperation> cancelRequestsForRole(RoleStatus role,
                                              int toCancel)

Cancel a number of outstanding requests for a role -that is, not actual containers, just requests for new ones.

Parameters:

role - role

toCancel - number to cancel

Returns:

a list of cancellable operations.