A database is a collection of data that can be organized, managed, modified, and retrieved using a computer. The system that helps with managing data in a database is called a database management system (DBMS).

In the 1950s and 1960s, several advancements were made in terms of processors, storage, memory, and networks. We also had our first programming languages, COBOL and FORTRAN. The development of hard disk drives for data storage further spurred the development of databases. Around the same time, the first notion of a modern-day computer with a mouse and graphical user interface came into existence, making it easy for the general public to consume it. In this section, we will discuss how various types of databases evolved.

SQL

The first database was designed by Charles William Bachman III, an American computer scientist. In 1963, he developed the Integrated Data Store (IDS), which gave rise to the concept of the navigational database. In navigational databases, we can find records by chasing references from other objects. For example, let's say that in a school database, you want to find all the students from a specific grade in a specific school. In a navigational database, first, you have to go to the group of students that belong to a particular school and then to the group that belongs to a particular grade. So, records can be accessed by hierarchical navigation. Based on IDS, Bachman later developed the CODASYL database model in 1969. CODASYL stands for Conference/Committee on Data Systems Languages, which was a consortium to guide the development of programming languages. Around the same time Edgar F. Codd, an IBM employee, developed the IBM Information Management System (IMS), which was based on the hierarchical database model. A hierarchical database model is a data model in which the data is designed in a tree-like structure. In 1970, Donald D. Chamberlin and Raymond F. Boyce developed Structured Query Language (SQL) based on what they'd learned about IMS. They initially called it Structured English Query Language (SEQUEL), which System R was later developed with by a group at the IBM San Jose research laboratory. In 1976, QUEL, which is a relational database query language designed by Michael Ralph Stonebraker, was developed as part of the Interactive Graphics Retrieval System (INGRES) database management system at the University of California, Berkeley.

Based on QUEL and SQL, several databases were implemented. Some of the most prominent ones include Post Ingres (Postgres), Sybase, Microsoft SQL, IBM DB2, Oracle, MariaDB, and MySQL.

Object-oriented databases

In the 1980s, object-oriented database systems (OODBMSes) grew in popularity. In OODBMSes, information is represented as objects compared to tables in relational databases. Some of the important ones include Gemstone/S, Objectivity/DB, InterSystems Cache, Perst, ZODB, Wakanda, ObjectDB, ODABA, and Realm.

NoSQL

The concept of non-SQL or non-relational databases has existed since the 1960s, but the term NoSQL became has much more popular in the last decade. NoSQL databases focus on performance and scaling and mostly rely on a non-relational data model such as a document, key-value, wide-column, or graph to organize the data. Some of the most popular ones in this category include Cassandra, MongoDB, Couchbase, Dynamo, FoundationDB, Neo4j, and Hbase.

NewSQL

With the introduction of the on-demand availability of compute, storage, and network resources and the pay-as-you-go model, which is collectively known as cloud computing, the amount of data that we collect, process, manage, and analyze has been growing exponentially. Although it was relatively easier for some of the NoSQL databases to adapt to the cloud, it is still much harder for traditional SQL databases to do so. Many of them are better suited for vertical scaling and do not consider geographically distributed data, the shared-nothing architecture, and enormous scale as part of their initial design. This created a void. We needed SQL databases that are cloud-native, scale well with data growth, and are easy to manage. Many companies developed in-house solutions on top of existing SQL databases:

• Facebook developed TAO, a NoSQL graph API built on top of sharded MySQL.
• YouTube developed Vitess to easily scale and manage MySQL clusters.
• Dropbox developed Edgestore, a metadata store to power their services and products, which again was built on top of MySQL.
• GreenPlum developed a massively parallel data platform by the same name for analytics, machine learning, and AI on top of Postgres.

However, it was still relatively hard and painful to manage the data as the underlying database was not built to scale.

In 2012, Google published a seminal paper on Google Spanner: a globally distributed database service and storage solution. Spanner essentially combined the important features of SQL databases such as ACID transactions, strongly consistent reads, and the SQL interface with some of the features that were only available with NoSQL databases, such as scaling across geographical locations, multi-site replication, and failover. It created a new category of databases called NewSQL, which is meant to indicate a combination of SQL features at NoSQL scale. YugabyteDB and CockroachDB were developed later, both of which got their inspiration from Google Spanner.

In this section, we will learn about some of the core database concepts, including cardinality, database models, and various processing models.

Cardinality

Before we discuss database models, it is important to know about cardinality. Cardinality refers to the relationship between two entities or tables. The most popular ones include one-to-many, many-to-one, and many-to-many.

One-to-one relationship

In the case of a one-to-one relationship, a row or entry in one entity or table can be related to only one row in another entity or table. For example, in a Department of Motor Vehicles database, let's say there are two tables called License Info and Driver Info, as shown in the following diagram:

Figure 1.1 – An example of a one-to-one relationship

Here, Driver ID can only be assigned to one driver as it has to uniquely identify a driver. Also, a driver can only be assigned one Driver ID. So, here, any row in the License Info table will be associated with a specific row in the Driver Info table.

One-to-many relationship

In a one-to-many relationship, a single row from one entity or table can be associated with multiple rows in another entity or table.

For example, let's consider the Driver Info and City Info tables shown in the following diagram::

Figure 1.2 – An example of a one-to-many relationship

Here, for every row in City Info, there will be multiple rows in Driver Info, as there can be many drivers that live in a particular city.

Many-to-many relationship

In a many-to-many relationship, a single row in one entity or table can be associated with multiple rows in another entity or table and vice versa.

For example, let's consider two tables: Vehicle Ownership History, where we are maintaining the history of ownership of a given vehicle, and Driver Ownership History, where we are maintaining the history of vehicles owned by a given driver:

Figure 1.3 – An example of a many-to-many relationship

Here, a driver can own multiple vehicles and a vehicle can have multiple owners over time. So, a given row in the Vehicle Ownership History table can be associated with multiple rows in the Driver Ownership History table. Similarly, a given row in the Driver Ownership History table can be associated with multiple rows in the Vehicle Ownership History table.

Now, let's take a look at some of the most important database models.

Overview of database models

A database model determines how the data is stored, organized, and modified. Databases are typically implemented based on a specific data model. It is also possible to borrow concepts from multiple database models when you are designing a new database. The relational database model happens to be the most widely known and has been popularized by databases such as Oracle, IBM DB2, and MySQL.

Hierarchical database model

In the hierarchical database model, the data is organized in the form of a tree. There is a root at the first level and multiple children at the subsequent levels. Since a single parent can have multiple children, one-to-many relationships can easily be represented here. A child cannot have multiple parents, so this results in the advantage of not being able to model many-to-many relationships.

IBM's Information Management System (IMS) was the first database that implemented this data model.

The following diagram shows an example of a hierarchical database model:

Figure 1.4 – An example of a hierarchical database model

Typically, the tree starts with a single root and the data is organized into this tree. Any node except the leaves can have multiple children, but a child can have only one parent.

Network model

The network model was developed as an enhancement of the hierarchical database model to accommodate many-to-many relationships. The network model relies on a graph structure to organize its data. So, there is no concept of a single root, and a child can have multiple parents and a parent can have multiple children. Integrated Data Store (IDS), Integrated Database Management Systems (IDMS), and Raima Database Manager (RDM) are some of the popular databases that use the network model.

As shown in the following diagram, there is no single root and a given child (for example, Object 2 can have multiple parents; that is, Object 1 and Object 3):

Figure 1.5 – An example of a network model

Relational model

Although the network model was an improvement over the hierarchical model, it was still a little restrictive when it came to representing data. In the relational model, any record can have a relationship with any other with the help of a common field. This drastically reduced the design's complexity and made it easier to independently add, update, and access records, without having to walk down the tree or traverse the graph. SQL was combined with the relational database model to provide a simple query interface to add and retrieve data.

All the popular traditional databases such as Oracle database, IBM DB2, MySQL, MariaDB, and Microsoft SQL Server implement relational data models.

Let's look at two tables called Employee and Employee Info:

Figure 1.6 – Employee tables showing the column names

Here, Employee ID is the common field or column between the Employee and Employee Info tables. The Employee table is responsible for ensuring that a given Employee ID is unique, while Employee Info is responsible for more detailed information about a given employee.

Object-relational model

The object-relational model, as the name suggests, combines the best of the relational and object data models. The concept of objects, classes, and inheritance are directly supported as first-class citizens as part of the database and in queries. SQL:1999, the fourth revision of SQL, introduced several features for embedding object concepts into the relational database. One of the main features was to create structured user-defined types with CREATE TYPE to define an object's structure.

Over time, relational databases have added more support for objects. There is a varying degree of support for object concepts in Oracle database, IBM DB2, PostgreSQL, and Microsoft SQL Server.

Given the scope of this book, we will not discuss the entity-relational model, object model, document model, star schema, snowflake schema, and many other less well-known models.

Now, let's look at how databases can be classified based on what kinds of workload they can be used for.

Processing models

Based on how you want to consume and process data, databases can be categorized into four different processing systems. Let's take a look.

Online transaction processing (OLTP)

OLTP systems support the concept of transactions. A transaction refers to the ability to atomically apply changes (insert, update, delete, and read) to a given system. One popular example is a bank, where withdrawing or depositing money to a given bank account must be done atomically to ensure data is not lost or incorrect. So, the main purpose here is to maintain data integrity and consistency. Also, these systems are generally suited for fast-running queries.

Online analytical processing (OLAP)

OLAP focuses mostly on running queries to analyze multi-dimensional data and to extract some intelligence or patterns from it. Typically, such systems support generating some sort of report that can be used for marketing, sales, financing, budgeting, management, and many more. Data mining and data analytics applications would typically have to have an OLAP system in some form. OLAP doesn't deal with transactions, and the emphasis is more on analyzing large amounts of data from different sources to extract business intelligence. Some databases also provide built-in support for MapReduce to run queries across a large set of data.

A data warehouse is a piece of software that's used for reporting and data analysis. Warehouses are typically developed for OLAP. It is also very common to retrieve the data from OLTP in batches or bulk, run it through an Extract, Load, and Transform (ELT) or Extract, Transform, and Load (ETL) data transformation pipeline, and store it in an OLAP system.

Online event processing (OLEP)

OLEP guarantees strong consistency without the traditional atomic commit protocols or distributed locking. OLEP also focuses on high performance, larger scales, and fault tolerance.

Hybrid transaction/analytical processing (HTAP)

As the name suggests, this system tries to provide the best of both transactions and analytical processing. Most of the NoSQL and NewSQL databases provide support for managing both transactional and analytical workloads. Vitess is a database clustering system that can be used to scale and shard MySQL instances. Vitess provides HTAP features on top of MySQL by allowing a given MySQL instance to be configured as master or read-only, where read-only can be used for analytical queries and MapReduce. It is possible to use CockroachDB as HTAP by propagating changes with the help of change data capture (CDC) in the OLTP cluster or primary cluster to a separate cluster, which is solely used for analytical processing.

Now, let's learn a bit about embedded and mobile databases, including why they exist and some of the most popular ones in this space.

Embedded and mobile databases

Embedded databases usually refer to databases that can be tightly integrated into an application, without needing separate hardware to support them. Also, they don't have to be managed separately. Some of the most popular embedded databases include SQLite, Berkeley DB from Oracle Corporation, and SQL Server Compact from Microsoft Corporation. Embedded databases are also very useful for testing purposes as they can be started within test suites.

Mobile database refers to the class of databases that work with very limited memory footprint and compute and can be deployed within a mobile device. They are typically used for storing user data for apps running on mobile devices. SQLite, SQL Server Compact, Oracle database Lite, Couchbase Lite, SQL Anywhere, SQL Server Express, and DB2 Everyplace belong to this category,

Database storage engines

A database storage engine is a component within a database management system that is responsible for Create, Read, Update, Delete (CRUD) operations and transferring data between disk and memory, without compromising data integrity. Some of the most popular ones include Apache Derby, HSQLDB, InfinityDB, LevelDB, RocksDB, and SQLite. CockroachDB initially started with RocksDB as its database engine, but from release 20.2 onward, Pebble will be the database engine by default. Pebble, as per Cockroach Labs, is a RocksDB-inspired and RocksDB-compatible key-value store focused on the needs of CockroachDB. RocksDB was implemented in C++, whereas Pebble was implemented in Golang. This makes it easier to manage and maintain as CockroachDB itself was written in Golang. This means that we only have to deal with one language now.

Eric A. Brewer gave a keynote talk in 2000 titled Towards Robust Distributed Systems at a symposium on Principles of Distributed Computing, summarizing his years of learning about distributed systems. Brewer talked about key aspects of a distributed system: consistency, availability, and tolerance toward network partition. Consistency refers to the fact that every read should see the data from the most recent write; otherwise, it should error out. Availability means every requested read or write should receive a non-error response. Partition tolerance indicates that the system should continue to serve, irrespective of delays and communication failures between nodes in the system. Consistency, Availability, and Partition Tolerance (CAP) theorem claims that, at most, you can only have two of these three properties in a distributed system.

Consistency and partition tolerance (CP)

A CP database provides consistency and partition tolerance but cannot provide availability. This is also called a CAP-consistent system. Let's understand this by looking at an example:

Figure 1.7 – CP system

Let's consider the system shown in the preceding diagram, where two servers are serving read and write traffic. For this example, let's say writes only land on Server 1 and reads only land on Server 2. So long as Server 1 can talk to Server 2, all the writes that come to Server 1 can be propagated synchronously to Server 2. This ensures that any reads that come to Server 2 are always consistent, which means they see the latest data written by the latest write in Server 1:

Figure 1.8 – CP system during a communication failure

Now, let's say that, as shown in the preceding diagram, the communication between Server 1 and Server 2 has broken down and now Server 1 is no longer able to propagate the writes synchronously. This results in partitioning. Since the data cannot be propagated between the two servers, read or write traffic cannot be served until we resolve the partition issue as we have to ensure data consistency.

Some of the most popular databases that have CP characteristics are HBase, Couchbase, and MongoDB. CockroachDB also falls into this category.

Availability and partition tolerance (AP)

In this case, a database is guaranteed to always be available and it can tolerate partitioning, but at the cost of consistency. This is also known as a CAP-available system. Here, the application is expected to deal with data consistency:

Figure 1.9 – AP system during a communication failure

Similar to the previous example, if the communication between Server 1 and Server 2 breaks down, Server 1 and Server 2 continue to serve the traffic but reads to Server 1 and Server 2 might return different versions of the data, based on when the communication has failed and whether there was any change to that data, after the communication failure. Cassandra, Riak, and CouchDB are popular examples of AP databases.

Consistency and availability (CA)

In the case of a CA database, the system cannot tolerate partitioning but can guarantee consistency and availability. Traditional databases with single-server deployments with no replication or slaves can be classified as CA. Now, many traditional RDBMS databases can be configured in various ways to have CA, CP, or AP as desired.

The name CockroachDB was inspired by the insect that goes by the same name. Just like how cockroaches have been surviving for millions of years and colonizing the entire planet and thriving, CockroachDB instances are supposed to replicate and repair data, spread naturally across multiple availability zones, and survive total regional failures. Also, once CockroachDB becomes part of a given software ecosystem, it's impossible to get rid of or replace it, just like cockroaches. Here, we will discuss why there is a need for yet another database, known as Inspiration, and provide a high-level overview of CockroachDB.

Why yet another database?

As more companies shift from on-premises to the cloud, they are looking for SQL datastores on various cloud platforms to manage their transactional data. Most of the traditional databases such as MySQL, Postgres, and Oracle are not built for the cloud. This necessitates a cloud-native, consistent, distributed SQL that can scale with the growth of data. CockroachDB fills this gap.

Inspiration

As we previously discussed in the NewSQL section, in 2012, Google published a seminal paper on Google Spanner: a globally distributed database service and storage solution. Although Google Spanner combined the best of both SQL and NoSQL and was very useful for a lot of applications, it was not available for public usage. Also, Google Spanner was and still is not an open source project and has only been available on Google Cloud Platform since 2017. So, this created a necessity for an open source Spanner-like database that can be used in different cloud providers and on-premises. Around 2012, Spencer Kimball, Peter Mattis, and Ben Darnell were working at Google on the Google File System and Google Reader projects. They also got acquainted with both Bigtable and Spanner during their tenure at Google. They decided to build something very similar to Spanner to make it available for everyone and started an open source project on GitHub in 2014. After a year, they decided to leave Google and founded Cockroach Labs in 2015 before officially working on CockroachDB in June 2015.

Key terms and concepts

Before we look at the various functional layers, let's look at some of the key concepts and terms. A CockroachDB cluster refers to a group of nodes that act as a single logical unit. A node is a single machine that runs an instance of CockroachDB. CockroachDB stores all the data as sorted key-value pairs. These keys are divided into ranges. CockroachDB replicates each range and stores each replica on a different node. For each range, there will be a leaseholder, which acts as a primary owner of a given range and receives and coordinates all the traffic for that range. For each range, one of the replicas acts as a leader for write requests and ensures that the majority of the replicas are in consensus, before committing a given write. For each range, there will be a time-ordered log of writes, called a raft log, for which the majority of replicas agreed upon.

High-level overview

CockroachDB is a cloud-native, consistent, highly scalable relational database. Some of the primary goals of CockroachDB are to provide strong consistency, geo-distribution of data, high availability, SQL support, easy deployment, and less maintenance. Since we will be dealing with CockroachDB internals in detail in subsequent chapters, we will just provide a high-level overview here:

Figure 1.10 – High-level overview of the CockroachDB architecture

CockroachDB exposes a SQL interface, using which clients can interact with the database. Client requests can land on any node within a given cluster and work just fine since all the nodes are symmetrical.

CockroachDB can be divided into five functional layers:

• SQL
• Transactional
• Distribution
• Replication
• Storage

The SQL layer is responsible for receiving SQL queries and converting them into key-value operations. The transactional layer ensures that all CRUD operations that happen on multiple key-value pairs are transactional. The distribution layer is responsible for ensuring ranges are evenly distributed among all the available nodes in a cluster. The replication layer ensures that ranges are replicated synchronously, whenever there is a change. Finally, the storage layer is responsible for managing key-value data on the disk.

In this chapter, we learned about the evolution of databases, how databases can be categorized based on various criteria, CAP theorem, and a brief introduction to CockroachDB. By now, you should also be familiar with database and processing models, what the CP, CA, and AP systems in CAP theorem offer, and the functional layers of CockroachDB.

In the next chapter, we will take a deep dive into CockroachDB's architecture and design concepts.