docs/src/dev/provider/index.asciidoc

////
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.
////

:docinfo: shared
:docinfodir: ../../
:toc-position: left

image::apache-tinkerpop-logo.png[width=500,link="https://tinkerpop.apache.org"]

*x.y.z*

= Introduction

image:tinkerpop-cityscape.png[]

This document discusses Apache TinkerPop™ implementation details that are most useful to developers who implement
TinkerPop interfaces and the Gremlin language. This document may also be helpful to Gremlin users who simply want a
deeper understanding of how TinkerPop works and what the behavioral semantics of Gremlin are. The
<<providers,Provider Section>> outlines the various integration and extension points that TinkerPop has while the
<<gremlin-semantics,Gremlin Semantics Section>> documents the Gremlin language itself.

Providers who rely on the TinkerPop execution engine generally receive the behaviors described in the Gremlin Semantics
section for free, but those who develop their own engine or extend upon the certain features should refer to that
section for the details required for a consistent Gremlin experience.

[[providers]]
= Provider Documentation

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 Java 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:

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`. `VertexProperty` should return empty
iterable for properties if meta-properties is not supported.

[[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[with()] 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 `GremlinLang` 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 GremlinLang gremlinLang) 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 GremlinLang and then submits it for remote processing on
Gremlin Server. Gremlin Server parses the script into a `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 traversal-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]
----
<!-- gremlin-tests contains test classes and gherkin tests -->
<dependency>
  <groupId>org.apache.tinkerpop</groupId>
  <artifactId>gremlin-test</artifactId>
  <version>x.y.z</version>
</dependency>
<!-- TinkerGraph is required for validating subgraph() tests -->
<dependency>
  <groupId>org.apache.tinkerpop</groupId>
  <artifactId>tinkergraph-gremlin</artifactId>
  <version>x.y.z</version>
</dependency>
----

TinkerPop provides `gremlin-test`, a technology test kit, that helps validate provider implementations. There is a
Gherkin-based test suite and a JVM-based test suite. The Gherkin suite is meant to test Gremlin operations and semantics
while the JVM-based suite is meant for lower-level `Graph` implementation validation and for Gremlin use cases that
apply to embedded operations (i.e. running Gremlin in the same JVM as the `Graph` implementation).

==== JVM Test Suite

IMPORTANT: 4.0.0-beta.1 Release - The final form of the TinkerPop test suite for 4.0 is not wholly settled, but going
forward providers should focus on implementing the <<gherkin-tests-suite>> as opposed to the JVM suite.

The JVM test suite is useful to graph system implementers who want to validate that their `Graph` implementation is
working properly and that Gremlin features specific to embedded use cases are behaving as expected. These tests do not
validate that all the semantics of the Gremlin query language are properly operating. Providers should implement the
<<gherkin-tests-suite>> for that type of validation.

If you are a remote graph database only or don't implement the `Graph` API but simply expose access to the Gremlin
language then it likely makes sense to skip implementing this suite and only implemented the Gherkin suite. On the other
hand, if you do implement the `Graph` API or `GraphComputer` API, it likely makes sense to implement both these tests
and the Gherkin suite.

To implement the JVM 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 which validate the core functionality of the Graph implementation.
// Start with this one because, generally speaking, if this suite passes then the others
// will mostly follow suit. Moreover, the structure tests operate on lower-level aspects
// of the Graph interfaces which are much easier to debug than Gremlin level tests.
@RunWith(StructureStandardSuite.class)
@GraphProviderClass(provider = XXXGraphProvider.class, graph = XXXGraph.class)
public class XXXStructureStandardTest {}

// Process API tests cover embedded Gremlin uses cases
@RunWith(ProcessEmbeddedStandardSuite.class)
@GraphProviderClass(provider = XXXGraphProvider.class, graph = XXXGraph.class)
public class XXXProcessStandardTest {}

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

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 that 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 Leaks

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".

[[gherkin-tests-suite]]
==== Gherkin Test Suite

The Gherkin Test Suite is a language agnostic set of tests that verify Gremlin semantics. It provides a unified set of
tests that validate many TinkerPop components internally. The tests themselves can be found in `gremlin-tests/features`
(link:https://github.com/apache/tinkerpop/tree/x.y.z/gremlin-test/src/main/resources/org/apache/tinkerpop/gremlin/test/features[here])
with their syntax described in the
link:https://tinkerpop.apache.org/docs/x.y.z/dev/developer/#gremlin-language-test-cases[TinkerPop Developer Documentation].

TinkerPop provides some infrastructure, for JVM based graphs, to help make it easier for providers to implement these
tests against their implementations. This infrastructure is built on `cucumber-java` which is a dependency of
`gremlin-test`. There are two main components to implementing the tests:

1. A `org.apache.tinkerpop.gremlin.features.World` implementation which is a class in `gremlin-test`.
2. A JUnit test class that will act as the runner for the tests with the appropriate annotations

TIP: It may be helpful to get familiar with link:https://cucumber.io/docs/installation/java/[Cucumber] before
proceeding with an implementation.

The `World` implementation provides context to the tests and allows providers to intercept test events that might be
important to proper execution specific to their implementations. The most important part of implementing `World` is
properly implementing the `GraphTraversalSource getGraphTraversalSource(GraphData)` method which provides to the test
the `GraphTraversalSource` to execute the test against.

The JUnit test class is really just the test runner. It is a simple class which must include some Cucumber annotations.
The following is just an example as taken from TinkerGraph:

[source,java]
----
@RunWith(Cucumber.class)
@CucumberOptions(
        tags = "not @GraphComputerOnly and not @AllowNullPropertyValues",
        glue = { "org.apache.tinkerpop.gremlin.features" },
        features = { "classpath:/org/apache/tinkerpop/gremlin/test/features" },
        plugin = {"progress", "junit:target/cucumber.xml",
        objectFactory = GuiceFactory.class})
----

The `@CucumberOptions` that are used are mostly implementation specific, so it will be up to the provider to make some
choices as to what is right for their environment. For TinkerGraph, it needed to ignore Gherkin tests with the
`@GraphComputerOnly` and `@AllowNullPropertyValues` tags (the full list of possible tags can be found
link:https://tinkerpop.apache.org/docs/x.y.z/dev/developer/#gherkin-tags[here]), since this configuration was for OLTP
tests and the graph was configured without support for storing `null` in the graph. The "glue" will be the same for all
test implementers as it refers to a package containing TinkerPop's test infrastructure in `gremlin-test` (unless of
course, a provider needs to develop their own infrastructure for some reason). The "features" is the path to the actual
Gherkin test files that should be made available locally. The files can be referenced on the classpath assuming
`gremlin-test` is a dependency. The "plugin" defines a JUnit style output, which happens to be understood by Maven.

The "objectFactory" is the last component. Cucumber relies on dependency injection to get a `World` implementation into
the test infrastructure. Providers may choose from multiple available implementations, but TinkerPop chose to use
Guice. To follow this approach include the following module:

[source,xml]
----
<dependency>
    <groupId>com.google.inject</groupId>
    <artifactId>guice</artifactId>
    <version>4.2.3</version>
    <scope>test</scope>
</dependency>
----

Following the Neo4jGraph implementation, there are two classes to construct:

[source,java]
----
public class ServiceModule extends AbstractModule {
    @Override
    protected void configure() {
        bind(World.class).to(Neo4jGraphWorld.class);
    }
}

public class WorldInjectorSource implements InjectorSource {
    @Override
    public Injector getInjector() {
        return Guice.createInjector(Stage.PRODUCTION, CucumberModules.createScenarioModule(), new ServiceModule());
    }
}
----

The key here is that the `Neo4jGraphWorld` implementation gets bound to `World` in the `ServiceModule` and there is
a `WorldInjectorSource` that specifies the `ServiceModule` to Cucumber. As a final step, the provider's test resources
needs a `cucumber.properties` file with an entry that specifies the `InjectorSource` so that Guice can find it. Here
is the example taken from TinkerGraph where the `WorldInjectorSource` is inner class of `TinkerGraphFeatureTest`
itself.

[source,text]
----
guice.injector-source=org.apache.tinkerpop.gremlin.neo4j.Neo4jGraphFeatureTest$WorldInjectorSource
----

In the event that a single `World` configuration is insufficient, it may be necessary to develop a custom
`ObjectFactory`. An easy way to do this is to create a class that extends from the `AbstractGuiceFactory` in
`gremlin-test` and provide that class to the `@CucumberOptions`. This approach does rely on the `ServiceLoader` which
means it will be important to include a `io.cucumber.core.backend.ObjectFactory` file in `META-INF/services` and an
entry that registers the custom implementation. Please see the TinkerGraph test code for further information on this
approach.

If implementing the Gherkin tests, providers can choose to opt-in to the slimmed down version of the normal JVM process
test suite to help alleviate test duplication between the two frameworks:

[source,java]
----
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_LIMITED_STANDARD)
@Graph.OptIn(Graph.OptIn.SUITE_PROCESS_LIMITED_COMPUTER)
----

=== 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.

Gremlin Server is distributed with a configuration that utilizes HTTP with a custom API.  Under this configuration,
Gremlin Server accepts requests containing a Gremlin script, evaluates that script and then streams back the results in
HTTP chunks.

Let's use the incoming request to process the Gremlin script of `g.V()` as an example.  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 into an HTTP chunk.  Each response is serialized given the requested serializer type (GraphBinary is
recommended) 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.

NOTE: TinkerPop provides a test server which may be useful for testing drivers. Details can be found
link:https://tinkerpop.apache.org/docs/current/dev/developer/#gremlin-socket-server-tests[here]

It is recommended but not required that a driver include a `User-Agent` header as part of any HTTP request to Gremlin
Server. Gremlin Server uses the user agent in building usage metrics as well as debugging. The standard format for
connection user agents is:

[[user-agent-format]]
`"[Application Name] [GLV Name].[Version] [Language Runtime Version] [OS].[Version] [CPU Architecture]"`
For example:
`"MyTestApplication Gremlin-Java.3.5.4 11.0.16.1 Mac_OS_X.12.6.1 aarch64"`

The following section provides an in-depth description of the TinkerPop HTTP API.  The HTTP API is used for
communicating the requests and responses that were described earlier.

=== HTTP API

This section describes the TinkerPop HTTP API which should be implemented by both graph system providers and graph
driver providers. There is only one endpoint that currently needs to be supported which is `POST /gremlin`. This
endpoint is a Gremlin evaluator which takes in a Gremlin script request and responds with the serialized results. The
formats below use a bit of pseudo-JSON to help represent request and response bodies. The actual format of the request
and response bodies will be determined by the serializers defined via the "Accept" and "Content-Type" headers. As a
result, a generic type definition in this document like "number" could translate to a "long" for a serializer that
supports types like GraphBinary.

==== HTTP Request

To formulate a request to Gremlin Server, a `RequestMessage` needs to be constructed.  The `RequestMessage` is a
generalized representation of a request. This message can be serialized in any fashion that is supported by Gremlin
Server, which by default is GraphBinary. An HTTP request that contains a `RequestMessage` has the following form:

[source,text]
----
POST /gremlin HTTP/1.1
Accept: <mimetype>
Content-Type: <mimetype>
Gremlin-Hints: <hints>

{
  "gremlin": string,
  "timeoutMs": number,
  "bindings": object,
  "g": string,
  "language" : string,
  "materializeProperties": string,
  "bulkResults": boolean
}
----

An actual, complete request might look like the following:

[source,text]
----
POST /gremlin HTTP/1.1
content-length: 61
host: 127.0.0.1
content-type: application/vnd.gremlin-v4.0+json
accept-encoding: deflate
accept: application/vnd.graphbinary-v4.0
user-agent: NotAvailable Gremlin-Java.4.0.0 11.0.25 Windows_11.10.0 amd64
{
    "gremlin": "g.V()",
    "language": "gremlin-lang"
}
----

===== Expected Request HTTP Headers

[width="100%",cols="3,10,3,3",options="header"]
|=========================================================
|Name |Description |Required |Default
|Accept |Serializer MIME types supported for the response. Must be a mimetype (see <<serializers>>). |No |`application/vnd.gremlin-v4.0+json;types=false`
|Accept-Encoding |The requested compression algorithm of the response. Valid values: `deflate`. |No |N/A
|Authorization |Header used with Basic authorization. |No |N/A
|Content-Length |The size of the payload |Yes |N/A
|Content-Type |The MIME type of the serialized body |No |None
|Gremlin-Hints |A semi-colon separated list of key/value pair metadata that could be helpful to the server in processing a particular request in some way. Must be a hints (see table below). |No |N/A
|User-Agent |The user agent. Follow the format specified by <<user-agent-format, user agent format>>. |No |<<user-agent-format, user agent format>>
|=========================================================

===== Request Header Value Options

[width="100%",cols="2,10",options="header"]
|=========================================================
|Name |Options
|mimetype |A MIME type listed in <<serializers>>.
|hints | mutations: yes, no, unknown - Indicates if the Gremlin contains steps that can mutate the graph.
|=========================================================

The body of the request should be a `RequestMessage` which is a `Map`. The `RequestMessage` should be serialized using
the serializer specified by the `Content-Type` header. The following are the key value pairs allowed in a
`RequestMessage`:

===== Request Message Format

[width="100%",cols="3,10,3,3",options="header"]
|=========================================================
|Key |Description |Value |Required
|gremlin |The Gremlin query to execute. |String containing script |Yes
|timeoutMs |The maximum time a query is allowed to execute in milliseconds. |Number between 0 and 2^31-1 |No
|bindings |A map used during query execution. Its usage depends on "language". For "gremlin-groovy", these are the variable bindings. For "gremlin-lang", these are the parameter bindings. |Object (Map) |No
|g |The name of the graph traversal source to which the query applies. Default: "g" |String containing traversal source name |No
|language |The name of the ScriptEngine to use to parse the gremlin query. Default: "gremlin-lang" |String containing ScriptEngine name |No
|materializeProperties |Whether to include all properties for results. One of "tokens" or "all". |String |No
|bulkResults |Whether the results should be bulked by the server (only applies to GraphBinary) |Boolean |No
|=========================================================

==== HTTP Response

When Gremlin Server receives that request, it will decode it given the "mime type", and execute it using the
`ScriptEngine` specified by the `language` field.  In this case, it will evaluate the script `g.V(x).out()` using the
`bindings` supplied in the `args` and stream back the results in HTTP chunks. When the chunks are combined, they will
form a single `ResponseMessage`.  The HTTP response containing the `ResponseMessage` has the following form:

[source,text]
----
HTTP/1.1 200
Content-type: <mimetype>
Transfer-Encoding: chunked
Gremlin-RequestId: <uuid>
{
  "result": list,
  "status": object
}
----

NOTE: While this response message is expected for all serialized responses, there may be some errors that are not
serialized. In that case, the `Content-Type` of the response should be `application/json` and the JSON should contain a
`message` key.

===== Response Message Format

[width="100%",cols="2,10a",options="header"]
|=========================================================
|Key |Description
|result |A map that contains the result data.
[width="100%",cols="3,10,3,3",options="header"]
!=========================================================
!Name !Description !Required !Default
!data !A list of result objects. !Array !Yes
!=========================================================
|status |A map that contains the status of the result.
[width="100%",cols="3,10,3,3",options="header"]
!=========================================================
!Name !Description !Required !Default
!code !The actual <<http-status-codes,status code>> of the result. !Number !Yes
!exception !A class of exception if an error occurred. !String !No
!message !The error message if an error occurred. !String !No
!=========================================================
|=========================================================

===== Expected Response HTTP Headers

[width="100%",cols="3,10,3,3",options="header"]
|=========================================================
|Name |Description |Required |Default
|Content-Type |The MIME type of the serialized body which is based on the request's `Accept` header. May also be "application/json". |Yes |N/A
|Gremlin-RequestId |The server generated UUID that is used as a request ID. |Yes |N/A
|Transfer-Encoding |The server should attempt to chunk all responses. |No |"chunked"
|=========================================================

===== Response Header Value Options

[width="100%",cols="3,10",options="header"]
|=========================================================
|Name |Options
|mimetype |A MIME type listed in <<serializers>>.
|uuid |A randomly generated UUID string.
|=========================================================

[[http-status-codes]]
===== Response Status Codes

The following table details the HTTP status codes that Gremlin Server will send:

[width="100%",cols="1,3,9",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.
|400 |BAD REQUEST |There was a problem with the HTTP request.
|401 |UNAUTHORIZED |The request attempted to access resources that the requesting user did not have access to.
|403 |FORBIDDEN |The server could authenticate the request, but will not fulfill it.
|404 |NOT FOUND |The server was unable to find the requested resource.
|405 |METHOD NOT ALLOWED |The request used an unsupported method. The server only supports POST.
|413 |REQUEST ENTITY TOO LARGE |The request was too large or the query could not be compiled due to size limitations.
|500 |INTERNAL SERVER ERROR |A general server error occurred that prevented the request from being processed.
|505 |HTTP VERSION NOT SUPPORTED |A server error indicating that an unsupported version of HTTP is being used. Only HTTP/1.1 is supported.
|=========================================================

===== Trailing Headers

Error responses will have trailing headers in addition to the status object in the response body. This information is
duplicated and should be the same, so graph driver providers should use whichever is easier for them. The trailers,
however, will only contain the `Status` and `Exception` without the `Message`.

==== HTTP Examples

For examples of actual requests and responses, take a look at the IO documentation for
link:https://tinkerpop.apache.org/docs/x.y.z/dev/io/#_requestmessage[GraphSON requests] and
link:https://tinkerpop.apache.org/docs/x.y.z/dev/io/#_responsemessage[GraphSON responses].

=== HTTP Request Interceptor

A graph driver may support HTTP request intercepting which provides a means for the user of your graph driver to update
the headers and body of the HTTP request before it is sent to the server.  This enables use cases where a graph system
provider's server implementation has additional capabilities that aren't included in the base Gremlin Server.  Although
every graph system provider is expected to support the protocol defined by the TinkerPop HTTP API, this doesn't
preclude them from including additional functionality.  Be aware that if you choose to not provide this functionality,
then your graph driver may not have access to some graph provider's features, or, possibly, it may not be able to
connect at all.

=== Authentication and Authorization

By default, Gremlin Server only supports
link:https://en.wikipedia.org/wiki/Basic_access_authentication[basic HTTP authentication].  This is handled by the
`HttpBasicAuthenticationHandler` which is the only `AbstractAuthenticationHandler` provided with the Gremlin Server.
Other common HTTP authentication schemes that are sent via an HTTP header can be supported by implementing a custom
`AbstractAuthenticationHandler`.  Because the communication protocol is HTTP/1.1, authentication should be header-based
and should not include negotiation.

When basic authentication is enabled, an incoming request is intercepted before it is evaluated by the `ScriptEngine`.
The request is examined for an `Authorization` header. If one doesn't exist then "401 Unauthorized" error response is
returned.

In addition to authenticating users at the start of a connection, Gremlin Server allows providers to authorize users on
a per request basis. If
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#_configuring_2[a java class is configured] that implements the
link:https://tinkerpop.apache.org/javadocs/x.y.z/full/org/apache/tinkerpop/gremlin/server/authz/Authorizer.html[Authorizer interface],
Gremlin Server passes each request to this `Authorizer`. The `Authorizer` can deny authorization for the request by
throwing an exception and Gremlin Server returns `UNAUTHORIZED` (status code `401`) to the client. The `Authorizer`
authorizes the request by returning the original request or the request with some additional constraints. Gremlin Server
proceeds with the returned request and on its turn returns the result of the request to the client. More details on
implementing authorization can be found in the
link:https://tinkerpop.apache.org/docs/x.y.z/reference/#security[reference documentation for Gremlin Server security].

NOTE: While Gremlin Server supports this authorization feature it is not a feature that TinkerPop requires of graph
providers as part of the agreement between client and server.

[[serializers]]
=== Serializers

In order to serialize and deserialize the requests and responses, your graph driver will need to implement
link:https://tinkerpop.apache.org/docs/x.y.z/dev/io/#graphbinary[GraphBinary]. The Gremlin Server is capable of
returning both GraphBinary and GraphSON, however, GraphBinary is a more compact format which can lead to increased
performance as fewer bytes need to be sent through the wire. For this reason, drivers only need to support GraphBinary.
link:https://tinkerpop.apache.org/docs/x.y.z/dev/io/#graphson[GraphSON] can be used by applications that only support JSON serialization.

The following table lists the serializers supported by the Gremlin Server and their MIME types. These MIME types should
be used in the `Content-Type` and `Accept` HTTP headers.

[width="100%",cols="3,5,5",options="header"]
|=========================================================
|Name |Description |MIME type
|Untyped GraphSON 4.0 |A JSON-based graph format |application/vnd.gremlin-v4.0+json;types=false
|Typed GraphSON 4.0 |A JSON-based graph format with embedded type information used for serialization |application/vnd.gremlin-v4.0+json;types=true
|GraphBinary 4.0 |A binary graph format |application/vnd.graphbinary-v4.0
|=========================================================

==== IO Tests

The IO test suite is a collection of files that contain the expected outcome of serialization of certain types. These
tests can be used to determine if a particular serializer has been correctly implemented. In general, a driver should
be able to "round trip" each of these types. That is, it should be able to both read from and write to those exact same
bytes. Not all programming languages provide library types that will match the specification of the corresponding type
defined by the serializer. In this case, it is not possible to completely round trip that type and you may skip that
test. The GraphBinary test files can be found
link:https://github.com/apache/tinkerpop/tree/x.y.z/gremlin-test/src/main/resources/org/apache/tinkerpop/gremlin/structure/io/graphbinary[here].
The link:https://github.com/apache/tinkerpop/blob/x.y.z/gremlin-util/src/test/java/org/apache/tinkerpop/gremlin/structure/io/AbstractTypedCompatibilityTest.java:[Java implementation]
can be used as a reference on how these files can be used and its
link:https://github.com/apache/tinkerpop/blob/x.y.z/gremlin-util/src/test/java/org/apache/tinkerpop/gremlin/structure/io/Model.java[model]
shows the Java representation of those files.

[[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).

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`.

include::gremlin-semantics.asciidoc[]

include::policies.asciidoc[]