COMPASS Talks

Abstract:

As enterprises collect ever more data from various sources for improved decision making and operational efficiency, data integration is becoming increasingly complex and critical. Even more, the disparate functionalities between different systems make it difficult for users to build and optimally run pipelines and workloads that address their needs. Advancements in reasoning LLMs enable such complex management tasks to be automated with the help of agents, to significantly reduce the effort and time otherwise required, and to optimize for cost and performance. In this talk, we introduce an AI agent-enabled composable data integration stack that assists data engineers in designing, productizing, and executing complex pipelines with unprecedented efficiency.

Abstract:

The need for ever more powerful GPUs, coupled with never-before-seen capital investments, motivates chipmakers to advance their ecosystem rapidly. The exponential boom of data used in training demands the fastest interconnect, the largest on-device memory possible, and so on. We explore this trend from a system perspective and discuss what this means for data processing workloads. We show how GPUs can be orders of magnitude faster for SELECT operations and look at some practical usage scenarios for the cloud and on-prem.

Abstract:

This talk will be focused on concurrency and design choices, in the context of concurrent applications and specifically for DBMS (Database Management Systems). Concurrency is hard; therefore it makes sense to be one of the top architectural considerations in software engineering. I hope to show you how tiny choices early in the design phase can have big consequences later down the road. And that there are certain concurrency abstractions where representing business logic is a lot easier than others. I'll show some specific examples, based on my personal experience in organization with thousands of developers. We'll go over the tricks we use to allow these developers to write correct concurrent and scalable applications, in a complex code base with hundreds of millions of lines. And we'll see some of the building blocks that have allowed us to build the world's highest scaling (single-node) DB engine, and how these have even more relevance in the context of CXL.

Bio:

Pedro works at Cisco Systems, on the core DB engine, designing and implementing scalable software. His current research is in Software Transactional Memory, and how to use concurrency abstractions to let software developers implement correct and scalable failure-resilient applications.