Goods: Organizing Google's Datasets
Total Page：16
File Type：pdf, Size：1020Kb
Goods: Organizing Google’s Datasets ∗ Alon Halevy2, Flip Korn1, Natalya F. Noy1, Christopher Olston1, Neoklis Polyzotis1, Sudip Roy1, Steven Euijong Whang1 1Google Research 2Recruit Institute of Technology [email protected], {flip, noy, olston, npolyzotis, sudipr, swhang}@google.com ABSTRACT exist for managing datasets. We argue that developing principled Enterprises increasingly rely on structured datasets to run their busi- and flexible approaches to dataset management has become imper- nesses. These datasets take a variety of forms, such as structured ative, lest companies run the risk of internal siloing of datasets, files, databases, spreadsheets, or even services that provide access which, in turn, results in significant losses in productivity and op- to the data. The datasets often reside in different storage systems, portunities, duplication of work, and mishandling of data. may vary in their formats, may change every day. In this paper, Enterprise Data Management (EDM) is one common way to or- we present Goods, a project to rethink how we organize structured ganize datasets in an enterprise setting. However, in the case of datasets at scale, in a setting where teams use diverse and often EDM, stakeholders in the company must embrace this approach, idiosyncratic ways to produce the datasets and where there is no using an EDM system to publish, retrieve, and integrate their datasets. centralized system for storing and querying them. Goods extracts An alternative approach is to enable complete freedom within the metadata ranging from salient information about each dataset (own- enterprise to access and generate datasets and to solve the problem ers, timestamps, schema) to relationships among datasets, such as of finding the right data in a post-hoc manner. This approach is similarity and provenance. It then exposes this metadata through similar in spirit to the concept of data lakes [4, 22], where the lake services that allow engineers to find datasets within the company, comprises and continuously accumulates all the datasets generated to monitor datasets, to annotate them in order to enable others to within the enterprise. The goal is then to provide methods to “fish” use their datasets, and to analyze relationships between them. We the right datasets out of the lake on the as-needed basis. discuss the technical challenges that we had to overcome in order In this paper, we describe Google Dataset Search (Goods), such to crawl and infer the metadata for billions of datasets, to main- a post-hoc system that we built in order to organize the datasets tain the consistency of our metadata catalog at scale, and to expose that are generated and used within Google. Specifically, Goods the metadata to users. We believe that many of the lessons that we collects and aggregates metadata about datasets after the datasets learned are applicable to building large-scale enterprise-level data- were created, accessed, or updated by various pipelines, without management systems in general. interfering with dataset owners or users. Put differently, teams and engineers continue to generate and access datasets using the tools of their choice, and Goods works in the background, in a non- 1. INTRODUCTION intrusive manner, to gather the metadata about datasets and their Most large enterprises today witness an explosion in the num- usage. Goods then uses this metadata to power services that en- ber of datasets that they generate internally for use in ongoing re- able Google engineers to organize and find their datasets in a more search and development. The reason behind this explosion is sim- principled manner. ple: by allowing engineers and data scientists to consume and gen- Figure 1 shows a schematic overview of our system. Goods con- erate datasets in an unfettered manner, enterprises promote fast de- tinuously crawls different storage systems and the production in- velopment cycles, experimentation, and ultimately innovation that frastructure (e.g., logs from running pipelines) to discover which drives their competitive edge. As a result, these internally gener- datasets exist and to gather metadata about each one (e.g., owners, ated datasets often become a prime asset of the company, on par time of access, content features, accesses by production pipelines). with source code and internal infrastructure. However, while enter- Goods aggregates this metadata in a central catalog and correlates prises have developed a strong culture on how to manage the lat- the metadata about a specific dataset with information about other ter, with source-code development tools and methodologies that we datasets. now consider “standard” in the industry (e.g., code versioning and Goods uses this catalog to provide Google engineers with ser- indexing, reviews, or testing), similar approaches do not generally vices for dataset management. To illustrate the types of services ∗Work done while at Google Research. powered by Goods, imagine a team that is responsible for devel- oping natural language understanding (NLU) of text corpora (say, news articles). The engineers on the team may be distributed across Permission to make digital or hard copies of part or all of this work for personal or the globe and they maintain several pipelines that add annotations classroom use is granted without fee provided that copies are not made or distributed to different text corpora. Each pipeline can have multiple stages for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. that add annotations based on various techniques including phrase For all other uses, contact the owner/author(s). chunking, part-of-speech tagging, and co-reference resolution. Other SIGMOD/PODS’16 June 26 - July 01, 2016, San Francisco, CA, USA teams can consume the datasets that the NLU team generates, and c 2016 Copyright held by the owner/author(s). the NLU team’s pipelines may consume datasets from other teams. ACM ISBN 978-1-4503-3531-7/16/06. DOI: http://dx.doi.org/10.1145/2882903.2903730 Figure 1: Overview of Google Dataset Search (Goods). The figure shows the Goods dataset catalog that collects metadata about datasets from various storage systems as well as other sources. We also infer metadata by processing additional sources such as logs and information about dataset owners and their projects, by analyzing content of the datasets, and by collecting input from the Goods users. We use the information in the catalog to build tools for search, monitoring, and visualizing flow of data. Based on the information in its catalog, Goods provides a dash- lar to the content of the current dataset. The similarity informa- board for the NLU team (in this case, dataset producers), which tion may enable novel combinations of datasets: for example, if displays all their datasets and enables browsing them by facets (e.g., two datasets share a primary key column, then they may provide owner, data center, schema). Even if the team’s datasets are in di- complementary information and are therefore a good candidate for verse storage systems, the engineers get a unified view of all their joining. datasets and dependencies among them. Goods can monitor fea- Goods allows users to expand the catalog with crowd-sourced tures of the dataset, such as its size, distribution of values in its metadata. For instance, dataset owners can annotate datasets with contents, or its availability, and then alert the owners if the features descriptions, in order to help users figure out which datasets are change unexpectedly. appropriate for their use (e.g., which analysis techniques are used Another important piece of information that Goods provides is in certain datasets and which pitfalls to watch out for). Dataset au- the dataset provenance: namely, the information about which datasets ditors can tag datasets that contain sensitive information and alert were used to create a given dataset (upstream datasets), and those dataset owners or prompt a review to ensure that the data is han- that rely on it (downstream datasets). Note that both the upstream dled appropriately. In this manner, Goods and its catalog become a and downstream datasets may be created by other teams. When an hub through which users can share and exchange information about engineer in the NLU team observes a problem with a dataset, she the generated datasets. Goods also exposes an API through which can examine the provenance visualization to determine whether a teams can contribute metadata to the catalog both for the teams own change in some upstream dataset had caused the problem. Simi- restricted use as well as to help other teams and users understand larly, if the team is about to make a significant change to its pipeline their datasets easily. or has discovered a bug in an existing dataset that other teams have As we discuss in the rest of the paper, we addressed many chal- consumed already, they can quickly notify those affected by the lenges in designing and building Goods, arising from the sheer problem. number of datasets (tens of billions in our case), the high churn in From the perspective of dataset consumers, such as those not terms of updates, the sizes of individual datasets (gigabytes or ter- part of the NLU team in our example, Goods provides a search en- abytes in many cases), the many different data formats and stores gine over all the datasets in the company, plus facets for narrowing they reside in, and the varying quality and importance of informa- search results, to find the most up-to-date or potentially important tion collected about each dataset. Many of the challenges that we datasets. Goods presents a profile page for every dataset, which addressed in Goods were precipitated by the scale and characteris- helps users unfamiliar with the data to understand its schema and tics of the data lake at Google.