All the Badges One Can Earn: Parsing Badges of CRAN Packages READMEs

A while ago we onboarded an exciting package, codemetar by Carl Boettiger. codemetar is an R specific information collector and parser for the CodeMeta project. In particular, codemetar can digest metadata about an R package in order to fill the terms recognized by CodeMeta. This means extracting information from DESCRIPTION but also from e.g. continuous integration¹ badges in the README! In this note, we’ll take advantage of codemetar::extract_badges function to explore the diversity of badges worn by the READMEs of CRAN packages.

🔗 Why codemetar::extract_badges(), and how

CodeMeta recognized terms include information we’ve been getting from badges:

• “contIntegration”, URLs to continuous integration services such as Travis, Appveyor, Codecov;

• “review”, information about review of the software if there was one. codemetar recognizes information from the peer-review badge we add to the README of onboarded packages thanks to work by Karthik Ram.

The list might get longer, so instead of using regular expressions on the README text, we extract and memoize² all badges at once to a data.frame that we then query. The badges extraction is based on (in the dev branch of codemetar):

• conversion to XML using commonmark (see my recent tech note) which works well for the Markdown badges e.g. [![Travis-CI Build Status](https://travis-ci.org/ropensci/codemetar.svg?branch=master)](https://travis-ci.org/ropensci/codemetar).

• extraction and parsing on the first HTML table there is if there’s any, which is necessary for badges in table, which we’ve been discussing for our guidelines.

Note that the CRAN version of codemetar already features extract_badges, but with a badges table creation based on regular expressions only (and a small encoding bug). Here is codemetar::extract_badges in action on the README of the drake package:

library("magrittr")
codemetar::extract_badges("https://raw.githubusercontent.com/ropensci/drake/master/README.md") %>%
  knitr::kable()

Quite handy for our metadata collection!

Read the source for extract_badges here and see how it’s used in this script. You can also compare codemetar’s README with its codemeta.json e.g. the lines for “contIntegration”.

Now since, codemetar::extract_badges exports a nice data.frame for any README with badges, and is exported, it’d be too bad not to use it to gain insights from many, many READMEs!

🔗 Extract badges from CRAN packages

In this exploration we shall concentrate on CRAN packages that indicate a GitHub repo link under the URL field of DESCRIPTION. By the way, if you don’t indicate such links in DESCRIPTION of your package yet, you can (and should) run usethis::use_github_links.

🔗 Get links to GitHub repos from CRAN information

I reckon that I could have also used BugReports like Steven M Mortimer did in his great analysis of CRAN downloads and GitHub stars. I am unsure of how one can get the link to and the content of the README of packages that don’t Rbuildignore their README, such as codemetar (see under Material). The imperfect sample I collected will do for this note.

Here’s how I got all the repo owners and names:

cran_db <- tools::CRAN_package_db()

# only packages that have a GitHub repo
github_cran <- dplyr::filter(cran_db[, c("Package", "URL")],
                             stringr::str_detect(URL, "github\\.com"))

# will need to keep only the URL to the repo
select_github_repo <- function(URL){
  URLs <- stringr::str_split(URL, pattern = ",", simplify = TRUE)
  github_repo <- URLs[stringr::str_detect(URLs, "github\\.com")][1]
  github_repo <- stringr::str_remove(github_repo, "\\#.*$")
  github_repo <- stringr::str_remove(github_repo, "\\#.*[ \\(.*\\)]")
  github_repo <- stringr::str_remove(github_repo, "/$")
  stringr::str_replace(github_repo,".*\\.com\\/", "")
}


github_cran <- dplyr::group_by(github_cran, Package)
github_cran <- dplyr::mutate(github_cran, github = select_github_repo(URL))
github_cran <- tidyr::separate(github_cran, github, "\\/", 
                               into = c("owner", "repo"))
github_cran <- dplyr::ungroup(github_cran)

# Not very general
github_cran$repo[which(github_cran$Package == "webp")] <- "webp"

I needed a bit of string cleaning mostly to deal with the URLs of Jeroen Ooms’ packages, see e.g. this one. I guess I could have cleaned even more, but it was good enough for this exploration.

🔗 Get all badges

Then for each repo I queried the download URL of the preferred README via GitHub’s V3 API, using the gh package. The preferred README is the one GitHub displays on the repo landing page. I used codemetar::extract_badges, of course. I rate-limited the basic function using ratelimitr.

library("magrittr")

github_cran <- readr::read_csv("data/github_cran_links.csv")

.get_badges <- function(owner, repo){
  message(paste(owner, repo, sep = "/"))
  readme <- try(gh::gh("GET /repos/:owner/:repo/readme",
                       owner = owner, repo = repo),
                silent = TRUE)
  if(inherits(readme, "try-error")){
    return(NULL)
  }else{
    
    badges <- codemetar::extract_badges(readme$download_url)
    
    if(nrow(badges)>0){
      badges$owner <- owner
      badges$repo <- repo
    }
    
    return(badges)
  }
  
  
}

.get_badges %>%
  ratelimitr::limit_rate(ratelimitr::rate(1, 1)) -> get_badges

purrr::map2_df(github_cran$owner, 
               github_cran$repo,
              get_badges) -> badges

🔗 Remove non badges from the sample

The way badges are recognized by codemetar::extract_badges is not specific enough, it can include images formatted like badges that aren’t badges but instead either local images or images whose credit is shown as URL. To remove them from the sample, I used a strategy in two steps:

• I first had a look at the most common domains. For the 17 most common of them, I accepted the images except for one, ropensci.org, included because the footer our packages get is formatted as a Markdown badge.

• For the remaining images, a bit more than 200, I used magick to obtain their width and height, and filtered actual badges based on their width/height ratio. Sometimes the link to the image wasn’t even valid, which was also a reason for exclusion, since it revealed the image was a local one.

# extract and parse URLs
badges %>%
  dplyr::pull(image_link) %>%
  purrr::map_df(urltools::url_parse) -> parsed_image_links

# count hits by domain

parsed_image_links %>%
  dplyr::count(domain, sort = TRUE) -> domain_count

# these were manually inspected
# as legit badge providers
ok_domain <- domain_count$domain[1:17]

# keep the badges needing a check
tbd <- dplyr::filter(parsed_image_links,
                     ! domain %in% ok_domain)
                     
# get their size ratio
get_size <- function(url){
  img <- try(magick::image_read(url),
             silent = TRUE)
  
  if(inherits(img, "try-error")){
    tibble::tibble(error = TRUE,
                   image_link = url)
  }else{
    info <- magick::image_info(img)
    info$error <- FALSE
    info$image_link <- url
    info
  }
  
  
}

img_info <- purrr::map_df(urltools::url_compose(tbd),
                          get_size)

img_info <- dplyr::mutate(img_info, ratio = width/height)

# filter badges from images
img_info <- dplyr::filter(img_info, 
                          ratio < 3|error)

# it'd have been wiser to use a row-wise workflow!
badges <- dplyr::filter(badges,
                        !image_link %in% img_info$image_link,
                        !image_link %in% stringr::str_remove(img_info$image_link,
                                                                      "/$"),
                        !tolower(image_link) %in% img_info$image_link,
                        !tolower(image_link) %in% stringr::str_remove(img_info$image_link,
                                                                      "/$"),
                        !stringr::str_detect(image_link,
                                             "ropensci\\.org\\/public\\_images\\/"))
readr::write_csv(badges, "data/aaall_badges.csv")

Don’t judge me by my filenaming skills. I was maybe a bit too enthusiastic!

🔗 Analyze badges from CRAN packages

I wanted to answer several questions about the badges of CRAN packages, beyond being just happy to have been able to collect so many of them.

🔗 How many repos have at least one badge?

github_cran <- readr::read_csv("data/github_cran_links.csv")
# the same repo can have been used by several packages!
badges <- readr::read_csv("data/aaall_badges.csv")
badges <- dplyr::distinct(badges)

nobadges <- dplyr::anti_join(github_cran, badges,
                             by = c("owner", "repo"))

There are 1277 packages without any badge (or rather said, without any badge that we identified) out of a sample of 3541 packages. That means 64% have at least one badge. As a reminder, there are more than 13,000 packages on CRAN so we’re only looking at a subset.

🔗 Among the repos with badges, how many badges?

library("magrittr")
badges %>%
  dplyr::count(repo, owner,
               sort = TRUE) -> badges_count

badges_count %>%
  dplyr::summarise(median = median(n))

## # A tibble: 1 x 1
##   median
##    <int>
## 1      4

library("ggplot2")
badges_count %>%
  ggplot() +
  geom_histogram(aes(n))+
  hrbrthemes::theme_ipsum(base_size = 12, 
                          axis_title_size = 12, 
                          axis_text_size = 12) +
  ggtitle("Number of badges per repo",
          subtitle = "Among repos with at least one badge")

The median number of badges is 4, which corresponds to my gut feeling that the answer would be “a few”. I have a new question, what are the repos with the most badges?

most_badges <- dplyr::filter(badges_count,
                             n == max(n)) 
most_badges

## # A tibble: 2 x 3
##   repo     owner               n
##   <chr>    <chr>           <int>
## 1 gpuR     cdeterman          13
## 2 psycho.R neuropsychology    13

You can browse them at https://github.com/cdeterman/gpuR, https://github.com/neuropsychology/psycho.R.

🔗 How many unique badges are there?

For counting types of badges, I’ll use the domain of image_link. This is an approximation, since e.g. www.r-pkg.org offers several badges.

badges %>%
  dplyr::pull(image_link) %>%
  purrr::map_df(urltools::url_parse) -> parsed_image_links

parsed_image_links %>%
  dplyr::pull(domain) %>%
  unique() %>%
  sort() -> unique_domains

length(unique_domains)

## [1] 50

Not that many after all, so I’ll print all of them! A special mention to https://github.com/ropensci/cchecksapi#badges maintained under our GitHub organization by Scott Chamberlain, to show the CRAN check status of your package!

Unique badge domains collapsed by glue::glue_collapse(unique_domains, sep = ", ", last = " and "): anaconda.org, api.codacy.com, api.travis-ci.org, app.wercker.com, assets.bcdevexchange.org, awesome.re, badge.fury.io, badges.frapsoft.com, badges.gitter.im, badges.herokuapp.com, badges.ropensci.org, bestpractices.coreinfrastructure.org, ci.appveyor.com, circleci.com, codeclimate.com, codecov.io, coveralls.io, cranchecks.info, cranlogs.r-pkg.org, depsy.org, dmlc.github.io, eddelbuettel.github.io, githubbadges.com, githubbadges.herokuapp.com, gitlab.com, hits.dwyl.io, i.imgur.com, img.shields.io, jhudatascience.org, joss.theoj.org, mybinder.org, popmodels.cancercontrol.cancer.gov, pro-pulsar-193905.appspot.com, raw.githubusercontent.com, readthedocs.org, saucelabs.com, semaphoreci.com, travis-ci.com, travis-ci.org, user-images.githubusercontent.com, usgs-r.github.io, www.nceas.ucsb.edu, www.ohloh.net, www.openhub.net, www.paypal.com, www.r-pkg.org, www.rdocumentation.org, www.repostatus.org, www.rpackages.io and zenodo.org.

🔗 What are the most common badges?

Note that this doesn’t take into account the fact that one domain can appear several times in a single README (Travis status for different branches for instance).

parsed_image_links %>%
  dplyr::count(domain, sort = TRUE) %>%
  head(n = 10) %>%
  knitr::kable()

The most common badges are Travis-CI badges, and METACRAN badges from www.r-pkg.org and cranlogs.r-pkg.org. Now, “img.shields.io” is a service for badges of other things… which?

badges %>%
  dplyr::filter(stringr::str_detect(image_link, "img\\.shields\\.io")) %>%
  dplyr::count(text, sort = TRUE)

## # A tibble: 135 x 2
##    text                  n
##    <chr>             <int>
##  1 Coverage Status     196
##  2 License              91
##  3 lifecycle            62
##  4 CoverageStatus       54
##  5 <NA>                 38
##  6 packageversion       37
##  7 Last-changedate      32
##  8 Licence              28
##  9 minimal R version    28
## 10 Github Stars         19
## # ... with 125 more rows

Diverse things, in particular the Tidyverse lifecycle badges. After some discussion, we at rOpenSci have adopted the repostatus.org status badges in our guidelines… but are actually open to repos using both types of badges since their nomenclature can complement each other!

🔗 Conclusion

In this tech note I presented and used one of codemetar’s tools for R package metadata munging, extract_badges. I extracted and analyzed badges information from the READMEs of all CRAN packages that indicate a GitHub repo in the URL field of DESCRIPTION. README badges are a way to show development status, test results, code coverage, peer-review merit, etc.; but can also be used as a machine-readable source of information about these same things.

Explore more of codemetar in its GitHub repo, and check out the CodeMeta project itself. Read our guidelines for package development and maintenance in this gitbook. And have fun adding pretty badges to your own package repos upping your package development & maintenance game!