Introduction

This report checks if the status of packages on CRAN are due to intermittent failures.

Failures defined as warnings, notes or errors without change on:

R version used (if not stable the same svn snapshot)
The package version (Note that CRAN might modify a package without changing the version)
Their dependencies

Reasons of these failures might be because the packages depend on:

Random generation numbers
Flacky external resources
Other ?

Why is this important?

Because package maintainers of dependencies of that package, R core and CRAN team need to check if the failures are false positives.

This report started because it was suggested as something that the R-repositories working group could help the CRAN team.

Retrieve data

It makes use of tools::CRAN_check_results to retrieve the data.

library("dplyr")
library("tools", include.only = c("package_dependencies", "CRAN_check_results"))
library("flextable", include.only = c("flextable", "autofit"))
# Use a LOCAL environment to check if files can be overwritten on my computer
local_build <- as.logical(Sys.getenv("LOCAL", "FALSE"))
yc <- readRDS("today.RDS")
tc <- CRAN_check_results()
# Added 2023/03/09: sometimes some flavors are reported without status: Omit those
tc <- tc[!is.na(tc$Status),]
if (!interactive() && !local_build) {
  message("Saving today's file.")
  saveRDS(tc, file = "today.RDS")
}

The checks are from multiple flavors release, devel, old release and patched on multiple machines and configurations.

old_flavors <- readRDS("flavors.RDS")
flavors <- unique(tc$Flavor)
# One flavor now present in all is the r-devel-windows-x86_64: skip
flavors <- setdiff(flavors, "r-devel-windows-x86_64")
proto <- data.frame(r_version = character(),
                    os = character(),
                    architecture = character(),
                    other = character())
flavors_df <- strcapture(
  pattern = "r-([[:alnum:]]+)-([[:alnum:]]+)-([[:alnum:]_\\+]+)-?(.*)", 
  x = flavors,
  proto = proto)

# Extract R version used and svn id
h <- "https://www.r-project.org/nosvn/R.check/%s/ggplot2-00check.html"
links <- sprintf(h, flavors)
extract_revision <- function(x) {
  r <- readLines(x, 12)[12]
  version <- strcapture(pattern = "([[:digit:]]\\.[[:digit:]]\\.[[:digit:]])",  
                        x = r, proto = data.frame(version = character()))
  revision <- strcapture(pattern = "(r[[:digit:]]+)",  x = r,
                         proto = data.frame(revision = character()))
  cbind(version, revision)
}
revision <- data.frame(version = character(),
                       revision = character())
for (i in links) {
  revision <- rbind(revision, extract_revision(i))
}

flavors_df <- cbind(flavors = flavors, flavors_df, revision)
if (!interactive() && !local_build) {
  saveRDS(flavors_df, "flavors.RDS")
}

m <- match(tc$Flavor, flavors_df$flavors)
tc_flavors <- cbind(tc, flavors_df[m, ])
flextable(flavors_df) |> 
  autofit()

flavors	r_version	os	architecture	other	version	revision
r-devel-linux-x86_64-debian-clang	devel	linux	x86_64	debian-clang		r87868
r-devel-linux-x86_64-debian-gcc	devel	linux	x86_64	debian-gcc		r87880
r-devel-linux-x86_64-fedora-clang	devel	linux	x86_64	fedora-clang		r87732
r-devel-linux-x86_64-fedora-gcc	devel	linux	x86_64	fedora-gcc		r87716
r-devel-macos-arm64	devel	macos	arm64			r87521
r-devel-macos-x86_64	devel	macos	x86_64			r87627
r-patched-linux-x86_64	patched	linux	x86_64		4.4.3	r87871
r-release-linux-x86_64	release	linux	x86_64		4.4.3
r-release-macos-arm64	release	macos	arm64		4.4.0	r86238
r-release-macos-x86_64	release	macos	x86_64		4.4.0
r-release-windows-x86_64	release	windows	x86_64		4.4.3
r-oldrel-macos-arm64	oldrel	macos	arm64		4.3.1
r-oldrel-macos-x86_64	oldrel	macos	x86_64		4.3.2
r-oldrel-windows-x86_64	oldrel	windows	x86_64		4.3.3

It assumes that the same configuration in one package is used for all. Or in other words that the reports of the configuration (svn revision and version) for the A3 package is the same as for all the other packages.

Warning: This assumption is not always true, but this would require to check each log file on each flavor to verify the R and svn id of each package (which could take too much time and resources).

Overview

Briefly an introduction of how much effort goes into checking

library("ggplot2")
theme_set(theme_minimal())
tc |> 
  filter(!is.na(T_install)) |> 
  ggplot() +
  geom_violin(aes(T_install, Flavor)) +
  scale_x_log10() +
  labs(x = "seconds", title = "Time to install", y = element_blank())

Machines (y axis) vs install time (seconds, x axis), violing plot usually around 10 seconds.

Distribution of install time on each machine.

This means that just to install all the packages on the multiple flavors with a single CPU would take 32 days.

tc |> 
  filter(!is.na(T_check)) |> 
  ggplot() +
  geom_violin(aes(T_check, Flavor), trim = FALSE) +
  scale_x_log10() +
  labs(x = "seconds", title = "Time to check", y = element_blank())

Machines (y axis) vs check time (seconds, x axis), violing plot usually around 100 seconds.

Distribution of checking time on each machine.

This means that to check all the packages on the multiple flavors with a single CPU would take 143 days.

tc |> 
  filter(!is.na(T_total)) |> 
  ggplot() +
  geom_violin(aes(T_total, Flavor)) +
  scale_x_log10() +
  labs(x = "seconds", title = "Time to check and install", y = element_blank())

Machines (y axis) vs total time (seconds, x axis), violing plot usually around 100 seconds.

Distribution of total time on each machine.

This means that to install and check all the packages with a single CPU would take 367 days.

I don’t know the computational cost of 266 days of CPU (every day), but a rough calculation of 2.5 cents per hour means 220.27 dollars daily dedicated to this.

tc |> 
  group_by(Package) |> 
  summarize(Versions = n_distinct(Version)) |> 
  ungroup() |> 
  count(Versions, name = "Packages", sort = TRUE) |> 
  flextable() |> 
  autofit()

Versions	Packages
1	21,608
2	545
3	3
4	1

This was surprising, but sometimes checks have multiple versions. Probably when a new version is added and the system don’t catch it for a certain machine.

tc |> 
  group_by(Package) |> 
  summarize(Flavors = n_distinct(Flavor)) |> 
  ungroup() |> 
  count(Flavors, name = "Packages", sort = TRUE) |> 
  flextable() |> 
  autofit()

Flavors	Packages
15	21,969
14	72
12	55
8	14
4	12
11	11
13	10
9	9
10	3
3	1
6	1

Similarly, often packages are only tested on few configurations.

Combining both we can have packages with few configurations that have multiple versions being tested.

tc |> 
  group_by(Package) |> 
  summarize(Versions = as.character(n_distinct(Version)),
            Flavors = n_distinct(Flavor)) |> 
  ungroup() |> 
  count(Flavors, Versions, name = "Packages") |> 
  ggplot() +
  geom_tile(aes(Flavors, Versions, fill = log10(Packages))) +
  scale_x_continuous(expand = expansion())

Flavors of machines and versions of packages

Most packages are just tested one version.

But focusing on those that have just one version of the package being tested, most of the machines have packages either OK or with some notes.

man_colors <- c("OK" = "green", "NOTE" = "darkgreen", 
                "WARNING" = "yellow", "ERROR" = "red", "FAILURE" = "black")
tc |> 
  group_by(Package) |> 
  filter(n_distinct(Version) == 1) |> 
  ungroup() |> 
  group_by(Flavor) |> 
  count(Status, name = "packages") |> 
  mutate(perc = packages/sum(packages),
         Status = forcats::fct_relevel(Status, names(man_colors))) |> 
  ggplot() + 
  geom_col(aes(perc, Flavor, fill = Status)) +
  scale_x_continuous(expand = expansion(), labels = scales::percent_format()) +
  scale_fill_manual(values = man_colors) +
  labs(title = "Packages check status", x = element_blank())

On the vertical axis the machine, on the horitzonal axis the packages colored by the status.

Most frequent status is OK or NOTE on all machines.

If we look at the most frequent status report for packages we can see this table:

ts <- tc |> 
  group_by(Package) |> 
  filter(n_distinct(Version) == 1) |> 
  count(Status, name = "flavors") |> 
  ungroup() |> 
  tidyr::pivot_wider(values_from = flavors, names_from = Status, 
                     values_fill = 0) |> 
  count(OK, NOTE, WARNING, ERROR, FAILURE, name = "packages", sort = TRUE)
download.file("https://cran.r-project.org/web/packages/packages.rds", 
              destfile = "packages.RDS") # From the help page
ap <- readRDS("packages.RDS") |> 
  as.data.frame() |> 
  distinct(Package, .keep_all = TRUE)
ap_bioc <- available.packages(repos = BiocManager::repositories()[1:5])
ap_bioc <- cbind(ap_bioc, Additional_repositories = NA)
ap_colm <- intersect(colnames(ap), colnames(ap_bioc))
ap <- rbind(ap[, ap_colm], ap_bioc[, ap_colm])
head(ts) |> 
  flextable() |> 
  autofit()

OK	NOTE	packages
15	0	12,710
7	8	1,555
3	12	1,176
12	3	1,016
0	15	936
9	6	935

We can see that the most common occurrences are some sort of OK and notes on checks. We can also check the official results on CRAN.

We can see that 0.55%, 0.54%, 0.16%, 0.06%, 0.05% of packages pass all checks without notes.

Now let’s see which of the notes or failures are due to intermittent issues.

Compare

First we need to make sure that we compare the right configurations. They must be the same machine, the same R version and the same svn revision between yesterday and today.

# Compare the previous flavor with today's
m_flavor <- which(flavors_df$flavors %in% old_flavors$flavors)
m_version <- which(flavors_df$version %in% old_flavors$version)
m_revision <- which(flavors_df$revision %in% old_flavors$revision)
tm <- table(c(m_flavor, m_version, m_revision))
compare <- flavors_df$flavors[tm == 3] # Only missing the packages version

All changes

Next, compare the status of the packages if the version of the package is the same.

# Find package on the flavors to compare that haven't changed versions
library("dplyr")
tcc <- filter(tc, Flavor %in% compare) |> 
  select(Flavor, Package, Version, Status) |> 
  arrange(Flavor, Package)
ycc <- filter(yc, Flavor %in% compare) |> 
  select(Flavor, Package, Version, Status) |> 
  arrange(Flavor, Package)

all_checks <- merge(tcc, ycc, by = c("Flavor", "Package"), 
                    suffixes = c(".t", ".y"), all = TRUE) 

possible_packages <- all_checks |> 
  filter(Version.t == Version.y & # Same version
           Status.t != Status.y & # Different status
           !is.na(Status.y) & # No new version or removed package
           !is.na(Status.t)) |> 
  rename(Today = Status.t, Yesterday = Status.y)
possible_packages |> 
  select(Package, Flavor, Today, Yesterday, -Version.t, -Version.y) |> 
  arrange(Package, Flavor) |> 
  flextable() |> 
  autofit()

Package	Flavor	Today	Yesterday
BayesChange	r-patched-linux-x86_64	ERROR	OK
BoomSpikeSlab	r-devel-linux-x86_64-fedora-clang	NOTE	OK
BoomSpikeSlab	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
CohortCharacteristics	r-devel-linux-x86_64-fedora-clang	NOTE	ERROR
CohortCharacteristics	r-devel-linux-x86_64-fedora-gcc	NOTE	ERROR
DataGraph	r-devel-linux-x86_64-fedora-clang	NOTE	OK
DataGraph	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
ade4	r-patched-linux-x86_64	WARNING	OK
cirls	r-patched-linux-x86_64	OK	ERROR
codebook	r-devel-linux-x86_64-fedora-gcc	OK	ERROR
curl	r-patched-linux-x86_64	ERROR	NOTE
epidatr	r-patched-linux-x86_64	ERROR	OK
exams.forge	r-patched-linux-x86_64	FAILURE	OK
httr2	r-devel-linux-x86_64-fedora-clang	NOTE	ERROR
httr2	r-devel-linux-x86_64-fedora-gcc	NOTE	ERROR
httr2	r-patched-linux-x86_64	OK	ERROR
huxtable	r-patched-linux-x86_64	ERROR	OK
iotarelr	r-patched-linux-x86_64	ERROR	OK
localScore	r-patched-linux-x86_64	OK	ERROR
maths.genealogy	r-devel-linux-x86_64-fedora-clang	ERROR	OK
maths.genealogy	r-devel-linux-x86_64-fedora-gcc	ERROR	OK
metaConvert	r-patched-linux-x86_64	OK	NOTE
modelbased	r-patched-linux-x86_64	ERROR	OK
oncoPredict	r-devel-linux-x86_64-fedora-clang	ERROR	OK
orcamentoBR	r-patched-linux-x86_64	OK	ERROR
paws.analytics	r-devel-linux-x86_64-fedora-clang	ERROR	OK
rempsyc	r-patched-linux-x86_64	OK	ERROR
restez	r-patched-linux-x86_64	ERROR	OK
scSpatialSIM	r-devel-linux-x86_64-fedora-gcc	ERROR	OK
teal.modules.clinical	r-oldrel-macos-arm64	FAILURE	ERROR
tidyr	r-devel-linux-x86_64-fedora-clang	NOTE	ERROR
tidyr	r-devel-linux-x86_64-fedora-gcc	NOTE	ERROR

If the machine and R versions is the same but the check of the package is different there might be some discrepancy between the dependencies.

# Extract dependencies
dependencies <- package_dependencies(unique(possible_packages$Package),
                                     # Should it check all the recursive dependencies or only direct?
                                     db = ap, # Only considering those dependencies on CRAN and Bioconductor but not any Additional_repositories. 
                                     recursive = TRUE, 
                                     which = c("Depends", "Imports", "LinkingTo", "Suggests"))

# Prepare to compare versions (as they are sorted by everything else we can compare directly)
intermittent_failures <- rep(FALSE, length(dependencies))
names(intermittent_failures) <- names(dependencies)
dep_0 <- lengths(dependencies) == 0
intermittent_failures[dep_0] <- TRUE

If they do not have any recursive dependency on Depends, Imports, LinkingTo and Suggests they might be have some intermittent problems on the packages. These is only on dependencies on CRAN and Bioconductor but not in other additional repositories (There are 160 packages with additional repositories).

If they have some dependencies and those dependencies didn’t change as far as we can tell then there might be some problems with random numbers or connectivity.

for (pkg in names(intermittent_failures[!intermittent_failures])) {
  dep <- dependencies[[pkg]]
  fl <- possible_packages$Flavor[possible_packages$Package == pkg]
  intermittent_failures[pkg] <- all_checks |> 
    filter(Package %in% dep,
           Flavor %in% fl,
           Version.t == Version.y,
           Status.t != Status.y) |> 
    nrow() == 0 # If packages outside || any(!dep %in% rownames(ap)) 
}
packages <- names(intermittent_failures)[intermittent_failures]

We finally show the differences on the status of those without any dependency change on version or status¹:

keep_files <- filter(possible_packages, Package %in% packages) |> 
  merge(y = flavors_df, by.x = "Flavor", by.y = "flavors", all.x = TRUE, all.y = FALSE) |> 
  select(Package, Flavor, Version = Version.t, R_version = r_version, OS = os, 
         architecture, other, version, revision) |> 
  mutate(Date = Sys.time())

if (nrow(keep_files >= 1)) {
  write.csv(keep_files, 
            paste0("cran-failing-", format(Sys.time(), "%Y%m%dT%H%M"), ".csv"),
            row.names = FALSE,
            quote = FALSE,
  )
}

filter(possible_packages, Package %in% packages) |> 
  select(Package, Flavor, Today, Yesterday, -Version.t, -Version.y) |> 
  flextable() |> 
  autofit()

Package	Flavor	Today	Yesterday
teal.modules.clinical	r-oldrel-macos-arm64	FAILURE	ERROR

Conclusion

cat("There are no packages detected with differences between yesterday and today attributable to intermittent failures.\n")
knitr::knit_exit()

cat("This suggests that these packages might have some problems with random numbers or connectivity:\n\n")

This suggests that these packages might have some problems with random numbers or connectivity:

if (any(dep_0)) {
  cat("\n## Packages with dependencies\n\n")
  cat(paste0(" - ", sort(intersect(packages, 
                                   names(dependencies)[dep_0])), "\n"), sep = "")
  cat("\n## Packages without dependencies\n\n")
  cat(paste0(" - ", sort(intersect(packages,
                                   names(dependencies)[!dep_0])), "\n"), sep = "")
  
} else {
  cat(paste0(" - ", sort(packages), "\n"), sep = "")
}

teal.modules.clinical

I think a new version might not propagate to check other packages until 24 hours later as checks might have already started for that day.↩︎

Packages intermittent failtures

Lluís Revilla Sancho

2025-03-05