Large language models (LLMs) have shown state-of-the-art results in translating natural language questions into SQL queries (Text-to-SQL), a long-standing challenge within the database community. However, security concerns remain largely unexplored, particularly the threat of backdoor attacks, which can introduce malicious behaviors into models through fine-tuning with poisoned datasets. In this work, we systematically investigate the vulnerabilities of LLM-based Text-to-SQL models and present ToxicSQL, a novel backdoor attack framework. Our approach leverages stealthy command-like and character-level triggers to make backdoors difficult to detect and remove, ensuring that malicious behaviors remain covert while maintaining high model accuracy on benign inputs. Furthermore, we propose leveraging SQL injection payloads as backdoor targets, enabling the generation of malicious yet executable SQL queries, which pose severe security and privacy risks in language model-based SQL development. We demonstrate that injecting only 0.44% of poisoned data can result in an attack success rate of 79.41%, posing a significant risk to database security. Additionally, we propose detection and mitigation strategies to enhance model reliability. Our findings highlight the urgent need for security-aware Text-to-SQL development, emphasizing the importance of robust defenses against backdoor threats.

Transformer-based large language models (LLMs) have demonstrated outstanding performance across diverse domains, particularly in the emerging pretrain-then-finetune paradigm. LoRA, a parameter-efficient fine-tuning method, is commonly used to adapt a base LLM to multiple downstream tasks. Further, LLM platforms enable developers to fine-tune multiple models and develop various domain-specific applications simultaneously. However, existing model parallelism schemes suffer from high communication overhead and inefficient GPU utilization. In this paper, we present mLoRA, a parallelism-efficient fine-tuning system designed for training multiple LoRA across GPUs and machines. mLoRA introduces a novel LoRA-aware pipeline parallelism scheme that efficiently pipelines LoRA adapters and their distinct fine-tuning stages across GPUs and machines, along with a new LoRA-efficient operator to enhance GPU utilization. Our extensive evaluation shows that mLoRA can significantly reduce average fine-tuning task completion time, e.g., by 30%, compared to state-of-the-art methods like FSDP. More importantly, mLoRA enables simultaneous fine-tuning of larger models, e.g., two Llama-2-13B models on four NVIDIA RTX A6000 48GB GPUs, which is not feasible for FSDP due to high memory requirements. Hence, mLoRA not only increases fine-tuning efficiency but also makes it more accessible on cost-effective GPUs.

Machine Learning (ML) has become ubiquitous, fueling data-driven applications across various organizations. Contrary to the traditional perception of ML in research, ML workflows can be complex, resource-intensive, and time-consuming. Expanding an ML workflow to encompass a wider range of data infrastructure and data types may lead to larger workloads and increased deployment costs. Currently, numerous workflow engines are available (with over ten being widely recognized). This variety poses a challenge for end-users in terms of mastering different engine APIs. While efforts have primarily focused on optimizing ML Operations (MLOps) for a specific workflow engine, current methods largely overlook workflow optimization across different engines. In this work, we design and implement Couler, a system designed for unified ML workflow optimization in the cloud. Our main insight lies in the ability to generate an ML workflow using natural language (NL) descriptions. We integrate Large Language Models (LLMs) into workflow generation, and provide a unified programming interface for various workflow engines. This approach alleviates the need to understand various workflow engines' APIs. Moreover, Couler enhances workflow computation efficiency by introducing automated caching at multiple stages, enabling large workflow auto-parallelization and automatic hyperparameters tuning. These enhancements minimize redundant computational costs and improve fault tolerance during deep learning workflow training. Couler is extensively deployed in real-world production scenarios at Ant Group, handling approximately 22k workflows daily, and has successfully improved the CPU/Memory utilization by more than 15% and the workflow completion rate by around 17%.

Modern database management systems (DBMS) expose hundreds of configurable knobs to control system behaviours. Determining the appropriate values for these knobs to improve DBMS performance is a long-standing problem in the database community. As there is an increasing number of knobs to tune and each knob could be in continuous or categorical values, manual tuning becomes impractical. Recently, automatic tuning systems using machine learning methods have shown great potentials. However, existing approaches still incur significant tuning costs or only yield sub-optimal performance. This is because they either ignore the extensive domain knowledge available (e.g., DBMS manuals and forum discussions) and only rely on the runtime feedback of benchmark evaluations to guide the optimization, or they utilize the domain knowledge in a limited way. Hence, we propose GPTuner, a manual-reading database tuning system that leverages domain knowledge extensively and automatically to optimize search space and enhance the runtime feedback-based optimization process. Firstly, we develop a Large Language Model (LLM)-based pipeline to collect and refine heterogeneous knowledge, and propose a prompt ensemble algorithm to unify a structured view of the refined knowledge. Secondly, using the structured knowledge, we (1) design a workload-aware and training-free knob selection strategy, (2) develop a search space optimization technique considering the value range of each knob, and (3) propose a Coarse-to-Fine Bayesian Optimization Framework to explore the optimized space. Finally, we evaluate GPTuner under different benchmarks (TPC-C and TPC-H), metrics (throughput and latency) as well as DBMS (PostgreSQL and MySQL). Compared to the state-of-the-art approaches, GPTuner identifies better configurations in 16x less time on average. Moreover, GPTuner achieves up to 30% performance improvement (higher throughput or lower latency) over the best-performing alternative.

Deep learning recommendation models (DLRM) rely on large embedding tables to manage categorical sparse features. Expanding such embedding tables can significantly enhance model performance, but at the cost of increased GPU/CPU/memory usage. Meanwhile, tech companies have built extensive cloud-based services to accelerate training DLRM models at scale. In this paper, we conduct a deep investigation of the DLRM training platforms at AntGroup and reveal two critical challenges: low resource utilization due to suboptimal configurations by users and the tendency to encounter abnormalities due to an unstable cloud environment. To overcome them, we introduce DLRover, an elastic training framework for DLRMs designed to increase resource utilization and handle the instability of a cloud environment. DLRover develops a resource-performance model by considering the unique characteristics of DLRMs and a three-stage heuristic strategy to automatically allocate and dynamically adjust resources for DLRM training jobs for higher resource utilization. Further, DLRover develops multiple mechanisms to ensure efficient and reliable execution of DLRM training jobs. Our extensive evaluation shows that DLRover reduces job completion times by 31%, increases the job completion rate by 6%, enhances CPU usage by 15%, and improves memory utilization by 20%, compared to state-of-the-art resource scheduling frameworks. DLRover has been widely deployed at AntGroup and processes thousands of DLRM training jobs on a daily basis. DLRover is open-sourced and has been adopted by 10+ companies.

In the cloud environment, different kinds of jobs (Flink, PyTorch, TensorFlow, AI-Serving) are running in the same cluster with different service-level agreements (SLA). To manage large amounts of jobs in a cloud environment efficiently, it is critical to build a system to optimize the job performance in consideration of multiple predefined objectives. For example, one kind of optimization target is improving the resource utilization of jobs, other kinds of objectives are to guarantee the system SLA (e.g., system throughput, response time, and so on). Currently, most of the existing frameworks are working on one aspect of optimization, and can not support different kinds of optimization targets via a uniform framework or system. In Antgroup, we have designed and implemented a general framework (named Cougar) to improve jobs and cluster performance to meet such requirements. Cougar provides the ability to support different optimization scenarios like the initial and runtime optimization for one job, and cross-job optimization for multiple jobs. Nowadays, Cougar has widely used in the production environment of Antgroup including 110,000 jobs and 800,000 Pods daily, and has successfully improved the CPU/Memory/GPU utilization by more than 20% and performance (i.e., throughput or completion time or latency) by around 10%. In the end, we also like to share our best practice on how to tune Flink and Deep Learning Job (GPU collocate) in the production environment.

In the era of big data, a data processing flow contains various types of tasks. It is nontrivial to discover the data flow/movement from its source to destination, such that monitoring different transformations and hops on its way in an enterprise environment. Therefore, data lineage or provenance is useful to learn how the data gets transformed along the way, how the representation and parameters change, and how the data splits or converges after each hop. However, existing systems offer limited support for such use cases in a distributed computing setup. To address this issue, we build Spark-Atlas-Connector (short as SAC), a new system to track data lineage in a distributed computation platform, e.g., Spark. SAC tracks different processes involved in the data flow and their dependencies, supporting different data storage (e.g., HBase, HDFS, and Hive) and data processing paradigms (e.g., SQL, ETL, machine learning, and streaming). SAC provides a visual representation of data lineage to track data from its origin to downstreams, and is deployed in a distributed production environment for demonstrating its efficiency and scalability.

Given a set of multidimensional data points, skyline queries retrieve those points that are not dominated by any other points in the set. Due to the ubiquitous use of skyline queries, such as in preference-based query answering and decision making, and the large amount of data that these queries have to deal with, enabling their scalable processing is of critical importance. However, there are several outstanding challenges that have not been well addressed. More specifically, in this paper, we are tackling the data straggler and data skew challenges introduced by distributed skyline query processing, as well as the ensuing high computation cost of merging skyline candidates. We thus introduce a new efficient three-phase approach for large scale processing of skyline queries. In the first preprocessing phase, the data is partitioned along the Z-order curve. We utilize a novel data partitioning approach that formulates data partitioning as an optimization problem to minimize the size of intermediate data. In the second phase, each compute node partitions the input data points into disjoint subsets, and then performs the skyline computation on each subset to produce skyline candidates in parallel. In the final phase, we build an index and employ an efficient algorithm to merge the generated skyline candidates. Extensive experiments demonstrate that the proposed skyline algorithm achieves more than one order of magnitude enhancement in performance compared to existing state-of-the-art approaches.

The use of large-scale machine learning~(ML) is becoming ubiquitous in various domains ranging from business intelligence to self-driving cars. Many companies are building ML pipelines in a unified data processing environment, and leveraging well-tuned numerical optimization packages for obtaining model parameters. However, most existing optimization tools are specifically designed for a single machine setup, and cannot handle vast volume of data. In this work, we build a distributed computing framework towards optimizing generalized linear models with billions of variables. We at first design a new distributed vector to represent data points from extremely large feature space. Then, we introduce an efficient and scalable approach to compute the second order derivatives of loss function, and optimizes model parameters with limited memory requirement. Experiments on real-world datasets demonstrate that our proposed techniques can scale up for ML models with billions of variables, and achieves better performance than state-of-the-art systems on a wide range of applications, e.g., ad CTR prediction and rideshare price bidding.

The widespread use of GPS-enabled smartphones along with the popularity of micro-blogging and social networking applications, e.g., Twitter and Facebook, has resulted in the generation of huge streams of geo-tagged textual data. Many applications require real-time processing of these streams. For example, location-based ad-targeting systems enable advertisers to register millions of ads to millions of users based on the users' location and textual profile. Existing streaming systems are either centralized or are not spatial-keyword aware, and hence these systems cannot efficiently support the processing of rapidly arriving spatial-keyword data streams. In this paper, we introduce a two-layered indexing scheme for the distributed processing of spatial-keyword data streams. We realize this indexing scheme in Tornado, a distributed spatial-keyword streaming system. The first layer, termed the routing layer, is used to fairly distribute the workload, and furthermore, co-locate the data objects and the corresponding queries at the same processing units. The routing layer uses the Augmented-Grid, a novel structure that is equipped with an efficient search algorithm for distributing the data objects and queries. The second layer, termed the evaluation layer, resides within each processing unit to reduce the processing overhead. The two-layered index adapts to changes in the workload by applying a cost formula that continuously represents the processing overhead at each processing unit. Extensive experimental evaluation using real Twitter data indicates that Tornado achieves high scalability and more than 2x improvement over the baseline approach in terms of the overall system throughput.

We introduce a simple data model to process non-relational data for relational operations, and SHC (Apache Spark - Apache HBase Connector), an implementation of this model in the cluster computing framework, Spark. SHC leverages optimization techniques of relational data processing over the distributed and column-oriented key-value store (i.e., HBase). Compared to existing systems, SHC makes two major contributions. At first, SHC offers a much tighter integration between optimizations of relational data processing and non-relational data store, through a plug-in implementation that integrates with Spark SQL, a distributed in-memory computing engine for relational data. The design makes the system maintenance relatively easy, and enables users to perform complex data analytics on top of key-value store. Second, SHC leverages the Spark SQL Catalyst engine for high performance query optimizations and processing, e.g., data partitions pruning, columns pruning, predicates pushdown and data locality. SHC has been deployed and used in multiple production environments with hundreds of nodes, and provides OLAP query processing on petabytes of data efficiently.

Given a set of multidimensional data points, skyline queries retrieve those points that are not dominated by any other points in the set. Due to the ubiquitous se of skyline queries, such as in preference-based query answering and decision making, and the large amount of data that these queries have to deal with, enabling their scalable processing is of critical importance. However, there are several outstanding challenges that have not been well addressed. More specifically, in this paper, we are tackling the data straggler and data skew challenges introduced by distributed skyline query processing, as well as the ensuing high computation cost of merging skyline candidates. We thus introduce a new efficient three-phase approach for large scale processing of skyline queries. In the first preprocessing phase, the data is partitioned along the Z-order curve. We utilize a novel data partitioning approach that formulates data partitioning as an optimization problem to minimize the size of intermediate data. In the second phase, each compute node partitions the input data points into disjoint subsets, and then performs the skyline computation on each subset to produce skyline candidates in parallel. In the final phase, we build an index and employ an efficient algorithm to merge the generated skyline candidates. Extensive experiments demonstrate that the proposed skyline algorithm achieves more than one order of magnitude enhancement in performance compared to existing state-of-the-art approaches.

The use of large-scale machine learning and data mining methods is becoming ubiquitous in many application domains ranging from business intelligence and bioinformatics to self-driving cars. These methods heavily rely on matrix computations, and it is hence critical to make these computations scalable and efficient. These matrix computations are often complex and involve multiple steps that need to be optimized and sequenced properly for efficient execution. This paper presents new efficient and scalable matrix processing and optimization techniques for in-memory distributed clusters. The proposed techniques estimate the sparsity of intermediate matrix-computation results and optimize communication costs. An evaluation plan generator for complex matrix computations is introduced as well as a distributed plan optimizer that exploits dynamic cost-based analysis and rule-based heuristics to optimize the cost of matrix computations in an in-memory distributed environment. The result of a matrix operation will often serve as an input to another matrix operation, thus defining the matrix data dependencies within a matrix program. The matrix query plan generator produces query execution plans that minimize memory usage and communication overhead by partitioning the matrix based on the data dependencies in the execution plan. We implemented the proposed matrix processing and optimization techniques in Spark, a distributed in-memory computing platform. Experiments on both real and synthetic data demonstrate that our proposed techniques achieve up to an order-of-magnitude performance improvement over state-of the-art distributed matrix computation systems on a wide range of applications.

The popularity of GPS-enabled cellular devices introduced numerous applications, e.g., social networks, micro-blogs, and crowd-powered reviews. These applications produce large amounts of geo-tagged textual data that need to be processed and queried. Nowadays, many complex spatio-textual operators and their matching complex indexing structures are being proposed in the literature to process this spatio-textual data. For example, there exist several complex variations of the spatio-textual group queries that retrieve groups of objects that collectively satisfy certain spatial and textual criteria. However, having complex operators is against the spirit of SQL and relational algebra. In contrast to these complex spatio-textual operators, in relational algebra, simple relational operators are offered, e.g., relational selects, projects, order by, and group by, that are composable to form more complex queries. In this paper, we introduce Atlas, an SQL extension to express complex spatial-keyword group queries. Atlas follows the philosophy of SQL and relational algebra in that it uses simple declarative spatial and textual building-block operators and predicates to extend SQL. Not only that Atlas can represent spatio-textual group queries from the literature, but also it can compose other important queries, e.g., retrieve spatio-textual groups from subsets of object datasets where the selected subset satisfies user-defined relational predicates and the groups of close-by objects contain miss-spelled keywords. We demonstrate that Atlas is able to represent a wide range of spatial-keyword queries that existing indexes and algorithms would not be able to address. The building- block paradigm adopted by Atlas creates room for query optimization, where multiple query execution plans can be formed.

Advances in location-based services (LBS) demand high-throughput processing of both static and streaming data. Recently, many systems have been introduced to support distributed main-memory processing to maximize the query throughput. However, these systems are not optimized for spatial data processing. In this demonstration, we showcase Cruncher, a distributed main-memory spatial data warehouse and streaming system. Cruncher extends Spark with adaptive query processing techniques for spatial data. Cruncher uses dynamic batch processing to distribute the queries and the data streams over commodity hardware according to an adaptive partitioning scheme. The batching technique also groups and orders the overlapping spatial queries to enable inter-query optimization. Both the data streams and the offline data share the same partitioning strategy that allows for data co-locality optimization. Furthermore, Cruncher uses an adaptive caching strategy to maintain the frequently-used location data in main memory. Cruncher maintains operational statistics to optimize query processing, data partitioning, and caching at runtime. We demonstrate two LBS applications over Cruncher using real datasets from OpenStreetMap and two synthetic data streams. We demonstrate that Cruncher achieves order(s) of magnitude throughput improvement over Spark when processing spatial data.

The SQL group-by operator plays an important role in summarizing and aggregating large datasets in a data analytics stack. The Similarity SQL-based Group-By operator (SGB, for short) extends the semantics of the standard SQL Group-by by grouping data with similar but not necessarily equal values. While existing similarity-based grouping operators efficiently realize these approximate semantics, they primarily focus on one-dimensional attributes and treat multi-dimensional attributes independently. However, correlated attributes, such as in spatial data, are processed independently, and hence, groups in the multi-dimensional space are not detected properly. To address this problem, we introduce two new SGB operators for multi-dimensional data. The first operator is the clique (or distance-to-all) SGB, where all the tuples in a group are within some distance from each other. The second operator is the distance-to-any SGB, where a tuple belongs to a group if the tuple is within some distance from any other tuple in the group. Since a tuple may satisfy the membership criterion of multiple groups, we introduce three different semantics to deal with such a case: (i) eliminate the tuple, (ii) put the tuple in any one group, and (iii) create a new group for this tuple. We implement and test the new SGB operators and their algorithms inside PostgreSQL. The overhead introduced by these operators proves to be minimal and the execution times are comparable to those of the standard Group-by. The experimental study, based on TPC-H and a social check-in data, demonstrates that the proposed algorithms can achieve up to three orders of magnitude enhancement in performance over baseline methods developed to solve the same problem.

We present LocationSpark, a spatial data processing system built on top of Apache Spark, a widely used distributed data processing system. LocationSpark offers a rich set of spatial query operators, e.g., range search, k NN, spatio-textual operation, spatial-join, and k NN-join. To achieve high performance, LocationSpark employs various spatial indexes for in-memory data, and guarantees that immutable spatial indexes have low overhead with fault tolerance. In addition, we build two new layers over Spark, namely a query scheduler and a query executor. The query scheduler is responsible for mitigating skew in spatial queries, while the query executor selects the best plan based on the indexes and the nature of the spatial queries. Furthermore, to avoid unnecessary network communication overhead when processing overlapped spatial data, We embed an efficient spatial Bloom filter into LocationSpark's indexes. Finally, LocationSpark tracks frequently accessed spatial data, and dynamically flushes less frequently accessed data into disk. We evaluate our system on real workloads and demonstrate that it achieves an order of magnitude performance gain over a baseline framework.

The SQL group-by operator plays an important role in summarizing and aggregating large datasets in a data analytics stack. While the standard group-by operator, which is based on equality, is useful in several applications, allowing similarity aware grouping provides a more realistic view on real-world data that could lead to better insights. The Similarity SQL-based Group-By operator (SGB, for short) extends the semantics of the standard SQL Group-by by grouping data with similar but not necessarily equal values. While existing similarity-based grouping operators efficiently realize these approximate semantics, they primarily focus on one-dimensional attributes and treat multi-dimensional attributes independently. However, correlated attributes, such as in spatial data, are processed independently, and hence, groups in the multi-dimensional space are not detected properly. To address this problem, we introduce two new SGB operators for multi-dimensional data. The first operator is the clique (or distance-to-all) SGB, where all the tuples in a group are within some distance from each other. The second operator is the distance-to-any SGB, where a tuple belongs to a group if the tuple is within some distance from any other tuple in the group. Since a tuple may satisfy the membership criterion of multiple groups, we introduce three different semantics to deal with such a case: (i) eliminate the tuple, (ii) put the tuple in any one group, and (iii) create a new group for this tuple. We implement and test the new SGB operators and their algorithms inside PostgreSQL. The overhead introduced by these operators proves to be minimal and the execution times are comparable to those of the standard Group-by. The experimental study, based on TPC-H and a social check-in data, demonstrates that the proposed algorithms can achieve up to three orders of magnitude enhancement in performance over baseline methods developed to solve the same problem.

Similarity search is crucial to many applications. Of particular interest are two flavors of the Hamming distance range query, namely, the Hamming select and the Hamming join (Hamming-select and Hamming-join, respectively). Hamming distance is widely used in approximate near neighbor search for high dimensional data, such as images and document collections. For example, using predefined similarity hash functions, high-dimensional data is mapped into one-dimensional binary codes that are, then linearly scanned to perform Hamming-distance comparisons. These distance comparisons on the binary codes are usually costly and, often involves excessive redundancies. This paper introduces a new index, termed the HA-Index, that speeds up distance comparisons and eliminates redundancies when performing the two flavors of Hamming distance range queries. An efficient search algorithm based on the HA-index is presented. A distributed version of the HA-index is introduced and algorithms for realizing Hamming distance-select and Hamming distance-join operations on a MapReduce platform are prototyped. Extensive experiments using real datasets demonstrates that the HA-index and the corresponding search algorithms achieve up to two orders of magnitude speedup over existing state-of-the-art approaches, while saving more than ten times in memory space.