Packages intermittent failtures

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		r89406
r-devel-linux-x86_64-debian-gcc	devel	linux	x86_64	debian-gcc		r89414
r-devel-linux-x86_64-fedora-clang	devel	linux	x86_64	fedora-clang		r89372
r-devel-linux-x86_64-fedora-gcc	devel	linux	x86_64	fedora-gcc		r89380
r-devel-macos-arm64	devel	macos	arm64			r89366
r-patched-linux-x86_64	patched	linux	x86_64		4.5.2	r89392
r-release-linux-x86_64	release	linux	x86_64		4.5.2
r-release-macos-arm64	release	macos	arm64		4.5.2
r-release-macos-x86_64	release	macos	x86_64		4.5.1
r-release-windows-x86_64	release	windows	x86_64		4.5.2
r-oldrel-macos-arm64	oldrel	macos	arm64		4.4.3
r-oldrel-macos-x86_64	oldrel	macos	x86_64		4.4.1
r-oldrel-windows-x86_64	oldrel	windows	x86_64		4.4.3

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())

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 68 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())

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 330 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())

Distribution of total time on each machine.

This means that to install and check all the packages with a single CPU would take 410 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 246.15 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	22,830
2	354
3	2

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
14	23,040
11	43
13	35
12	27
3	12
6	8
8	6
9	6
10	5
7	4

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())

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())

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	WARNING	ERROR	FAILURE	packages
14	0	0	0	0	13,148
12	2	0	0	0	4,430
11	3	0	0	0	1,487
0	14	0	0	0	1,112
9	5	0	0	0	796
10	4	0	0	0	318

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 0.79%, 0.66%, 0.06%, 0.04%, 0.00% 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
BWStest	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
DIscBIO	r-oldrel-windows-x86_64	ERROR	OK
FLORAL	r-oldrel-windows-x86_64	NOTE	ERROR
Grouphmap	r-oldrel-windows-x86_64	OK	ERROR
HMDA	r-devel-linux-x86_64-fedora-gcc	WARNING	OK
HMDA	r-release-linux-x86_64	WARNING	OK
IVPP	r-oldrel-windows-x86_64	OK	ERROR
IVPP	r-release-linux-x86_64	OK	ERROR
Luminescence	r-oldrel-windows-x86_64	ERROR	NOTE
MASS	r-oldrel-windows-x86_64	OK	ERROR
MarkowitzR	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
MexicoDataAPI	r-oldrel-windows-x86_64	OK	ERROR
MixStable	r-release-linux-x86_64	ERROR	OK
NetworkComparisonTest	r-oldrel-windows-x86_64	OK	ERROR
NetworkComparisonTest	r-release-linux-x86_64	OK	ERROR
OmopViewer	r-oldrel-windows-x86_64	OK	ERROR
PDQutils	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
QuickJSR	r-oldrel-windows-x86_64	NOTE	ERROR
REDCapExporter	r-oldrel-windows-x86_64	ERROR	OK
REDCapExporter	r-release-linux-x86_64	ERROR	OK
RcppParallel	r-oldrel-windows-x86_64	NOTE	ERROR
RcppThread	r-release-linux-x86_64	ERROR	OK
SharpeR	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
SimMultiCorrData	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
SuperLearner	r-oldrel-windows-x86_64	FAILURE	OK
TCGAretriever	r-release-linux-x86_64	OK	ERROR
TaxaNorm	r-oldrel-windows-x86_64	OK	ERROR
attrib	r-release-linux-x86_64	ERROR	OK
box.lsp	r-release-linux-x86_64	OK	ERROR
camtrapR	r-devel-linux-x86_64-fedora-gcc	OK	NOTE
cli	r-oldrel-windows-x86_64	OK	ERROR
cocktailApp	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
coreheat	r-oldrel-windows-x86_64	NOTE	ERROR
curl	r-oldrel-windows-x86_64	NOTE	ERROR
dplyr	r-oldrel-windows-x86_64	ERROR	OK
epsiwal	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
filibustr	r-oldrel-windows-x86_64	OK	ERROR
fromo	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
gap	r-devel-linux-x86_64-fedora-gcc	ERROR	OK
ggallin	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
git2rdata	r-release-linux-x86_64	OK	ERROR
glue	r-oldrel-windows-x86_64	OK	ERROR
gtools	r-devel-linux-x86_64-fedora-gcc	OK	NOTE
gtools	r-oldrel-windows-x86_64	OK	ERROR
gtools	r-release-linux-x86_64	NOTE	OK
jsonlite	r-oldrel-windows-x86_64	OK	ERROR
klassR	r-oldrel-windows-x86_64	OK	ERROR
lessSEM	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
madness	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
magrittr	r-oldrel-windows-x86_64	OK	ERROR
matrixStats	r-oldrel-windows-x86_64	OK	ERROR
mazealls	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
mixKernel	r-oldrel-windows-x86_64	OK	ERROR
mlr3viz	r-oldrel-windows-x86_64	ERROR	NOTE
networkDynamic	r-oldrel-windows-x86_64	NOTE	OK
networkDynamic	r-release-linux-x86_64	NOTE	OK
nnmf	r-release-linux-x86_64	ERROR	OK
occCite	r-devel-linux-x86_64-fedora-gcc	OK	ERROR
occCite	r-release-linux-x86_64	ERROR	OK
ohenery	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
optimall	r-release-linux-x86_64	ERROR	OK
osmapiR	r-oldrel-windows-x86_64	ERROR	OK
osmapiR	r-release-linux-x86_64	ERROR	OK
otelsdk	r-devel-linux-x86_64-fedora-gcc	NOTE	OK
photon	r-oldrel-windows-x86_64	ERROR	OK
photon	r-release-linux-x86_64	ERROR	OK
pkgdiff	r-release-linux-x86_64	OK	ERROR
pmparser	r-oldrel-windows-x86_64	ERROR	OK
pmparser	r-release-linux-x86_64	ERROR	OK
popdemo	r-release-linux-x86_64	ERROR	OK
powerly	r-oldrel-windows-x86_64	OK	ERROR
powerly	r-release-linux-x86_64	OK	ERROR
rbrsa	r-oldrel-windows-x86_64	ERROR	OK
rbrsa	r-release-linux-x86_64	ERROR	OK
restfulr	r-oldrel-windows-x86_64	OK	ERROR
rlang	r-oldrel-windows-x86_64	OK	ERROR
rodeo	r-oldrel-windows-x86_64	ERROR	OK
scPOEM	r-oldrel-windows-x86_64	OK	ERROR
simRestore	r-release-linux-x86_64	ERROR	OK
sovereign	r-release-linux-x86_64	OK	ERROR
spinebil	r-oldrel-windows-x86_64	ERROR	OK
statnet	r-release-linux-x86_64	OK	NOTE
tibble	r-oldrel-windows-x86_64	ERROR	OK
tidyREDCap	r-release-linux-x86_64	ERROR	OK
winch	r-oldrel-windows-x86_64	ERROR	WARNING
xfun	r-oldrel-windows-x86_64	OK	ERROR
xml2	r-oldrel-windows-x86_64	OK	ERROR
yaml	r-oldrel-windows-x86_64	OK	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 185 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()

Conclusion

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

There are no packages detected with differences between yesterday and today attributable to intermittent failures.

knitr::knit_exit()

---

1. 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.↩︎