////
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements.  See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License.  You may obtain a copy of the License at

  http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
////
image::apache-tinkerpop-logo.png[width=500,link="https://tinkerpop.apache.org"]

*x.y.z*

:toc-position: left

= Provider Documentation

Apache TinkerPop™ exposes a set of interfaces, protocols, and tests that make it possible for third-parties to build
libraries and systems that plug-in to the TinkerPop stack.  TinkerPop refers to those third-parties as "providers" and
this documentation is designed to help providers understand what is involved in developing code on these lower levels
of the TinkerPop API.

This document attempts to address the needs of the different providers that have been identified:

* Graph System Provider
** Graph Database Provider
** Graph Processor Provider
* Graph Driver Provider
* Graph Language Provider
* Graph Plugin Provider

[[graph-system-provider-requirements]]
== Graph System Provider Requirements

image:tinkerpop-enabled.png[width=140,float=left] At the core of TinkerPop 3.x is a Java8 API. The implementation of this
core API and its validation via the `gremlin-test` suite is all that is required of a graph system provider wishing to
provide a TinkerPop-enabled graph engine. Once a graph system has a valid implementation, then all the applications
provided by TinkerPop (e.g. Gremlin Console, Gremlin Server, etc.) and 3rd-party developers (e.g. Gremlin-Scala,
Gremlin-JS, etc.) will integrate properly. Finally, please feel free to use the logo on the left to promote your
TinkerPop implementation.

[[graph-structure-api]]
=== Graph Structure API

The graph structure API of TinkerPop provides the interfaces necessary to create a TinkerPop enabled system and
exposes the basic components of a property graph to include `Graph`, `Vertex`, `Edge`, `VertexProperty` and `Property`.
The structure API can be used directly as follows:

[source,java]
----
Graph graph = TinkerGraph.open(); <1>
Vertex marko = graph.addVertex(T.label, "person", T.id, 1, "name", "marko", "age", 29); <2>
Vertex vadas = graph.addVertex(T.label, "person", T.id, 2, "name", "vadas", "age", 27);
Vertex lop = graph.addVertex(T.label, "software", T.id, 3, "name", "lop", "lang", "java");
Vertex josh = graph.addVertex(T.label, "person", T.id, 4, "name", "josh", "age", 32);
Vertex ripple = graph.addVertex(T.label, "software", T.id, 5, "name", "ripple", "lang", "java");
Vertex peter = graph.addVertex(T.label, "person", T.id, 6, "name", "peter", "age", 35);
marko.addEdge("knows", vadas, T.id, 7, "weight", 0.5f); <3>
marko.addEdge("knows", josh, T.id, 8, "weight", 1.0f);
marko.addEdge("created", lop, T.id, 9, "weight", 0.4f);
josh.addEdge("created", ripple, T.id, 10, "weight", 1.0f);
josh.addEdge("created", lop, T.id, 11, "weight", 0.4f);
peter.addEdge("created", lop, T.id, 12, "weight", 0.2f);
----

<1> Create a new in-memory `TinkerGraph` and assign it to the variable `graph`.
<2> Create a vertex along with a set of key/value pairs with `T.label` being the vertex label and `T.id` being the vertex id.
<3> Create an edge along with a  set of key/value pairs with the edge label being specified as the first argument.

In the above code all the vertices are created first and then their respective edges. There are two "accessor tokens":
`T.id` and `T.label`. When any of these, along with a set of other key value pairs is provided to
`Graph.addVertex(Object...)` or `Vertex.addEdge(String,Vertex,Object...)`, the respective element is created along
with the provided key/value pair properties appended to it.

Below is a sequence of basic graph mutation operations represented in Java 8.

image:basic-mutation.png[width=240,float=right]
[source,java]
----
// create a new graph
Graph graph = TinkerGraph.open();
// add a software vertex with a name property
Vertex gremlin = graph.addVertex(T.label, "software",
                             "name", "gremlin"); <1>
// only one vertex should exist
assert(IteratorUtils.count(graph.vertices()) == 1)
// no edges should exist as none have been created
assert(IteratorUtils.count(graph.edges()) == 0)
// add a new property
gremlin.property("created",2009) <2>
// add a new software vertex to the graph
Vertex blueprints = graph.addVertex(T.label, "software",
                                "name", "blueprints"); <3>
// connect gremlin to blueprints via a dependsOn-edge
gremlin.addEdge("dependsOn",blueprints); <4>
// now there are two vertices and one edge
assert(IteratorUtils.count(graph.vertices()) == 2)
assert(IteratorUtils.count(graph.edges()) == 1)
// add a property to blueprints
blueprints.property("created",2010) <5>
// remove that property
blueprints.property("created").remove() <6>
// connect gremlin to blueprints via encapsulates
gremlin.addEdge("encapsulates",blueprints) <7>
assert(IteratorUtils.count(graph.vertices()) == 2)
assert(IteratorUtils.count(graph.edges()) == 2)
// removing a vertex removes all its incident edges as well
blueprints.remove() <8>
gremlin.remove() <9>
// the graph is now empty
assert(IteratorUtils.count(graph.vertices()) == 0)
assert(IteratorUtils.count(graph.edges()) == 0)
// tada!
----

The above code samples are just examples of how the structure API can be used to access a graph. Those APIs are then
used internally by the process API (i.e. Gremlin) to access any graph that implements those structure API interfaces
to execute queries. Typically, the structure API methods are not used directly by end-users.

=== Implementing Gremlin-Core

The classes that a graph system provider should focus on implementing are itemized below. It is a good idea to study
the link:https://tinkerpop.apache.org/docs/x.y.z/reference/#tinkergraph-gremlin[TinkerGraph] (in-memory OLTP and OLAP
in `tinkergraph-gremlin`), link:https://tinkerpop.apache.org/docs/x.y.z/reference/#neo4j-gremlin[Neo4jGraph]
(OLTP w/ transactions in `neo4j-gremlin`) and/or
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#hadoop-gremlin[HadoopGraph] (OLAP in `hadoop-gremlin`)
implementations for ideas and patterns.

. Online Transactional Processing Graph Systems (*OLTP*)
 .. Structure API: `Graph`, `Element`, `Vertex`, `Edge`, `Property` and `Transaction` (if transactions are supported).
 .. Process API: `TraversalStrategy` instances for optimizing Gremlin traversals to the provider's graph system (i.e. `TinkerGraphStepStrategy`).
. Online Analytics Processing Graph Systems (*OLAP*)
 .. Everything required of OLTP is required of OLAP (but not vice versa).
 .. GraphComputer API: `GraphComputer`, `Messenger`, `Memory`.

Please consider the following implementation notes:

* Use `StringHelper` to ensuring that the `toString()` representation of classes are consistent with other
implementations.
* Ensure that your implementation's `Features` (Graph, Vertex, etc.) are correct so that test cases handle particulars
accordingly.
* Use the numerous static method helper classes such as `ElementHelper`, `GraphComputerHelper`, `VertexProgramHelper`, etc.
* There are a number of default methods on the provided interfaces that are semantically correct. However, if they are
not efficient for the implementation, override them.
* Implement the `structure/` package interfaces first and then, if desired, interfaces in the `process/` package
interfaces.
* `ComputerGraph` is a `Wrapper` system that ensure proper semantics during a GraphComputer computation.
* The link:https://tinkerpop.apache.org/javadocs/x.y.z/core/[javadoc] is often a good resource in understanding
expectations from both the user's perspective as well as the graph provider's perspective. Also consider examining
the javadoc of TinkerGraph which is often well annotated and the interfaces and classes of the test suite itself.

[[oltp-implementations]]
==== OLTP Implementations

image:pipes-character-1.png[width=110,float=right] The most important interfaces to implement are in the `structure/`
package. These include interfaces like `Graph`, `Vertex`, `Edge`, `Property`, `Transaction`, etc. The
`StructureStandardSuite` will ensure that the semantics of the methods implemented are correct. Moreover, there are
numerous `Exceptions` classes with static exceptions that should be thrown by the graph system so that all the
exceptions and their messages are consistent amongst all TinkerPop implementations.

The following bullets provide some tips to consider when implementing the structure interfaces:

* `Graph`
** Be sure the `Graph` implementation is named as `XXXGraph` (e.g. TinkerGraph, Neo4jGraph, HadoopGraph, etc.).
** This implementation needs to be `GraphFactory` compatible which means that the implementation should have a static
`Graph open(Configuration)` method where the `Configuration` is an Apache Commons class of that name. Alternatively, the
`Graph` implementation can have the `GraphFactoryClass` annotation which specifies a class with that static
`Graph open(Configuration)` method.
* `VertexProperty`
** This interface is both a `Property` and an `Element` as `VertexProperty` is a first-class graph element in that it
can have its own properties (i.e. meta-properties). Even if the implementation does not intend to support
meta-properties, the `VertexProperty` needs to be implemented as an `Element`.

[[olap-implementations]]
==== OLAP Implementations

image:furnace-character-1.png[width=110,float=right] Implementing the OLAP interfaces may be a bit more complicated.
Note that before OLAP interfaces are implemented, it is necessary for the OLTP interfaces to be, at minimal,
implemented as specified in <<oltp-implementations,OLTP Implementations>>. A summary of each required interface
implementation is presented below:

. `GraphComputer`: A fluent builder for specifying an isolation level, a VertexProgram, and any number of MapReduce jobs to be submitted.
. `Memory`: A global blackboard for ANDing, ORing, INCRing, and SETing values for specified keys.
. `Messenger`: The system that collects and distributes messages being propagated by vertices executing the VertexProgram application.
. `MapReduce.MapEmitter`: The system that collects key/value pairs being emitted by the MapReduce applications map-phase.
. `MapReduce.ReduceEmitter`: The system that collects key/value pairs being emitted by the MapReduce applications combine- and reduce-phases.

NOTE: The VertexProgram and MapReduce interfaces in the `process/computer/` package are not required by the graph
system. Instead, these are interfaces to be implemented by application developers writing VertexPrograms and MapReduce jobs.

IMPORTANT: TinkerPop provides two OLAP implementations:
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#tinkergraph-gremlin[TinkerGraphComputer] (TinkerGraph),
and
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#sparkgraphcomputer[SparkGraphComputer] (Hadoop).
Given the complexity of the OLAP system, it is good to study and copy many of the patterns used in these reference
implementations.

===== Implementing GraphComputer

image:furnace-character-3.png[width=150,float=right] The most complex method in GraphComputer is the `submit()`-method. The method must do the following:

. Ensure the GraphComputer has not already been executed.
. Ensure that at least there is a VertexProgram or 1 MapReduce job.
. If there is a VertexProgram, validate that it can execute on the GraphComputer given the respectively defined features.
. Create the Memory to be used for the computation.
. Execute the VertexProgram.setup() method once and only once.
. Execute the VertexProgram.execute() method for each vertex.
. Execute the VertexProgram.terminate() method once and if true, repeat VertexProgram.execute().
. When VertexProgram.terminate() returns true, move to MapReduce job execution.
. MapReduce jobs are not required to be executed in any specified order.
. For each Vertex, execute MapReduce.map(). Then (if defined) execute MapReduce.combine() and MapReduce.reduce().
. Update Memory with runtime information.
. Construct a new `ComputerResult` containing the compute Graph and Memory.

===== Implementing Memory

image:gremlin-brain.png[width=175,float=left] The Memory object is initially defined by `VertexProgram.setup()`.
The memory data is available in the first round of the `VertexProgram.execute()` method. Each Vertex, when executing
the VertexProgram, can update the Memory in its round. However, the update is not seen by the other vertices until
the next round. At the end of the first round, all the updates are aggregated and the new memory data is available
on the second round. This process repeats until the VertexProgram terminates.

===== Implementing Messenger

The Messenger object is similar to the Memory object in that a vertex can read and write to the Messenger. However,
the data it reads are the messages sent to the vertex in the previous step and the data it writes are the messages
that will be readable by the receiving vertices in the subsequent round.

===== Implementing MapReduce Emitters

image:hadoop-logo-notext.png[width=150,float=left] The MapReduce framework in TinkerPop is similar to the model
popularized by link:http://hadoop.apache.org[Hadoop]. The primary difference is that all Mappers process the vertices
of the graph, not an arbitrary key/value pair. However, the vertices' edges can not be accessed -- only their
properties. This greatly reduces the amount of data needed to be pushed through the MapReduce engine as any edge
information required, can be computed in the VertexProgram.execute() method. Moreover, at this stage, vertices can
not be mutated, only their token and property data read. A Gremlin OLAP system needs to provide implementations for
to particular classes: `MapReduce.MapEmitter` and `MapReduce.ReduceEmitter`. TinkerGraph's implementation is provided
below which demonstrates the simplicity of the algorithm (especially when the data is all within the same JVM).

[source,java]
----
public class TinkerMapEmitter<K, V> implements MapReduce.MapEmitter<K, V> {

    public Map<K, Queue<V>> reduceMap;
    public Queue<KeyValue<K, V>> mapQueue;
    private final boolean doReduce;

    public TinkerMapEmitter(final boolean doReduce) { <1>
        this.doReduce = doReduce;
        if (this.doReduce)
            this.reduceMap = new ConcurrentHashMap<>();
        else
            this.mapQueue = new ConcurrentLinkedQueue<>();
    }

    @Override
    public void emit(K key, V value) {
        if (this.doReduce)
            this.reduceMap.computeIfAbsent(key, k -> new ConcurrentLinkedQueue<>()).add(value); <2>
        else
            this.mapQueue.add(new KeyValue<>(key, value)); <3>
    }

    protected void complete(final MapReduce<K, V, ?, ?, ?> mapReduce) {
        if (!this.doReduce && mapReduce.getMapKeySort().isPresent()) { <4>
            final Comparator<K> comparator = mapReduce.getMapKeySort().get();
            final List<KeyValue<K, V>> list = new ArrayList<>(this.mapQueue);
            Collections.sort(list, Comparator.comparing(KeyValue::getKey, comparator));
            this.mapQueue.clear();
            this.mapQueue.addAll(list);
        } else if (mapReduce.getMapKeySort().isPresent()) {
            final Comparator<K> comparator = mapReduce.getMapKeySort().get();
            final List<Map.Entry<K, Queue<V>>> list = new ArrayList<>();
            list.addAll(this.reduceMap.entrySet());
            Collections.sort(list, Comparator.comparing(Map.Entry::getKey, comparator));
            this.reduceMap = new LinkedHashMap<>();
            list.forEach(entry -> this.reduceMap.put(entry.getKey(), entry.getValue()));
        }
    }
}
----

<1> If the MapReduce job has a reduce, then use one data structure (`reduceMap`), else use another (`mapList`). The
difference being that a reduction requires a grouping by key and therefore, the `Map<K,Queue<V>>` definition. If no
reduction/grouping is required, then a simple `Queue<KeyValue<K,V>>` can be leveraged.
<2> If reduce is to follow, then increment the Map with a new value for the key. `MapHelper` is a TinkerPop class
with static methods for adding data to a Map.
<3> If no reduce is to follow, then simply append a KeyValue to the queue.
<4> When the map phase is complete, any map-result sorting required can be executed at this point.

[source,java]
----
public class TinkerReduceEmitter<OK, OV> implements MapReduce.ReduceEmitter<OK, OV> {

    protected Queue<KeyValue<OK, OV>> reduceQueue = new ConcurrentLinkedQueue<>();

    @Override
    public void emit(final OK key, final OV value) {
        this.reduceQueue.add(new KeyValue<>(key, value));
    }

    protected void complete(final MapReduce<?, ?, OK, OV, ?> mapReduce) {
        if (mapReduce.getReduceKeySort().isPresent()) {
            final Comparator<OK> comparator = mapReduce.getReduceKeySort().get();
            final List<KeyValue<OK, OV>> list = new ArrayList<>(this.reduceQueue);
            Collections.sort(list, Comparator.comparing(KeyValue::getKey, comparator));
            this.reduceQueue.clear();
            this.reduceQueue.addAll(list);
        }
    }
}
----

The method `MapReduce.reduce()` is defined as:

[source,java]
public void reduce(final OK key, final Iterator<OV> values, final ReduceEmitter<OK, OV> emitter) { ... }

In other words, for the TinkerGraph implementation, iterate through the entrySet of the `reduceMap` and call the
`reduce()` method on each entry. The `reduce()` method can emit key/value pairs which are simply aggregated into a
`Queue<KeyValue<OK,OV>>` in an analogous fashion to `TinkerMapEmitter` when no reduce is to follow. These two emitters
are tied together in `TinkerGraphComputer.submit()`.

[source,java]
----
...
for (final MapReduce mapReduce : mapReducers) {
    if (mapReduce.doStage(MapReduce.Stage.MAP)) {
        final TinkerMapEmitter<?, ?> mapEmitter = new TinkerMapEmitter<>(mapReduce.doStage(MapReduce.Stage.REDUCE));
        final SynchronizedIterator<Vertex> vertices = new SynchronizedIterator<>(this.graph.vertices());
        workers.setMapReduce(mapReduce);
        workers.mapReduceWorkerStart(MapReduce.Stage.MAP);
        workers.executeMapReduce(workerMapReduce -> {
            while (true) {
                final Vertex vertex = vertices.next();
                if (null == vertex) return;
                workerMapReduce.map(ComputerGraph.mapReduce(vertex), mapEmitter);
            }
        });
        workers.mapReduceWorkerEnd(MapReduce.Stage.MAP);

        // sort results if a map output sort is defined
        mapEmitter.complete(mapReduce);

        // no need to run combiners as this is single machine
        if (mapReduce.doStage(MapReduce.Stage.REDUCE)) {
            final TinkerReduceEmitter<?, ?> reduceEmitter = new TinkerReduceEmitter<>();
            final SynchronizedIterator<Map.Entry<?, Queue<?>>> keyValues = new SynchronizedIterator((Iterator) mapEmitter.reduceMap.entrySet().iterator());
            workers.mapReduceWorkerStart(MapReduce.Stage.REDUCE);
            workers.executeMapReduce(workerMapReduce -> {
                while (true) {
                    final Map.Entry<?, Queue<?>> entry = keyValues.next();
                    if (null == entry) return;
                        workerMapReduce.reduce(entry.getKey(), entry.getValue().iterator(), reduceEmitter);
                    }
                });
            workers.mapReduceWorkerEnd(MapReduce.Stage.REDUCE);
            reduceEmitter.complete(mapReduce); // sort results if a reduce output sort is defined
            mapReduce.addResultToMemory(this.memory, reduceEmitter.reduceQueue.iterator()); <1>
        } else {
            mapReduce.addResultToMemory(this.memory, mapEmitter.mapQueue.iterator()); <2>
        }
    }
}
...
----

<1> Note that the final results of the reducer are provided to the Memory as specified by the application developer's
`MapReduce.addResultToMemory()` implementation.
<2> If there is no reduce stage, the map-stage results are inserted into Memory as specified by the application
developer's `MapReduce.addResultToMemory()` implementation.

==== Hadoop-Gremlin Usage

Hadoop-Gremlin is centered around `InputFormats` and `OutputFormats`. If a 3rd-party graph system provider wishes to
leverage Hadoop-Gremlin (and its respective `GraphComputer` engines), then they need to provide, at minimum, a
Hadoop2 `InputFormat<NullWritable,VertexWritable>` for their graph system. If the provider wishes to persist computed
results back to their graph system (and not just to HDFS via a `FileOutputFormat`), then a graph system specific
`OutputFormat<NullWritable,VertexWritable>` must be developed as well.

Conceptually, `HadoopGraph` is a wrapper around a `Configuration` object. There is no "data" in the `HadoopGraph` as
the `InputFormat` specifies where and how to get the graph data at OLAP (and OLTP) runtime. Thus, `HadoopGraph` is a
small object with little overhead. Graph system providers should realize `HadoopGraph` as the gateway to the OLAP
features offered by Hadoop-Gremlin. For example, a graph system specific `Graph.compute(Class<? extends GraphComputer>
graphComputerClass)`-method may look as follows:

[source,java]
----
public <C extends GraphComputer> C compute(final Class<C> graphComputerClass) throws IllegalArgumentException {
  try {
    if (AbstractHadoopGraphComputer.class.isAssignableFrom(graphComputerClass))
      return graphComputerClass.getConstructor(HadoopGraph.class).newInstance(this);
    else
      throw Graph.Exceptions.graphDoesNotSupportProvidedGraphComputer(graphComputerClass);
  } catch (final Exception e) {
    throw new IllegalArgumentException(e.getMessage(),e);
  }
}
----

Note that the configurations for Hadoop are assumed to be in the `Graph.configuration()` object. If this is not the
case, then the `Configuration` provided to `HadoopGraph.open()` should be dynamically created within the
`compute()`-method. It is in the provided configuration that `HadoopGraph` gets the various properties which
determine how to read and write data to and from Hadoop. For instance, `gremlin.hadoop.graphReader` and
`gremlin.hadoop.graphWriter`.

===== GraphFilterAware Interface

link:https://tinkerpop.apache.org/docs/x.y.z/reference/#graph-filter[Graph filters] by OLAP processors to only pull a subgraph of the full graph from the graph data source. For instance, the
example below constructs a `GraphFilter` that will only pull the "knows"-graph amongst people into the `GraphComputer`
for processing.

[source,java]
----
graph.compute().vertices(hasLabel("person")).edges(bothE("knows"))
----

If the provider has a custom `InputRDD`, they can implement `GraphFilterAware` and that graph filter will be provided to their
`InputRDD` at load time. For providers that use an `InputFormat`, state but the graph filter can be accessed from the configuration
as such:

[source,java]
----
if (configuration.containsKey(Constants.GREMLIN_HADOOP_GRAPH_FILTER))
  this.graphFilter = VertexProgramHelper.deserialize(configuration, Constants.GREMLIN_HADOOP_GRAPH_FILTER);
----

===== PersistResultGraphAware Interface

A graph system provider's `OutputFormat` should implement the `PersistResultGraphAware` interface which
determines which persistence options are available to the user. For the standard file-based `OutputFormats` provided
by Hadoop-Gremlin (e.g. link:++https://tinkerpop.apache.org/docs/x.y.z/reference/#gryo-io-format++[`GryoOutputFormat`], link:++https://tinkerpop.apache.org/docs/x.y.z/reference/#graphson-io-format++[`GraphSONOutputFormat`],
and link:++https://tinkerpop.apache.org/docs/x.y.z/reference/#script-io-format++[`ScriptInputOutputFormat`]) `ResultGraph.ORIGINAL` is not supported as the original graph
data files are not random access and are, in essence, immutable. Thus, these file-based `OutputFormats` only support
`ResultGraph.NEW` which creates a copy of the data specified by the `Persist` enum.

[[io-implementations]]
==== IO Implementations

If a `Graph` requires custom serializers for IO to work properly, implement the `Graph.io` method.  A typical example
of where a `Graph` would require such a custom serializers is if their identifier system uses non-primitive values,
such as OrientDB's `Rid` class.  From basic serialization of a single `Vertex` all the way up the stack to Gremlin
Server, the need to know how to handle these complex identifiers is an important requirement.

The first step to implementing custom serializers is to first implement the `IoRegistry` interface and register the
custom classes and serializers to it. Each `Io` implementation has different requirements for what it expects from the
`IoRegistry`:

* *GraphML* - No custom serializers expected/allowed.
* *GraphSON* - Register a Jackson `SimpleModule`.  The `SimpleModule` encapsulates specific classes to be serialized,
so it does not need to be registered to a specific class in the `IoRegistry` (use `null`).
* *Gryo* - Expects registration of one of three objects:
** Register just the custom class with a `null` Kryo `Serializer` implementation - this class will use default "field-level" Kryo serialization.
** Register the custom class with a specific Kryo `Serializer' implementation.
** Register the custom class with a `Function<Kryo, Serializer>` for those cases where the Kryo `Serializer` requires the `Kryo` instance to get constructed.

This implementation should provide a zero-arg constructor as the stack may require instantiation via reflection.
Consider extending `AbstractIoRegistry` for convenience as follows:

[source,java]
----
public class MyGraphIoRegistry extends AbstractIoRegistry {
    public MyGraphIoRegistry() {
        register(GraphSONIo.class, null, new MyGraphSimpleModule());
        register(GryoIo.class, MyGraphIdClass.class, new MyGraphIdSerializer());
    }
}
----

In the `Graph.io` method, provide the `IoRegistry` object to the supplied `Builder` and call the `create` method to
return that `Io` instance as follows:

[source,java]
----
public <I extends Io> I io(final Io.Builder<I> builder) {
    return (I) builder.graph(this).registry(myGraphIoRegistry).create();
}}
----

In this way, `Graph` implementations can pre-configure custom serializers for IO interactions and users will not need
to know about those details. Following this pattern will ensure proper execution of the test suite as well as
simplified usage for end-users.

IMPORTANT: Proper implementation of IO is critical to successful `Graph` operations in Gremlin Server.  The Test Suite
does have "serialization" tests that provide some assurance that an implementation is working properly, but those
tests cannot make assertions against any specifics of a custom serializer.  It is the responsibility of the
implementer to test the specifics of their custom serializers.

TIP: Consider separating serializer code into its own module, if possible, so that clients that use the `Graph`
implementation remotely don't need a full dependency on the entire `Graph` - just the IO components and related
classes being serialized.

There is an important implication to consider when the addition of a custom serializer. Presumably, the custom
serializer was written for the JVM to be deployed with a `Graph` instance. For example, a graph may expose a
geographical type like a `Point` or something similar. The library that contains `Point` assuming users expected to
deserialize back to a `Point` would need to have the library with `Point` and the "`PointSerializer`" class available
to them. In cases where that deployment approach is not desirable, it is possible to coerce a class like `Point` to
a type that is already in the list of types supported in TinkerPop. For example, `Point` could be coerced one-way to
`Map` of keys "x" and "y". Of course, on the client side, users would have to construct a `Map` for a `Point` which
isn't quite as user-friendly.

If doing a type coercion is not desired, then it is important to remember that writing a `Point` class and related
serializer in Java is not sufficient for full support of Gremlin, as users of non-JVM Gremlin Language Variants (GLV)
will not be able to consume them. Getting full support would mean writing similar classes for each GLV. While
developing  those classes is not hard, it also means more code to support.

===== Supporting Gremlin-Python IO

The serialization system of Gremlin-Python provides ways to add new types by creating serializers and deserializers in
Python and registering them with the `RemoteConnection`.

[source,python]
----
class MyType(object):
  GRAPHSON_PREFIX = "providerx"
  GRAPHSON_BASE_TYPE = "MyType"
  GRAPHSON_TYPE = GraphSONUtil.formatType(GRAPHSON_PREFIX, GRAPHSON_BASE_TYPE)

  def __init__(self, x, y):
    self.x = x
    self.y = y

  @classmethod
  def objectify(cls, value, reader):
    return cls(value['x'], value['y'])

  @classmethod
  def dictify(cls, value, writer):
    return GraphSONUtil.typedValue(cls.GRAPHSON_BASE_TYPE,
                                  {'x': value.x, 'y': value.y},
                                  cls.GRAPHSON_PREFIX)

graphson_reader = GraphSONReader({MyType.GRAPHSON_TYPE: MyType})
graphson_writer = GraphSONWriter({MyType: MyType})

connection = DriverRemoteConnection('ws://localhost:8182/gremlin', 'g',
                                     graphson_reader=graphson_reader,
                                     graphson_writer=graphson_writer)
----

===== Supporting Gremlin.Net IO

The serialization system of Gremlin.Net provides ways to add new types by creating serializers and deserializers in
any .NET language and registering them with the `GremlinClient`.

[source,csharp]
----
include::../../../../gremlin-dotnet/test/Gremlin.Net.IntegrationTest/Docs/Dev/Provider/IndexTests.cs[tags=myTypeSerialization]

include::../../../../gremlin-dotnet/test/Gremlin.Net.IntegrationTest/Docs/Dev/Provider/IndexTests.cs[tags=supportingGremlinNetIO]
----

[[remoteconnection-implementations]]
==== RemoteConnection Implementations

A `RemoteConnection` is an interface that is important for usage on traversal sources configured using the
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#connecting-via-drivers[withRemote()] option. A `Traversal`
that is generated from that source will apply a `RemoteStrategy` which will inject a `RemoteStep` to its end. That
step will then send the `Bytecode` of the `Traversal` over the `RemoteConnection` to get the results that it will
iterate.

There is one method to implement on `RemoteConnection`:

[source,java]
public <E> CompletableFuture<RemoteTraversal<?, E>> submitAsync(final Bytecode bytecode) throws RemoteConnectionException;

Note that it returns a `RemoteTraversal`. This interface should also be implemented and in most cases implementers can
simply extend the `AbstractRemoteTraversal`.

TinkerPop provides the `DriverRemoteConnection` as a useful and
link:https://github.com/apache/tinkerpop/blob/x.y.z/gremlin-driver/src/main/java/org/apache/tinkerpop/gremlin/driver/remote[example implementation].
`DriverRemoteConnection` serializes the `Traversal` as Gremlin bytecode and then submits it for remote processing on
Gremlin Server. Gremlin Server rebinds the `Traversal` to a configured `Graph` instance and then iterates the results
back as it would normally do.

Implementing `RemoteConnection` is not something routinely done for those implementing `gremlin-core`. It is only
something required if there is a need to exploit remote traversal submission. If a graph provider has a "graph server"
similar to Gremlin Server that can accept bytecode-based requests on its own protocol, then that would be one example
of a reason to implement this interface.

[[bulk-import-export]]
==== Bulk Import Export

When it comes to doing "bulk" operations, the diverse nature of the available graph databases and their specific
capabilities, prevents TinkerPop from doing a good job of generalizing that capability well. TinkerPop thus maintains
two positions on the concept of import and export:

1. TinkerPop refers users to the bulk import/export facilities of specific graph providers as they tend to be more
efficient and easier to use than the options TinkerPop has tried to generalize in the past.
2. TinkerPop encourages graph providers to expose those capabilities via `g.io()` and the `IoStep` by way of a
`TraversalStrategy`.

That said, for graph providers that don't have a special bulk loading feature, they can either rely on the default
OLTP (single-threaded) `GraphReader` and `GraphWriter` options that are embedded in `IoStep` or get a basic bulk loader
from TinkerPop using the link:https://tinkerpop.apache.org/docs/x.y.z/reference/#clonevertexprogram[CloneVertexProgram].
Simply provide a `InputFormat` and `OutputFormat` that can be referenced by a `HadoopGraph` instance as discussed
in the link:https://tinkerpop.apache.org/docs/x.y.z/reference/#clonevertexprogram[Reference Documentation].

[[validating-with-gremlin-test]]
=== Validating with Gremlin-Test

image:gremlin-edumacated.png[width=225]

[source,xml]
----
<dependency>
  <groupId>org.apache.tinkerpop</groupId>
  <artifactId>gremlin-test</artifactId>
  <version>x.y.z</version>
</dependency>
----

The operational semantics of any OLTP or OLAP implementation are validated by `gremlin-test`. To implement these tests,
provide test case implementations as shown below, where `XXX` below denotes the name of the graph implementation (e.g.
TinkerGraph, Neo4jGraph, HadoopGraph, etc.).

[source,java]
----
// Structure API tests
@RunWith(StructureStandardSuite.class)
@GraphProviderClass(provider = XXXGraphProvider.class, graph = XXXGraph.class)
public class XXXStructureStandardTest {}

// Process API tests
@RunWith(ProcessComputerSuite.class)
@GraphProviderClass(provider = XXXGraphProvider.class, graph = XXXGraph.class)
public class XXXProcessComputerTest {}

@RunWith(ProcessStandardSuite.class)
@GraphProviderClass(provider = XXXGraphProvider.class, graph = XXXGraph.class)
public class XXXProcessStandardTest {}
----

The above set of tests represent the minimum test suite set to implement.  There are other "integration" and
"performance" tests that should be considered optional.  Implementing those tests requires the same pattern as shown
above.

IMPORTANT: It is as important to look at "ignored" tests as it is to look at ones that fail.  The `gremlin-test`
suite utilizes the `Feature` implementation exposed by the `Graph` to determine which tests to execute.  If a test
utilizes features that are not supported by the graph, it will ignore them.  While that may be fine, implementers
should validate that the ignored tests are appropriately bypassed and that there are no mistakes in their feature
definitions.  Moreover, implementers should consider filling gaps in their own test suites, especially when
IO-related tests are being ignored.

TIP: If it is expensive to construct a new `Graph` instance, consider implementing `GraphProvider.getStaticFeatures()`
which can help by caching a static feature set for instances produced by that `GraphProvider` and allow the test suite
to avoid that construction cost if the test is ignored.

The only test-class that requires any code investment is the `GraphProvider` implementation class. This class is a
used by the test suite to construct `Graph` configurations and instances and provides information about the
implementation itself.  In most cases, it is best to simply extend `AbstractGraphProvider` as it provides many
default implementations of the `GraphProvider` interface.

Finally, specify the test suites that will be supported by the `Graph` implementation using the `@Graph.OptIn`
annotation.  See the `TinkerGraph` implementation below as an example:

[source,java]
----
@Graph.OptIn(Graph.OptIn.SUITE_STRUCTURE_STANDARD)
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_STANDARD)
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_COMPUTER)
public class TinkerGraph implements Graph {
----

Only include annotations for the suites the implementation will support.  Note that implementing the suite, but
not specifying the appropriate annotation will prevent the suite from running (an obvious error message will appear
in this case when running the mis-configured suite).

There are times when there may be a specific test in the suite that the implementation cannot support (despite the
features it implements) or should not otherwise be executed.  It is possible for implementers to "opt-out" of a test
by using the `@Graph.OptOut` annotation. This annotation can be applied to either a `Graph` instance or a
`GraphProvider` instance (the latter would typically be used for "opting out" for a particular `Graph` configuration
that was under test). The following is an example of this annotation usage as taken from `HadoopGraph`:

[source,java]
----
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_STANDARD)
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_COMPUTER)
@Graph.OptOut(
        test = "org.apache.tinkerpop.gremlin.process.graph.step.map.MatchTest$Traversals",
        method = "g_V_matchXa_hasXname_GarciaX__a_inXwrittenByX_b__a_inXsungByX_bX",
        reason = "Hadoop-Gremlin is OLAP-oriented and for OLTP operations, linear-scan joins are required. This particular tests takes many minutes to execute.")
@Graph.OptOut(
        test = "org.apache.tinkerpop.gremlin.process.graph.step.map.MatchTest$Traversals",
        method = "g_V_matchXa_inXsungByX_b__a_inXsungByX_c__b_outXwrittenByX_d__c_outXwrittenByX_e__d_hasXname_George_HarisonX__e_hasXname_Bob_MarleyXX",
        reason = "Hadoop-Gremlin is OLAP-oriented and for OLTP operations, linear-scan joins are required. This particular tests takes many minutes to execute.")
@Graph.OptOut(
        test = "org.apache.tinkerpop.gremlin.process.computer.GraphComputerTest",
        method = "shouldNotAllowBadMemoryKeys",
        reason = "Hadoop does a hard kill on failure and stops threads which stops test cases. Exception handling semantics are correct though.")
@Graph.OptOut(
        test = "org.apache.tinkerpop.gremlin.process.computer.GraphComputerTest",
        method = "shouldRequireRegisteringMemoryKeys",
        reason = "Hadoop does a hard kill on failure and stops threads which stops test cases. Exception handling semantics are correct though.")
public class HadoopGraph implements Graph {
----

The above examples show how to ignore individual tests.  It is also possible to:

* Ignore an entire test case (i.e. all the methods within the test) by setting the `method` to "*".
* Ignore a "base" test class such that test that extend from those classes will all be ignored.
* Ignore a `GraphComputer` test based on the type of `GraphComputer` being used.  Specify the "computer" attribute on
the `OptOut` (which is an array specification) which should have a value of the `GraphComputer` implementation class
that should ignore that test. This attribute should be left empty for "standard" execution and by default all
`GraphComputer` implementations will be included in the `OptOut` so if there are multiple implementations, explicitly
specify the ones that should be excluded.

Also note that some of the tests in the Gremlin Test Suite are parameterized tests and require an additional level of
specificity to be properly ignored.  To ignore these types of tests, examine the name template of the parameterized
tests.  It is defined by a Java annotation that looks like this:

[source,java]
@Parameterized.Parameters(name = "expect({0})")

The annotation above shows that the name of each parameterized test will be prefixed with "expect" and have
parentheses wrapped around the first parameter (at index 0) value supplied to each test.  This information can
only be garnered by studying the test set up itself.  Once the pattern is determined and the specific unique name of
the parameterized test is identified, add it to the `specific` property on the `OptOut` annotation in addition to
the other arguments.

These annotations help provide users a level of transparency into test suite compliance (via the
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#describe-graph[describeGraph()] utility function). It also
allows implementers to have a lot of flexibility in terms of how they wish to support TinkerPop.  For example, maybe
there is a single test case that prevents an implementer from claiming support of a `Feature`.  The implementer could
choose to either not support the `Feature` or to support it but "opt-out" of the test with a "reason" as to why so
that users understand the limitation.

IMPORTANT: Before using `OptOut` be sure that the reason for using it is sound and it is more of a last resort.
It is possible that a test from the suite doesn't properly represent the expectations of a feature, is too broad or
narrow for the semantics it is trying to enforce or simply contains a bug.  Please consider raising issues in the
developer mailing list with such concerns before assuming `OptOut` is the only answer.

IMPORTANT: There are no tests that specifically validate complete compliance with Gremlin Server.  Generally speaking,
a `Graph` that passes the full Test Suite, should be compliant with Gremlin Server.  The one area where problems can
occur is in serialization.  Always ensure that IO is properly implemented, that custom serializers are tested fully
and ultimately integration test the `Graph` with an actual Gremlin Server instance.

WARNING: Configuring tests to run in parallel might result in errors that are difficult to debug as there is some
shared state in test execution around graph configuration.  It is therefore recommended that parallelism be turned
off for the test suite (the Maven SureFire Plugin is configured this way by default).  It may also be important to
include this setting, `<reuseForks>false</reuseForks>`, in the SureFire configuration if tests are failing in an
unexplainable way.

WARNING: For graph implementations that require a schema, take note that TinkerPop tests were originally developed
without too much concern for these types of graphs. While most tests utilize the standard toy graphs there are
instances where tests will utilize their own independent schema that stands alone from all other tests. It may be
necessary to create schemas specific to certain tests in those situations.

TIP: When running the `gremlin-test` suite against your implementation, you may need to set `build.dir` as an
environment variable, depending on your project layout. Some tests require this to find a writable directory for
creating temporary files. The value is typically set to the project build directory. For example using the Maven
SureFire Plugin, this is done via the configuration argLine with `-Dbuild.dir=${project.build.directory}`.

==== Checking resource leak

The TinkerPop query engine retrieves data by interfacing with the provider using iterators. These iterators (depending
on the provider) may hold up resources in the underlying storage layer and hence, it is critical to close them after
the query is finished.

TinkerPop provides you with the ability to test for such resource leaks by checking for leaks when you run the
Gremlin-Test suites against your implementation. To enable this leak detection, providers should increment the
`StoreIteratorCounter` whenever a resource is opened and decrement it when it is closed. A reference implementation
is provided with TinkerGraph as `TinkerGraphIterator.java`.

Assertions for leak detection are enabled by default when running the test suite. They can be temporarily disabled
by way of a system property - simply set `-DtestIteratorLeaks=false".

=== Accessibility via GremlinPlugin

image:gremlin-plugin.png[width=100,float=left] The applications distributed with TinkerPop do not distribute with
any graph system implementations besides TinkerGraph. If your implementation is stored in a Maven repository (e.g.
Maven Central Repository), then it is best to provide a <<gremlin-plugins,`GremlinPlugin`>> implementation so the respective jars can be
downloaded according and when required by the user. Neo4j's GremlinPlugin is provided below for reference.

[source,java]
----
include::{basedir}/neo4j-gremlin/src/main/java/org/apache/tinkerpop/gremlin/neo4j/jsr223/Neo4jGremlinPlugin.java[]
----

With the above plugin implementations, users can now download respective binaries for Gremlin Console, Gremlin Server, etc.

[source,groovy]
gremlin> g = Neo4jGraph.open('/tmp/neo4j')
No such property: Neo4jGraph for class: groovysh_evaluate
Display stack trace? [yN]
gremlin> :install org.apache.tinkerpop neo4j-gremlin x.y.z
==>loaded: [org.apache.tinkerpop, neo4j-gremlin, …]
gremlin> :plugin use tinkerpop.neo4j
==>tinkerpop.neo4j activated
gremlin> g = Neo4jGraph.open('/tmp/neo4j')
==>neo4jgraph[EmbeddedGraphDatabase [/tmp/neo4j]]

=== In-Depth Implementations

image:gremlin-painting.png[width=200,float=right] The graph system implementation details presented thus far are
minimum requirements necessary to yield a valid TinkerPop implementation. However, there are other areas that a
graph system provider can tweak to provide an implementation more optimized for their underlying graph engine. Typical
areas of focus include:

* Traversal Strategies: A link:https://tinkerpop.apache.org/docs/x.y.z/reference/#traversalstrategy[TraversalStrategy]
can be used to alter a traversal prior to its execution. A typical example is converting a pattern of
`g.V().has('name','marko')` into a global index lookup for all vertices with name "marko". In this way, a `O(|V|)`
lookup becomes an `O(log(|V|))`. Please review `TinkerGraphStepStrategy` for ideas.
* Step Implementations: Every link:https://tinkerpop.apache.org/docs/x.y.z/reference/#graph-traversal-steps[step] is
ultimately referenced by the `GraphTraversal` interface. It is possible to extend `GraphTraversal` to use a graph
system specific step implementation. Note that while it is sometimes possible to develop custom step implementations
by extending from a TinkerPop step (typically, `AddVertexStep` and other `Mutating` steps), it's important to
consider that doing so introduces some greater risk for code breaks on upgrades as opposed to other areas of the code
base. As steps are more internal features of TinkerPop, they might be subject to breaking API and behavioral changes
that would be less likely to be accepted by more public facing interfaces.

== Graph Driver Provider Requirements

image::gremlin-server-protocol.png[width=325]

One of the roles for link:https://tinkerpop.apache.org/docs/x.y.z/reference/#gremlin-server[Gremlin Server] is to
provide a bridge from TinkerPop to non-JVM languages (e.g. Go, Python, etc.).  Developers can build language bindings
(or driver) that provide a way to submit Gremlin scripts to Gremlin Server and get back results.  Given the
extensible nature of Gremlin Server, it is difficult to provide an authoritative guide to developing a driver.
It is however possible to describe the core communication protocol using the standard out-of-the-box configuration
which should provide enough information to develop a driver for a specific language.

image::gremlin-server-flow.png[width=300,float=right]

Gremlin Server is distributed with a configuration that utilizes link:http://en.wikipedia.org/wiki/WebSocket[WebSocket]
with a custom sub-protocol.  Under this configuration, Gremlin Server accepts requests containing a Gremlin script,
evaluates that script and then streams back the results.  The notion of "streaming" is depicted in the diagram to the
right.

The diagram shows an incoming request to process the Gremlin script of `g.V()`.  Gremlin Server evaluates that script,
getting an `Iterator` of vertices as a result, and steps through each `Vertex` within it.  The vertices are batched
together given the `resultIterationBatchSize` configuration.  In this case, that value must be `2` given that each
"response" contains two vertices.  Each response is serialized given the requested serializer type (JSON is likely
best for non-JVM languages) and written back to the requesting client immediately.  Gremlin Server does not wait for
the entire result to be iterated, before sending back a response.  It will send the responses as they are realized.

This approach allows for the processing of large result sets without having to serialize the entire result into memory
for the response.  It places a bit of a burden on the developer of the driver however, because it becomes necessary to
provide a way to reconstruct the entire result on the client side from all of the individual responses that Gremlin
Server returns for a single request.  Again, this description of Gremlin Server's "flow" is related to the
out-of-the-box configuration.  It is quite possible to construct other flows, that might be more amenable to a
particular language or style of processing.

To formulate a request to Gremlin Server, a `RequestMessage` needs to be constructed.  The `RequestMessage` is a
generalized representation of a request that carries a set of "standard" values in addition to optional ones that are
dependent on the operation being performed.  A `RequestMessage` has these fields:

[width="100%",cols="3,10",options="header"]
|=========================================================
|Key |Description
|requestId |A link:http://en.wikipedia.org/wiki/Globally_unique_identifier[UUID] representing the unique identification for the request.
|op |The name of the "operation" to execute based on the available `OpProcessor` configured in the Gremlin Server.  To evaluate a script, use `eval`.
|processor |The name of the `OpProcessor` to utilize. The default `OpProcessor` for evaluating scripts is unnamed and therefore script evaluation purposes, this value can be an empty string.
|args |A `Map` of arbitrary parameters to pass to Gremlin Server.  The requirements for the contents of this `Map` are dependent on the `op` selected.
|=========================================================

This message can be serialized in any fashion that is supported by Gremlin Server.  New serialization methods can
be plugged in by implementing a `ServiceLoader` enabled `MessageSerializer`, however Gremlin Server provides for
JSON serialization by default which will be good enough for purposes of most developers building drivers.
A `RequestMessage` to evaluate a script with variable bindings looks like this in JSON:

[source,js]
----
{ "requestId":"1d6d02bd-8e56-421d-9438-3bd6d0079ff1",
  "op":"eval",
  "processor":"",
  "args":{"gremlin":"g.traversal().V(x).out()",
          "bindings":{"x":1},
          "language":"gremlin-groovy"}}
----

The above JSON represents the "body" of the request to send to Gremlin Server. When sending this "body" over
WebSocket, Gremlin Server can accept a packet frame using a "text" (1) or a "binary" (2) opcode. Using "text"
is a bit more limited in that Gremlin Server will always process the body of that request as JSON.  Generally speaking
"text" is just for testing purposes.

The preferred method for sending requests to Gremlin Server is to use the "binary" opcode.  In this case, a "header"
will need be sent in addition to to the "body".  The "header" basically consists of a "mime type" so that Gremlin
Server knows how to deserialize the `RequestMessage`. So, the actual byte array sent to Gremlin Server would be
formatted as follows:

image::gremlin-server-request.png[]

The first byte represents the length of the "mime type" string value that follows.  Given the default configuration of
Gremlin Server, this value should be set to `application/json`.  The "payload" represents the JSON message above
encoded as bytes.

NOTE: Gremlin Server will only accept masked packets as it pertains to a WebSocket packet header construction.

When Gremlin Server receives that request, it will decode it given the "mime type", pass it to the requested
`OpProcessor` which will execute the `op` defined in the message.  In this case, it will evaluate the script
`g.traversal().V(x).out()` using the `bindings` supplied in the `args` and stream back the results in a series of
`ResponseMessages`.  A `ResponseMessage` looks like this:

[width="100%",cols="3,10",options="header"]
|=========================================================
|Key |Description
|requestId |The identifier of the `RequestMessage` that generated this `ResponseMessage`.
|status | The `status` contains a `Map` of three keys: `code` which refers to a `ResultCode` that is somewhat analogous to an link:http://www.w3.org/Protocols/rfc2616/rfc2616-sec10.html[HTTP status code], `attributes` that represent a `Map` of protocol-level information, and `message` which is just a human-readable `String` usually associated with errors.
|result | The `result` contains a `Map` of two keys: `data` which refers to the actual data returned from the server (the type of data is determined by the operation requested) and `meta` which is a `Map` of meta-data related to the response.
|=========================================================

In this case the `ResponseMessage` returned to the client would look something like this:

[source,js]
----
{"result":{"data":[{"id": 2,"label": "person","type": "vertex","properties": [
  {"id": 2, "value": "vadas", "label": "name"},
  {"id": 3, "value": 27, "label": "age"}]},
  ], "meta":{}},
 "requestId":"1d6d02bd-8e56-421d-9438-3bd6d0079ff1",
 "status":{"code":206,"attributes":{},"message":""}}
----

Gremlin Server is capable of streaming results such that additional responses will arrive over the WebSocket connection until
the iteration of the result on the server is complete.  Each successful incremental message will have a `ResultCode`
of `206`. Termination of the stream will be marked by a final `200` status code.  Note that all messages without a
`206` represent terminating conditions for a request.  The following table details the various status codes that
Gremlin Server will send:

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Code |Name |Description
|200 |SUCCESS |The server successfully processed a request to completion - there are no messages remaining in this stream.
|204 |NO CONTENT |The server processed the request but there is no result to return (e.g. an `Iterator` with no elements) - there are no messages remaining in this stream.
|206 |PARTIAL CONTENT |The server successfully returned some content, but there is more in the stream to arrive - wait for a `SUCCESS` to signify the end of the stream.
|401 |UNAUTHORIZED |The request attempted to access resources that the requesting user did not have access to.
|407 |AUTHENTICATE |A challenge from the server for the client to authenticate its request.
|497 |CLIENT SERIALIZATION ERROR |The request message contained an object that was not serializable.
|498 |MALFORMED REQUEST |The request message was not properly formatted which means it could not be parsed at all or the "op" code was not recognized such that Gremlin Server could properly route it for processing.  Check the message format and retry the request.
|499 |INVALID REQUEST ARGUMENTS |The request message was parseable, but the arguments supplied in the message were in conflict or incomplete. Check the message format and retry the request.
|500 |SERVER ERROR |A general server error occurred that prevented the request from being processed.
|597 |SCRIPT EVALUATION ERROR |The script submitted for processing evaluated in the `ScriptEngine` with errors and could not be processed.  Check the script submitted for syntax errors or other problems and then resubmit.
|598 |SERVER TIMEOUT |The server exceeded one of the timeout settings for the request and could therefore only partially responded or did not respond at all.
|599 |SERVER SERIALIZATION ERROR |The server was not capable of serializing an object that was returned from the script supplied on the request. Either transform the object into something Gremlin Server can process within the script or install mapper serialization classes to Gremlin Server.
|=========================================================

NOTE: Please refer to the link:https://tinkerpop.apache.org/docs/current/dev/io[IO Reference Documentation] for more
examples of `RequestMessage` and `ResponseMessage` instances.

=== OpProcessors Arguments

The following sections define a non-exhaustive list of available operations and arguments for embedded `OpProcessors`
(i.e. ones packaged with Gremlin Server).

==== Common

All `OpProcessor` instances support these arguments.

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|batchSize |Int |When the result is an iterator this value defines the number of iterations each `ResponseMessage` should contain - overrides the `resultIterationBatchSize` server setting.
|=========================================================

==== Standard OpProcessor

The "standard" `OpProcessor` handles requests for the primary function of Gremlin Server - executing Gremlin.
Requests made to this `OpProcessor` are "sessionless" in the sense that a request must encapsulate the entirety
of a transaction.  There is no state maintained between requests.  A transaction is started when the script is first
evaluated and is committed when the script completes (or rolled back if an error occurred).

[width="100%",cols="3,10a",options="header"]
|=========================================================
|Key |Description
|processor |As this is the default `OpProcessor` this value can be set to an empty string.
|op |[width="100%",cols="3,10",options="header"]
!=========================================================
!Key !Description
!`authentication` !A request that contains the response to a server challenge for authentication.
!`eval` !Evaluate a Gremlin script provided as a `String`.
!=========================================================
|=========================================================

**`authentication` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|sasl |String | *Required* The response to the server authentication challenge.  This value is dependent on the SASL authentication mechanism required by the server and is Base64 encoded.
|saslMechanism |String | The SASL mechanism: `PLAIN` or `GSSAPI`. Note that it is up to the server implementation to use or disregard this setting (default implementation in Gremlin Server ignores it).
|=========================================================

**`eval` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|gremlin |String | *Required* The Gremlin script to evaluate.
|bindings |Map |A map of key/value pairs to apply as variables in the context of the Gremlin script.
|language |String |The flavor of Gremlin used (e.g. `gremlin-groovy`).
|aliases |Map |A map of key/value pairs that allow globally bound `Graph` and `TraversalSource` objects to
be aliased to different variable names for purposes of the current request.  The value represents the name of the
global variable and its key represents the new binding name as it will be referenced in the Gremlin query.  For
example, if the Gremlin Server defines two `TraversalSource` instances named `g1` and `g2`, it would be possible
to send an alias pair with key of "g" and value of "g2" and thus allow the script to refer to "g2" simply as "g".
|evaluationTimeout |Long |An override for the server setting that determines the maximum time to wait for a script to execute on the server.
|=========================================================

==== Session OpProcessor

The "session" `OpProcessor` handles requests for the primary function of Gremlin Server - executing Gremlin. It is
like the "standard" `OpProcessor`, but instead maintains state between sessions and allows the option to leave all
transaction management up to the calling client.  It is important that clients that open sessions, commit or roll
them back, however Gremlin Server will try to clean up such things when a session is killed that has been abandoned.
It is important to consider that a session can only be maintained with a single machine.  In the event that multiple
Gremlin Server are deployed, session state is not shared among them.

[width="100%",cols="3,10a",options="header"]
|=========================================================
|Key |Description
|processor |This value should be set to `session`
|op |
[cols="3,10",options="header"]
!=========================================================
!Key !Description
!`authentication` !A request that contains the response to a server challenge for authentication.
!`eval` !Evaluate a Gremlin script provided as a `String`.
!`close` !Close the specified session. Will return a `NO CONTENT` message as confirmation of the close being completed.
|=========================================================

NOTE: The "close" message is deprecated as of 3.3.11 as servers at this version are required to automatically interrupt
running processes on the close of the connection and release resources such as sessions. Servers wishing to be
compatible with older versions of the driver need only send back a `NO_CONTENT` for this message. Drivers wishing to
be compatible with servers prior to 3.3.11 may continue to send the message on calls to `close()` otherwise such code
can be removed.

**`authentication` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|saslMechanism |String | The SASL mechanism: `PLAIN` or `GSSAPI`. Note that it is up to the server implementation to use or disregard this setting (default implementation in Gremlin Server ignores it).
|sasl |String | *Required* The response to the server authentication challenge.  This value is dependent on the SASL authentication mechanism required by the server and is Base64 encoded.
|=========================================================

**`eval` operation arguments**

[width="100%",options="header"]
|=========================================================
|Key |Type |Description
|gremlin |String | *Required* The Gremlin script to evaluate.
|session |String | *Required* The session identifier for the current session - typically this value should be a UUID (the session will be created if it doesn't exist).
|manageTransaction |Boolean |When set to `true` the transaction for the current request is auto-committed or rolled-back as are done with sessionless requests - defaulted to `false`.
|bindings |Map |A map of key/value pairs to apply as variables in the context of the Gremlin script.
|evaluationTimeout |Long |An override for the server setting that determines the maximum time to wait for a script to execute on the server.
|language |String |The flavor of Gremlin used (e.g. `gremlin-groovy`)
|aliases |Map |A map of key/value pairs that allow globally bound `Graph` and `TraversalSource` objects to
be aliased to different variable names for purposes of the current request.  The value represents the name the
global variable and its key represents the new binding name as it will be referenced in the Gremlin query.  For
example, if the Gremlin Server defines two `TraversalSource` instances named `g1` and `g2`, it would be possible
to send an alias pair with key of "g" and value of "g2" and thus allow the script to refer to "g2" simply as "g".
|=========================================================

**`close` operation arguments**
[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|session |String | *Required* The session identifier for the session to close.
|force |Boolean | Determines if the session should be force closed when the client is closed. Force closing will not
attempt to close open transactions from existing running jobs and leave it to the underlying graph to decided how to
proceed with those orphaned transactions. Setting this to `true` tends to lead to faster close operation and release
of resources which can be desirable if Gremlin Server has a long session timeout and a long script evaluation timeout
as attempts to close long run jobs can occur more rapidly. If not provided, this value is `false`.
|=========================================================

==== Traversal OpProcessor

Both the Standard and Session OpProcessors allow for Gremlin scripts to be submitted to the server. The
`TraversalOpProcessor` however allows Gremlin `Bytecode` to be submitted to the server. Supporting this `OpProcessor`
makes it possible for a link:https://tinkerpop.apache.org/docs/current/tutorials/gremlin-language-variants/[Gremlin Language Variant]
to submit a `Traversal` directly to Gremlin Server in the native language of the GLV without having to use a script in
a different language.

Unlike Standard and Session OpProcessors, the Traversal OpProcessor does not simply return the results of the
`Traversal`. It instead returns `Traverser` objects which allows the client to take advantage of
link:https://tinkerpop.apache.org/docs/current/reference/#barrier-step[bulking]. To describe this interaction more
directly, the returned `Traverser` will represent some value from the `Traversal` result and the number of times it
is represented in the full stream of results. So, if a `Traversal` happens to return the same vertex twenty times
it won't return twenty instances of the same object. It will return one in `Traverser` with the `bulk` value set to
twenty. Under this model, the amount of processing and network overhead can be reduced considerably.

To demonstrate consider this example:

[gremlin-groovy]
----
cluster = Cluster.open()
client = cluster.connect()
aliased = client.alias("g")
g = org.apache.tinkerpop.gremlin.structure.util.empty.EmptyGraph.instance().traversal()     <1>
rs = aliased.submit(g.V().both().barrier().both().barrier()).all().get()                    <2>
aliased.submit(g.V().both().barrier().both().barrier().count()).all().get().get(0).getInt() <3>
rs.collect{[value: it.getObject().get(), bulk: it.getObject().bulk()]}                      <4>
----

<1> All commands through this step are just designed to demonstrate bulking with Gremlin Server and don't represent
a real-world way that this feature would be used.
<2> Submit a `Traversal` that happens to ensure that the server uses bulking. Note that a `Traverser` is returned
and that there are only six results.
<3> In actuality, however, if this same `Traversal` is iterated there are thirty results. Without bulking, the previous
request would have sent back thirty traversers.
<4> Note that the sum of the bulk of each `Traverser` is thirty.

The full iteration of a `Traversal` is thus left to the client. It must interpret the bulk on the `Traverser` and
unroll it to represent the actual number of times it exists when iterated. The unrolling is typically handled
directly within TinkerPop's remote traversal implementations.

One of the important aspects of the Traversal OpProcessor is that it enables the user to not only get the results of
the `Traversal` but also the side-effects that it produces. When the `Traversal` is submitted with the "bytecode"
operation, the results are iterated back as usual, but any side-effects are retained on the server in a cache keyed by
the identifier of the request that originally submitted the `Traversal`. The client will want to retain that identifier
if it intends to later get side-effects. The Traversal OpProcessor supplies the "keys" and "gather" operations to get
the keys stored in the side-effect and to get the value of a specific side-effect respectively. Finally, a "close"
operation is available to clear the cache of a specific side-effect.

[width="100%",cols="3,10a",options="header"]
|=========================================================
|Key |Description
|processor |This value should be set to `traversal`
|op |
[cols="3,10",options="header"]
!=========================================================
!Key !Description
!`authentication` !A request that contains the response to a server challenge for authentication.
!`bytecode` !A request that contains the `Bytecode` representation of a `Traversal`.
!`close` !Releases side-effects held in cache for a particular `Traversal`.
!`gather` !Gets a particular side-effect as produced by a previously executed `Traversal`.
!`keys` !Gets all the keys of all side-effects as produced by a previously executed `Traversal`.
|=========================================================

**`authentication` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|sasl |String | *Required* The response to the server authentication challenge.  This value is dependent on the SASL authentication mechanism required by the server and is Base64 encoded.
|=========================================================

**`bytecode` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|gremlin |String | *Required* The `Bytecode` representation of a `Traversal`.
|aliases |Map | *Required* A map with a single key/value pair that refers to a globally bound `TraversalSource` object
to be aliased to different variable names for purposes of the current request.  The value represents the name of the
global variable and its key represents the new binding name as it will be referenced in the Gremlin query.  For
example, if the Gremlin Server defines two `TraversalSource` instances named `g1`, it would be possible
to send an alias pair with key of "g" and value of "g1" and thus allow the script to refer to "g1" simply as "g". Note
that unlike users of `alias` in other contexts, in this case, the key can *only* be set to "g" and there can be only
one key value pair present (since only one `Traversal` is being submitted, there is no sense to having more than a
single alias).
|=========================================================

**`close` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|sideEffect |UUID | *Required* The unique identifier for the request that original submitted the traversal (side-effects are keyed by that value)
|=========================================================

**`gather` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|sideEffect |UUID | *Required* The unique identifier for the request that original submitted the traversal (side-effects are keyed by that value)
|sideEffectKey |String | *Required* The key for a specific side-effect.
|aliases |Map | *Required* A map with a single key/value pair that refers to a globally bound `TraversalSource` object
to be aliased to different variable names for purposes of the current request.  The value represents the name of the
global variable and its key represents the new binding name as it will be referenced in the Gremlin query.  For
example, if the Gremlin Server defines two `TraversalSource` instances named `g1`, it would be possible
to send an alias pair with key of "g" and value of "g1" and thus allow the script to refer to "g1" simply as "g". Note
that unlike users of `alias` in other contexts, in this case, the key can *only* be set to "g" and there can be only
one key value pair present (since only one `Traversal` is being submitted, there is no sense to having more than a
single alias).
|=========================================================

When using "gather" it is important to note the metadata that is returned on the `ResponseMessage`. It returns both the
"sideEffectKey" that was requested as well as a value keyed as "aggregateTo". The "aggregateTo" field describes how the
streamed side-effect data should be treated on the client. It provides a hint as to whether or not the data should be
rolled back up into a single object or simply left as-is. There are four values for "aggregateTo": `bulkset`, `list`,
`map`, `set` and `none`.

**`keys` operation arguments**

[width="100%",cols="2,2,9",options="header"]
|=========================================================
|Key |Type |Description
|sideEffect |UUID | *Required* The unique identifier for the request that original submitted the traversal (side-effects are keyed by that value)
|=========================================================

=== Authentication

Gremlin Server supports link:https://en.wikipedia.org/wiki/Simple_Authentication_and_Security_Layer[SASL-based]
authentication.  A SASL implementation provides a series of challenges and responses that a driver must comply with
in order to authenticate.  By default, Gremlin Server only supports the "PLAIN" SASL mechanism, which is a cleartext
password system.  When authentication is enabled, an incoming request is intercepted before it is evaluated by the
`ScriptEngine`.  The request is saved on the server and a `AUTHENTICATE` challenge response (status code `407`) is
returned to the client.

The client will detect the `AUTHENTICATE` and respond with an `authentication` for the `op` and an `arg` named `sasl`
that contains the password.  The password should be either, an encoded sequence of UTF-8 bytes, delimited by 0
(US-ASCII NUL), where the form is : `<NUL>username<NUL>password`, or a Base64 encoded string of the former (which
in this instance would be `AHVzZXJuYW1lAHBhc3N3b3Jk`).  Should Gremlin Server be able to authenticate with the
provided credentials, the server will return the results of the original request as it normally does without
authentication.  If it cannot authenticate given the challenge response from the client, it will return `UNAUTHORIZED`
(status code `401`).

NOTE: Gremlin Server does not support the "authorization identity" as described in link:https://tools.ietf.org/html/rfc4616[RFC4616].

[[gremlin-plugins]]
== Gremlin Plugins

image:gremlin-plugin.png[width=125]

Plugins provide a way to expand the features of a `GremlinScriptEngine`, which stands at that core of both Gremlin
Console and Gremlin Server. Providers may wish to create plugins for a variety of reasons, but some common examples
include:

* Initialize the `GremlinScriptEngine` application with important classes so that the user doesn't need to type their
own imports.
* Place specific objects in the bindings of the `GremlinScriptEngine` for the convenience of the user.
* Bootstrap the `GremlinScriptEngine` with custom functions so that they are ready for usage at startup.

The first step to developing a plugin is to implement the link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/GremlinPlugin.html[GremlinPlugin]
interface:

[source,java]
----
include::{basedir}/gremlin-core/src/main/java/org/apache/tinkerpop/gremlin/jsr223/GremlinPlugin.java[]
----

The most simple plugin and the one most commonly implemented will likely be one that just provides a list of classes
for import.  This type of plugin is the easiest way for implementers of the TinkerPop Structure and Process APIs to
make their implementations available to users.  The TinkerGraph implementation has just such a plugin:

[source,java]
----
include::{basedir}/tinkergraph-gremlin/src/main/java/org/apache/tinkerpop/gremlin/tinkergraph/jsr223/TinkerGraphGremlinPlugin.java[]
----

This plugin extends from the abstract base class of link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/AbstractGremlinPlugin.html[AbstractGremlinPlugin]
which provides some default implementations of the `GremlinPlugin` methods. It simply allows those who extend from it
to be able to just supply the name of the module and a list of link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/Customizer.html[Customizer]
instances to apply to the `GremlinScriptEngine`. In this case, the TinkerGraph plugin just needs an
link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/ImportCustomizer.html[ImportCustomizer]
which describes the list of classes to import when the plugin is activated and applied to the `GremlinScriptEngine`.

The `ImportCustomizer` is just one of several provided `Customizer` implementations that can be used in conjunction
with plugin development:

* link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/BindingsCustomizer.html[BindingsCustomizer] - Inject a key/value pair into the global bindings of the `GremlinScriptEngine` instances
* link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/ImportCustomizer.html[ImportCustomizer] - Add imports to a `GremlinScriptEngine`
* link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/ScriptCustomizer.html[ScriptCustomizer] - Execute a script on a `GremlinScriptEngine` at startup

Individual `GremlinScriptEngine` instances may have their own `Customizer` instances that can be used only with that
engine - e.g. `gremlin-groovy` has some that are specific to controlling the Groovy compiler configuration. Developing
a new `Customizer` implementation is not really possible without changes to TinkerPop, as the framework is not designed
to respond to external ones. The base `Customizer` implementations listed above should cover most needs.

A `GremlinPlugin` must support one of two instantiation models so that it can be instantiated from configuration files
for use in various situations - e.g. Gremlin Server. The first option is to use a static initializer given a method
with the following signature:

[source,java]
----
public static GremlinPlugin instance()
----

The limitation with this approach is that it does not provide a way to supply any configuration to the plugin so it
tends to only be useful for fairly simplistic plugins. The more advanced approach is to provide a "builder" given a
method with the following signature:

[source,java]
----
public static Builder build()
----

It doesn't really matter what kind of class is returned from `build` so long as it follows a "Builder" pattern, where
methods on that object return an instance of itself, so that builder methods can be chained together prior to calling
a final `create` method as follows:

[source,java]
----
public GremlinPlugin create()
----

Please see the link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/jsr223/ImportGremlinPlugin.html[ImportGremlinPlugin]
for an example of what implementing a `Builder` might look like in this context.

Note that the plugin provides a unique name for the plugin which follows a namespaced pattern as _namespace_._plugin-name_
(e.g. "tinkerpop.hadoop" - "tinkerpop" is the reserved namespace for TinkerPop maintained plugins).

For plugins that will work with Gremlin Console, there is one other step to follow to ensure that the `GremlinPlugin`
will work there.  The console loads `GremlinPlugin` instances via link:http://docs.oracle.com/javase/8/docs/api/java/util/ServiceLoader.html[ServiceLoader]
and therefore need a resource file added to the jar file where the plugin exists.  Add a file called
`org.apache.tinkerpop.gremlin.jsr223.GremlinPlugin` to `META-INF/services`.  In the case of the TinkerGraph
plugin above, that file will have this line in it:

[source,java]
----
include::{basedir}/tinkergraph-gremlin/src/main/resources/META-INF/services/org.apache.tinkerpop.gremlin.jsr223.GremlinPlugin[]
----

Once the plugin is packaged, there are two ways to test it out:

. Copy the jar and its dependencies to the Gremlin Console path and start it. It is preferrable that the plugin is
copied to the `/ext/_plugin_name_` directory.
. Start Gremlin Console and try the `:install` command: `:install com.company my-plugin 1.0.0`.

In either case, once one of these two approaches is taken, the jars and their dependencies are available to the
Console.  The next step is to "activate" the plugin by doing `:plugin use my-plugin`, where "my-plugin" refers to the
name of the plugin to activate.

NOTE: When `:install` is used logging dependencies related to link:http://www.slf4j.org/[SLF4J] are filtered out so as
not to introduce multiple logger bindings (which generates warning messages to the logs).

A plugin can do much more than just import classes.  One can expand the Gremlin language with new functions or steps,
provide useful commands to make repetitive or complex tasks easier to execute, or do helpful integrations with other
systems.  The secret to doing so lies in the `PluginAcceptor`.  As mentioned earlier, the `PluginAcceptor` provides
access to the host of the plugin.  It provides several important methods for doing so:

. `addBinding` - These two function allow the plugin to inject whatever context it wants to the host.  For example,
doing `addBinding('x',1)` would place a variable of `x` with a value of 1 into the console at the time of the plugin load.
. `eval` - Evaluates a script in the context of the host at the time of plugin startup.  For example, doing
`eval("sum={x,y->x+y}")` would create a `sum` function that would be available to the user of the Console after the
load of the plugin.
. `environment` - Provides context from the host environment.  For the console, the environment will return a `Map`
containing a reference to the `IO` stream and the `Groovysh` instance.  These classes represent very low-level access
to the underpinnings of the console.  Access to `Groovysh` allows for advanced features such as registering new
commands (e.g. like the `:plugin` or `:remote` commands).

Plugins can also tie into the `:remote` and `:submit` commands.  Recall that a `:remote` represents a different
context within which Gremlin is executed, when issued with `:submit`.  It is encouraged to use this integration point
when possible, as opposed to registering new commands that can otherwise follow the `:remote` and `:submit` pattern.
To expose this integration point as part of a plugin, implement the `RemoteAcceptor` interface:

TIP: Be good to the users of plugins and prevent dependency conflicts. Maintaining a conflict free plugin is most
easily done by using the link:http://maven.apache.org/enforcer/maven-enforcer-plugin/[Maven Enforcer Plugin].

TIP: Consider binding the plugin's minor version to the TinkerPop minor version so that it's easy for users to figure
out plugin compatibility.  Otherwise, clearly document a compatibility matrix for the plugin somewhere that users can
find it.

[source,java]
----
include::{basedir}/gremlin-core/src/main/java/org/apache/tinkerpop/gremlin/jsr223/console/RemoteAcceptor.java[]
----

The `RemoteAcceptor` can be bound to a `GremlinPlugin` by adding a `ConsoleCustomizer` implementation to the list of
`Customizer` instances that are returned from the `GremlinPlugin`. The `ConsoleCustomizer` will only be executed when
in use with the Gremlin Console plugin host.  Simply instantiate and return a `RemoteAcceptor` in the
`ConsoleCustomizer.getRemoteAcceptor(GremlinShellEnvironment)` method.  Generally speaking, each call to
`getRemoteAcceptor(GremlinShellEnvironment)` should produce a new instance of a `RemoteAcceptor`.