Replit — Speeding up Deployments with Lazy Image Streaming

Replit Deployments is our new offering that allows you to quickly go from idea, to code, to production. To make the experience as seamless as possible, we built tooling to convert a Repl into a container imager which can be deployed to either a Google Cloud Virtual Machine or to Cloud Run. Early on, we started to hit some issues with large images taking too long to deploy to a virtual machine. It could take minutes to pull and unpack the container image before it could be started. There’s two angles of attack: reduce the image size or speed up the pulling of images. It's preferable to shrink the container image size; however, that is not always possible. In this post we’ll go into some of the technologies and approaches used to speed up image pulling/booting. What is a container image? First, we’ll need to establish some baseline knowledge around container images. If you already know the details, you can skip this section. At a high level, a container image provides both a root filesystem and configuration for running a containerized workload. Inside the container, the filesystem is mounted to the root directory /. The root filesystem is stored as a list of multiple compressed tarballs, called layers, which are overlaid on top of each other. That is, if two layers have the same file, layers later in the list have higher precedence and their files replace the files from lower layers.