October 2023 Summaries

Felt, a web platform for creating and collaborating on maps, emphasizes rapid deployment as a core aspect of its engineering strategy, often executing more than 10 deployments a day to remain responsive to customer feedback. This rapid iteration is facilitated by using Render's preview environments, which allow each pull request to become a standalone infrastructure, enabling comprehensive product testing and development. Key principles such as "Day 1 Deploys" and "PR Reviews" are integral to their approach, ensuring new engineers can contribute quickly and efficiently, while human-in-the-loop automation balances automation with manual oversight to maintain data security and reliability. The use of feature flags and a structured pull request etiquette further supports their fast-paced deployment environment, fostering a culture where engineers are encouraged to make manageable changes and test them thoroughly. Felt's commitment to speed aims to enhance its position as a leading map-making platform by quickly delivering innovations to users and reducing the gap between user needs and product offerings.

Render has introduced two new features aimed at enhancing service builds: the pre-deploy command and the Performance build pipeline. The pre-deploy command allows users to run specific tasks, such as database migrations, just before each deploy, even for services deploying a prebuilt Docker image, thus providing a more organized workflow. Logs for these commands are included in deploy logs, and failures result in canceled deployments. Meanwhile, the Performance build pipeline offers increased memory and CPU resources for build tasks, beneficial for apps with heavy-duty builds, and is available at a higher cost than the Starter tier. These features are designed to streamline infrastructure management, allowing developers to focus on application development and customer satisfaction while Render handles the underlying cloud infrastructure.

In November 2021, Render introduced a free tier for hobbyist developers, which saw a significant increase in adoption after Heroku discontinued its free offering, leading to infrastructure challenges. Render used Knative to manage the scalability of their free-tier services, which scale to zero when inactive to save resources. However, a surge in free-tier usage quadrupled the number of apps, straining networking components like Calico and kube-proxy due to increased updates and CPU consumption, causing latency issues. Render addressed these by removing unnecessary Kubernetes Services managed by Knative, such as the Placeholder, Public, and Private Services, optimizing traffic routing directly through the activator. This reduced CPU usage and improved network stability and latency but highlighted the need for a more tailored solution, suggesting future plans to develop a custom system better suited to Render's specific requirements.

Render has developed a strategy to reduce the networking overhead in its free-tier infrastructure by minimizing the number of Kubernetes (K8s) Services tied to each Revision to zero, thereby minimizing the burden on K8s networking machinery. This approach involves modifying the Knative Activator to track Pods directly instead of relying on Endpoints updates, allowing for more efficient handling of requests by utilizing the Pod's IP. The plan to roll out this change includes running the new Pod-based flow in parallel with the existing Endpoint-based flow, using specific labeling and environment variables to manage deployment and ensure a smooth transition. This change is expected to enhance system stability, lower CPU usage for calico and kube-proxy, and enable clusters to support more paid-tier customers without impacting free-tier services. The rollout will be conducted gradually, starting with Render-owned free-tier services and moving onto new and existing free-tier services across clusters, with the ability to roll back if necessary.