Schedule

The talk examines the path from AI-powered operation to self-managed infrastructure: levels of autonomy, agent-based control loops, dynamic sandboxes, operational security, verification, and the changing role of the human operator.

Neural networks and LLMs are often perceived as impenetrable "magic," but in reality, their perimeter is less secure than that of classic IT systems.

If you want to feel like a DevVibeOps, then come to me! :)

An ML team has come to you: "The LLM is responding in the container, now give it to the users and maintain it."

Let's explore what can go wrong on the way to production: weights and GPUs, the first token, keys and accesses, updates, metrics, and crashes. And how to handle it with minimal losses.

I will tell you how we built MagnitGPT, what problems we solved during development, and how it helps hundreds and thousands of users every day.

How to build an agent system where an agent has control over your infrastructure.

I will show you how to make running Spark tasks convenient for DE/ML engineers and maintain cluster management.

I'll tell you how we migrated Kafka from VM to Kubernetes.

How our fault-tolerant K8s clusters are structured in the infrastructure, and how we have automated their management through CAPI on a scale of hundreds of servers.

Using examples from RTLabs practice, I will show you how to find errors in CI/CD that are not caught by linting Helm charts.

Let's compare the capabilities of ArgoCD and Flux, kubectl, Helm, and Nelm/werf in organizing deployments, monitoring readiness, and managing resource lifecycles.

How to find the causes of complex problems in GPU clusters when they occur at the junction of different layers of the system and server hardware, and why the real causes of such problems are often not where you would expect to find them.

I'll tell you about working with PCI devices, including under high load conditions, and about some bugs that we found in Kubelet.

A talk on building a large "hard" infrastructure and how we at K2 Cloud have gone from complete chaos to a standardized yet flexible system that allows us to quickly adapt to the dynamic pace of cloud platform development.

Real experience in evolving the approach to managing centralized configuration of Gitlab, Hashicorp Vault, and other projects in 2GIS: from simple bash scripts to a combination of Terraform and Terragrunt.

How the Cluster API update almost destroyed our clusters, how we mitigated it, what we changed, and how we're working now.

The main idea of the talk is simple: a good migration does not require heroics, but becomes a predictable and boring engineering task.

Modern powerful server GPUs are liquid-cooled by default. I will tell you what options are possible when creating a GPU computing cluster or data center for AI, as well as what benefits we receive and how we pay.

The talk will cover onboarding, learning mechanisms, knowledge transfer, technologies used, and working conditions. You will learn about the perks and benefits you pay for.

We will tell you about the difficulties we faced and show you how we turned chaos into a manageable system: what went wrong at the start, which tools actually worked, and how we measured the effect.

I'll describe how we created the DevOps Cookbook.

Every other incoming resume lies (17 years of Kubernetes for a thing that is 12 years old), and the model that catches it either bankrupts you on bills, lies itself, or makes a hiring call you cannot later explain to lawyers. I show how to fix this with a three-tier model cascade with the expensive one last, three guard layers, and real before-and-after numbers — on a live collection of caught fabrications.

I'll show you how to build a cascade deployment at FluxCD and not to spread the erroneous change to all environments at once.

I will explain how the Middleware Service Hub transforms disparate data into a connected technology context, evolves from a component catalog to an organization's Digital Twin, and becomes the foundation for Architecture as Code, API governance, and AI Harness.

How to build a scalable system of dynamic test environments and avoid the problems associated with shared test benches.

The story of DORA metrics implementation: from decomposition and value formulation to implementation results and error analysis.

Let’s illustrate how our tracing platform grew, how the tools evolved — from Tempo to VictoriaTraces, what challenges we faced, what problems we solved, and how we enhanced the service guarantees.

I want to tell you how and why we implemented OpenTelemetry and Observability-as-Service in Dodo Engineering, what traps we fell into, what hypotheses we tested, and why you need or don't need it all.

I will explain how NGFW import substitution in a bank turned into a redesign of the production Kubernetes network model: namespace-level access control, cloud migration, and moving away from the familiar syslog-based authorization flow.

The talk is divided into two parts: network isolation of infected infrastructure and how to do it for engineers. And where to run if you need to deploy new infrastructure.

Coming soon.

We'll talk about the schedule, sessions and share information.

A team workshop where you solve a real-time combat incident using board game rules.

The main idea of my session is to show in practice the possibilities of improving the "standard" Kubernetes.

Are we ready to put LLM into a closed loop, knowing about hallucinations and unpredictability?

The hottest hiring trends questions from big-tech interviewers. Come to feel the engineering culture and refresh your knowledge!

Fail Talks is a series of sincere and useful stories about failures in DevOps practices.

Why no one needs to release 70 times more often, and will engineers still be needed in 2027?

Let's explore popular open-source tools that can enhance Kubernetes security at runtime.

Summing up the results of the conference, remembering the highlights and talking about plans.