April 2024 Summaries

Koyeb has launched Serverless GPUs in public preview, offering an innovative solution for AI inference workloads through their platform. These GPUs, designed to handle both generative AI and computer vision models, feature high-end specifications such as up to 48GB of vRAM, 733 TFLOPS, and 900GB/s memory bandwidth, with pricing starting at $0.50 per hour. The platform supports seamless deployment with features like one-click Docker container deployment, built-in load balancing, horizontal autoscaling, zero downtime, auto-healing, and real-time monitoring. Users can pause GPU Instances to optimize costs and deploy services using the Koyeb CLI or Dashboard, either by using pre-built containers or linking to a GitHub repository. Koyeb is actively working with early users to expand their offerings, including more GPUs and accelerators, and invites feedback from users requiring specific configurations. The company fosters community engagement through their serverless community and social media presence.

This blog post explores the concept of template databases in PostgreSQL, specifically focusing on template0 and template1, which serve as the basis for creating new databases by simply copying them. The author highlights the utility of these templates in automating the process of adding extensions like uuid-ossp, which can streamline database creation by embedding such extensions into template1. This approach eliminates the need for repetitive installation of extensions each time a new database is created. Additionally, the post discusses a method for creating quick local backups using template databases, which offers a convenient alternative to more time-consuming pg_dump and pg_restore processes, although it is not suitable for production databases due to storage on the same physical hard drive. The article encourages readers to explore further readings and engage with the authors through various social media platforms.

Understanding and mastering key concepts in artificial intelligence, particularly generative AI and large language models (LLMs), is crucial for keeping up with technological advancements. LLMs, a subset of AI and machine learning, are designed to understand and generate human language by being trained on vast datasets. The lifecycle of these models includes training, where the model learns from data, and inference, where it applies this learning to new data. These models are characterized by parameters, weights, and tokens, each playing a unique role in their architecture and function. The size of an LLM, often measured by its parameter count, memory footprint, and compute requirements, influences its performance and accuracy. While larger models tend to excel in complex tasks, they also demand more computational resources and risk overfitting. To mitigate these challenges, techniques like quantization and sparsity are employed to reduce model size and maintain efficiency. The ongoing trend in AI involves developing increasingly larger models, such as OpenAI's GPT-3, pushing the boundaries of computational capabilities.

Koyeb is introducing a series of short blog posts to share interesting technical insights and ideas, some of which may relate to their work while others are simply topics they find intriguing. These posts will be brief, published three times a week, and will include "further reading" links for those interested in exploring the topics deeper. One featured topic is Drizzle, a lightweight, typesafe ORM for TypeScript that simplifies SQL queries, making complex query construction more accessible. The blog provides a practical example of setting up a PostgreSQL database and using Drizzle to perform queries, highlighting Drizzle's promise that if you know SQL, you know Drizzle. The posts encourage readers to share content they find valuable and provide their feedback or suggestions for future topics through various social media platforms.

Koyeb, a distributed team with a strong European presence, has welcomed Paweł as a new software engineer to help develop its state-of-the-art serverless engine, which aims to simplify deploying full-stack applications, AI workloads, APIs, and databases globally. Paweł, who joined Koyeb to tackle infrastructure and distributed systems challenges, shares insights from his programming journey, which began in high school during competitive math and programming courses. He finds problem formalization and solution design the most enjoyable aspects of technology, and his favorite side project, FashionLens, combined his interests in computer vision and neural nets to create a mobile app for finding clothes. Outside of work, Paweł enjoys active hobbies like lifting, wakeboarding, and climbing, and fondly recalls a sailing trip to Greece as his best vacation. He expresses his movie preferences, shifting from the Harry Potter saga to Dune, and invites connections through LinkedIn, GitHub, and Twitter. Koyeb offers a free tier for deploying initial services and databases, emphasizing ease of deployment with its One-Click Apps.

toddl.co is an all-in-one booking platform for children's activities in Spain, offering over 2,000 classes and events to 17,000 monthly visitors, and aims to serve 15,000 businesses and 50,000 families this year. The platform faced challenges with long build times and manual processes, prompting a switch to server-side rendering and a search for a more suitable hosting provider. The adoption of Koyeb, a managed infrastructure provider, allowed toddl.co to automate its CI/CD pipeline, reducing deployment times and infrastructure costs while scaling its business more effectively. The transition to Koyeb was completed in weeks, enhancing deployment speed, cost optimization, and operational efficiency, with features such as automatic continuous deployments, global load balancing, and seamless GitHub integration. This strategic move not only streamlined toddl.co's workflow but also significantly saved time and resources, enabling a faster push of changes to production and providing a robust platform for future growth.