org.apache.tinkerpop.gremlin.process.computer

Interface MapReduce<MK,MV,RK,RV,R>

All Superinterfaces:

Cloneable

public interface MapReduce<MK,MV,RK,RV,R>
extends Cloneable

A MapReduce is composed of map(), combine(), and reduce() stages. The map() stage processes the vertices of the Graph in a logically parallel manner. The combine() stage aggregates the values of a particular map emitted key prior to sending across the cluster. The reduce() stage aggregates the values of the combine/map emitted keys for the keys that hash to the current machine in the cluster. The interface presented here is nearly identical to the interface popularized by Hadoop save the map() is over the vertices of the graph.

Author:

Marko A. Rodriguez (http://markorodriguez.com)

Nested Class Summary

Nested Classes

Modifier and Type	Interface and Description
static interface	MapReduce.MapEmitter<K,V> The MapEmitter is used to emit key/value pairs from the map() stage of the MapReduce job.
static class	MapReduce.NullObject A convenience singleton when a single key is needed so that all emitted values converge to the same combiner/reducer.
static interface	MapReduce.ReduceEmitter<OK,OV> The ReduceEmitter is used to emit key/value pairs from the combine() and reduce() stages of the MapReduce job.
static class	MapReduce.Stage MapReduce is composed of three stages: map, combine, and reduce.

Field Summary

Fields

Modifier and Type	Field and Description
static String	MAP_REDUCE

Method Summary

Modifier and Type	Method and Description
default void	addResultToMemory(Memory.Admin memory, Iterator<KeyValue<RK,RV>> keyValues) The final result can be generated and added to Memory and accessible via DefaultComputerResult.
MapReduce<MK,MV,RK,RV,R>	clone() When multiple workers on a single machine need MapReduce instances, it is possible to use clone.
default void	combine(MK key, Iterator<MV> values, MapReduce.ReduceEmitter<RK,RV> emitter) The combine() method is logically executed at all "machines" in parallel.
static <M extends MapReduce> M	createMapReduce(Graph graph, org.apache.commons.configuration.Configuration configuration) A helper method to construct a MapReduce given the content of the supplied configuration.
boolean	doStage(MapReduce.Stage stage) A MapReduce job can be map-only, map-reduce-only, or map-combine-reduce.
R	generateFinalResult(Iterator<KeyValue<RK,RV>> keyValues) The key/value pairs emitted by reduce() (or map() in a map-only job) can be iterated to generate a local JVM Java object.
default Optional<Comparator<MK>>	getMapKeySort() If a Comparator is provided, then all pairs leaving the MapReduce.MapEmitter are sorted.
String	getMemoryKey() The results of the MapReduce job are associated with a memory-key to ultimately be stored in Memory.
default Optional<Comparator<RK>>	getReduceKeySort() If a Comparator is provided, then all pairs leaving the MapReduce.ReduceEmitter are sorted.
default void	loadState(Graph graph, org.apache.commons.configuration.Configuration configuration) When it is necessary to load the state of a MapReduce job, this method is called.
void	map(Vertex vertex, MapReduce.MapEmitter<MK,MV> emitter) The map() method is logically executed at all vertices in the graph in parallel.
default void	reduce(MK key, Iterator<MV> values, MapReduce.ReduceEmitter<RK,RV> emitter) The reduce() method is logically on the "machine" the respective key hashes to.
default void	storeState(org.apache.commons.configuration.Configuration configuration) When it is necessary to store the state of a MapReduce job, this method is called.
default void	workerEnd(MapReduce.Stage stage) This method is called at the end of the respective MapReduce.Stage for a particular "chunk of vertices." The set of vertices in the graph are typically not processed with full parallelism.
default void	workerStart(MapReduce.Stage stage) This method is called at the start of the respective MapReduce.Stage for a particular "chunk of vertices." The set of vertices in the graph are typically not processed with full parallelism.

Field Detail

MAP_REDUCE

static final String MAP_REDUCE

See Also:

Constant Field Values

Method Detail

storeState

default void storeState(org.apache.commons.configuration.Configuration configuration)

When it is necessary to store the state of a MapReduce job, this method is called. This is typically required when the MapReduce job needs to be serialized to another machine. Note that what is stored is simply the instance state, not any processed data.

Parameters:

configuration - the configuration to store the state of the MapReduce job in.

loadState

default void loadState(Graph graph,
                       org.apache.commons.configuration.Configuration configuration)

When it is necessary to load the state of a MapReduce job, this method is called. This is typically required when the MapReduce job needs to be serialized to another machine. Note that what is loaded is simply the instance state, not any processed data.

It is important that the state loaded from loadState() is identical to any state created from a constructor. For those GraphComputers that do not need to use Configurations to migrate state between JVMs, the constructor will only be used.

Parameters:

graph - the graph the MapReduce job will run against

configuration - the configuration to load the state of the MapReduce job from.

doStage

boolean doStage(MapReduce.Stage stage)

A MapReduce job can be map-only, map-reduce-only, or map-combine-reduce. Before executing the particular stage, this method is called to determine if the respective stage is defined. This method should return true if the respective stage as a non-default method implementation.

Parameters:

stage - the stage to check for definition.

Returns:

whether that stage should be executed.

map

void map(Vertex vertex,
         MapReduce.MapEmitter<MK,MV> emitter)

The map() method is logically executed at all vertices in the graph in parallel. The map() method emits key/value pairs given some analysis of the data in the vertices (and/or its incident edges). All MapReduce classes must at least provide an implementation of MapReduce#map(Vertex, MapEmitter).

Parameters:

vertex - the current vertex being map() processed.

emitter - the component that allows for key/value pairs to be emitted to the next stage.

combine

default void combine(MK key,
                     Iterator<MV> values,
                     MapReduce.ReduceEmitter<RK,RV> emitter)

The combine() method is logically executed at all "machines" in parallel. The combine() method pre-combines the values for a key prior to propagation over the wire. The combine() method must emit the same key/value pairs as the reduce() method. If there is a combine() implementation, there must be a reduce() implementation. If the MapReduce implementation is single machine, it can skip executing this method as reduce() is sufficient.

Parameters:

key - the key that has aggregated values

values - the aggregated values associated with the key

emitter - the component that allows for key/value pairs to be emitted to the reduce stage.

reduce

default void reduce(MK key,
                    Iterator<MV> values,
                    MapReduce.ReduceEmitter<RK,RV> emitter)

The reduce() method is logically on the "machine" the respective key hashes to. The reduce() method combines all the values associated with the key and emits key/value pairs.

Parameters:

key - the key that has aggregated values

values - the aggregated values associated with the key

emitter - the component that allows for key/value pairs to be emitted as the final result.

workerStart

default void workerStart(MapReduce.Stage stage)

This method is called at the start of the respective MapReduce.Stage for a particular "chunk of vertices." The set of vertices in the graph are typically not processed with full parallelism. The vertex set is split into subsets and a worker is assigned to call the MapReduce methods on it method. The default implementation is a no-op.

Parameters:

stage - the stage of the MapReduce computation

workerEnd

default void workerEnd(MapReduce.Stage stage)

This method is called at the end of the respective MapReduce.Stage for a particular "chunk of vertices." The set of vertices in the graph are typically not processed with full parallelism. The vertex set is split into subsets and a worker is assigned to call the MapReduce methods on it method. The default implementation is a no-op.

Parameters:

stage - the stage of the MapReduce computation

getMapKeySort

default Optional<Comparator<MK>> getMapKeySort()

If a Comparator is provided, then all pairs leaving the MapReduce.MapEmitter are sorted. The sorted results are either fed sorted to the combine/reduce-stage or as the final output. If sorting is not required, then Optional.empty() should be returned as sorting is computationally expensive. The default implementation returns Optional.empty().

Returns:

an Optional of a comparator for sorting the map output.

getReduceKeySort

default Optional<Comparator<RK>> getReduceKeySort()

If a Comparator is provided, then all pairs leaving the MapReduce.ReduceEmitter are sorted. If sorting is not required, then Optional.empty() should be returned as sorting is computationally expensive. The default implementation returns Optional.empty().

Returns:

an Optional of a comparator for sorting the reduce output.

generateFinalResult

R generateFinalResult(Iterator<KeyValue<RK,RV>> keyValues)

The key/value pairs emitted by reduce() (or map() in a map-only job) can be iterated to generate a local JVM Java object.

Parameters:

keyValues - the key/value pairs that were emitted from reduce() (or map() in a map-only job)

Returns:

the resultant object formed from the emitted key/values.

getMemoryKey

String getMemoryKey()

The results of the MapReduce job are associated with a memory-key to ultimately be stored in Memory.

Returns:

the memory key of the generated result object.

addResultToMemory

default void addResultToMemory(Memory.Admin memory,
                               Iterator<KeyValue<RK,RV>> keyValues)

The final result can be generated and added to Memory and accessible via DefaultComputerResult. The default simply takes the object from generateFinalResult() and adds it to the Memory given getMemoryKey().

Parameters:

memory - the memory of the GraphComputer

keyValues - the key/value pairs emitted from reduce() (or map() in a map only job).

clone

MapReduce<MK,MV,RK,RV,R> clone()

When multiple workers on a single machine need MapReduce instances, it is possible to use clone. This will provide a speedier way of generating instances, over the storeState(org.apache.commons.configuration.Configuration) and loadState(org.apache.tinkerpop.gremlin.structure.Graph, org.apache.commons.configuration.Configuration) model. The default implementation simply returns the object as it assumes that the MapReduce instance is a stateless singleton.

Returns:

A clone of the MapReduce object

createMapReduce

static <M extends MapReduce> M createMapReduce(Graph graph,
                                               org.apache.commons.configuration.Configuration configuration)

A helper method to construct a MapReduce given the content of the supplied configuration. The class of the MapReduce is read from the MAP_REDUCE static configuration key. Once the MapReduce is constructed, loadState(org.apache.tinkerpop.gremlin.structure.Graph, org.apache.commons.configuration.Configuration) method is called with the provided configuration.

Parameters:

graph - The graph that the MapReduce job will run against

configuration - A configuration with requisite information to build a MapReduce

Returns:

the newly constructed MapReduce