KubeCon + CloudNativeCon India 2024

Cloud Load Balancers (CLBs) have evolved significantly, now featuring advanced capabilities such as proxying, SSL/TLS termination, and auto-scaling. The underlying data-path technologies have advanced from traditional Linux networking to include eBPF, DPDK, VPP, and P4. Each of these technologies offers unique advantages and trade-offs: eBPF excels in real-time processing but lacks some comprehensive features, DPDK provides high performance with reduced transparency, VPP offers modular packet handling with strong Linux compatibility, and P4-based solutions, while still evolving, promise enhanced flexibility and performance. This session will delve into these data-path technologies, examining their maturity, limitations, and suitability for modern data centers. As specialized hardware units become increasingly integral to large data centers, we will highlight how these units effectively leverage data-path technologies to perform numerous tasks, thereby enhancing overall efficiency.

As EarnIn, one of the leading fintech companies, rapidly scaled in the number of microservices and user base, we recognized the need to ensure our network layer could keep pace with our scale. Initially relying on an enterprise service mesh that eventually became a bottleneck for our applications as well as our operators, we explored all the available service mesh out in the market, ultimately choosing Linkerd. This talk will delve into our decision-making process, focusing on why we chose LinkerD and how it has transformed our network and application architectures. We'll cover the crucial aspects of Linkerd that enhanced our system's performance, robustness, and security, all while simplifying operations and ensuring reliabilityᅳensuring our operators can rest easy. Join us to learn how LinkerD can elevate your service mesh strategy to meet the demands of substantial growth while keeping your architecture simple.

This talk will explain new options for proxyless service mesh to enable secure authentication with serverless-based backends. Vanity URLs and reverse proxies have historically posed an issue for proxyless service mesh, making it difficult to use them, and the speaker will talk about upcoming new features in gRPC that make this possible, while also ensuring robust security and preventing misrouted traffic. The audience will discover how proxyless service meshes secure authentication with serverless backends, opening up new possibilities for modern architectures. The audience will also come to know how the client can intelligently determine the control plane's support for these new features, enhancing adaptability and streamlining deployment.

While anyone working with Kubernetes knows about Controllers and Custom Resources, most don't know that there are many ways of designing a Controller, but just one good practice. That practice is "single resource": your controller should only manage a single resource, instead of having a single controller for multiple resources or multiple controllers for a single resource. This talk guides you on implementation of single-resource Controllers, showcases examples of a well-designed controller and good practice of designing Single-Resource controllers by comparing it with discouraged alternative designs. This talk focuses on practical insights, on aforementioned design’s pros and cons, and how its implementation affects the efficiency of Operators. After the talk, attendees will be able to design and implement single-resource controller, and have a deeper understanding of good practices of designing a kubernetes controller

The rapidly evolving landscape of Machine Learning and Large Language Models demands efficient scalable ways to run distributed workloads to train, fine-tune and serve models. Ray is an Open Source framework that simplifies distributed machine learning, and Kubernetes streamlines deployment. In this introductory talk, we'll uncover how to combine Ray and Kubernetes for your ML projects. You will learn about: - Basic Ray concepts (actors, tasks) and their relevance to ML - Setting up a simple Ray cluster within Kubernetes - Running your first distributed ML training job

Traditionally, search engines and vector DBs (the boundaries got blurry lately) are deployed as monoliths in a cluster, managed by a StatefulSet and a lot of technology-specific logic. Newer use-cases (e.g. RAG, semantic search in general) implicitly push for more stateless work: e.g. generate embeddings on the fly, re-rank, local LLMs. Decoupling this work and managing it separately makes more sense now, but is it worth the complexity? This session takes a deep dive into the architectures of search engines (mainly Vespa, but Elasticsearch, Solr and others will be in the picture, too) to highlight both pros and cons of separating concerns in information retrieval. And how much separation is good for which scenario. In the end, you should have a better idea about both the technology and the deployment type that fits your use-case best.