We’re overhauling Dgraph’s docs to make them clearer and more approachable. If you notice any issues during this transition or have suggestions, please let us know.

When building an app, you might wonder which database is the best choice. A traditional relational database that you can query using SQL is a familiar choice, but does a relational database really provide a natural fit to your data model, and the performance that you need if your app goes viral and needs to scale up rapidly?

This tutorial takes a deeper look at data modeling using relational databases compared to graph databases like Dgraph, to give you a better understanding of the advantages of using a graph database to power your app. If you aren’t familiar with graph data models or graph databases, this tutorial was written for you.

​
Learning goals

In this tutorial, you’ll learn about graphs, and how a graph database is different from a database built on a relational data model. This tutorial doesn’t include any code or syntax, but rather a comparison of graphs and relational data models. By the end of this tutorial, you should be able to answer the following questions:

  • What’s a graph?
  • How are graphs different from relational models?
  • How’s data modeled in a graph?
  • How’s data queried from a graph?

Along the way, you might find that a graph is the right fit for the data model used by your app. Any data model that tracks lots of different relationships (or edges) between various data types is a good candidate for a graph model.

Whether this is the first time you are learning about graphs or looking to deepen your understanding of graphs with some concrete examples, this tutorial should help you along your journey.

If you are already familiar with graphs, you can jump right into our coding example for React.

​
Graphs and natural data modeling

Graphs provide an alternative to tabular data structures, allowing for a more natural way to store and retrieve data.

For example, you could imagine that we’re modeling a conversation within a family:

  • A father, who starts a conversation about going to get ice cream.
  • A mother, who comments that she would also like ice cream.
  • A child, who likes the idea of the family going to get ice cream.

This conversation could easily occur in the context of a modern social media or messaging app, so you can imagine the data model for such an app as follows:

For the remainder of this guide, we use this as our example app: a basic social media or messaging app, with a data model that includes people, posts, comments, and reactions.

A graph data model is different from a relational model. A graph focuses on the relationships between information, whereas a relational model focuses on storing similar information in a list. The graph model received its name because it resembles a graph when illustrated.

  • Data objects are called nodes and are illustrated with a circle.
  • Properties of nodes are called predicates and are illustrated as a panel on the node.
  • Relationships between nodes are called edges and are illustrated as connecting lines. Edges are named to describe the relationship between two nodes. A reaction is an example of an edge, in which a person reacts to a post.

Some illustrations omit the predicates panel and show only the nodes and edges.

Referring back to the example app, the father, mother, child, post, and comment are nodes. The name of the people, the post’s title, and text of the comment are the predicates. The natural relationships between the authors of the posts, authors of the comments, and the comments’ topics are edges.

As you can see, a graph models data in a natural way that shows the relationships (edges) between the entities (nodes) that contain predicates.

​
Relational Data Modeling

This section considers the example social media app introduced in the previous section and discusses how it could be modeled with a traditional relational data model, such as those used by SQL databases.

With relational data models, you create lists of each type of data in tables, and then add columns in those tables to track the attributes of that table’s data. Looking back on our data, we remember that there are three main types, People, Posts, and Comments

To define relationships between records in two tables, a relational data model uses numeric identifiers called foreign keys, that take the form of table columns. Foreign keys can only model one-to-many relationship types, such as the following:

  • The relationship from Posts to People, to track contributors (authors, editors, etc.) of a Post
  • The relationship from Comments to People, to track the author of the comment
  • The relationship from Comments to Posts, to track on which post comments were made
  • The relationship between rows in the Comments table, to track comments made in reply to other comments (a self-reference relationship)

The limitations of foreign keys become apparent when your app requires you to model many-to-many relationships. In our example app, a person can like many posts or comments, and posts and comments can be liked by many people. The only way to model this relationship in a relational database is to create a new table. This so-called pivot table usually doesn’t store any information itself, it just stores links between two other tables.

In our example app, we decided to limit the number of tables by having a single “Likes” table instead of having people_like_posts and people_like_comments tables. None of these solutions is perfect, though, and there is a trade-off between having a lower table count or having more empty fields in our tables (also known as “sparse data”).

Because foreign keys can’t be added in reference to entities that don’t exist, adding new posts and authors requires additional work. To add a new post and a new author at the same time (in the Posts and People tables), we must first add a row to the People table and then retrieve their primary key and associate it with the new row in the Posts table.

By now, you might ask yourself: how does a relational model expand to handle new data, new types of data, and new data relationships?

When new data is added to the model, the model changes to accept the data. The simplest type of change is when you add a new row to a table. The new row adopts all of the columns from the table. When you add a new property to a table, the model changes and adds the new property as a column on every existing and future row for the table. And when you add a new data type to the database, you create a new table with its own pre-defined columns. This new data type might link to existing tables or need more pivot tables for a new many-to-many relationship. So, with each data type added to your relational data model, the need to add foreign keys and pivot tables increases, making support for querying every potential data relationship increasingly unwieldy.

Properties are stored as new columns and relationships require new columns and sometimes new pivot tables. Changing the schema in a relational model directly effects the data that’s held by the model, and can impact database query performance.

​
Graph Data Modeling

In this section we take our example social media app and see how it could be modeled in a graph.

The concept of modeling data in a graph starts by placing dots, which represent nodes. Nodes can have one or more predicates (properties). A person may have predicates for their name, age, and gender. A post might have a predicate value showing when it was posted, and a value containing the contents of the post. A comment would most likely have a predicate containing the comment string. However, any one node could have other predicates that aren’t contained on any other node. Each node represents an individual item, hence the singular naming structure.

As graphs naturally resemble the data you are modeling, the individual nodes can be moved around this conceptual space to clearly show the relationships between these data nodes. Relationships are formed in graphs by creating an edge between them. In our app, a post has an author, a post can have comments, a comment has an author, and a comment can have a reply.

For sake of illustration we also show the family tree information. The father and the mother are linked together with a spouse edge, and both parents are related to the child along a child edge.

With a graph, you can also name the inverse relations. From here we can quickly see the inverse relationships. A Post has an Author and a Person has Posts. A Post has Comments and a Comment is on a Post. A Comment has an Author, and a Person has Comments. A Parent has a Child, and a Child has a Parent.

You create many-to-many relationships in the same way that you make one-to-many relationships, with an edge between nodes.

Adding groups of related data occurs naturally within a graph. The data is sent as a complete object instead of separate pieces of information that needs to be connected afterwards. Adding a new person and a new post to our graph is a one-step process. New data coming in doesn’t have to be related to any existing data. You can insert this whole data object with 3 people, a post, and a comment all in one step.

When new data is added to the model, the model changes to accept the data. Every change to a graph model is received naturally. When you add a new node with a data type, you are simply creating a new dot in space and applying a type to it. The new node doesn’t include any predicates or relationships other than what you define for it. When you want to add a new predicate onto an existing data type, the model changes and adds the new property onto the items that you define. Other items not specifically given the new property type aren’t changed. When you add a new data type to the database, a new node is created, ready to receive new edges and predicates.

The key to remember when modeling a graph is to focus on the relationships between nodes. In a graph you can change the model without affecting the underlying data. Because the graph is stored as individual nodes, you can adjust predicates of individual nodes, create edges between sets of nodes, and add new node types without affecting any of the other nodes.

​
Query Data in a Relational Model

Storing our data is great, but the best data model would be useless without the ability to query the data our app requires. So, how does information get retrieved in a relational model compared to a graph model?

In a relational model, tables are stored in files. To support the sample social media app described in this tutorial, you would need four files: People, Posts, Comments, and Likes.

When you request data from a file, one of two things happens: either a table scan takes place or an index is invoked. To find this data, the whole file must be read until the data is found or the end of the file is reached. In our example app there is a post titled, “Ice Cream?”. If the title isn’t indexed, every post in the file would need to be read until the database finds the post entitled, “Ice Cream?”. This method would be like reading the entire dictionary to find the definition of a single word: very time-consuming. This process could be optimized by creating an index on the post’s title column. Using an index speeds up searches for data, but it can still be time-consuming.

​
What’s an index?

An index is an algorithm used to find the location of data. Instead of scanning an entire file looking for a piece of data, an index is used to aggregate the data into “chunks” and then create a decision tree pointing to the individual chunks of data. Such a decision tree could look like the following:

Relational data models rely heavily on indexes to quickly find the requested data. Because the data required to answer a single question usually lives in multiple tables, you must use multiple indexes each time that related data is joined together. And because you can’t index every column, some types of queries won’t benefit from indexing.

​
How data is joined in a relational model

In a relational model, the request’s response must be returned as a single table consisting of columns and rows. To form this single table response, data from multiple tables must be joined together. In our app example, we found the post entitled “Ice Cream?” and also found the comments, “Yes!”, “When?”, and “After Lunch”. Each of these comments also has a corresponding author: Mother, Child, and Father. Because there is only one post as the root of the join, the post is duplicated to join to each comment.

Flattening query results can lead to many duplicate rows. Consider the case where you also want to query which people liked the comments on this example post. This query requires mapping a many-to-many relationship, which invokes two additional index searches to get the list of likes by person.

Joining all of this together would form a single table containing many duplicates: duplicate posts and duplicate comments. Another side effect of this response approach is that it’s likely that empty data exists in the response.

In the next section, you’ll see that querying a graph data model avoids the issues that you would face when querying a relational data model.

​
Query Data in a Graph Model

As you’ll see in this section, the data model we use determines the ease with which we can query for different types of data. The more your app relies on queries about the relationships between different types of data, the more you benefit from querying data using a graph data model.

In a graph data model, each record (a person, post or comment) is stored as a data object (sometimes also called a node). In the example social media app described in this tutorial, there are objects for individual people, posts, and comments.

When data is requested from the graph, a root function determines which nodes are selected for the starting points. This root function uses indexes to determine which nodes match quickly. In our app example, we want to start with the root being the post with the title “Ice Cream?”. This type of lookup evokes an index on the post’s title, much like indexes work in a relational model. The indexes at the root of the graph use the full index tree to find the data.

Connecting edges together to form a connected graph is called traversal. After arriving at our post, “Ice Cream?”, we traverse the graph to arrive at the post’s comments. To find the post’s author, we traverse the next step to arrive at the people who authored the comment. This process follows the natural progression of related data, and graph data models allow us to query our data to follow this progression efficiently.

What do we mean by efficiently? A graph data model lets you traverse from one node to a distantly related node without the need for anything like pivot tables. This means that queries based on edges can be updated easily, with no need to change the schema to support new many-to-many relationships. And, with no need to build tables specifically for query optimization, you can adjust your schema quickly to accommodate new types of data without adversely impacting existing queries.

A feature of a graph model is that related edges can be filtered anywhere within the graph’s traversing. When you want to know the most recent comment on your post or the last person to like the comment, filters can be applied to the edge.

When filters get applied along an edge, only the nodes that match the edge are filtered - not all of the nodes in the graph. Applying this logic reduces the size of the graph and makes index trees smaller. The smaller an index tree is, the faster that it can be resolved.

In a graph model, data is returned in an object-oriented format. Any related data is joined to its parent within the object in a nested structure.

{
  "title": "IceCream?",
  "comments": [
    {
      "title": "Yes!",
      "author": {
        "name": "Mother"
      }
    },
    {
      "title": "When?",
      "author": {
        "name": "Child"
      }
    },
    {
      "title": "After Lunch",
      "author": {
        "name": "Father"
      }
    }
  ]
}

This object-oriented structure allows data to be joined without duplication.

​
Conclusion

Congratulations on finishing the Dgraph learn course Graph Data Models 101!

Now that you have an overview and understanding of

  • what a graph is
  • how a graph differs from a relational model
  • how to model a graph
  • and how to query a graph

you are ready to jump into using Dgraph, the only truly native distributed graph database.