CWI Lectures & Dijkstra Fellowship for Marcin Żukowski

Title: Architecting the Snowflake Data Cloud

Abstract
In 2014, Snowflake emerged from stealth mode and announced a "data warehouse, built for the cloud". The system was built from the ground up, using a new architecture called the "multi-cluster shared data" architecture, which separates compute from storage, in order to allow users to take advantage of the elasticity of the cloud. In the decade since then, our system has evolved into a global, multi-cloud platform that supports data collaboration and many data processing workloads beyond data warehousing. In this talk, I will explain why the multi-cluster shared data architecture is well suited for the cloud, and how this architecture has enabled capabilities like massive scalability, automatic optimizations, and global secure data collaboration. Finally, I'll outline how the architecture has evolved over time as our capabilities have grown, and a few lessons learned along the way.

Biography
Allison Lee has spent the last two decades building commercial database systems. She currently leads the database engineering team at Snowflake, which owns the core query processing stack, Snowflake's internal key-value metadata store, and the data warehousing and hybrid transactional-analytic workloads. Allison joined Snowflake in 2013 as a founding engineer, and contributed to many of Snowflake's foundational components, including the query optimizer, metadata management, language extensibility, and data protection.

Prior to Snowflake, Allison worked on the Oracle database, with a focus on query optimizations for analytical workloads and bridging the gap between query optimization and execution with adaptive optimization techniques.

Allison holds dozens of patents in query optimization, data protection, and secure data collaboration.

Allison holds a Bachelor of Science and a Master of Engineering in electrical engineering and computer science from the Massachusetts Institute of Technology.

Title: What Goes Around Comes Around... And Around...

Abstract
Doesn't it feel like there is always a new crop of database management systems (DBMSs) pushing the idea that the relational model (RM) is outdated and SQL is dying? Proponents of vector databases have recently taken up this mantle fueled by interest in AI/ML technologies. Before that, NoSQL and MapReduce users claimed that RM/SQL was insufficient for "webscale" applications. And then, back in the 1990s, the object-oriented database vendors wanted every developer to switch to their non-RM, non-SQL systems. Database history does not repeat itself, but it indeed does rhyme.

In this talk, I will present the 60-year history of data modeling research and demonstrate why RM/SQL is the preferred default choice for database applications of any size. All efforts to completely replace the data model or query language have failed. Instead, SQL absorbed the best ideas from these alternative approaches and remains relevant for modern applications.

Biography
Andy Pavlo is an Associate Professor with Indefinite Tenure of Databaseology in the Computer Science Department at Carnegie Mellon University. He knows some pile about databases.

Title: Leaving The Two Tier Architecture Behind

Abstract
Analytical data management systems have long been monolithic monsters far removed from the action by ancient protocols. The common two- or three-tier architecture firmly placed data engines on dedicated, expensive hardware. Clients were reduced to politely asking for answers to their questions and then patiently waiting for an answer.

High-efficiency data processing methods like vectorized processing and morsel-driven parallelism greatly increase the capabilities of data engines. This coincides with both a staggering growth in hardware capabilities and a leveling-off of useful dataset growth. Together, those developments allow moving analytical engines everywhere, if properly designed.

The new class of in-process analytical engines allows them to move into any application process, which greatly streamlines data transfer, deployment, and management. This new class of systems also enables a whole new list of use cases, for example in-browser or edge OLAP, running SQL queries in lambdas and Big Data on laptops.

Biography
Prof. Dr. Hannes Mühleisen is a creator of the DuckDB database management system and Co-founder and CEO of DuckDB Labs, a consulting company providing services around DuckDB. Hannes is also Professor of Data Engineering at Radboud Universiteit Nijmegen and a senior researcher at the CWI Database Architectures Group. His main interest are - shockingly - analytical data management systems.

Title: What Table Representation Learning Brings to Data Systems

Abstract
We all observe the impressive capabilities of representation learning and generative models for text, videos, and images on a daily basis. Structured data such as tables in relational databases, however, have long been overlooked despite their prevalence in the organizational data landscape and critical use in high-value applications and decision-making processes. Learned representations, or embeddings, that capture the semantics of structured data can play a key role in making data systems more efficient, robust and accurate, at scale. Models that generalize to real-world databases are critical to make this work. In this context, I will discuss how rather compact and specialized column embeddings can be more effective than using GPT-something for table understanding, and reflect on the importance of capturing the core properties of relational databases in the embedding space. I will close by illustrating the value of embeddings for table retrieval to make LLM-powered query interfaces to structured data truly useful.

Biography
Dr. ir. Madelon Hulsebos is a tenure track researcher at CWI in Amsterdam. Prior to that, she was a postdoctoral fellow at UC Berkeley, and received her PhD from the University of Amsterdam for which she did research at MIT and Sigma Computing. Her general research interest is on the intersection of machine learning and data management, currently focusing on Table Representation Learning to democratize insights from structured data. Madelon founded the Table Representation Learning workshop at NeurIPS, and leads various other efforts in this space. She was awarded a BIDS-Accenture fellowship for her postdoctoral research on retrieval systems for structured data at UC Berkeley as well as a 5-year AiNed fellowship grant.

Title: The Future Of Cloud Database Systems

Abstract

Cloud computing is transforming the technology landscape, with database systems at the forefront of this change. A striking example is an online bookstore that has grown to dominate the database market. The appeal of cloud computing for IT users lies in several key factors: a reduced total cost of ownership through economies of scale and advanced services that minimize the burden of "undifferentiated heavy lifting". More broadly, cloud computing reflects a civilizational trend toward increased technological and economic specialization.

However, the current state of cloud computing often falls short of these promises. Hyperscalers are evolving into vertically integrated oligopolies, controlling everything from basic server rentals to high-level services. This trend is only accelerating, potentially leading to a future where hyperscalers establish software standards and design their own hardware, making it impossible to compete. Moreover, despite differences in branding, the major cloud providers are fundamentally similar, lacking interoperability and fostering vendor lock-in. As a result, we risk returning to the monopolistic conditions of the IBM and Wintel eras and ultimately technological stagnation due to limited competition.

Yet there is cause for optimism. Great technology can still succeed, as the multi-cloud data warehouse Snowflake has shown. The rise of data lakes and open standards, such as Parquet and Iceberg, further underscores the potential for interoperability and innovation. Additionally, there are orders-of-magnitude gaps between the price of existing cloud services and what is theoretically achievable, creating opportunities for disruption. These price gaps persist because cloud services are inherently complex to build, requiring redundant efforts and leading to high barriers to entry. For example, a DBMS might need a highly available control plane, a write-ahead log service, and distributed storage servers. None of these abstractions is available as a read-to-use service, which makes it difficult to enter the cloud database market. The current cloud landscape is more a result of historical circumstances than optimal design, leaving ample room for disruption.

In this talk, I will outline a blueprint for reinventing the cloud by focusing on three key areas: First, we need a unified multi-cloud abstraction over virtualized hardware. Second, we should establish new open standards for existing low-level cloud services. Third, we need abstractions that simplify the creation of new cloud services, such as reusable control planes and foundational components like log services and page servers. Together, this will make it significantly easier to build, deploy, and monetize new cloud services. Increased competition would commoditize foundational services and spur technological innovation.

Biography
Viktor Leis is a Professor in the Computer Science Department at the Technical University of Munich (TUM). His research focuses on designing cost-efficient data systems for the cloud, encompassing core database systems topics such as query processing, query optimization, transaction processing, index structures, and storage. He received his doctoral degree in 2016 from TUM and held professorships at the universities of Jena and Erlangen before returning to TUM in 2022.