Comparison with containers

Another comparison that is quite popular for FaaS offerings is with containers.

Let's look at what containers are, as we haven't discussed them so far. Containerization of the infrastructure has been around for a long time, but it was made popular in 2013 by Docker when they combined operation system level virtualization with filesystem images on Linux. With this, it became easy to build and deploy containerized applications. The biggest benefit that Docker-containers provides is that, unlike Virtual Machines (VMs),  they share the same operating system as the host, which tremendously reduces the size of the image that is required to run a containerized application. As the footprint of Docker containers is less, you can run multiple Docker containers on the same host without significant impact on the host system. 

As container runtimes gained popularity, container platforms started to evolve. Some of the notable ones are:

• CloudFoundry
• Apache Mesos

The next step in the evolution of containerization was providing an API that let developers deploy Docker images across a fleet of compute infrastructures. The most popular container orchestration and scheduling platform that is out there right now is Kubernetes. It started as an internal project within Google in 2004, called Borg. Later, Google open sourced the solution as Kubernetes and it is right now the most commonly used container platform out there. There are other solutions as well, such as Mesosphere.

And even further advancement in container solutions was cloud-hosted container platforms, such as Amazon EC2 Container Service (ECS) and Google Container Engine (GKE), which, like serverless offerings, eliminates the need to provision and manage the master nodes in the cluster. However, you still have to provision and manage the nodes on which the containers will be scheduled to run. There are a lot of solutions that are out there that help in the maintenance of nodes within the cluster. AWS introduced their new container offering called Fargate, which is completely serverless, where you don't have to provision the nodes within the cluster ahead of time. It is still early days for Fargate, but if it offers what it promises, it will certainly help a lot of developers who don't have to spend time provisioning and managing nodes within the cluster.

Fundamentally, the same argument that I made for PaaS still holds true with containers as well. For serverless offerings or FaaS offerings, scaling is automatically managed and is transparent. The scaling also has fine grain control. Container platforms do not yet offer such a solution. There are scaling solutions within container platforms, such as Kubernetes Horizontal Pod Autoscaling, which do scale based on the load on the system. However, at the moment, they don't offer the same level of control that FaaS offerings provide.

As the gap of scaling and management between FaaS offerings and hosted container solutions narrow, the choice between the two may come down to the type of the application. For example, you would choose FaaS for applications that are event driven and containers for request-driven components with a lot of entry points. I expect FaaS offerings and container offerings to merge in the coming years.