On-Demand Videos
Real-time OLAP databases are optimized for speed and often rely on tightly coupled storage-compute architectures using disks or SSDs. Decoupled architectures, which use cloud object storage, introduce an unavoidable tradeoff: cost efficiency at the expense of performance. This makes them unsuitable for databases that need to provide low-latency, real-time analytics, especially the new wave of LLM-powered dashboards, retrieval-augmented generation (RAG), and vector-embedding searches that thrive only when fresh data is milliseconds away. Can we achieve both cost efficiency and performance?
In this talk, we’ll explore the engineering challenges of extending Apache Pinot—a real-time OLAP system—onto cloud object storage while still maintaining sub-second P99 latencies.
We’ll dive into how we built an abstraction in Apache Pinot to make it agnostic to the location of data. We’ll explain how we can query data directly from the cloud (without needing to download the entire dataset, as with lazy-loading) while achieving sub-second latencies. We’ll cover the data fetch and optimization strategies we implemented, such as pipelining fetch and compute, prefetching, selective block fetches, index pinning, and more. We'll also share our latest work about integration with open table formats like iceberg, and how we will continue to achieve fast analytics directly on parquet files by implementing all the same techniques that apply to tiered storage.
The data lake is a fantastic, low-cost place to put data at rest for offline analytics, but we've built it under the terms of a terrible bargain: all that cheap storage at scale was a great thing, but we gave up schema management and transactions along the way. Apache Iceberg has emerged as king of the Open Table Formats to fix this very problem.
Built on the foundation of Parquet files, Iceberg adds a simple yet flexible metadata layer and integration with standard data catalogs to provide robust schema support and ACID transactions to the once ungoverned data lake. In this talk, we'll build Iceberg up from the basics, see how the read and write path work, and explore how it supports streaming data sources like Apache Kafka™. Then we'll see how Confluent's Tableflow brings Kafka together with open table formats like Iceberg and Delta Lake to make operational data in Kafka topics instantly visible to the data lake without the usual ETL—unifying the operational/analytical divide that has been with us for decades.
Storing data as Parquet files on S3 is increasingly used not just as a data lake but also as a lightweight feature store for ML training/inference or a document store for RAG. However, querying petabyte- to exabyte-scale data lakes directly from cloud object storage remains notoriously slow (e.g., latencies ranging from hundreds of milliseconds to several seconds on AWS S3).
In this talk, we show how architecture co-design, system-level optimizations, and workload-aware engineering can deliver over 1000× performance improvements for these workloads—without changing file formats, rewriting data paths, or provisioning expensive hardware.
We introduce a high-performance, low-latency S3 proxy layer powered by Alluxio, deployed atop hyperscale data lakes. This proxy delivers sub-millisecond Time-to-First-Byte (TTFB)—on par with Amazon S3 Express—while preserving compatibility with standard S3 APIs. In real-world benchmarks, a 50-node Alluxio cluster sustains over 1 million S3 queries per second, offering 50× the throughput of S3 Express for a single account, with no compromise in latency.
Beyond accelerating access to Parquet files byte-to-byte, we also offload partial Parquet processing from query engines via a pluggable interface into Alluxio. This eliminates the need for costly index scans and file parsing, enabling point queries with 0.3 microseconds latency and up to 3,000 QPS per instance (measured using a single-thread)—a 100× improvement over traditional query paths.
.png)
Uber builds one of the biggest data lakes in the industry, which stores exabytes of data. In this talk, we will introduce the evolution of our data storage architecture, and delve into multiple key initiatives during the past several years.
Specifically, we will introduce:
- Our on-prem HDFS cluster scalability challenges and how we solved them
- Our efficiency optimizations that significantly reduced the storage overhead and unit cost without compromising reliability and performance
- The challenges we are facing during the ongoing Cloud migration and our solutions
Shengxuan Liu from ByteDance presents the new ByteDance’s native Parquet Reader. The talk covers the architecture and key features of the Reader, and how the new Reader is able to facilitate data processing efficiency.
In this session, cloud optimization specialists Chunxu and Siyuan break down the challenges and present a fresh architecture designed to optimize I/O across the data pipeline, ensuring GPUs function at peak performance. The integrated solution of PyTorch/Ray + Alluxio + S3 offers a promising way forward, and the speakers delve deep into its practical applications. Attendees will not only gain theoretical insights but will also be treated to hands-on instructions and demonstrations of deploying this cutting-edge architecture in Kubernetes, specifically tailored for Tensorflow/PyTorch/Ray workloads in the public cloud.
In this session, Jingwen presents an overview of using Alluxio Edge caching to accelerate Trino or Presto queries. She offers practical best practices for using distributed caching with compute engines. In addition, this session also features insights from real-world examples.
As a cache eviction algorithm, FIFO has a lot of attractive properties, such as simplicity, speed, scalability, and flash-friendliness. The most prominent criticism of FIFO is its low efficiency (high miss ratio). In this talk, Juncheng Yangb describes a simple, scalable FIFO-based algorithm with three static queues (S3-FIFO). Evaluated on 6594 cache traces from 14 datasets, we show that S3- FIFO has lower miss ratios than state-of-the-art algorithms across traces. Moreover, S3-FIFO’s efficiency is robust — it has the lowest mean miss ratio on 10 of the 14 datasets. FIFO queues enable S3-FIFO to achieve good scalability with 6× higher throughput compared to optimized LRU at 16 threads. Our insight is that most objects in skewed workloads will only be accessed once in a short window, so it is critical to evict them early (also called quick demotion). The key of S3-FIFO is a small FIFO queue that filters out most objects from entering the main cache, which provides a guaranteed demotion speed and high demotion precision.
Many companies are working with development architectures for AI platforms but have concerns about efficiency at scale as data volumes increase. They use centralized cloud data lakes, like S3, to store training data for AI platforms. However, GPU shortages add more complications. Storage and compute can be separate, or even remote, making data loading slow and expensive:
- Optimizing a developmental setup can include manual copies, which are slow and error-prone
- Directly transferring data across regions or from cloud to on-premises can incur expensive egress fees
This webinar covers solutions to improve data loading for model training. You will learn:
- The data loading challenges with distributed infrastructure
- Typical solutions, including NFS/NAS on object storage, and why they are not the best options
- Common architectures that can improve data loading and cost efficiency
- Using Alluxio to accelerate model training and reduce costs
As the AI landscape rapidly evolves, the advancements in generative AI technologies, such as ChatGPT, are driving a need for a robust AI infra stack. This opening keynote will explore the key trends of the AI infra stack in the generative AI era.
In this session, Adit Madan, Director of Product Management at Alluxio, presents an overview of using distributed caching to accelerate model training and serving. He explores the requirements of data access patterns in the ML pipeline and offer practical best practices for using distributed caching in the cloud. This session features insights from real-world examples, such as AliPay, Zhihu, and more.
Machine learning models power Uber’s everyday business. However, developing and deploying a model is not a one-time event but a continuous process that requires careful planning, execution, and monitoring. In this session, Sally (Mihyong) Lee, Senior Staff Engineer & TLM @ Uber, highlights Uber’s practice on the machine learning lifecycle to ensure high model quality.
In this talk, Wanchao Liang, Software Engineer at Meta Pytorch Team, explores the technology advancements of PyTorch Distributed, and dives into the details of how multi-dimensional parallelism is made possible to train Large Language Models by composing different PyTorch native distributed training APIs.
ChatGPT and other massive models represents an amazing step forward in AI, yet they do not solve real-world business problems. In this session, Jordan Plawner, Global Director of Artificial Intelligence Product Manager and Strategy at Intel, surveys how the AI ecosystem has worked non-stop over this last year to take these all-purpose multi-task models and optimize them to they can be used by organizations to address domain specific problems. He explains these new AI-for-the-real world techniques and methods such as fine tuning and how they can be applied to deliver results which are highly performant with state-of-the-art accuracy while also being economical to build and deploy everywhere to enhance products and services.
This hands-on session discusses best practices for using PyTorch and Alluxio during model training on AWS. Shawn and Lu provide a step-by-step demonstration of how to use Alluxio on EKS as a distributed cache to accelerate computer vision model training jobs that read datasets from S3. This architecture significantly improves the utilization of GPUs from 30% to 90%+, archives ~5x faster training, and lower cloud storage costs.