Research
AnHai Doan
Data
integration (DI) has been a long-standing challenge in the data management
community. So far the vast majority of DI works have focused on developing DI algorithms.
Going forward, we argue that far more efforts should be devoted to building DI systems, in
order to advance the field. DI is engineering by nature. We cannot just keep developing DI
algorithms in a vacuum. At some point we must build end-to-end systems to evaluate the
algorithms, to integrate research and development efforts, and to make practical impacts.
The question then is what kind of DI systems we should build, and how? In this direction we
focus on identifying problems with current DI systems, then developing a radically new
agenda for building DI systems. These new kinds of DI system have the following
distinguishing characteristics:
1. They guide the user through the end-to-end DI workflow, step by step.
2. For each step, they provide automated or semi-automated tools to address the "pain points" of the step.
3. Tools seek to cover the entire DI workflow, not just a few steps as current DI systems often do.
4. Tools are being built on top of a data science and big data eco-system. Today the two most popular such eco-systems build on R and Python. We currently target the Python data science and big data eco-system.
1. They guide the user through the end-to-end DI workflow, step by step.
2. For each step, they provide automated or semi-automated tools to address the "pain points" of the step.
3. Tools seek to cover the entire DI workflow, not just a few steps as current DI systems often do.
4. Tools are being built on top of a data science and big data eco-system. Today the two most popular such eco-systems build on R and Python. We currently target the Python data science and big data eco-system.
Paris Koutris
Pricing Relational Data: Data is sold and bought everywhere in the web. This project
aims to study the theoretical principles of pricing data, and design and implement a
practical and flexible framework that will support pricing relational queries and other
tasks (graph analytics, learning ML models).
Models for Large-scale Parallelism: There exists a wide spectrum of large-scale data management systems that exploit parallelism to speed up computation. The overarching goal of this project is to develop theoretical models that help us understand the limits and tradeoffs of parallelizing data processing tasks. Some of the key questions we study: How much synchronization is necessary to perform a certain task? What is the minimum amount of communication needed to compute queries over relational data? How does data skew effect the design of load-balanced algorithms?
Models for Large-scale Parallelism: There exists a wide spectrum of large-scale data management systems that exploit parallelism to speed up computation. The overarching goal of this project is to develop theoretical models that help us understand the limits and tradeoffs of parallelizing data processing tasks. Some of the key questions we study: How much synchronization is necessary to perform a certain task? What is the minimum amount of communication needed to compute queries over relational data? How does data skew effect the design of load-balanced algorithms?
Tianyu Li
New data processing paradigms like serverless analytics and agentic exploration challenge
performance assumptions and abstraction boundaries of legacy systems that today's infrastructure
is built on. My research aims to identify the resulting inefficiencies and build novel
abstractions to bridge the gap.
Resilient Composition for Serverless Compute: Modern cloud applications are composed from distributed components, and the "glue" between them — messaging, retries, recovery — is where fault tolerance breaks and developer time goes. We designed the Composable Resilient Step (CReSt) abstraction, in which applications are structured as atomic steps that consume messages, update state, and send messages exactly once, and built DARQ (SIGMOD '23), a runtime that executes steps efficiently using database techniques such as group commit and optimistic concurrency control. DARQ applications outperform industry-standard solutions like Kafka Streams and Azure Durable Functions by up to 2× in throughput and asymptotically in latency, without user-written glue code. Companion work includes DPR (SIGMOD '21), a protocol for distributed speculative execution that hides fault-tolerance latency, and EPVS (Best of DaMoN '22), a concurrency framework for almost-latch-free workloads that ships in Microsoft's FASTER. Ideas from this line have influenced next-generation system designs at Microsoft and Temporal.
Transactions for AI Agents: LLM agents are crossing from analyzing enterprise data to modifying it: refunding orders, rerouting shipments, settling insurance claims. These workflows require strong isolation and recovery guarantees provided by transactions, but their workload patterns are highly suboptimal for existing concurrency control schemes: agents interleave millisecond database actions with seconds of model reasoning, and frequently rely on agentic speculation to play out different scenarios and committing all but one at the end. This project aims to redesign the OLTP stack to accommodate long-running and speculation-heavy agentic workloads.
The Semantic Data Warehouse: Enterprises increasingly ask questions of their data in natural language, and each answer performs expensive semantic work — extracting fields from documents, resolving entities, classifying records — that today's semantic processing systems throw away the moment the answer is returned. We are building a semantic data warehouse that keeps this work: each question's extractions persist as incrementally accumulated structure, so the warehouse organizes itself around the questions people actually ask. Recurring questions compile progressively down to structured queries over semantically extracted state, novel questions pay for inference exactly once, and no one designs a schema up front — the workload is the schema. The research agenda spans classical and new territory: when an extraction answers a differently-phrased question, how accumulated structure is maintained as underlying data changes, and how a warehouse reasons about structure whose correctness is probabilistic rather than exact.
Resilient Composition for Serverless Compute: Modern cloud applications are composed from distributed components, and the "glue" between them — messaging, retries, recovery — is where fault tolerance breaks and developer time goes. We designed the Composable Resilient Step (CReSt) abstraction, in which applications are structured as atomic steps that consume messages, update state, and send messages exactly once, and built DARQ (SIGMOD '23), a runtime that executes steps efficiently using database techniques such as group commit and optimistic concurrency control. DARQ applications outperform industry-standard solutions like Kafka Streams and Azure Durable Functions by up to 2× in throughput and asymptotically in latency, without user-written glue code. Companion work includes DPR (SIGMOD '21), a protocol for distributed speculative execution that hides fault-tolerance latency, and EPVS (Best of DaMoN '22), a concurrency framework for almost-latch-free workloads that ships in Microsoft's FASTER. Ideas from this line have influenced next-generation system designs at Microsoft and Temporal.
Transactions for AI Agents: LLM agents are crossing from analyzing enterprise data to modifying it: refunding orders, rerouting shipments, settling insurance claims. These workflows require strong isolation and recovery guarantees provided by transactions, but their workload patterns are highly suboptimal for existing concurrency control schemes: agents interleave millisecond database actions with seconds of model reasoning, and frequently rely on agentic speculation to play out different scenarios and committing all but one at the end. This project aims to redesign the OLTP stack to accommodate long-running and speculation-heavy agentic workloads.
The Semantic Data Warehouse: Enterprises increasingly ask questions of their data in natural language, and each answer performs expensive semantic work — extracting fields from documents, resolving entities, classifying records — that today's semantic processing systems throw away the moment the answer is returned. We are building a semantic data warehouse that keeps this work: each question's extractions persist as incrementally accumulated structure, so the warehouse organizes itself around the questions people actually ask. Recurring questions compile progressively down to structured queries over semantically extracted state, novel questions pay for inference exactly once, and no one designs a schema up front — the workload is the schema. The research agenda spans classical and new territory: when an extraction answers a differently-phrased question, how accumulated structure is maintained as underlying data changes, and how a warehouse reasons about structure whose correctness is probabilistic rather than exact.
Xiangyao Yu
1000-Core Database: Computer architectures are moving towards many-core machines with
dozens or even hundreds of cores on a single chip, which the current database management
systems (DBMSs) are not designed for. We performed an evaluation of concurrency control for
on-line transaction processing (OLTP) workloads on many-core chips. Our Analysis (VLDB'14) shows that all algorithms
fail to scale to this level of parallelism. For each algorithm, we identified artificial and
fundamental bottlenecks. We conclude that rather than pursuing incremental solutions,
many-core chips may require a completely redesigned DBMS architecture that is built from
ground up and is tightly coupled with the hardware. Our DBMS is open source on github
(DBx1000).
IMP: Indirect Memory Prefetcher: Important applications like machine learning, graph analytics, and sparse linear algebra are dominated by irregular memory accesses which have little temporal or spatial locality and are difficult to prefetcher using traditional techniques. A majority of these irregular accesses come from indirect patterns of the form A[B[i]]. We propose an efficient hardware indirect memory prefetcher (IMP) to hide memory latency of this access pattern. We also propose a partial cacheline accessing mechanism to reduce the network and DRAM bandwidth pressure from the lack of spatial locality. Evaluated on seven applications, IMP showed 56% speedup on average (up to 2.3x) compared to a baseline streaming prefetchers on a 64 core system.
Near Cloud Storage Computing: Modern cloud platforms disaggregate computation and storage into separate services. In this project, we explored the idea of using limited computation inside the simple storage service (S3) offered by AWS to accelerate data analytics. We use the existing S3 Select feature to accelerate not only simple database operators like select and project, but also complex operators like join, group-by, and top-K. We propose optimization techniques for each individual operator and demonstrated more than 6x performance improvement over a set of representative queries.
IMP: Indirect Memory Prefetcher: Important applications like machine learning, graph analytics, and sparse linear algebra are dominated by irregular memory accesses which have little temporal or spatial locality and are difficult to prefetcher using traditional techniques. A majority of these irregular accesses come from indirect patterns of the form A[B[i]]. We propose an efficient hardware indirect memory prefetcher (IMP) to hide memory latency of this access pattern. We also propose a partial cacheline accessing mechanism to reduce the network and DRAM bandwidth pressure from the lack of spatial locality. Evaluated on seven applications, IMP showed 56% speedup on average (up to 2.3x) compared to a baseline streaming prefetchers on a 64 core system.
Near Cloud Storage Computing: Modern cloud platforms disaggregate computation and storage into separate services. In this project, we explored the idea of using limited computation inside the simple storage service (S3) offered by AWS to accelerate data analytics. We use the existing S3 Select feature to accelerate not only simple database operators like select and project, but also complex operators like join, group-by, and top-K. We propose optimization techniques for each individual operator and demonstrated more than 6x performance improvement over a set of representative queries.