PDFs are one of the largest sources of unstructured data, but most pipelines treat them as trusted input. They're not. PDFs are arbitrary code, and parsing them is a real attack surface. This talk presents a rasterize-first architecture: sandbox the PDF, render to images, then OCR. I'll cover the threat model, the three-stage pipeline, accuracy benchmarks against traditional extraction, and lessons from scale. You'll leave with a practical approach to processing PDFs safely without sacrificing quality.

Table services—including Compaction, Cleaning, and Clustering—are critical to balancing ingestion latency with query performance. In this talk, we share our learnings from operating these services at Uber scale. We will dive into the technical architecture that powers our background maintenance and discuss how we decouple these operations to ensure data freshness and high-performance analytics on a massive scale.

Data and AI workloads at Walmart are evolving rapidly and require data streaming at millions per second scale. Every day thousands of pipelines ingest data at trillion scale. This talk will focus on how we built a multi-cloud, multi-tenant self service Streaming Engines (Kafka Connect, Spark, Flink). Lessons learned and how we optimized a Petabyte worth compute reservation to few hundred terabytes to improve multi-tenancy and scalability.

In this talk, I will dive into the transformative journey of Conductor, transitioning from feature teams to platform teams. I will showcase the implementation of updated data pipelines and the creation of a Data Platform utilizing Apache Hudi and Onehouse. This platform enables us to fully harness the potential of our data, thereby enhancing user value significantly.The new architecture facilitates the rapid integration of new data sources and pipelines, enabling the development of innovative products based on AI search results from ChatGPT, Perplexity, and other AI engines. I will illustrate how open table formats have empowered us to experiment with consumption-optimized tools and demonstrate the efficient utilization of Starrocks for the dynamic generation of analytical reports within seconds.

Attendees will gain insights into the process of integrating a Data Lake with Apache Hudi and Onehouse into existing architectures, identifying successful strategies and areas for improvement, and several optimization techniques will be presented.

This presentation holds significant relevance for OpenXData, as it demonstrates the tangible benefits that businesses can derive from technical solutions grounded in open architectures such as Apache Hudi with Onehouse.

This session provides an overview of the current state of data lake at JD.com. We will delve into our team's latest innovations within the Apache Hudi core, detailing the technical characteristics. We will also share practical business implementations that leverage these newly developed capabilities, followed by insights into our future roadmap.

Agentic AI is rapidly moving from “assistive” to “autonomous”: natural-language requests are now translated into SQL and Spark programs that can execute end-to-end with little or no oversight. That shift changes the risk profile of modern data platforms. A single hallucinated join, an unbounded scan, a silent Cartesian product, or a poorly placed UDF can explode costs, violate access policies, or publish incorrect results—at machine speed.

This talk explores where governance should live when the generator is probabilistic and the runtime is powerful. We’ll compare three patterns: (1) “AI with AI” approaches using LLM judges and automated policy checks, (2) traditional human-in-the-loop review, and (3) a pragmatic middle path that combines deterministic controls with selective human escalation. We’ll outline a governance architecture that sits before execution and before productionization: a pre-flight layer that validates intent, enforces policy, and optimizes queries and code artifacts prior to deployment.

You’ll see how to implement layered guardrails—schema and lineage checks, access control verification, cost and performance estimation, semantic validation against known metrics, and regression tests on representative data—paired with explainability and approval workflows when confidence is low. We’ll also discuss how LLM judges can add value (e.g., intent consistency, anti-pattern detection, test generation) while keeping them bounded by verifiable rules and measurable acceptance criteria.

By the end, you’ll leave with a blueprint for a system that turns “prompt-to-prod” into a controlled pipeline: safer execution, predictable cost, better performance, and auditable accountability—without killing the productivity gains that made agentic AI attractive in the first place.

This talk will cover the Vortex, a Linux Foundation project building the fastest open-source file format on both CPUs and GPUs. We'll briefly cover Vortex's performance, and then break down how we built the format to support efficient late materialization on both GPUs & CPUs. The key insight of the talk is that designing for GPUs results in better CPU vectorization as well as random access.

What is a Data Lakehouse and why did I decide to build one for myself?  This talk covers the process of building a personal Data Lakehouse with open source tooling and explores data management and the architectural considerations for a data management system and how system requirements and characteristics differ in enterprise, corporate, and personal contexts.

At Booking.com, we serve millions of real-time ML predictions every minute powering ranking, fraud detection, and personalization. To meet strict latency requirements, we built an ultra-low latency feature platform capable of serving features with p99.9 latency under 25ms at 200K requests per second.
In this talk, we share how we designed a high-throughput, self-service feature platform built on Amazon ElastiCache, enabling teams across Booking.com to serve ML features reliably at scale while maintaining strict latency SLOs.

Most data architectures were designed around a simple assumption: queries are deterministic. You send a query, you get a result. Agents break that assumption. The same input, the same data, a different output — sometimes subtly, sometimes catastrophically. In this talk, I'll walk through what that non-determinism actually means for open data architectures at scale: why your existing governance model doesn't account for it, why MCP changes the integration calculus but doesn't solve the trust problem, and what three specific architectural decisions teams should make before their agent layer is in production. No frameworks. No vision. Just the problems showing up in real systems right now and the patterns that are working.

At Preset, we don't have a dedicated data team. Instead, we've built an AI agent, internally called "DatAgor," that handles a surprising amount of what a data team would do: answering ad-hoc analytics questions, building charts and dashboards, debugging pipeline issues, and helping anyone in the company self-serve on data. It's not magic, and it doesn't get everything right, but it's genuinely useful, and adoption across the team is growing fast.

This talk is an honest walkthrough of what we built, what works well, and where the gaps still are.

The foundation is an Agor Assistant, a persistent AI agent with access to our full data stack. Through Superset MCP, it can query data, build charts, and create dashboards, the same operations a human would do in the UI. We've layered in access to our dbt models and Airflow pipelines for semantic context and data freshness, a custom CLI skill for Superset operations, and internal documentation. When tasks get complex, it can dispatch sub-agents to work in parallel. And through Slack integration, anyone at Preset, engineers, PMs, execs, can just ask a question and get a useful answer back.

The honest version: it handles the bulk of routine analytics work remarkably well. Ad-hoc questions, straightforward dashboards, pipeline investigations, these are its sweet spot. But it still needs human review. It doesn't always have full business context. The dashboards it builds are functional, not beautiful. Complex analysis still requires back-and-forth and sometimes a human taking the wheel. The good news is that it learns. Through a memory and skill system, it gets better at the workflows your team actually uses.

We'll demo real workflows, share the architecture, and talk candidly about what it takes to run this in practice, not as a replacement for data expertise, but as a force multiplier that makes a small team punch well above its weight.

Key takeaways:
- Architecture of a production AI data agent: Superset MCP, pipeline context, sub-agent orchestration, and Slack access
- Where agentic analytics works great today and where it still needs a human in the loop
- How memory and skill systems help the agent improve over time on your team's specific workflows