Search

Computational notebooks (e.g., Jupyter, Google Colab) are widely used by data scientists. A key feature of notebooks is the interactive computing model of iteratively executing cells (i.e., a set of statements) and observing the result (e.g., model or plot). Unfortunately, existing notebook systems do not offer time-traveling to past states: when the user executes a cell, the notebook session state consisting of user-defined variables can be irreversibly modified - e.g., the user cannot 'un-drop' a dataframe column. This is because, unlike DBMS, existing notebook systems do not keep track of the session state. Existing techniques for checkpointing and restoring session states, such as OS-level memory snapshot or application-level session dump, are insufficient: checkpointing can incur prohibitive storage costs and may fail, while restoration can only be inefficiently performed from scratch by fully loading checkpoint files. In this paper, we introduce a new notebook system, Kishu, that offers time-traveling to and from arbitrary notebook states using an efficient and fault-tolerant incremental checkpoint and checkout mechanism. Kishu creates incremental checkpoints that are small and correctly preserve complex inter-variable dependencies at a novel Co-variable granularity. Then, to return to a previous state, Kishu accurately identifies the state difference between the current and target states to perform incremental checkout at sub-second latency with minimal data loading. Kishu is compatible with 146 object classes from popular data science libraries (e.g., Ray, Spark, PyTorch), and reduces checkpoint size and checkout time by up to 4.55x and 9.02x, respectively, on a variety of notebooks.

Computational notebooks (e.g., Jupyter, Google Colab) are widely used for interactive data science and machine learning. In those frameworks, users can start a session, then execute cells (i.e., a set of statements) to create variables, train models, visualize results, etc. Unfortunately, existing notebook systems do not offer live migration: when a notebook launches on a new machine, it loses its state, preventing users from continuing their tasks from where they had left off. This is because, unlike DBMS, the sessions directly rely on underlying kernels (e.g., Python/R interpreters) without an additional data management layer. Existing techniques for preserving states, such as copying all variables or OS-level checkpointing, are unreliable (often fail), inefficient, and platform-dependent. Also, re-running code from scratch can be highly time-consuming. In this paper, we introduce a new notebook system, ElasticNotebook, that offers live migration via checkpointing/restoration using a novel mechanism that is reliable, efficient, and platform-independent. Specifically, by observing all cell executions via transparent, lightweight monitoring, ElasticNotebook can find a reliable and efficient way (i.e., replication plan) for reconstructing the original session state, considering variable-cell dependencies, observed runtime, variable sizes, etc. To this end, our new graph-based optimization problem finds how to reconstruct all variables (efficiently) from a subset of variables that can be transferred across machines. We show that ElasticNotebook reduces end-to-end migration and restoration times by 85%-98% and 94%-99%, respectively, on a variety (i.e., Kaggle, JWST, and Tutorial) of notebooks with negligible runtime and memory overheads of <2.5% and <10%., Comment: Accepted to VLDB 2024

The end-to-end lookup latency of a hierarchical index -- such as a B-tree or a learned index -- is determined by its structure such as the number of layers, the kinds of branching functions appearing in each layer, the amount of data we must fetch from layers, etc. Our primary observation is that by optimizing those structural parameters (or designs) specifically to a target system's I/O characteristics (e.g., latency, bandwidth), we can offer a faster lookup compared to the ones that are not optimized. Can we develop a systematic method for finding those optimal design parameters? Ideally, the method must have the potential to generate almost any existing index or a novel combination of them for the fastest possible lookup. In this work, we present new data and an I/O-aware index builder (called AirIndex) that can find high-speed hierarchical index designs in a principled way. Specifically, AirIndex minimizes an objective function expressing the end-to-end latency in terms of various designs -- the number of layers, types of layers, and more -- for given data and a storage profile, using a graph-based optimization method purpose-built to address the computational challenges rising from the inter-dependencies among index layers and the exponentially many candidate parameters in a large search space. Our empirical studies confirm that AirIndex can find optimal index designs, build optimal indexes within the times comparable to existing methods, and deliver up to 4.1x faster lookup than a lightweight B-tree library (LMDB), 3.3x--46.3x faster than state-of-the-art learned indexes (RMI/CDFShop, PGM-Index, ALEX/APEX, PLEX), and 2.0 faster than Data Calculator's suggestion on various dataset and storage settings., Comment: 13 pages, 3 appendices, 19 figures, to appear at SIGMOD 2024

In machine learning (ML), Python serves as a convenient abstraction for working with key libraries such as PyTorch, scikit-learn, and others. Unlike DBMS, however, Python applications may lose important data, such as trained models and extracted features, due to machine failures or human errors, leading to a waste of time and resources. Specifically, they lack four essential properties that could make ML more reliable and user-friendly -- durability, atomicity, replicability, and time-versioning (DART). This paper presents our vision of Transactional Python that provides DART without any code modifications to user programs or the Python kernel, by non-intrusively monitoring application states at the object level and determining a minimal amount of information sufficient to reconstruct a whole application. Our evaluation of a proof-of-concept implementation with public PyTorch and scikit-learn applications shows that DART can be offered with overheads ranging 1.5%--15.6%., Comment: 5 pages, 5 figures, to appear at DEEM 2023

With data pipeline tools and the expressiveness of SQL, managing interdependent materialized views (MVs) are becoming increasingly easy. These MVs are updated repeatedly upon new data ingestion (e.g., daily), from which database admins can observe performance metrics (e.g., refresh time of each MV, size on disk) in a consistent way for different types of updates (full vs. incremental) and for different systems (single node, distributed, cloud-hosted). One missed opportunity is that existing data systems treat those MV updates as independent SQL statements without fully exploiting their dependency information and performance metrics. However, if we know that the result of a SQL statement will be consumed immediately after for subsequent operations, those subsequent operations do not have to wait until the early results are fully materialized on storage because the results are already readily available in memory. Of course, this may come at a cost because keeping results in memory (even temporarily) will reduce the amount of available memory; thus, our decision should be careful. In this paper, we introduce a new system, called S/C, which tackles this problem through efficient creation and update of a set of MVs with acyclic dependencies among them. S/C judiciously uses bounded memory to reduce end-to-end MV refresh time by short-circuiting expensive reads and writes; S/C's objective function accurately estimates time savings from keeping intermediate data in memory for particular periods. Our solution jointly optimizes an MV refresh order, what data to keep in memory, and when to release data from memory. At a high level, S/C still materializes all data exactly as defined in MV definitions; thus, it doesn't impact any service-level agreements. In our experiments with TPC-DS datasets (up to 1TB), we show S/C's optimization can speedup end-to-end runtime by 1.04x-5.08x with 1.6GB memory., Comment: Accepted to ICDE 2023

For exploratory data analysis, it is often desirable to know what answers you are likely to get before actually obtaining those answers. This can potentially be achieved by designing systems to offer the estimates of a data operation result -- say op(data) -- earlier in the process based on partial data processing. Those estimates continuously refine as more data is processed and finally converge to the exact answer. Unfortunately, the existing techniques -- called Online Aggregation (OLA) -- are limited to a single operation; that is, we cannot obtain the estimates for op(op(data)) or op(...(op(data))). If this Deep OLA becomes possible, data analysts will be able to explore data more interactively using complex cascade operations. In this work, we take a step toward Deep OLA with evolving data frames (edf), a novel data model to offer OLA for nested ops -- op(...(op(data))) -- by representing an evolving structured data (with converging estimates) that is closed under set operations. That is, op(edf) produces yet another edf; thus, we can freely apply successive operations to edf and obtain an OLA output for each op. We evaluate its viability with Wake, an edf-based OLA system, by examining against state-of-the-art OLA and non-OLA systems. In our experiments on TPC-H dataset, Wake produces its first estimates 4.93x faster (median) -- with 1.3x median slowdown for exact answers -- compared to conventional systems. Besides its generality, Wake is also 1.92x faster (median) than existing OLA systems in producing estimates of under 1% relative errors., Comment: 16 pages, 13 figures, 3 appendices, to appear at SIGMOD 2023

Existing learned indexes (e.g., RMI, ALEX, PGM) optimize the internal regressor of each node, not the overall structure such as index height, the size of each layer, etc. In this paper, we share our recent findings that we can achieve significantly faster lookup speed by optimizing the structure as well as internal regressors. Specifically, our approach (called AirIndex) expresses the end-to-end lookup time as a novel objective function, and searches for optimal design decisions using a purpose-built optimizer. In our experiments with state-of-the-art methods, AirIndex achieves 3.3x-7.7x faster lookup for the data stored on local SSD, and 1.4x-3.0x faster lookup for the data on Azure Cloud Storage., Comment: 5 pages, to be published in AIDB at VLDB 2022

Modern data warehouses can scale compute nodes independently of storage. These systems persist their data on cloud storage, which is always available and cost-efficient. Ad-hoc compute nodes then fetch necessary data on-demand from cloud storage. This ability to quickly scale or shrink data systems is highly beneficial if query workloads may change over time. We apply this new architecture to search engines with a focus on optimizing their latencies in cloud environments. However, simply placing existing search engines (e.g., Apache Lucene) on top of cloud storage significantly increases their end-to-end query latencies (i.e., more than 6 seconds on average in one of our studies). This is because their indexes can incur multiple network round-trips due to their hierarchical structure (e.g., skip lists, B-trees, learned indexes). To address this issue, we develop a new statistical index (called IoU Sketch). For lookup, IoU Sketch makes multiple asynchronous network requests in parallel. While IoU Sketch may fetch more bytes than existing indexes, it significantly reduces the index lookup time because parallel requests do not block each other. Based on IoU Sketch, we build an end-to-end search engine, called Airphant; we describe how Airphant builds, optimizes, and manages IoU Sketch; and ultimately, supports keyword-based querying. In our experiments with four real datasets, Airphant's average end-to-end latencies are between 13 milliseconds and 300 milliseconds, being up to 8.97x faster than Apache Lucence and 113.39x faster than Elasticsearch., Comment: 17 pages, to be published in ICDE 2022

The rising volume of datasets has made training machine learning (ML) models a major computational cost in the enterprise. Given the iterative nature of model and parameter tuning, many analysts use a small sample of their entire data during their initial stage of analysis to make quick decisions (e.g., what features or hyperparameters to use) and use the entire dataset only in later stages (i.e., when they have converged to a specific model). This sampling, however, is performed in an ad-hoc fashion. Most practitioners cannot precisely capture the effect of sampling on the quality of their model, and eventually on their decision-making process during the tuning phase. Moreover, without systematic support for sampling operators, many optimizations and reuse opportunities are lost. In this paper, we introduce BlinkML, a system for fast, quality-guaranteed ML training. BlinkML allows users to make error-computation tradeoffs: instead of training a model on their full data (i.e., full model), BlinkML can quickly train an approximate model with quality guarantees using a sample. The quality guarantees ensure that, with high probability, the approximate model makes the same predictions as the full model. BlinkML currently supports any ML model that relies on maximum likelihood estimation (MLE), which includes Generalized Linear Models (e.g., linear regression, logistic regression, max entropy classifier, Poisson regression) as well as PPCA (Probabilistic Principal Component Analysis). Our experiments show that BlinkML can speed up the training of large-scale ML tasks by 6.26x-629x while guaranteeing the same predictions, with 95% probability, as the full model., Comment: 22 pages, SIGMOD 2019

Estimating the selectivity of a query is a key step in almost any cost-based query optimizer. Most of today's databases rely on histograms or samples that are periodically refreshed by re-scanning the data as the underlying data changes. Since frequent scans are costly, these statistics are often stale and lead to poor selectivity estimates. As an alternative to scans, query-driven histograms have been proposed, which refine the histograms based on the actual selectivities of the observed queries. Unfortunately, these approaches are either too costly to use in practice---i.e., require an exponential number of buckets---or quickly lose their advantage as they observe more queries. In this paper, we propose a selectivity learning framework, called QuickSel, which falls into the query-driven paradigm but does not use histograms. Instead, it builds an internal model of the underlying data, which can be refined significantly faster (e.g., only 1.9 milliseconds for 300 queries). This fast refinement allows QuickSel to continuously learn from each query and yield increasingly more accurate selectivity estimates over time. Unlike query-driven histograms, QuickSel relies on a mixture model and a new optimization algorithm for training its model. Our extensive experiments on two real-world datasets confirm that, given the same target accuracy, QuickSel is 34.0x-179.4x faster than state-of-the-art query-driven histograms, including ISOMER and STHoles. Further, given the same space budget, QuickSel is 26.8%-91.8% more accurate than periodically-updated histograms and samples, respectively.

Despite 25 years of research in academia, approximate query processing (AQP) has had little industrial adoption. One of the major causes of this slow adoption is the reluctance of traditional vendors to make radical changes to their legacy codebases, and the preoccupation of newer vendors (e.g., SQL-on-Hadoop products) with implementing standard features. Additionally, the few AQP engines that are available are each tied to a specific platform and require users to completely abandon their existing databases---an unrealistic expectation given the infancy of the AQP technology. Therefore, we argue that a universal solution is needed: a database-agnostic approximation engine that will widen the reach of this emerging technology across various platforms. Our proposal, called VerdictDB, uses a middleware architecture that requires no changes to the backend database, and thus, can work with all off-the-shelf engines. Operating at the driver-level, VerdictDB intercepts analytical queries issued to the database and rewrites them into another query that, if executed by any standard relational engine, will yield sufficient information for computing an approximate answer. VerdictDB uses the returned result set to compute an approximate answer and error estimates, which are then passed on to the user or application. However, lack of access to the query execution layer introduces significant challenges in terms of generality, correctness, and efficiency. This paper shows how VerdictDB overcomes these challenges and delivers up to 171$\times$ speedup (18.45$\times$ on average) for a variety of existing engines, such as Impala, Spark SQL, and Amazon Redshift, while incurring less than 2.6% relative error. VerdictDB is open-sourced under Apache License., Comment: Extended technical report of the paper that appeared in Proceedings of the 2018 International Conference on Management of Data, pp. 1461-1476. ACM, 2018

In today's databases, previous query answers rarely benefit answering future queries. For the first time, to the best of our knowledge, we change this paradigm in an approximate query processing (AQP) context. We make the following observation: the answer to each query reveals some degree of knowledge about the answer to another query because their answers stem from the same underlying distribution that has produced the entire dataset. Exploiting and refining this knowledge should allow us to answer queries more analytically, rather than by reading enormous amounts of raw data. Also, processing more queries should continuously enhance our knowledge of the underlying distribution, and hence lead to increasingly faster response times for future queries. We call this novel idea---learning from past query answers---Database Learning. We exploit the principle of maximum entropy to produce answers, which are in expectation guaranteed to be more accurate than existing sample-based approximations. Empowered by this idea, we build a query engine on top of Spark SQL, called Verdict. We conduct extensive experiments on real-world query traces from a large customer of a major database vendor. Our results demonstrate that Verdict supports 73.7% of these queries, speeding them up by up to 23.0x for the same accuracy level compared to existing AQP systems., Comment: This manuscript is an extended report of the work published in ACM SIGMOD conference 2017

Interactive visualizations are crucial in ad hoc data exploration and analysis. However, with the growing number of massive datasets, generating visualizations in interactive timescales is increasingly challenging. One approach for improving the speed of the visualization tool is via data reduction in order to reduce the computational overhead, but at a potential cost in visualization accuracy. Common data reduction techniques, such as uniform and stratified sampling, do not exploit the fact that the sampled tuples will be transformed into a visualization for human consumption. We propose a visualization-aware sampling (VAS) that guarantees high quality visualizations with a small subset of the entire dataset. We validate our method when applied to scatter and map plots for three common visualization goals: regression, density estimation, and clustering. The key to our sampling method's success is in choosing tuples which minimize a visualization-inspired loss function. Our user study confirms that optimizing this loss function correlates strongly with user success in using the resulting visualizations. We also show the NP-hardness of our optimization problem and propose an efficient approximation algorithm. Our experiments show that, compared to previous methods, (i) using the same sample size, VAS improves user's success by up to 35% in various visualization tasks, and (ii) VAS can achieve a required visualization quality up to 400 times faster.