Modeling designs

As with any database, the data model that you design is important in determining the logic your queries and the structure of data in storage. This practice extends to graph databases, with one exception. Neo4j is schema-free, which means that your data model can adapt and change easily with your business.

Need to start collecting a new field and capture new analysis? Or need to change the way you interpret a customer or other entity and modify its definition? Or regulation requires systems to capture less information or restrict readability (change data format/types)?

You may have worked for a company where each area or department defines a domain differently. Take, for instance, a generic customer domain. To different areas within the business, a customer can be defined as different types of individuals. These definitions may also change over time or the company may decide to unify the meaning of a customer across departments.

If you have worked with other types of databases, you will already be familiar with the development and administrative work that any of these scenarios entail. However, Neo4j allows you to effortlessly adjust detailed and broad changes across pieces or the entirety of the graph. Whether it is small changes over time or a broad definition that includes a variety of needed information about your entities, the database is able to handle it. It is simply up to the developers and architects to determine the structure of the data model and how to define entities for queries.

In the next few paragraphs, we will introduce a few different ways to look at different data sets and show how each impacts queries and performance for traversing graph data.

One of the earliest decisions you may encounter is whether to model something as a property on a node or as a relationship to a separate node. Take, for example, the data below modeling a movie genre as a property on the Movie node.

To write a query finding the genre(s) of a particular movie is very simple. It would find the Movie node it wants to know about, then return the values listed in the genre property. However, to find out which movies share genres, you would need a much more complex query to find each Movie node, loop through each of the genres in the property array, and compare with each value in the second movie’s property array of genres. This would take a toll on performance (nested looping and comparison of node properties), and the query would be much more complicated, as well.

The code block below is what the syntax would look like for each query. You can see the shift in logic and complexity of the loop in the second query.

Now, instead, if you were to model movies and their genres as separate nodes and create a relationship between them, you would come up with a model like the Figure 2.

This creates a completely separate entity (node) for the genre, allowing you to connect all the movies with a shared genre to that Genre node. Let us see how this changes our queries. To find the genres of a particular movie, it first needs to find the Movie node it is looking for (in this case, 'The Matrix'), then find the node that is connected to that movie through the IN_GENRE relationship.

The biggest difference is in the syntax for the second query to find which movies share genres. It is much simpler than our earlier version because it uses a natural, graph pattern (entity-relationship-entity) to find the information needed. First, Cypher^® finds a movie and the genre it is related to, then looks for a second movie that is in that same genre.

Neither version of the data model is worse or better, but the 'best' option highly depends on the types of queries you intend to run against your data.

If you plan to do analysis on individual items and return only details about that entity (like genres on a particular movie), then the first data model would serve perfectly well for your needs. However, if you need to run analysis to find common ground between entities or look at a group of nodes, then the second data model would definitely improve performance of those types of queries.

As many of us can probably agree, not all data models are simple and straightforward. Data is messy, and the model must attempt to better-organize it to help us see patterns and make decisions.

One excellent example of a complex data structure that is difficult to model is Marvel comic data. In the Marvel universe, there are comics that have characters who make appearances or play lead roles. Comics can be organized into a series of particular storylines or narratives for a certain time, and major events can take place in a comic that define a character path or series. Creators (including writers, illustrators, etc) are the authors of comics, defining storyline, character adaptations, and events that happen. Multiple creators can also participate interchangeably to create a comic or series.

This dataset already seems complicated, with several entities and relationships at work. It adds a new layer of complexity when trying to model the hierarchies and intermediate entities that exist here.

If you have some time, you can view the full video link to Peter’s presentation on Vimeo, but we want to highlight two key challenges that Peter discusses in the data set.

First, he found that comic characters tend to be extremely dynamic. Many characters cannot be identified by name or costume or any particular property, as all of those change often.

Second, Peter identified the issue of chronology. For those new to the comic universe, some might want to determine where to start or what comic(s) come next. However, comic issues are not always sequentially numbered, and there are even some storylines that appear across multiple series and back again. This makes it incredibly difficult to separate certain blocks of stories or events, along with renditions of characters.

One way to model time-specific data and relationships is by including data in the relationship type. Because Neo4j is optimized specifically for traversing relationships between entities, you can often improve query performance by specifying a date as the relationship type and only traversing particular dated relationships.

A common example is for modeling airline flights. An airline has a particular flight on a certain day from and to a specific location. We might start with a model like the Figure 4 below to show how flights travel from airport to airport.

We would soon realize that we need to model a Flight entity that exists between two destinations because multiple planes can travel between two destinations several times in one day.

However, your queries probably still show the model’s weakness in filtering through all of the flights at a specific airport - especially for London and other major cities that have hundreds of flights connected to an Airport node over any span of time. Inspecting the several properties of each Flight node could be expensive on resources.

If we were to create a node for a particular airport day and a relationship with a date in the type, then we could write queries to find flights from an airport on any specified date (or date range). This way, you wouldn’t need to check each flight relationship to an airport. Instead, you would only look at the relationships for the dates you cared about. This model turns out like the one below.

For the full walkthrough of the modeling process for airline flights, see Blog post: Modeling Airline Flights in Neo4j.

Sometimes, you might find that one model works really well for one scenario you need, but another model is better for something else. For instance, some models will perform better with write queries and other models handle read queries better. Both capabilities are important to your use case, so what do you do?

In these cases, you can combine both models and use the benefits of each. Yes, you can use more than one data model in your graph!

The tradeoff is that now you will need to maintain two models. Each time you create a new node or relationship or update pieces of the graph, you will need to make changes to accommodate both models. This can also impact query performance, as you might have double the syntax needed to update each model.

While this is definitely a possible option, you should know the maintenance costs and evaluate whether those costs are overcome by the performance improvements you will see for each needed query. If so, being able to use more than one data model is a great solution!