DiNoDB: Efficient Large-Scale Raw Data Analytics

Modern big data workflows, found in e.g., machine learning use cases, often involve iterations of cycles of batch analytics and interactive analytics on temporary data. Whereas batch analytics solutions for large volumes of raw data are well established (e.g., Hadoop, MapReduce), state-of-the-art interactive analytics solutions (e.g., distributed shared nothing RDBMSs) require data loading and/or transformation phase, which is inherently expensive for temporary data. In this paper, we propose a novel scalable distributed solution for in-situ data analytics, that offers both scalable batch and interactive data analytics on raw data, hence avoiding the loading phase bottleneck of RDBMSs. Our system combines a MapReduce based platform with the recently proposed NoDB paradigm, which optimizes traditional centralized RDBMSs for in-situ queries of raw files. We revisit the NoDB's centralized design and scale it out supporting multiple clients and data processing nodes to produce a new distributed data analytics system we call Distributed NoDB (DiNoDB). DiNoDB leverages MapReduce batch queries to produce critical pieces of metadata (e.g., distributed positional maps and vertical indices) to speed up interactive queries without the overheads of the data loading and data movement phases allowing users to quickly and efficiently exploit their data. Our experimental analysis demonstrates that DiNoDB significantly reduces the data-to-query latency with respect to comparable state-of-the-art distributed query engines, like Shark, Hive and HadoopDB.

Toward timely, predictable and cost-effective data analytics

Renata Borovica-Gajic

Modern industrial, government, and academic organizations are collecting massive amounts of data at an unprecedented scale and pace. The ability to perform timely, predictable and cost-effective analytical processing of such large data sets in order to extract deep insights is now a key ingredient for success. Traditional database systems (DBMS) are, however, not the first choice for servicing these modern applications, despite 40 years of database research. This is due to the fact that modern applications exhibit different behavior from the one assumed by DBMS: a) timely data exploration as a new trend is characterized by ad-hoc queries and a short user interaction period, leaving little time for DBMS to do good performance tuning, b) accurate statistics representing relevant summary information about distributions of ever increasing data are frequently missing, resulting in suboptimal plan decisions and consequently poor and unpredictable query execution performance, and c) cloud service providers - a major winner in the data analytics game due to the low cost of (shared) storage - have shifted the control over data storage from DBMS to the cloud providers, making it harder for DBMS to optimize data access. This thesis demonstrates that database systems can still provide timely, predictable and cost-effective analytical processing, if they use an agile and adaptive approach. In particular, DBMS need to adapt at three levels (to workload, data and hardware characteristics) in order to stabilize and optimize performance and cost when faced with requirements posed by modern data analytics applications. Workload-driven data ingestion is introduced with NoDB as a means to enable efficient data exploration and reduce the data-to-insight time (i.e., the time to load the data and tune the system) by doing these steps lazily and incrementally as a side-effect of posed queries as opposed to mandatory first steps. Data-driven runtime access path decision making introduced with Smooth Scan alleviates suboptimal query execution, postponing the decision on access paths from query optimization, where statistics are heavily exploited, to query execution, where the system can obtain more details about data distributions. Smooth Scan uses access path morphing from one physical alternative to another to fit the observed data distributions, which removes the need for a priori access path decisions and substantially improves the predictability of DBMS. Hardware-driven query execution introduced with Skipper enables the usage of cold storage devices (CSD) as a cost-effective solution for storing the ever increasing customer data. Skipper uses an out-of-order CSD-driven query execution model based on multi-way joins coupled with efficient cache and I/O scheduling policies to hide the non-uniform access latencies of CSD. This thesis advocates runtime adaptivity as a key to dealing with raising uncertainty about workload characteristics that modern data analytics applications exhibit. Overall, the techniques introduced in this thesis through the three levels of adaptivity (workload, data and hardware-driven adaptivity) increase the usability of database systems and the user satisfaction in the case of big data exploration, making low-cost data analytics reality.

EPFL2016

Adaptive Query Processing on Raw Data Files

Ioannis Alagiannis

Nowadays, business and scientific applications accumulate data at an increasing pace. This growth of information has already started to outgrow the capabilities of database management systems (DBMS). In a typical DBMS usage scenario, the user should define a schema, load the data and tune the system for an expected workload before submitting any queries. Copying data into a database is a significant investment in terms of time and resources, and in many cases unnecessary or even no longer feasible in practice due to the explosive data growth. Additionally, the way DBMS store and organize data during data loading defines how data will be accessed for a given workload and thus, the maximum performance. Selecting the underlying data layout (row-store or column-store) is a critical first tuning decision which cannot change. Nevertheless, today query analysis is not static; it evolves as queries change. Hence, static design decisions can be suboptimal. In this thesis, we advocate in situ query processing as the principal way to manage data in a database. We reconsider the data loading phase and redesign traditional query processing architectures to work efficiently over raw data files to address the heavy initialization cost that comes with data loading. We present adaptive data loading as an alternative to traditional full a priori data loading. We explore the potential of in situ query processing in the context of current DBMS architectures. We identify performance bottlenecks specific for in situ processing and we introduce an adaptive indexing mechanism (positional map) that maintains positional information to provide efficient access to raw data files, together with a flexible caching structure and techniques for collecting statistics over raw data files. Moreover, we design a flexible query engine that is not built around a single storage layout but it can exploit different storage layouts and data execution strategies in a single engine. It decides during query processing, which design fits the input queries and properly adapts the underlying data storage. By applying code generation techniques, we dynamically generate access operators tailored for specific classes of queries. This thesis revises the traditional paradigm of loading, tuning and then querying by using in situ query processing as the principal way to minimize data-to-query time. We show that raw data files should not be considered ``outside'' the DBMS and full data loading should not be a requirement to exploit database technology. On the contrary, proper techniques specifically tailored to overcome limitations that come with accessing raw data files can eliminate the data loading overhead making, therefore, raw data files a first-class citizen, fully integrated with the query engine. The proposed roadmap can provide guidance on how to convert any traditional DBMS into an efficient in situ query engine.

EPFL2015

Robustness and invariance properties of image classifiers

Apostolos Modas

Deep neural networks have achieved impressive results in many image classification tasks. However, since their performance is usually measured in controlled settings, it is important to ensure that their decisions remain correct when deployed in noisy environments. In fact, deep networks are not robust to a large variety of semantic-preserving image modifications, even to imperceptible image changes -- known as adversarial perturbations -- that can arbitrarily flip the prediction of a classifier. The poor robustness of image classifiers to small data distribution shifts raises serious concerns regarding their trustworthiness. To build reliable machine learning models, we must design principled methods to analyze and understand the mechanisms that shape robustness and invariance. This is exactly the focus of this thesis.First, we study the problem of computing sparse adversarial perturbations, and exploit the geometry of the decision boundaries of image classifiers for computing sparse perturbations very fast. We evaluate the robustness of deep networks to sparse adversarial perturbations in high-dimensional datasets, and reveal a qualitative correlation between the location of the perturbed pixels and the semantic features of the images. Such correlation suggests a deep connection between adversarial examples and the data features that image classifiers learn.To better understand this connection, we provide a geometric framework that connects the distance of data samples to the decision boundary, with the features existing in the data. We show that deep classifiers have a strong inductive bias towards invariance to non-discriminative features, and that adversarial training exploits this property to confer robustness. We demonstrate that the invariances of robust classifiers are useful in data-scarce domains, while the improved understanding of the data influence on the inductive bias of deep networks can be exploited to design more robust classifiers. Finally, we focus on the challenging problem of generalization to unforeseen corruptions of the data, and we propose a novel data augmentation scheme that relies on simple families of max-entropy image transformations to confer robustness to common corruptions. We analyze our method and demonstrate the importance of the mixing strategy on synthesizing corrupted images, and we reveal the robustness-accuracy trade-offs arising in the context of common corruptions. The controllable nature of our method permits to easily adapt it to other tasks and achieve robustness to distribution shifts in data-scarce applications.Overall, our results contribute to the understanding of the fundamental mechanisms of deep image classifiers, and pave the way for building more reliable machine learning systems that can be deployed in real-world environments.

EPFL2022