From Tables to Graphs: How to Model Entities in a Graph Database
Fluree is a graph database, and graph databases use a different paradigm for modeling data than you're used to with relational databases. This guide will lead you from the relational world to the graph world, pointing out the traps you might encounter when trying to translate your relational intuition to this new environment. By the end of the guide, you should be able to start creating data models for your systems that take advantage of the possibilities that a graph database affords.
Breaking out of the rectangle: modeling data as graphs
To start, we must separate data modeling from schema design.
Data modeling, by my definition, is the process of figuring out what parts of the world we care about, and how to represent them. It involves identifying what entity types we want to record information about, and what attributes we want to use to describe those entities. It can also include the logical rules governing the lifecycles of those entities, the relationships among them, and any security concerns.
Data modeling takes place at the conceptual level, and is meant to convey your understanding of the business domain. We often record and share this understanding using entity relationship diagrams (ERDs). Here's an example ERD which shows how you model a system where customers place orders:
Schema design is the process of translating a data model into an implementation that your data store understands. It's about specifying the columns your tables will contain, as well as metadata like their indexes. Schema design is also where you think about how to enforce the constraints of your data model, e.g. by choosing the correct column data types, adding uniqueness constraints, and so on.
You can also use ERDs to convey schema designs. Notice in this ERD it contains information on primary keys and foreign keys, information that's not needed for a data model:
In the relational world, you must define table schemas to store data. These schemas implement our understanding of our domain's data model. They give us something concrete to refer to to make sure that the data we're recording correctly matches our understanding of the domain, and they enforce constraints on the data, giving us some guardrails to make sure we don't record invalid data. Table schemas are where your conceptual understanding of how you want to represent the world meets the cold hard reality of how you're actually storing data.
In the graph world, you don't have to define any schemas to start storing data. You just start storing data. In the examples we've seen so far, we just throw together any triples we feel like. This feels one hundred percent exactly like crossing a high wire one thousand feet above the earth -- without a safety net -- while juggling chainsaws. How do you cope?
When migrating from the relational world to the graph world, it's unclear how to transfer your skillset over. What applies, and what doesn't? Are your approaches to data modeling still relevant? What about schema design?
To give you just a couple examples:
- In the relational world, you have to create a join table to capture many-to-many relationships. In the graph world, nodes can just directly reference each other without an intermediary.
- In the relational world, you must store data in tables and tables necessarily constrain what attributes can be associated with an entity. Not so in the graph world - by default, entities can have any attributes.
While schema design and data modeling are interconnected, this document focuses on data modeling. It touches on schema design to the extent that it points out how our table-oriented notions of schema design percolate up into our data modeling, but it doesn't get into any specifics about how to implement schemas for your graph data store; that's a large enough topic that it deserves its own, separate explanation.
With that understanding of what we mean by "data modeling", let's look at the different ways we can record the data that we're attempting to model, and how the structure we use (graph or table), influences our approach to data modeling.
Representing Entities with Tables
If you're an experienced database practitioner, your brain is now wired at a fundamental level to think of data as rows in tables, so to rewire it we'll need to revisit the fundamentals. It's like making the transition from imperative programming to functional programming, or even learning programming at all. What you thought was the way of representing the world was really only a way of representing the world.
So, fundamentals: To record information about the world, we organize symbols using some kind of structure. The text that you're reading, for example, is written English, which is a phonetic system for mapping symbols to spoken language. It organizes letter symbols into word symbols, and then organizes those into sentences using punctuation symbols and white space.
While it's possible touse full English sentences to record all the information we care about, it would be hard to write programs that can query that information and make connections across datasets. Instead we rely on additional structures to encode information.
A table is a visual structure for organizing information that lets us consistently and compactly describe an entity's properties. I'm talking about tables in general here, not just database tables: tables that you draw on paper, or excel sheets, or even nutrition labels. Take this table of tea information:
|Uji Gyokuro Yume no Ukihashi
|Uji Gyokuro Fujitsubo
Each row represents some entity, some thing in the universe that we want to
describe. Each column represents an attribute of that entity, some property of
the entity that we're interested in. The intersection of row and column is a
cell, and cells record a specific value for that entity and attribute. For
example, the third row describes an entity that has the
Shirakawa Gyokuro, a
Brew Temp of
50 (celsius), and a
Brew Time of
The table is probably one of the oldest structures we humans use for organizing information. As far as technology goes, the humble table is wildly successful. If you could buy stock in "table" back in like 10,000 b.c. you would be very rich by now.
Its longevity is due in part, I think, to its intuitive appeal. The visual processing systems in our brains are able to quickly parse and navigate the information stored in tables.
But the information itself is not inherently tabular. Our brains are just used to thinking about it that way. Tables are just one way of representing it. Anything you can represent with a table, you can also represent with a graph. Let's have a look.
Representing Entities with Graphs
Here's how you could organize the same data into a graph:
This graph encodes the same information as the table above, capturing the
relationships between entities, attributes, and values. The nodes
t3 correspond to rows in a database. Each of these nodes has three
directed edges labeled
brew temp, and
brew time, and pointing to
values which correspond to the cell values in a table.
A directed edge indicates that the relationship exists only one way:
t2 has a
brew time of
150 does not have a
brew time of
starting node is called the tail and the ending node is called the head.
So tables and graphs are two different ways of organizing symbols to record information. There's no information in the graph example that's not contained in the table example, and vice versa.
Still, there are some notable differences between how we're able to use these two representations, including:
- Tables contain less repetition than graphs
- The visual structure of tables makes it easier for us as humans to answer some kinds of questions by scanning the table. For example, it's easier to answer the question, "Do we have information about Shirakawa Gyokuro?"
- All entities are explicitly named in graphs, whereas with tables entities are implied by the row. You can refer to an entity by e.g. row number, or by the value of one of the columns, but this kind of naming is ad hoc
There are other differences between the representations, and we'll talk about those more later. The main takeaway for now, though, is that it's possible to structure the information stored in tables as graphs.
Entities, Attributes, and Values
Tables and graphs are two different ways of structuring entities, attributes, and values. Their different organization systems are key to understanding the nuances in how we approach data modeling between table world and graph worrld.
The combination of entity, attribute, and value is the atomic unit of information. This combination is how we say something about something. "Daniel has a cat", "the cat is named Potato", "Potato is yelling" - all of these statements express a relationship between two things, and these relationships can all be captured using EAV (entity, attribute, value) triples. (If you're familiar with RDF, EAV triples are another, more generic way of talking about subject, predicate, object triples.) Triples are also called statements or facts.
A triple is an ordered sequence of three phrases:
person, name, Danielperson, has, catcat, name, Potatocat, yelling?, true
Or, to take our tea example:
t1, name, Uji Gyokuro Yume no Ukihashit1, brew temp, 60t1, brew time, 150-180
Tables organize triples by displaying them in a grid. While we tend to think of "rows" as the atomic unit of information, really you can conceptually break down rows further into aggregates of cells that describe the attributes and values for a single entity.
Graphs organize triples by displaying them as nodes with directed edges. Each combination of head, directed edge, tail captures a triple. With graphs you don't inherently need to constrain the possible attributes.
If you're recording information purely as EAV triples, there is no inherent notion of "entity type". It's possible and reasonable to record information as collections of EAV triples, without introducing constraints or trying to categorize groups of triples.
When you define a table, however, you must define its columns. Therefore, tables by their very nature constrain what attributes are associated with the entities recorded in that table. If the table doesn't have a column for an attribute, you can't record that attribute in that table.
This isn't the case with graphs! This understanding of how tables and graphs represent EAV triples, and the limits that tables introduce, will help us understand how to model data outside of the world of tables.
Data modeling skills still apply, but the world is expanded
How do we do data modeling with graph data? Good news: you can still rely on the data modeling techniques that you're used to. Tools like Entity Relationship Diagrams are just as useful for graph data as they are for relational data. You might have other implementation-agnostic tools at your disposal like Domain Driven Design to discover the relevant entity types in your system; you can still use those too.
The difference is that with graph data, the notions of "entity," "entity type," and "attribute" are more generic and more expansive than with tables. Take our Daniel / cat example:
person, name, Danielperson, has, catcat, name, Potatocat, yelling?, true
How does this differ from the way you would store this in a relational database?
- There is no explicitly defined entity type. Database systems can allow you to define entity types which they then use to validate your data and enforce constraints, but this is optional - entities don't have to be an instance of an entity type to exist.
catare both able to use the same attribute,
personentity is able to directly refer to the
catentity. With tables, you can only indirectly refer to other entities via foreign keys.
This example illustrates the ways that "thinking in tables" influences our data modeling process. To get used to data modeling with graphs, step one is disentangling table-specific design details from your conceptual models. Lower-level table-specific implementation considerations (like creating primary and foreign keys) can exert influence upward, shaping your conceptual designs. It's like how the building materials you use might affect the way you design a house, even though the principles of architecture remain constant. (I think. I have never designed a house. Or studied architecture.)
Graphs let you capture any relationships between any two things
How do we go about disentangling table-specific design details from conceptual models? Let's start by recalling that an EAV triple captures a relationship between two things, and "a relationship between two things" is the basic building block of information. Tables structure relationships (captured as columns) between entities (captured as rows) and values (captured as the value in a cell). Graphs structure relationships as two nodes with a directed edge pointing from one to the other.
When you're in data modeling mode, you want to consider the direct relationships that can exist between the entities in your system, without concern for how you might implement those relationships in a table. Tables constrain our thinking around entity relationships in a few ways:
- Entities can't directly refer to each other.
- Entities can't exist without the supporting structures of rows and tables
- Entities are only allowed to have a limited set of attributes
- There's a strict correspondence between entity type and attribute; if an attribute isn't in a table schema, the entity can't have it. Likewise, entities can't have attributes from other table schemas.
These constraints don't exist with graphs. In fact, graphs allow you to capture
nearly any relationships between any two entities, and tables don't. The
notion of "entity" is also broader than in the relational world. For example,
you can even treat an attribute as an entity, in that you can create EAV triples
that describe attributes by relating them to other entities. The graph below
shows EAV triples we've recorded for the
attributes, along with a new
name, describes, identityhas, describes, ownershipyelling?, describes, vocal statedescribes, describes, meaning
This might feel strange to you. How can
name be both an entity and an
attribute? For that matter, it might even feel strange to you to talk about
name as an entity at all; we're used to thinking of entities as anonymous
containers for attribute/value pairs, not as referenceable things that exist in
the same data space as all other data. With graph databases, there are no
containers for entities; there's just nodes and the edges that connect them with
no end or beginning. Thus we can even have
describes used as both an entity
and an attribute. It feels strange, but over time you'll get used to it.
Anyway, if we return to our definition of information as consisting of statements about how things are related, we can see that graph databases are both more permissive and more direct in the relationships that you're allowed to capture, and you'll want to take this permissiveness and directness into account when you're constructing your data models. When working with tables you must necessarily constrain the attributes that an entity is allowed to have, and entities can't refer directly to each other. When working with graph data, you don't need to define any constraints, and entities can have any attributes unless you introduce constraints. Entities can also refer to each other directly.
These differences have further implications for data modeling, and we're going to keep exploring them below.
Entities can refer to each other directly
In the relational world you cannot construct an EAV triple where one row directly references another row; instead, you must indirectly reference other rows via foreign keys. You cannot construct your data like this:
Instead, you must construct a
person table and a
cat table, where the person
catId column whose value matches the
id column in
The relationship between
cat has to be inferred, usually via the
WHERE clause of a SQL query.
You don't need join tables for many-to-many relationships
An extension of this concept is that you don't have to create the equivalent of a join table to capture many-to-many relationships. You can just capture them directly:
But if you need to you can introduce an entity that represents a relationship.
The graph below includes nodes named
which serve the same function as a record in a join table. We introduce these
intermediary nodes in this case because there's additional information we want
to include about the feelings these entities have for each other: the type of
hate, etc) and the intensity of the feeling:
Here's how you might capture this as an ERD:
Attributes are independent of entity type
In the relational world, attribute names are scoped to table definitions. Two
tables might use the same name for a column, and it might have the same meaning
or it might not. For example, a
poem table and a
book table both have a column
author, and it's likely they both name the same thing. If they do,
however, that is a matter of convention, a matter of agreement among the people
using that database. The fact that
author is meant identically in the two
tables cannot be captured in the database itself.
By the same token, you might have a
locations table with a
column, and your colleague might have a
places table with a
coordinates mean the same thing in these two tables? Maybe, maybe
not. Maybe you're both using latitude and longitude, or maybe one of you is and
the other is using the Universal Transverse Mercator (UTM) coordinate system.
There's no way to know without additional contextual information (like docs)
outside the databases themselves.
With graph data, the meaning is clear. If you have triples like this, you know
that the meaning of
author is the same in both:
The Leash, author, Ada LimonDemon Copperhead, author, Barbara Kingsolver
And if you have triples like this, you know that there are two distinct attributes being used:
"durham nc", lat_lng_coordinates, "35.9,78.9""durham nc", utm_coordinates, "689419.80,3985328.70"
Table constraints artificially limit the way you think about and communicate data modeling. Because it's not possible to use the same attribute in two different tables, we don't think about what that means for our data or how to communicate it.
With graphs, the same attribute can belong to multiple entity types, and you can therefore capture that in your data models. You can communicate this by using the same attribute names in your ERDs when the meaning of the attribute is the same, and using different attribute names when the meaning is different.
With relational databases, we attempt to normalize our schemas to reduce duplication and minimize the likelihood of data drift, where the same information is recorded in multiple places and those places are not all kept in sync.
Normalizing your data can introduce inefficiencies, however, as querying the data can involve expensive JOINs and UNIONs. Normalization is therefore balanced with intentional denormalization, duplicating data across table rows to improve performance.
With graph databases, you generally normalize your data without concern for denormalization. You can apply the same heuristics for figuring out when to normalize your graph data: for example, if you have an e-commerce site and are recording orders, you might record your data in a "denormalized" way by recording a billing address for each order:
Or you might introduce an
address entity that orders can refer to:
The latter is generally preferred, where possible. It more clearly reflects your application's data model and it's better for data integrity. Because of how graph databases store data, you don't have to worry about the efficiency of querying normalized data like you do with relational databases.
Entities can have multiple types
One of the biggest differences between the table approach and graph approach is
that, with graphs, entities can have multiple types. This is because an entity
type is defined in terms of attributes an entity has, whereas with tables the
"type" of the entity is rigidly bound to the table that contains it. Sure, it's
possible to create tables that have a
type column as a kind of workaround, but
even then graph systems exhibit a degree of flexibility that's simply not
possible with tables.
With a graph system, you could say that an entity has a
PERSON type and an
PERSON entities should have
EMPLOYEE entities should have an
Here's how you might capture this as an ERD:
The following triples would define an entity that has both types:
entity-1, name, "Ash Williams"entity-1, birth-date, "1958-06-22"entity-1, employee-number, "333"
This article doesn't describe the practical implications of this fact or the ways that you can leverage it. The point here is that it's at least possible, whereas it's not possible to do something like this with a relational database.