Skip to main content
Category

R Language

ISC PROPOSAL SUBMISSION FAILURE

By Blog, News, R Language

by Joseph Rickert

The ISC has determined that an error in the ISC proposal submission process has caused us to lose some, but not all proposals. If you submitted a project proposal to the ISC and have received a confirmation email then you are fine. However, if you have not received a confirmation email, please email your proposal as a pdf attachment to proposal@r-consortium.org

If you do not receive a confirmation within 24 hours please email hadley@rstudio.com. Once again, if you have already received a confirmation that your proposal was received you do not need to take further action.

The revised deadline for submitting proposals is now midnight PST, Sunday October 15th.

We apologize for the inconvenience.

Take the R Consortium’s Survey on R!

By Announcement, Blog, News, R Consortium Project, R Language

by Joseph Rickert and Hadley Wickham

Help us keep the conversation going: Take the R Consortium’s Survey. Let us know: What are you thinking? What do you make of the way R is developing? How do you use R? What is important to you? How could life be better? What issues should we be addressing? What does the big picture look like? We are looking for a few clues and we would like to hear from the entire R Community.

    

The R Consortium exists to promote R as a language, environment and community. In order to answer some of the questions above and to help us understand our mission better we have put together the first of what we hope will be an annual survey of R users. This first attempt is a prototype. We don’t have any particular hypothesis or point of view. We would like to reach everyone who is interested in participating. So please, take a few minutes to take the survey yourself and help us get the word out. The survey will adapt depending on your answers, but will take about 10 minutes to complete.

The anonymized results of the survey will be made available to the community for analysis. Thank you for participating.

                                                                                   Take the survey now!      

现在进行调查!     今すぐ調査をしてください!    Participez à l’enquête en ligne!    ¡Tome la encuesta ahora!

 

 

Code Coverage Tool for R Working Group Achieves First Release

By Blog, News, R Consortium Project, R Language

by Mark Hornick, Code Coverage Working Group Leader

The “Code Coverage Tool for R” project, proposed by Oracle and approved by the R Consortium Infrastructure Steering Committee, started just over a year ago. Project goals included providing an enhanced tool that determines code coverage upon execution of a test suite, and leveraging such a tool more broadly as part of the R ecosystem.

What is code coverage?

As defined in Wikipedia, “code coverage is a measure used to describe the degree to which the source code of a program is executed when a particular test suite runs. A program with high code coverage, measured as a percentage, has had more of its source code executed during testing which suggests it has a lower chance of containing undetected software bugs compared to a program with low code coverage.”

Why code coverage?

Code coverage is an essential metric for understanding software quality. For R, developers and users alike should be able to easily see what percent of an R package’s code has been tested and the status of those tests. By knowing code is well-tested, users have greater confidence in selecting CRAN packages. Further, automating test suite execution with code coverage analysis helps ensure new package versions don’t unknowingly break existing tests and user code.

Approach and main features in release

After surveying the available code coverage tools in the R ecosystem, the working group decided to use the covr package, started by Jim Hester in December 2014, as a foundation and continue to build on its success. The working group has enhanced covr to support even more R language aspects and needed functionality, including:

  • R6 methods support
  • Address parallel code coverage
  • Enable compiling R with Intel compiler ICC
  • Enhanced documentation / vignettes
  • Provide tool for benchmarking and defining canonical test suite for covr
  • Clean up dependent package license conflicts and change covr license to GPL-3

CRAN Process

Today, code coverage is an optional part of R package development. Some package authors/maintainers provide test suites and leverage code coverage to assess code quality. As noted above, code coverage has significant benefits for the R community to help ensure correct and robust software. One of the goals of the Code Coverage project is to incorporate code coverage testing and reporting into the CRAN process. This will involve working with the R Foundation and the R community on the following points:

  • Encourage package authors and maintainers to develop, maintain, and expand test suites with their packages, and use the enhanced covr package to assess coverage
  • Enable automatic execution of provided test suites as part of the CRAN process, just as binaries of software packages are made available, test suites would be executed and code coverage computed per package
  • Display on each packages CRAN web page its code coverage results, e.g., the overall coverage percentage and a detailed report showing coverage per line of source code.

Next Steps

The working group will assess additional enhancements for covr that will benefit the R community. In addition, we plan to explore with the R Foundation the inclusion of code coverage results in the CRAN process.

Acknowledgements

The following individuals are members of the Code Coverage Working Group:

  • Shivank Agrawal
  • Chris Campbell
  • Santosh Chaudhari
  • Karl Forner
  • Jim Hester
  • Mark Hornick
  • Chen Liang
  • Willem Ligtenberg
  • Andy Nicholls
  • Vlad Sharanhovich
  • Tobias Verbeke
  • Qin Wang
  • Hadley Wickham – ISC Sponsor

Improving DBI: A Retrospect

By Blog, News, R Consortium Project, R Language

by Kirill Müller

The “Improving DBI” project, funded by the R consortium and started about a year ago includes the definition and implementation of a testable specification for DBI and making RSQLite fully compliant to the new specification. Besides the established DBI and RSQLite packages, I have spent a lot of time on the new DBItest package. Final updates to these packages will be pushed to CRAN at the end of May. This should give downstream maintainers some time to make accommodations. The follow-up project “Establishing DBI” will focus on fully DBI-compliant backends for MySQL/MariaDB and PostgreSQL, and on minor updates to the specs where appropriate.

DBItest: Specification

The new DBItest package provides a comprehensive backend-agnostic test suite for DBI backends. When the project started, it was merely a collection of test cases. I have considerably expanded the test cases and provided a human-readable description for each, using literate programming techniques powered by roxygen2. The DBI package weaves these chunks of text to a single document that describes all test cases covered by the test suite, the textual DBI specification. This approach ensures that further updates to the specification are reflected in both the automatic tests and the text.

This package is aimed at backend implementers, who now can programmatically check with very little effort if their DBI backend conforms to the DBI specification. The verification can be integrated in the automated tests which are run as part of R’s package check mechanism in R CMD check. The odbc package, a new DBI-compliant interface to the ODBC interface, has been using DBItest from day one to enable test-driven development. The bigrquery package is another user of DBItest.

Because not all DBMS support all aspects of DBI, the DBItest package allows developers to restrict which parts of the specification are tested, and “tweak” certain aspects of the tests, e.g., the format of placeholders in parameterized queries. Adapting to other DBMS may require more work due to subtle differences in the implementation of SQL between various DBMS.

DBI: Definition

This package has been around since 2001, it defines the actual DataBase Interface in R.

I have taken over maintenance, and released versions 0.4-1, 0.5-1, and 0.6-1, with release of version 0.7 pending. The most prominent change in this package is, of course, the textual DBI specification, which is included as an HTML vignette in the package. The documentation for the various methods defined by DBI is obtained directly from the specification. These help topics are combined in a sensible order to a single, self-contained document. This format is useful for both DBI users and implementers: users can look up the behavior of a method directly from its help page, and implementers can browse a comprehensive document that describes all aspects of the interface. I have also revised the description and the examples for all help topics. Other changes include:

  • the definition of new generics dbSendStatement() and dbExecute(), for backends that distinguish between queries that return a table and statements that manipulate data,
  • the new dbWithTransaction() generic and the dbBreak() helper function, thanks Barbara Borges Ribero,
  • improved or new default implementations for methods like dbGetQuery(), dbReadTable(), dbQuoteString(), dbQuoteIdentifier(),
  • internal changes that allow methods that don’t have a meaningful return value to return silently,
  • translation of a helper function from C++ to R, to remove the dependency on Rcpp (thanks Hannes Mühleisen).

Fortunately, none of the changes seemed to have introduced any major regression issues with downstream packages. The news contain a comprehensive list of changes.

RSQLite: Implementation

RSQLite 1.1-2 is a complete rewrite of the original C implementation. Before focusing on compliance to the new DBI specification, it was important to assert compatibility to more than 100 packages on CRAN and Bioconductor that use RSQLite. These packages revealed many usage patterns that were difficult to foresee. Most of these usage patterns are supported in version 1.1-2, the more esoteric ones (such as supplying an integer where a logical is required) trigger a warning.

Several rounds of “revdep checking” were necessary before most packages showed no difference in their check output compared to the original implementation. The downstream maintainers and the Bioconductor team were very supportive, and helped spotting functional and performance regressions during the release process. Two point releases were necessary to finally achieve a stable state.

Supporting 64-bit integers also was trickier than anticipated. There is no built-in way to represent 64-bit integers in R. The bit64 package works around this limitation by using a numeric vector as storage, which also happens to use 8 bytes per element, and providing coercion functions. But when an integer column is fetched, it cannot be foreseen if a 64-bit value will occur in the result, and smaller integers must use R’s built-in integer type. For this purpose, an efficient data structure for collecting vectors, which is capable of changing the data type on the fly, has been implemented in C++. This data structure will be useful for many other DBI backends that need support for a 64-bit integer data type, and will be ported to the RKazam package in the follow-up project.

Once the DBI specification was completed, the process of making RSQLite compliant was easy: enable one of the disabled tests, fix the code, make sure all tests pass, rinse, and repeat. If you haven’t tried it, I seriously recommend test-driven development, especially when the tests are already implemented.

The upcoming release of RSQLite 2.0 will require stronger adherence to the DBI specification also from callers. Where possible, I tried to maintain backward compatibility, but in some cases breaks were inevitable because otherwise I’d have had to introduce far too many exceptions and corner cases in the DBI spec. For instance, row names are no longer included by default when writing or reading tables. The original behavior can be re-enabled by calling pkgconfig::set_config(), so that packages or scripts that rely on row names continue to work as before. (The setting is active for the duration of the session, but only for the caller that has called pkgconfig::set_config().) I’m happy to include compatibility switches for other breaking changes if necessary and desired, to achieve both adherence to the specs and compatibility with existing behavior.

A comprehensive list of changes can be found in the news.

Other bits and pieces

The RKazam package is a ready-to-use boilerplate for a DBI backend, named after the hypothetical DBMS used as example in a DBI vignette. It already “passes” all tests of the DBItest package, mostly by calling a function that skips the current test. Starting a DBI backend from scratch requires only copying and renaming the package’s code.

R has limited support for time-of-day data. The hms package aims at filling this gap. It will be useful especially in the follow-up project, because SQLite doesn’t have an intrinsic type for time-of-day data, unlike many other DBMS.

Next steps

The ensemble CRAN release of the three packages DBI, DBItest and RSQLite will occur in parallel to the startup phase for the “Establishing DBI” follow-up project. This project consists of:

  • Fully DBI compatible backends for MySQL/MariaDB and Postgres
  • A backend-agnostic C++ data structure to collect column data in the RKazam package
  • Support for spatial data

In addition, it will contain an update to the DBI specification, mostly concerning support for schemas and for querying the structure of the table returned for a query. Targeting three DBMS instead of one will help properly specify these two particularly tricky parts of DBI. I’m happy to take further feedback from users and backend implementers towards further improvement of the DBI specification.

Acknowledgments

Many thanks to the R Consortium, which has sponsored this project, and to the many contributors who have spotted problems, suggested improvements, submitted pull requests, or otherwise helped make this project a great success. In particular, I’d like to thank Hadley Wickham, who suggested the idea, supported initial development of the DBItest package, and provided helpful feedback; and Christoph Hösler, Hannes Mühleisen, Imanuel Costigan, Jim Hester, Marcel Boldt, and @thrasibule for using it and contributing to it. I enjoyed working on this project, looking forward to “Establishing DBI”!

Simple Features Now on CRAN

By Blog, R Consortium Project, R Language

by Edzer Pebesma

Support for handling and analyzing spatial data in R goes back a long way. In 2003, a group of package developers sat together and decided to adopt a shared understanding of how spatial data should be organized in R. This led to the development of the package sp and its helper packages rgdal and rgeos. sp offers simple classes for points, lines, polygons and grids, which may be associated with further properties (attributes), and takes care of coordinate reference systems. The sp package has helped many users and has made it attractive for others to develop new packages that share sp’s conventions for organizing spatial data by reusing its classes. Today, approximately 350 packages directly depend on sp and many more are indirectly dependent.

After 2003, the rest of the world has broadly settled on adopting a standard for so-called “features”, which can be thought of as “things” in the real world that have a geometry along with other properties. A feature geometry is called simple when it consists of points connected by straight line pieces, and does not intersect itself. Simple feature access is a standard for accessing and exchanging spatial data (points, lines, polygons) as well as for operations defined on them that has been adopted widely over the past ten years, not only by spatial databases such as PostGIS, but also more recent standards such as GeoJSON. The sp package and supporting packages such as rgdal and rgeos predate this standard, which complicates exchange and handling of simple feature data.

The “Simple Features for R” project, one of the projects supported by the R Consortium in its first funding round, addresses these problems by implementing simple features as native R data. The resulting package, sf provides functionality similar to the sp, rgdal for vector data, and rgeos packages together, but for simple features. Instead of S4 classes used by the sp family, it extends R’s data.frame directly, adding a list-column for geometries. This makes it easier to manipulate them with other tools that assume all data objects are data.frames, such as dplyr and tidyverse. Package sf links to the GDAL, PROJ.4 and GEOS libraries, three major geospatial “swiss army knives” for respectively input/output, cartographic (re)projections, and geometric operations (e.g. unions, buffers, intersections and topological relations). sf can be seen as a successor to sp, rgdal (for vector data), and rgeos.

The simple feature standard describes two encodings: well-known text, a human readable form that looks like “POINT(10 12)” or “LINESTRING(4 0,3 2,5 1)”, and well-known binary, a simple binary serialization. The sf package can read and write both. Exchange routines for binary encodings were written in Rcpp, to allow for very fast exchange of data with the linked GDAL and GEOS libraries, but also with other data formats or spatial databases.

The sf project on GitHub has received a considerable attention. Over 100 issues have been raised, many of which received dozens of valuable contributions, and several projects currently under development (mapview, tmap, stplanr) are experimenting with the new data classes. Several authors have provided useful pull requests, and efforts have begun to implement spatial analysis in pipe-based workflows, support dplyr-style verbs and integrate with ggplot.

Besides using data.frames and offering considerably simpler data structures for spatial geometries, advantages of sf over the sp family include: simpler handling of coordinate reference systems (using either EPSG code or PROJ.4 string), the ability to return distance or area values with proper units (meter, feet or US feet), and support for geosphere functions to compute distances or areas for longitude/latitude data, using datum-dependent values for the Earth’s radius and flattening.

The sf package is now available from CRAN, both in source form as in binary form for Windows and MacOSX platforms. The authors are grateful to the CRAN team for their strong support in getting the sf package compiled on all platforms. Support from the R Consortium has helped greatly to give this project priority, draw attention in a wider community, and facilitate travel and communication events.

For additional technical information about sf, look here on my website.

 

Halfway through “Improving DBI”

By Blog, R Consortium Project, R Language

by Kirill Müller

In early 2016 the R Consortium partially accepted my “Improving DBI” proposal. An important part is the design and implementation of a testable DBI specification. Initially I also proposed to make three DBI backends to open-source databases engines (RSQLite, RMySQL, and RPostgres) compatible to the new DBI specification, but funding allows to work on only one DBI backend. I chose RSQLite for a number of reasons:

  • It is a very important package, judging by the number of reverse CRAN and Bioconductor dependencies
  • It’s easy to work with, because everything (including the database engine) is bundled with the package
  • It seemed to be the most advanced package, closest to the (yet to be completed) DBI specification
  • An informal Twitter poll supports this decision by a tiny margin

The project has reached an important milestone, with the release of RSQLite 1.1. This post reports the progress achieved so far, and outlines the next steps.

RSQLite

While the RSQLite API has changed very little (hence the minor version update), it includes a complete rewrite of the original 1.0.0 sources in C++. This has considerably simplified the code, which makes future maintenance easier, and allows us to take advantage of the more sophisticated memory management tools available in Rcpp, which help protect against memory leaks and crashes.

RSQLite 1.1 brings a number of improvements:

  • New strategy for prepared queries: Create a prepared query with dbSendQuery() or dbSendStatement() and bind values with dbBind(). This allows you to efficiently re-execute the same query/statement with different parameter values iteratively (by calling dbBind() several times) or in a batch (by calling dbBind() once with a data-frame-like object).
  • Support for inline parametrised queries via the param argument to dbSendQuery(), dbGetQuery(), dbSendStatement() and dbExecute(), to protect from SQL injection.
  • The existing methods dbSendPreparedQuery() and dbGetPreparedQuery() have been soft-deprecated, because the new API is more versatile, more consistent and stricter about parameter validation.
  • Using UTF8 for queries and parameters: this mean that non-English data should just work without any additional intervention.
  • Improved mapping between SQLite’s cell-types and R’s column-types.

See the release notes for further changes.

The rewrite was implemented by Hadley Wickham before the “Improving DBI” project started, and has been available for a long time on GitHub. Nevertheless, the CRAN release has proven much more challenging than anticipated, because so many CRAN and Bioconductor packages import it. (Maintainers of reverse dependencies might remember multiple e-mails where I was threatening to release RSQLite “for real”.) My aim was to break as little existing code as possible. After numerous rounds of revdep-checking and improving RSQLite, I’m proud to report that the vast majority of reverse dependencies pass their checks just as well (and as quickly!) as they did with v1.0.0. Most tests from v1.0.0 are still present in the current codebase. This means that non-packaged code also has a good chance to work unchanged. I’m happy to work with package maintainers or users whose code breaks after the update.

DBI

I have also released several DBI updates to CRAN, mostly to introduce new generics such as dbBind() (for parametrized/prepared queries) or dbSendStatement() and dbExecute() (for statements which don’t return data). The definition of a formal DBI specification is part of the project, a formatted version is updated continuously.

DBItest

In addition to the textual specification in the DBI package, the DBItest package provides backend independent tests for DBI packages. It can be easily used by package authors to ensure that they follow the DBI specification. This is important because it allows you to take code that works with one DBI backend and easily switch to a different backend (providing that they both support the same SQL dialect). Literate programming techniques using advanced features of roxygen2 help keeping both code and textual specifications in close proximity, so that amendments to the text can be easily tracked back to changes of the test code, and vice versa.

Next steps

The rest of the project will focus on finalizing the specification in both code and text (mostly discussed on GitHub in the issue trackers for the DBI and DBItest projects). At least one new helper package (to handle 64-bit integer types) will be created, and DBI, DBItest, and RSQLite will see yet another release: The first two will finalize the DBI specification, and RSQLite will fully conform to it.

The development happens entirely on GitHub in repositories of the rstats-db organization. Feel free to try out development versions of the packages found there, and to report any problems or ideas at the issue trackers.