Schedule

In this talk, I will analyze a practical approach to measuring self-hosted LLM performance.

The talk is about practical experience in optimizing inference and ML-serving based on GPUStack in the production environment of the corporate AI Portal.

I will tell you how we built a single knowledge graph on top of dozens of disparate corporate datasets — an infrastructure where an AI agent doesn't guess an answer based on similar chunks, but consciously navigates the structure and relationships of data.

I'll show you how to count memory and KV-cache, how inference layer solutions change the load profile, and then we'll move on to our implementation in Deckhouse.

A talk on how to build a production video analytics system for RTSP cameras: run an ML model, track objects, turn detections into business events, save results, and provide them via API and interface.

Let's explore how to create a good semantic search engine.

We will show how we built a scalable ML platform for detecting hackers using open-source tools (Airflow, Trino, Iceberg, and MLflow).

We'll dive into the architectural decisions behind CubeFS that make it possible to build exabyte-scale storage for ML and analytics workloads. Topics include its high-performance, horizontally scalable metadata service, local and distributed caching, transparent data movement across storage tiers, and other key design features.

Let's talk about what important functions are needed to manage Iceberg tables and the role of REST Catalog in this.

How to separate metadata storage and business logic in S3 object storage while maintaining performance and all supported functionality.

We are going to discuss the real experience of migrating data marts from a monolithic solution based on Greenplum 6 to the Data Lakehouse stack, paying attention to how to make this process the least painful for users. You will learn what non-obvious problems you will have to face and how to build processes so that the new architecture is more efficient than the legacy solution, rather than its less productive copy.

Our Trino storage hit the performance ceiling of a single Ceph cluster — so we started spreading every table across several clusters at once, hiding all the sharding logic in the HAProxy sidecars on our compute nodes, without adding a single new component to the architecture. Reads sped up from 20 to 60–80 GB/s, and GET latency dropped from minutes to 1–2 seconds.

Using a real‑world example, we’ll examine how a format breaks down without metadata about its own structure, and then, step by step, at the byte level, we’ll put together a new one.

Let's talk about testing, finding and debugging problems in highly loaded software, as well as support for storage with third-party vendor solutions.

How to teach an LLM not just to write plausible SQL, but to actually work with corporate data: find the right sources, understand metrics, write ETL, and validate your own answers.

I will analyze the components of success and failure and provide a practical checklist that will help you quickly decide whether you need an agent or a classic AutoML model to generate a baseline model.

LLM agents confidently hallucinate in business reports, and the accuracy of Text-to-SQL is clearly insufficient for regulatory and management reporting. I will show you how a semantic layer based on MetricFlow can increase accuracy to 90% or higher, and how to deploy this solution on-prem to ensure that your reports can be trusted.

How Vectorless helps you deal with the problem of losing data hierarchy.

Our experience of building a Metric Store.

Let's explore what actually happens between a request and a result.

The talk focuses on practical PostgreSQL performance diagnostics for backend developers who maintain their databases independently and do not have a dedicated DBA.

How pgvector works: vector storage, HNSW and IVFFlat algorithms, performance degradation points. An honest breakdown of where the solution holds up and where it doesn't.

I will explain how we implemented an atomic commit of distributed transactions at the PostgreSQL core level, based on the processing of 2PC/XA mechanisms, and show the results of its testing.

I will tell you about the experience of implementing and using YTsaurus in Chestny Znak.

Parallel reading from Kafka topics, KRaft, server balancing, and tiered storage.

In the talk, I will review the current state of the data transformation ecosystem, as well as alternative tools and promising projects that may replace dbt.

How we learned to use Python, K8s, and S3 to efficiently count data in the cloud.

I will tell you about implementing a review and deployment process for NiFi threads in a team with many developers, where changes to the threads are made several times a day.

A talk about why a collection of fragmented tools stops working at the scale of a large Data Platform, and why the platform should be viewed as a unified ADLC rather than a set of separate services. I will show how this affects ETL, ad hoc development, Data Governance, Data Quality, and metrics, and why AI and the agent-based approach are becoming the main drivers of new platform requirements.

I will tell you how we built the Magnit Data ecosystem, where the catalog, glossary, DQ engine, dashboards, and chatbot work as one mechanism.

A classic MDM system often assumes that data needs to be brought together in one place: loaded, normalized, matched, assigned a golden record, and then managed centrally as master data. But what do you do when, due to security or regulatory requirements, the system is not allowed to store data within its own perimeter?

On the production pipeline, we will show how one merge triggers validation, compatibility checks, ingestion generation, data publishing, and catalog updates.

Coming soon.

Let's discuss again State of Data.

Without getting boring and without slides, let’s discuss whether it’s worth using Greenplum (or some fork of it) for a new project in 2026.

AI is already helping to solve mathematical problems, write code, and conduct research. What remains a human task in this reality? Do we still need the old knowledge and skills, what should a university teach, and what will the scientist of the future be like?

20‑minute presentations on professional or semi‑professional topics and lively discussions.

Let’s talk about what remains valuable in a world of information overload, why attention becomes part of artistic value, and how conscious perception influences what art emerges and gains recognition.

Become a detective and try to solve the mysterious disappearance of the artist and his painting.

In this panel discussion, experts will pit three camps against each other: proponents of clean code, advocates of Low/No‑Code, and AI evangelists.

What real threats modern devices pose to an ordinary person, and whether you need to protect yourself from them.

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