Learning Redis – Part 1: Introduction to Key-Value Databases

Want to bring more performance, speed, and scalability to your website? Or scale your sites for real-time services or message passing? Learn how, and get practical real-world tips in this exploration of Redis, part of a series on choosing the right data storage.

Steven Edouard and I (Rami Sayar) show you how to get up and running with Redis, a powerful key-value cache and store. In this tutorial series, you can check out a number of practical and advanced use cases for Redis as cache, queue, and publish/subscribe (pub/sub) tool, look at NoSQL and data structures, see how to create list sets and sorted sets in the cache, and much more. You can watch the course online on Microsoft Virtual Academy.

Level: Beginner to Intermediate.


By the end of this installment, you will be able to understand the difference between key-value stores, structured databases and SQL databases.


Key-Value Stores are a type of data store that organize data differently from your traditional SQL store. The fundamental data model of a key-value store is the associative array (a.k.a. a map, a dictionary or a hash). It is a collection of key-value pairs where the key is unique in the collection. A key can be an ID or a name or anything you want to use as an identifier. The value can be anything. Rather than storing data into a variety of tables and columns like in SQL stores, key-value stores split a data model into a collection of data structures such as key-value strings, lists, hashes, sets, etc… Although it may sound simplistic, you can build more complex data structures on top of key-values.

Redis focuses on high performance and a simple querying language that is just a set of data retrieval commands. Unlike SQL, there are no worries about writing complex queries.

The nature of key-value stores makes them best suited to operate as caches or data structure stores and in situations that are performance sensitive. As previously mentioned, you can build more advanced data structures on top of key-value pairs. You can also use the high performance to build queues or publish-subscribe mechanisms.

Key-value stores fall into the NoSQL family of databases, they don’t use SQL and have a flexible schema. Your application defines the key-value pairs and can change the definition at any time. You decide how to store your data!

Tabular Structured Data vs Key-Value Structure

Lets look at an example of a bank, which keeps track of the following entities:

  • Person
  • Account

The basic relation here is that a Person can have many Accounts and therefore an Account has only one Person.

Tabular Structure

Here’s what this data may look like in a SQL database


0 Steven Edouard
1 Sam Brightwood


0 Investment 80000.00 USD 0
1 Savings 70400.00 USD 0
2 Checking 4500.00 USD 0
3 Checking 4500.00 YEN 1
4 Investment 4500.00 YEN 1
5 Savings 4500.00 YEN 1

In the above data schema, the HOLDER field in the Accounts table is a Foreign Key into the Accounts table. This key is what creates the association between the two data entities. SQL requires that in order to create an Account, you must have a valid Foreign key that points to an existing person, which can cause some loss in flexibility.

With the schema above its easy to answer the question:

Who is the account holder for the account with ID = 3?

Because we have a foreign key HOLDER for the account we can quickly and easily look up that the account holder isSam Brightwood. This is because there is a many:one relationship from Account entities and Person entities.

Now, how about the question:

Which accounts does person Steven Edouard hold?

To answer this question, it requires us to do a join which essentially is a set operation to find all of the rows in the Accounts table which match the Person ID = 0. This operation can become an expensive as the data in the tables grow in larger and larger numbers.

Key-value Structure

In the key-value data structure, we want to reduce the Person table into a collection of keys and values that are identifiable by a Person ID. The key is an index in the key-value store, but we can add a second index embedded in the same key. For example, we want to take Person with ID=0 and store their first name, we can name our key:0:first_name or maybe person:0:first_name by using the ‘:’ as an index separator in the key. If we follow this nomenclature of table-name:key:property, the above data can be flattened to the following key-value pairs.

"person:0:first_name" = "Steven"
"person:0:last_name" = "Edouard"

"person:1:first_name" = "Sam"
"person:1:last_name" = "Brightwood"

"account:0:type" = "Investment"
"account:0:balance" = "80000.00"
"account:0:currency" = "USD"

"account:1:type" = "Savings"
"account:1:balance" = "70400.00"
"account:1:currency" = "USD"

"account:2:type" = "Checking"
"account:2:balance" = "80000.00"
"account:2:currency" = "USD"

"account:3:type" = "Checking"
"account:3:balance" = "4500.00"
"account:3:currency" = "YEN"

"account:4:type" = "Investment"
"account:4:balance" = "4500.00"
"account:4:currency" = "YEN"

"account:5:type" = "Savings"
"account:5:balance" = "4500.00"
"account:5:currency" = "YEN"

The data structure gets more interesting when we want to describe relationships like foreign keys. Our application layer will decide how to interpret the values and conceivably we can instruct the application to interpret values as a key:

"account:0:holder" = "person:0"
"account:1:holder" = "person:0"
"account:2:holder" = "person:0"
"account:3:holder" = "person:1"
"account:4:holder" = "person:1"
"account:5:holder" = "person:1"

Now, let’s answer the original two questions again:

Who is the account holder for account with ID = 3?

To answer this question we can just look up the account holder with the unique ID 3, account:3:holder and by using the value person:1, we can easily pull out that user’s name by appending :first_name to the value.

Which accounts does person Steven Edouard hold?

This is slightly more difficult. First, we will have to scan through all the person keys until we find a key that has bothfirst_name="Steven" and `last_name="Edouard". Secondly, once we locate that person, we will have to scan through all the account:*:holder keys until we find all those that are associated with the person:id key of Steven Edouard.

Alternatively, we can create two reverse indices to speed up those two operations and not require any scans at all. We can add:

"person:name:edouard_steven" = "person:0"
"person:name:sam_brightwood" = "person:1"

"person:0:accounts" = "[account:0, account:1, account:2]"
"person:1:accounts" = "[account:3, account:4, account:5]"

Adding those two reverse indices, one to have a mapping between first_name & last_name and the key, second to have a mapping between a person and their accounts, allows us to answer our original question significantly faster than our initial approach. Fundamentally, how we store our data affects data retrieval in the key-value scenario.

Note: I am defining a scan here as reading through an array, an O(n) operation.

It seems like a lot more work to use a key-value database to store tabular data. In practice, the above operations are already abstracted away in a database like Redis through data types like hashes and lists.

Where Key-Value Stores Shine?

The previous example was using a data model that was relational and shows key-value stores in not the best light. Well, all that doesn’t matter when your data is by nature key-value pairs. A cache is a component that stores results so that they can be served faster in the future.

Let’s say that you have to perform a long-running computation based on a key, for example: an encryption or a mathematical operation. You have calculated this computation once already; you want to keep it for the next time you need it, but you don’t want to hold on to it for too long. This situation is where a key-value stores shine! The simplicity and the lack of overhead from a query system result in a very high-performance system.

Stay Tuned!

Stay tuned for next installment of this tutorial series. You can stay up to date by following @ramisayar and @sedouard.

Comments (0)

Skip to main content