arXiv subjects

Florin Rusu

Publications and source records attributed to Florin Rusu.

At least 19 recordsLinked to original sources

arXiv2026

Filtered Approximate Nearest Neighbor Search in Vector Databases: System Design and Performance Analysis

Retrieval-Augmented Generation (RAG) applications increasingly rely on Filtered Approximate Nearest Neighbor Search (FANNS) to combine semantic retrieval with metadata constraints. While algorithmic innovations for FANNS have been proposed, there remains a lack of understanding regarding how generic filtering strategies perform within Vector Databases. In this work, we systematize the taxonomy of filtering strategies and evaluate their integration into FAISS, Milvus, and pgvector. To provide a robust benchmarking framework, we introduce a new relational dataset, \textit{MoReVec}, consisting of two tables, featuring 768-dimensional text embeddings and a rich schema of metadata attributes. We further propose the \textit{Global-Local Selectivity (GLS)} correlation metric to quantify the relationship between filters and query vectors. Our experiments reveal that algorithmic adaptations within the engine often override raw index performance. Specifically, we find that: (1) \textit{Milvus} achieves superior recall stability through hybrid approximate/exact execution; (2) \textit{pgvector}'s cost-based query optimizer frequently selects suboptimal execution plans, favoring approximate index scans even when exact sequential scans would yield perfect recall at comparable latency; and (3) partition-based indexes (IVFFlat) outperform graph-based indexes (HNSW) for low-selectivity queries. To facilitate this analysis, we extend the widely-used \textit{ANN-Benchmarks} to support filtered vector search and make it available online. Finally, we synthesize our findings into a set of practical guidelines for selecting index types and configuring query optimizers for hybrid search workloads.

cs.DB↗

arXiv2025

Sketched Sum-Product Networks for Joins

Sketches have shown high accuracy in multi-way join cardinality estimation, a critical problem in cost-based query optimization. Accurately estimating the cardinality of a join operation -- analogous to its computational cost -- allows the optimization of query execution costs in relational database systems. However, although sketches have shown high efficacy in query optimization, they are typically constructed specifically for predefined selections in queries that are assumed to be given a priori, hindering their applicability to new queries. As a more general solution, we propose for Sum-Product Networks to dynamically approximate sketches on-the-fly. Sum-Product Networks can decompose and model multivariate distributions, such as relations, as linear combinations of multiple univariate distributions. By representing these univariate distributions as sketches, Sum-Product Networks can combine them element-wise to efficiently approximate the sketch of any query selection. These approximate sketches can then be applied to join cardinality estimation. In particular, we implement the Fast-AGMS and Bound Sketch methods, which have successfully been used in prior work, despite their costly construction. By accurately approximating them instead, our work provides a practical alternative to apply these sketches to query optimization.

cs.DB↗

arXiv2025

PLANSIEVE: Real-time Suboptimal Query Plan Detection Through Incremental Refinements

Cardinality estimation remains a fundamental challenge in query optimization, often resulting in sub-optimal execution plans and degraded performance. While errors in cardinality estimation are inevitable, existing methods for identifying sub-optimal plans -- such as metrics like Q-error, P-error, or L1-error -- are limited to post-execution analysis, requiring complete knowledge of true cardinalities and failing to prevent the execution of sub-optimal plans in real-time. This paper introduces PLANSIEVE, a novel framework that identifies sub-optimal plans during query optimization. PLANSIEVE operates by analyzing the relative order of sub-plans generated by the optimizer based on estimated and true cardinalities. It begins with surrogate cardinalities from any third-party estimator and incrementally refines these surrogates as the system processes more queries. Experimental results on the augmented JOB-LIGHT-SCALE and STATS-CEB-SCALE workloads demonstrate that PLANSIEVE achieves an accuracy of up to 88.7\% in predicting sub-optimal plans.

cs.DB↗

arXiv2024

Spanning Tree-based Query Plan Enumeration

In this work, we define the problem of finding an optimal query plan as finding spanning trees with low costs. This approach empowers the utilization of a series of spanning tree algorithms, thereby enabling systematic exploration of the plan search space over a join graph. Capitalizing on the polynomial time complexity of spanning tree algorithms, we present the Ensemble Spanning Tree Enumeration (ESTE) strategy. ESTE employs two conventional spanning tree algorithms, Prim's and Kruskal's, together to enhance the robustness of the query optimizer. In ESTE, multiple query plans are enumerated exploring different areas of the search space. This positions ESTE as an intermediate strategy between exhaustive and heuristic enumeration strategies. We show that ESTE is more robust in identifying efficient query plans for large queries. In the case of data modifications and workload demand increase, we believe that our approach can be a cheaper alternative to maintain optimizer robustness by integrating additional spanning tree algorithms rather than completely changing the optimizer to another plan enumeration algorithm. The experimental evaluation shows that ESTE achieves better consistency in plan quality and optimization time than existing solutions while identifying similarly optimal plans.

cs.DB↗

arXiv2023

Analyzing Query Optimizer Performance in the Presence and Absence of Cardinality Estimates

Most query optimizers rely on cardinality estimates to determine optimal execution plans. While traditional databases such as PostgreSQL, Oracle, and Db2 utilize many types of synopses -- including histograms, samples, and sketches -- recent main-memory databases like DuckDB and Heavy.AI often operate with minimal or no estimates, yet their performance does not necessarily suffer. To the best of our knowledge, no analytical comparison has been conducted between optimizers with and without cardinality estimates to understand their performance characteristics in different settings, such as indexed, non-indexed, and multi-threaded. In this paper, we present a comparative analysis between optimizers that use cardinality estimates and those that do not. We use the Join Order Benchmark (JOB) for our evaluation and true cardinalities as the baseline. Our investigation reveals that cardinality estimates have marginal impact in non-indexed settings. Meanwhile, when indexes are available, inaccurate estimates may lead to sub-optimal physical operators -- even with an optimal join order. Furthermore, the impact of cardinality estimates is less significant in highly-parallel main-memory databases.

cs.DB↗

arXiv2021

Simpli-Squared: A Very Simple Yet Unexpectedly Powerful Join Ordering Algorithm Without Cardinality Estimates

The Join Order Benchmark (JOB) has become the de facto standard to assess the performance of relational database query optimizers due to its complexity and completeness. In order to compute the optimal execution plan -- join order -- existing solutions employ extensive data synopses and correlations -- functional dependencies -- between table attributes. These structures incur significant overhead to design, build, and maintain. In this paper, we present \textit{Simpli}city \textit{Simpli}fied (\textit{Simpli-Squared}), a very simple join ordering algorithm that achieves unexpectedly good results. Simpli-Squared computes the join order without using any statistics or cardinality estimates. It takes as input only the referential integrity constraints declared at schema definition and the number of tuples (size) in the base tables. The join order of a given query is computed by splitting the join graph along the many-to-many joins and sorting the tables based on their size. The tables involved in one-to-many joins are greedily included based on size and the query join graph. The resulting plan can be efficiently generated by a lightweight query rewriting procedure. Experiments on the JOB benchmark in PostgreSQL show that Simpli-Squared achieves runtimes having an increase of only up to 16\% -- and sometimes even a reduction -- compared to four state-of-the-art solutions that are considerably more intricate. Based on these results, we question whether JOB adequately tests query optimizers or if accurate cardinality estimation is such a fundamental requirement for performing well on the JOB benchmark.

cs.DB↗

arXiv2021

Adaptive Elastic Training for Sparse Deep Learning on Heterogeneous Multi-GPU Servers

Motivated by extreme multi-label classification applications, we consider training deep learning models over sparse data in multi-GPU servers. The variance in the number of non-zero features across training batches and the intrinsic GPU heterogeneity combine to limit accuracy and increase the time to convergence. We address these challenges with Adaptive SGD, an adaptive elastic model averaging stochastic gradient descent algorithm for heterogeneous multi-GPUs that is characterized by dynamic scheduling, adaptive batch size scaling, and normalized model merging. Instead of statically partitioning batches to GPUs, batches are routed based on the relative processing speed. Batch size scaling assigns larger batches to the faster GPUs and smaller batches to the slower ones, with the goal to arrive at a steady state in which all the GPUs perform the same number of model updates. Normalized model merging computes optimal weights for every GPU based on the assigned batches such that the combined model achieves better accuracy. We show experimentally that Adaptive SGD outperforms four state-of-the-art solutions in time-to-accuracy and is scalable with the number of GPUs.

cs.DC↗

arXiv2021

Online Sketch-based Query Optimization

Cost-based query optimization remains a critical task in relational databases even after decades of research and industrial development. Query optimizers rely on a large range of statistical synopses -- including attribute-level histograms and table-level samples -- for accurate cardinality estimation. As the complexity of selection predicates and the number of join predicates increase, two problems arise. First, statistics cannot be incrementally composed to effectively estimate the cost of the sub-plans generated in plan enumeration. Second, small errors are propagated exponentially through join operators, which can lead to severely sub-optimal plans. In this paper, we introduce COMPASS, a novel query optimization paradigm for in-memory databases based on a single type of statistics -- Fast-AGMS sketches. In COMPASS, query optimization and execution are intertwined. Selection predicates and sketch updates are pushed-down and evaluated online during query optimization. This allows Fast-AGMS sketches to be computed only over the relevant tuples -- which enhances cardinality estimation accuracy. Plan enumeration is performed over the query join graph by incrementally composing attribute-level sketches -- not by building a separate sketch for every sub-plan. We prototype COMPASS in MapD -- an open-source parallel database -- and perform extensive experiments over the complete JOB benchmark. The results prove that COMPASS generates better execution plans -- both in terms of cardinality and runtime -- compared to four other database systems. Overall, COMPASS achieves a speedup ranging from 1.35X to 11.28X in cumulative query execution time over the considered competitors.

cs.DB↗

arXiv2020

DJEnsemble: On the Selection of a Disjoint Ensemble of Deep Learning Black-Box Spatio-Temporal Models

In this paper, we present a cost-based approach for the automatic selection and allocation of a disjoint ensemble of black-box predictors to answer predictive spatio-temporal queries. Our approach is divided into two parts -- offline and online. During the offline part, we preprocess the predictive domain data -- transforming it into a regular grid -- and the black-box models -- computing their spatio-temporal learning function. In the online part, we compute a DJEnsemble plan which minimizes a multivariate cost function based on estimates for the prediction error and the execution cost -- producing a model spatial allocation matrix -- and run the optimal ensemble plan. We conduct a set of extensive experiments that evaluate the DJEnsemble approach and highlight its efficiency. We show that our cost model produces plans with performance close to the actual best plan. When compared against the traditional ensemble approach, DJEnsemble achieves up to $4X$ improvement in execution time and almost $9X$ improvement in prediction accuracy. To the best of our knowledge, this is the first work to solve the problem of optimizing the allocation of black-box models to answer predictive spatio-temporal queries.

cs.AI↗

arXiv2020

Heterogeneous CPU+GPU Stochastic Gradient Descent Algorithms

The widely-adopted practice is to train deep learning models with specialized hardware accelerators, e.g., GPUs or TPUs, due to their superior performance on linear algebra operations. However, this strategy does not employ effectively the extensive CPU and memory resources -- which are used only for preprocessing, data transfer, and scheduling -- available by default on the accelerated servers. In this paper, we study training algorithms for deep learning on heterogeneous CPU+GPU architectures. Our two-fold objective -- maximize convergence rate and resource utilization simultaneously -- makes the problem challenging. In order to allow for a principled exploration of the design space, we first introduce a generic deep learning framework that exploits the difference in computational power and memory hierarchy between CPU and GPU through asynchronous message passing. Based on insights gained through experimentation with the framework, we design two heterogeneous asynchronous stochastic gradient descent (SGD) algorithms. The first algorithm -- CPU+GPU Hogbatch -- combines small batches on CPU with large batches on GPU in order to maximize the utilization of both resources. However, this generates an unbalanced model update distribution which hinders the statistical convergence. The second algorithm -- Adaptive Hogbatch -- assigns batches with continuously evolving size based on the relative speed of CPU and GPU. This balances the model updates ratio at the expense of a customizable decrease in utilization. We show that the implementation of these algorithms in the proposed CPU+GPU framework achieves both faster convergence and higher resource utilization than TensorFlow on several real datasets and on two computing architectures -- an on-premises server and a cloud instance.

cs.DC↗

arXiv2019

Progressive Data Science: Potential and Challenges

Data science requires time-consuming iterative manual activities. In particular, activities such as data selection, preprocessing, transformation, and mining, highly depend on iterative trial-and-error processes that could be sped-up significantly by providing quick feedback on the impact of changes. The idea of progressive data science is to compute the results of changes in a progressive manner, returning a first approximation of results quickly and allow iterative refinements until converging to a final result. Enabling the user to interact with the intermediate results allows an early detection of erroneous or suboptimal choices, the guided definition of modifications to the pipeline and their quick assessment. In this paper, we discuss the progressiveness challenges arising in different steps of the data science pipeline. We describe how changes in each step of the pipeline impact the subsequent steps and outline why progressive data science will help to make the process more effective. Computing progressive approximations of outcomes resulting from changes creates numerous research challenges, especially if the changes are made in the early steps of the pipeline. We discuss these challenges and outline first steps towards progressiveness, which, we argue, will ultimately help to significantly speed-up the overall data science process.

cs.HC↗

arXiv2019

In-Depth Benchmarking of Graph Database Systems with the Linked Data Benchmark Council (LDBC) Social Network Benchmark (SNB)

In this study, we present the first results of a complete implementation of the LDBC SNB benchmark -- interactive short, interactive complex, and business intelligence -- in two native graph database systems---Neo4j and TigerGraph. In addition to thoroughly evaluating the performance of all of the 46 queries in the benchmark on four scale factors -- SF-1, SF-10, SF-100, and SF-1000 -- and three computing architectures -- on premise and in the cloud -- we also measure the bulk loading time and storage size. Our results show that TigerGraph is consistently outperforming Neo4j on the majority of the queries---by two or more orders of magnitude (100X factor) on certain interactive complex and business intelligence queries. The gap increases with the size of the data since only TigerGraph is able to scale to SF-1000---Neo4j finishes only 12 of the 25 business intelligence queries in reasonable time. Nonetheless, Neo4j is generally faster at bulk loading graph data up to SF-100. A key to our study is the active involvement of the vendors in the tuning of their platforms. In order to encourage reproducibility, we make all the code, scripts, and configuration parameters publicly available online.

cs.DB↗

arXiv2019

Exact Selectivity Computation for Modern In-Memory Database Query Optimization

Selectivity estimation remains a critical task in query optimization even after decades of research and industrial development. Optimizers rely on accurate selectivities when generating execution plans. They maintain a large range of statistical synopses for efficiently estimating selectivities. Nonetheless, small errors -- propagated exponentially -- can lead to severely sub-optimal plans---especially, for complex predicates. Database systems for modern computing architectures rely on extensive in-memory processing supported by massive multithread parallelism and vectorized instructions. However, they maintain the same synopses approach to query optimization as traditional disk-based databases. We introduce a novel query optimization paradigm for in-memory and GPU-accelerated databases based on \textit{exact selectivity computation (ESC)}. The central idea in ESC is to compute selectivities exactly through queries during query optimization. In order to make the process efficient, we propose several optimizations targeting the selection and materialization of tables and predicates to which ESC is applied. We implement ESC in the MapD open-source database system. Experiments on the TPC-H and SSB benchmarks show that ESC records constant and less than 30 milliseconds overhead when running on GPU and generates improved query execution plans that are as much as 32X faster.

cs.DB↗

arXiv2018

Distributed Caching for Complex Querying of Raw Arrays

As applications continue to generate multi-dimensional data at exponentially increasing rates, fast analytics to extract meaningful results is becoming extremely important. The database community has developed array databases that alleviate this problem through a series of techniques. In-situ mechanisms provide direct access to raw data in the original format---without loading and partitioning. Parallel processing scales to the largest datasets. In-memory caching reduces latency when the same data are accessed across a workload of queries. However, we are not aware of any work on distributed caching of multi-dimensional raw arrays. In this paper, we introduce a distributed framework for cost-based caching of multi-dimensional arrays in native format. Given a set of files that contain portions of an array and an online query workload, the framework computes an effective caching plan in two stages. First, the plan identifies the cells to be cached locally from each of the input files by continuously refining an evolving R-tree index. In the second stage, an optimal assignment of cells to nodes that collocates dependent cells in order to minimize the overall data transfer is determined. We design cache eviction and placement heuristic algorithms that consider the historical query workload. A thorough experimental evaluation over two real datasets in three file formats confirms the superiority -- by as much as two orders of magnitude -- of the proposed framework over existing techniques in terms of cache overhead and workload execution time.

cs.DB↗

arXiv2018

Stochastic Gradient Descent on Highly-Parallel Architectures

There is an increased interest in building data analytics frameworks with advanced algebraic capabilities both in industry and academia. Many of these frameworks, e.g., TensorFlow and BIDMach, implement their compute-intensive primitives in two flavors---as multi-thread routines for multi-core CPUs and as highly-parallel kernels executed on GPU. Stochastic gradient descent (SGD) is the most popular optimization method for model training implemented extensively on modern data analytics platforms. While the data-intensive properties of SGD are well-known, there is an intense debate on which of the many SGD variants is better in practice. In this paper, we perform a comprehensive study of parallel SGD for training generalized linear models. We consider the impact of three factors -- computing architecture (multi-core CPU or GPU), synchronous or asynchronous model updates, and data sparsity -- on three measures---hardware efficiency, statistical efficiency, and time to convergence. In the process, we design an optimized asynchronous SGD algorithm for GPU that leverages warp shuffling and cache coalescing for data and model access. We draw several interesting findings from our extensive experiments with logistic regression (LR) and support vector machines (SVM) on five real datasets. For synchronous SGD, GPU always outperforms parallel CPU---they both outperform a sequential CPU solution by more than 400X. For asynchronous SGD, parallel CPU is the safest choice while GPU with data replication is better in certain situations. The choice between synchronous GPU and asynchronous CPU depends on the task and the characteristics of the data. As a reference, our best implementation outperforms TensorFlow and BIDMach consistently. We hope that our insights provide a useful guide for applying parallel SGD to generalized linear models.

cs.DB↗

arXiv2017

OLA-RAW: Scalable Exploration over Raw Data

In-situ processing has been proposed as a novel data exploration solution in many domains generating massive amounts of raw data, e.g., astronomy, since it provides immediate SQL querying over raw files. The performance of in-situ processing across a query workload is, however, limited by the speed of full scan, tokenizing, and parsing of the entire data. Online aggregation (OLA) has been introduced as an efficient method for data exploration that identifies uninteresting patterns faster by continuously estimating the result of a computation during the actual processing---the computation can be stopped as early as the estimate is accurate enough to be deemed uninteresting. However, existing OLA solutions have a high upfront cost of randomly shuffling and/or sampling the data. In this paper, we present OLA-RAW, a bi-level sampling scheme for parallel online aggregation over raw data. Sampling in OLA-RAW is query-driven and performed exclusively in-situ during the runtime query execution, without data reorganization. This is realized by a novel resource-aware bi-level sampling algorithm that processes data in random chunks concurrently and determines adaptively the number of sampled tuples inside a chunk. In order to avoid the cost of repetitive conversion from raw data, OLA-RAW builds and maintains a memory-resident bi-level sample synopsis incrementally. We implement OLA-RAW inside a modern in-situ data processing system and evaluate its performance across several real and synthetic datasets and file formats. Our results show that OLA-RAW chooses the sampling plan that minimizes the execution time and guarantees the required accuracy for each query in a given workload. The end result is a focused data exploration process that avoids unnecessary work and discards uninteresting data.

cs.DB↗

arXiv2017

Dot-Product Join: An Array-Relation Join Operator for Big Model Analytics

Big Model analytics tackles the training of massive models that go beyond the available memory of a single computing device, e.g., CPU or GPU. It generalizes Big Data analytics which is targeted at how to train memory-resident models over out-of-memory training data. In this paper, we propose an in-database solution for Big Model analytics. We identify dot-product as the primary operation for training generalized linear models and introduce the first array-relation dot-product join database operator between a set of sparse arrays and a dense relation. This is a constrained formulation of the extensively studied sparse matrix vector multiplication (SpMV) kernel. The paramount challenge in designing the dot-product join operator is how to optimally schedule access to the dense relation based on the non-contiguous entries in the sparse arrays. We prove that this problem is NP-hard and propose a practical solution characterized by two technical contributions---dynamic batch processing and array reordering. We devise three heuristics -- LSH, Radix, and K-center -- for array reordering and analyze them thoroughly. We execute extensive experiments over synthetic and real data that confirm the minimal overhead the operator incurs when sufficient memory is available and the graceful degradation it suffers as memory becomes scarce. Moreover, dot-product join achieves an order of magnitude reduction in execution time over alternative in-database solutions.

cs.DB↗

arXiv2015

Workload-Driven Vertical Partitioning for Effective Query Processing over Raw Data

Traditional databases are not equipped with the adequate functionality to handle the volume and variety of "Big Data". Strict schema definition and data loading are prerequisites even for the most primitive query session. Raw data processing has been proposed as a schema-on-demand alternative that provides instant access to the data. When loading is an option, it is driven exclusively by the current-running query, resulting in sub-optimal performance across a query workload. In this paper, we investigate the problem of workload-driven raw data processing with partial loading. We model loading as fully-replicated binary vertical partitioning. We provide a linear mixed integer programming optimization formulation that we prove to be NP-hard. We design a two-stage heuristic that comes within close range of the optimal solution in a fraction of the time. We extend the optimization formulation and the heuristic to pipelined raw data processing, scenario in which data access and extraction are executed concurrently. We provide three case-studies over real data formats that confirm the accuracy of the model when implemented in a state-of-the-art pipelined operator for raw data processing.

cs.DB↗