This article is exceprted from the book, Big Data, published by Manning.
The properties you should strive for in Big Data systems are as much about complexity as they are about scalability. Not only must a Big Data system perform well and be resource—efficient, it must be easy to reason about as well. Let's go over each property one by one.
Robustness and fault tolerance
Building systems that "do the right thing" is difficult in the face of the challenges of distributed systems. Systems need to behave correctly despite machines going down randomly, the complex semantics of consistency in distributed databases, duplicated data, concurrency, and more. These challenges make it difficult even to reason about what a system is doing. Part of making a Big Data system robust is avoiding these complexities so that you can easily reason about the system.
It's imperative for systems to be human-fault tolerant. This is an oft-overlooked property of systems that we're not going to ignore. In a production system, it's inevitable that someone will make a mistake sometime, such as by deploying incorrect code that corrupts values in a database. If you build immutability and recomputation into the core of a Big Data system, the system will be innately resilient to human error by providing a clear and simple mechanism for recovery.
Low latency reads and updates
The vast majority of applications require reads to be satisfied with very low latency, typically between a few milliseconds to a few hundred milliseconds. On the other hand, the update latency requirements vary a great deal between applications. Some applications require updates to propagate immediately, but in other applications a latency of a few hours is fine. Regardless, you need to be able to achieve low latency updates when you need them in your Big Data systems. More importantly, you need to be able to achieve low latency reads and updates without compromising the robustness of the system.
Scalability is the ability to maintain performance in the face of increasing data or load by adding resources to the system. The Lambda Architecture is horizontally scalable across all layers of the system stack: scaling is accomplished by adding more machines.
A general system can support a wide range of applications. Because the Lambda Architecture is based on functions of all data, it generalizes to all applications, whether financial management systems, social media analytics, scientific applications, social networking, or anything else.
You don't want to have to reinvent the wheel each time you add a related feature or make a change to how your system works. Extensible systems allow functionality to be added with a minimal development cost.
Often a new feature or a change to an existing feature requires a migration of old data into a new format. Part of making a system extensible is making it easy to do large-scale migrations. Being able to do big migrations quickly and easily is core to the approach you'll learn.
Ad hoc queries
Being able to do ad hoc queries on your data is extremely important. Nearly every large dataset has unanticipated value within it. Being able to mine a dataset arbitrarily gives opportunities for business optimization and new applications. Ultimately, you can't discover interesting things to do with your data unless you can ask arbitrary questions of it.
Maintenance is a tax on developers. Maintenance is the work required to keep a system running smoothly. This includes anticipating when to add machines to scale, keeping processes up and running, and debugging anything that goes wrong in production.
An important part of minimizing maintenance is choosing components that have as little implementation complexity as possible. You want to rely on components that have simple mechanisms underlying them. In particular, distributed databases tend to have very complicated internals. The more complex a system, the more likely something will go wrong, and the more you need to understand about the system to debug and tune it.
You combat implementation complexity by relying on simple algorithms and simple components. A trick employed in the Lambda Architecture is to push complexity out of the core components and into pieces of the system whose outputs are discardable after a few hours. The most complex components used, like read/write distributed databases, are in this layer where outputs are eventually discardable.
A Big Data system must provide the information necessary to debug the system when things go wrong. The key is to be able to trace, for each value in the system, exactly what caused it to have that value.
"Debuggability" is accomplished in the Lambda Architecture through the functional nature of the batch layer and by preferring to use recomputation algorithms when possible.
Achieving all these properties together in one system may seem like a daunting challenge. But by starting from first principles, as the Lambda Architecture does, these properties emerge naturally from the resulting system design.
In my book, Big Data, I go into much more detail on each of these properties.
By Nathan Marz