Kubernetes, Cloud Native and Platform Engineering Meetup Munich - November Edition
Join us for an exciting evening of discussions around Kubernetes and Cloud Native technologies! Whether you're a seasoned DevOps engineer, a platform architect, or just starting your cloud-native journey, this meetup is designed to foster learning, networking, and innovation.
We’ll have two industry-leading speakers who will share their experiences. Expect deep dives into real-world use cases, best practices, and the latest advancements in the cloud-native ecosystem.
Sponsor: Thanks to our sponsor x-cellent!
Agenda:
6:00 PM - Arrival & Networking
6:30 PM - From Snapshots to Streams: A Live State Graph for LLM-Based Root Cause Analysis in Kubernetes - Oliver Stempel, Cloud Engineer @ x-cellent
7:15 PM - kubectl apply -f llm.yaml - and Then What? - Lukas Rammelmüller, Senior Consultant @ TNG Technology Consulting
8:20 PM - Networking & Refreshments
9:00 PM - Closing
Why Attend: Learn about the latest trends and updates in Kubernetes and Cloud Native development. Gain insights from an expert community with hands-on industry experience. Network with other professionals in the Kubernetes and Cloud Native space. Stay up-to-date on tools, best practices, and strategies to accelerate your Cloud Native journey. Whether you’re building modern applications or working on optimizing cloud infrastructure, this meetup will provide valuable takeaways for everyone involved in the Kubernetes ecosystem.
Talk 1: From Snapshots to Streams: A Live State Graph for LLM-Based Root Cause Analysis in Kubernetes
Root cause analysis (RCA) in Kubernetes is complicated by an implicit dependency structure spread across many short-lived resources and in constant flux. Existing LLM-based approaches rely on periodic cluster snapshots and LLM-generated Cypher queries, which introduce temporal blind spots, make results fragile, and yield conclusions operators cannot trace.
We propose a continuously updated live-state graph as an alternative foundation for LLM-based RCA. Kubernetes controllers stream normalized resource deltas through Redis Streams to a planner, which incrementally maintains a temporal property graph in Neo4j that separates stable entity identity from versioned state history. RCA is triggered automatically by Prometheus Alertmanager alerts and performed by a ReAct-style LLM agent restricted to a minimal, fixed interface of two tools: graph neighborhood traversal and structured report submission.
This replaces query generation with reproducible, inspectable interactions and makes each conclusion traceable to the graph evidence it rests on. We further contribute an extensible catalog of synthetic fault scenarios with ground truth, on which we evaluate ingestion throughput, freshness, and scalability as well as diagnostic correctness, efficiency, and interpretability.
Talk 2: kubectl apply -f llm.yaml - and Then What?
On the surface, serving large language models sounds easy: make GPUs available to a Pod, download the model weights and plug them into an inference engine. In a way, that's it. But once real traffic arrives, many familiar Kubernetes patterns stop working: simple load balancing ignores request cost and cache locality, CPU-based autoscaling misses the actual bottleneck, and pulling hundreds of gigabytes of weights turns a scale-up into a coffee break.
In this talk, we'll go through the ingredients for serving LLMs on Kubernetes: GPU cluster management, some bare-metal hosting and the inference-engineering techniques that keep latency and cost under control. You'll leave with a mental model of how LLM serving differs from typical stateless workloads, and which building blocks you'll need to serve your own models.
We want YOU! We are always looking for speakers - submit here! Want to host/sponsor/support us? This way :)