August 2025 Summaries

Koyeb has introduced a significant enhancement to its serverless platform by reducing cold start times to around 200ms for CPU workloads using a feature called Light Sleep. This was achieved by transitioning from Firecracker to Cloud Hypervisor for better hardware support, implementing Kata Containers for flexibility, and leveraging eBPF for efficient idle detection. The company faced challenges with snapshotting, notably with virtio-fs and network restoration, which were overcome by modifying Cloud Hypervisor processes and integrating a scaletozero-agent. This agent, combined with eBPF, allows for precise idle detection and efficient sleep and wake mechanisms for applications, ensuring seamless scaling to zero without disrupting services. The new setup, now available in public preview, aims to extend similar capabilities to GPU-based services, addressing the complexities of VRAM preservation.

Koyeb has announced the public preview of Light Sleep, an enhancement to their Scale-to-Zero feature, designed to minimize latency for CPU workloads by reducing cold start times to around 200 milliseconds. This improvement builds on the original Scale-to-Zero iteration, which introduced Deep Sleep to enhance compute efficiency but resulted in cold starts that took between one to five seconds. Light Sleep enables serverless applications to wake up nearly instantaneously, making it ideal for large-scale deployments and real-time workloads. Available on Koyeb's Starter, Pro, and Scale plans, Light Sleep allows users to specify idle periods before transitioning to Deep Sleep, offering a more efficient use of resources. During the public preview, Light Sleep is free, but will later cost 15% of the normal per-second instance price. Interested users can activate Light Sleep through the Koyeb console by adjusting settings under the Scaling options.

The text recounts the author's experiences with conference talk submissions and evaluations, particularly in the context of organizing the AI Engineer Paris event. Despite encountering challenges like the lack of feedback on rejections and the overwhelming volume of submissions, the author offers insights on the importance of company reputation, role relevance, and social media presence in the selection process. The author underscores the significance of including a video showcasing speaking abilities, as it heavily influences the evaluation of proposals. Additionally, proofreading submissions and subtly integrating company endorsements into talks are emphasized to avoid appearing overly promotional. The narrative acknowledges the competitive nature of conference selections and encourages persistence despite potential rejections, while also expressing gratitude to those who submitted proposals and urging continued engagement with AI Engineer events.