tldr - powered by Generative AI

• Resource management in Kubernetes involves setting CPU and memory requests and limits for containers
• Memory limits result in termination of pods when exceeded, while CPU limits can lead to throttling or termination
• Setting appropriate requests and limits is crucial for efficient scheduling and avoiding noisy neighbors
• Developers should be aware of namespace limits and available resources when setting requests and limits
• In general, it is advisable to set memory requests equal to memory limits and avoid setting CPU limits
• Exceptions include cases where consistent workloads or overcommitment of memory are preferred

tldr - powered by Generative AI

• Ebpf technology was used to achieve cni networking with Mizar and XDP without changing any lines of kernel code
• Future plans include enhancing the technology with a tx hook for XDP, proposing a formal EPA change to Kubernetes, and improving performance measurement
• The speaker also discusses the need for a management plane and multi-tenant networking
• The presentation includes a demo of the technology using four virtual machines

tldr - powered by Generative AI

• Cloud and Kubernetes can solve problems of legacy infrastructure by providing on-demand, elastic, and scalable resources with strong resource isolation and cutting-edge security techniques.
• Design principles include fully embracing public cloud and cognitive way of thinking, containerization for elasticity and reproducibility, and decoupling compute and storage for independent scaling.
• The architecture of the cloud-native Spark on Kubernetes platform involves multiple Spark Kubernetes clusters, a Spark service gateway, and a multi-tenant platform with advanced features such as physical isolation and min/max capacity setting.
• Auto-scaling is necessary for cost-saving and elasticity, and the presentation discusses the design of reactive auto-scaling and its productionization.
• The platform has been running in production for a year, supporting many business-critical workloads for Apple AML.

tldr - powered by Generative AI

• The challenge faced by the customer was to optimize their application while keeping on releasing application updates to introduce new business functionalities and align to new regulations.
• The tuning practice in place was manual and took almost two months to tune one single macro service.
• The AI-powered optimization methodology works in five steps: applying new configuration suggested by AI, applying workload to target system, collecting KPIs, analyzing results, and producing new configuration to be tested in the next iteration.
• The methodology allows setting constraints and goals, such as minimizing application cost and ensuring service reliability.
• The presentation provides an anecdote of how the methodology was used to optimize a customer's authentication service on Kubernetes, resulting in a 49% improvement on cost efficiency compared to the baseline configuration.

tldr - powered by Generative AI

• Factors to consider when sizing your Kubernetes application
• Horizontal vs Vertical autoscaling
• Selecting the right auto scaling metric for your application
• A Turing-complete autoscaler demo

tldr - powered by Generative AI

• Telemetry Aware Scheduling (TAS) is an open-source project that extends Kubernetes' scheduling paradigm to use knowledge of resources to impact scheduling decisions.
• TAS uses telemetry to help make scheduling decisions and is an extender of the Kubernetes scheduler.
• TAS allows for filtering and scoring nodes and utilizes node affinity rules via fixed and custom labels.
• TAS uses telemetry where scheduling policies that are structurally based on rules which are based on metrics that come from the cluster.
• TAS requires a metrics pipeline to expose, collect, store, and make metrics available to the Kubernetes custom metrics API.
• TAS works together with the default scheduler and returns a suggested outcome of pod placement to the default scheduler.
• TAS supports multi-metric rules that contain multiple metrics and can link them together with operators such as any off or all of.

tldr - powered by Generative AI

• Parka is a tool for profiling and analyzing performance in software applications
• The storage architecture of Parka is designed to handle stack traces as a first-class citizen
• Write requests are ingested and validated using protobuf and metadata label sets
• The metadata store is implemented in SQLite and can be used with any SQL database
• The end result of a write request is a set of location IDs and corresponding sample values