Chapter 10. Things to Consider in a Multi-Node JanusGraph Cluster

JanusGraph is a distributed graph database, which means it can be setup in a multi-node cluster. However, when working in such an environment, there are important things to consider. Furthermore, if configured properly, JanusGraph handles some of these special considerations for the user.

10.1. Dynamic Graphs

JanusGraph supports dynamically creating graphs. This is deviation from the way in which standard Gremlin Server implementations allow one to access a graph. Traditionally, users create bindings to graphs at server-start, by configuring the gremlin-server.yaml file accordingly. For example, if the graphs section of your yaml file looks like this:

graphs {
  graph1: conf/graph1.properties,
  graph2: conf/graph2.properties
}

then you will access your graphs on the Gremlin Server using the fact that the String graph1 will be bound to the graph opened on the server as per its supplied properties file, and the same holds true for graph2.

However, if we use the ConfiguredGraphFactory to dynamically create graphs, then those graphs are managed by the JanusGraphManager and the graph configurations are managed by the ConfigurationManagementGraph. This is especially useful because it 1. allows you to define graph configurations post-server-start and 2. allows the graph configurations to be managed in a persisted and distributed nature across your JanusGraph cluster.

To properly use the ConfiguredGraphFactory, you must configure every Gremlin Server in your cluster to use the JanusGraphManager and the ConfigurationManagementGraph. This procedure is explained in detail here.

10.1.1. Graph Reference Consistency

If you configure all your JanusGraph servers to use the ConfiguredGraphFactory, JanusGraph will ensure all graph representations are-up-to-date across all JanusGraph nodes in your cluster.

For example, if you update or delete the configuration to a graph on one JanusGraph node, then we must evict that graph from the cache on every JanusGraph node in the cluster. Otherwise, we may have inconsistent graph representations across your cluster. JanusGraph automatically handles this eviction using a messaging log queue through the backend system that the graph in question is configured to use.

If one of your servers is configured incorrectly, then it may not be able to successfully remove the graph from the cache.

[Important]Important

Any updates to your TemplateConfiguration will not result in the updating of graphs/graph configurations previously created using said template configuration. If you want to update the individual graph configurations, you must do so using the available update APIs. These update APIs will then result in the graphe cache eviction across all JanusGraph nodes in your cluster.

10.1.2. Dynamic Graph and Traversal Bindings

JanusGraph has the ability to bind dynamically created graphs and their traversal references to <graph.graphname> and <graph.graphname>_traversal, respectively, across all JanusGraph nodes in your cluster, with a maximum of a 20s lag for the binding to take effect on any node in the cluster. Read more about this here.

JanusGraph accomplishes this by having each node in your cluster poll the ConfigurationManagementGraph for all graphs for which you have created configurations. The JanusGraphManager will then open said graph with its persisted configuration, store it in its graph cache, and bind the <graph.graphname> to the graph reference on the GremlinExecutor as well as bind <graph.graphname>_traversal to the graph’s traversal reference on the GremlinExecutor.

This allows you to access a dynamically created graph and its traversal reference by their string bindings, on every node in your JanusGraph cluster. This is particularly important to be able to work with Gremlin Server clients and use TinkerPops’s withRemote functionality.

10.1.2.1. Set Up

To set up your cluster to bind dynamically created graphs and their traversal references, you must:

  1. Configure each node to use the ConfiguredGraphFactory.
  2. Configure each node to use a JanusGraphChannelizer, which injects lower-level Gremlin Server components, like the GremlinExecutor, into the JanusGraph project, giving us greater control of the Gremlin Server.

To configure each node to use a JanusGraphChannelizer, we must update the gremlin-server.yaml to do so:

channelizer: org.janusgraph.channelizers.JanusGraphWebSocketChannelizer

There are a few channelizers you can choose from:

  1. org.janusgraph.channelizers.JanusGraphWebSocketChannelizer
  2. org.janusgraph.channelizers.JanusGraphHttpChannelizer
  3. org.janusgraph.channelizers.JanusGraphNioChannelizer
  4. org.janusgraph.channelizers.JanusGraphWsAndHttpChannelizer

All of the channelizers share the exact same functionality as their TinkerPop counterparts.

10.1.2.2. Using TinkerPop’s withRemote Functionality

Since traversal references are bound on the JanusGraph servers, we can make use of TinkerPop’s withRemote functionality. This will allow one to run gremlin queries locally, against a remote graph reference. Traditionally, one runs queries against remote Gremlin Servers by sending String script representations, which are processed on the remote server and the response serialized and sent back. However, TinkerPop also allows for the use of remoteGraph, which could be useful if you are building a TinkerPop compliant graph infrastructure that is easily transferable to multiple implementations.

To use this functionality in JanusGraph, we must first ensure we have created a graph on the remote JanusGraph cluster:

ConfiguredGraphFactory.create("graph1");

Next, we must wait 20 seconds to ensure the traversal reference is bound on every JanusGraph node in the remote cluster.

Finally, we can locally make use of the withRemote method to access a local reference to a remote graph:

gremlin> cluster = Cluster.open('conf/remote-objects.yaml')
==>localhost/127.0.0.1:8182
gremlin> graph = EmptyGraph.instance()
==>emptygraph[empty]
gremlin> g = graph.traversal().withRemote(DriverRemoteConnection.using(cluster, "graph1_traversal"))
==>graphtraversalsource[emptygraph[empty], standard]

For completion, the above conf/remote-objects.yaml should tell the Cluster API how to access the remote JanusGraph servers; for example, it may look like:

hosts: [remoteaddress1.com, remoteaddress2.com]
port: 8182
username: admin
password: password
connectionPool: { enableSsl: true }
serializer: { className: org.apache.tinkerpop.gremlin.driver.ser.GryoMessageSerializerV3d0, config: { ioRegistries: [org.janusgraph.graphdb.tinkerpop.JanusGraphIoRegistry] }}