pkgdown consumed a huge amount (10+GB) of RAM on build_site()

[Robinlovelace]

I'm not sure how or why this happened but I'm confident it's pkgdown because I tried it twice (both times it crashed my R session).

Here's the crash message:

<simpleError in process_initialize(self, private, command, args, stdin, stdout,     stderr, connections, poll_connection, env, cleanup, cleanup_tree,     wd, echo_cmd, supervise, windows_verbatim_args, windows_hide_window,     encoding, post_process): Cannot fork Unknown error -12 at unix/processx.c:333>
Warning message:
system call failed: Cannot allocate memory 
Error in process_initialize(self, private, command, args, stdin, stdout,  : 
  Cannot fork Unknown error -12 at unix/processx.c:333

Here's the dramatic memory use log on Ubuntu:

With:

packageVersion("pkgdown")
#> [1] '1.1.0'

Created on 2018-08-13 by the reprex package (v0.2.0).

[Robinlovelace]

Live update: now trying with

pkgdown::build_site(new = F)

In a terminal to see if that solves it - seems to be something to do with running in parallel. Update - crazy RAM use happened again - didn't solve it.

[Robinlovelace]

In the name of reproducibility, I think others will be able to reproduce this by cloning this repo and running build_site() from in there: https://github.com/geocompr/geocompkg

I suspect the culprit is https://github.com/geocompr/geocompkg/blob/master/vignettes/point-pattern-heatmap.Rmd which has some leaflet code.

[Robinlovelace]

Update on the last attempt run from a terminal with new_process = F - it was definitely the R session itself that was munching the session because when it stopped R with quit() this happened:

[Robinlovelace]

Another update: thinking maybe running it on a bigger computer with a different OS might help I tried it on my work desktop and get an error - the error message should shed more light on it though, hope this helps:

[Robinlovelace]

In any case, for users, all the vignettes can be built by running build_articles(), that seems to run fine.

[Nowosad]

I also got this problem today. Vignette works well when knitted, but there is a memory leak while using build_site().

[jayhesselberth]

Could you please build a minimal package that recreates the issue? The geocompkg package has too many dependencies to debug this effectively.

[Robinlovelace]

I think it was the number of vignettes, not the dependencies of the package the website of which we were trying to build, that caused the issue. Agree hard to debug.

[Robinlovelace]

FYI for anyone else reading this issue: run

build_article()

one-by-one and it seems to be fine. All good!

[samvaltenbergs]

@Robinlovelace did you ever resolve this issue (other than having to build each article individually, which I don't want to do as part of our CI/CD pipeline)? I see @hadley closed the issue, but I don't see details of a resolution. I'm having the same issue: vignettes knit by themselves NP and I can build each article manually by itself, but when pkgdown::build_site() runs and builds all the articles at once it crashes because it "cannot allocate vector of size 11.0 Gb". The behavior is the same on my local machine in RStudio as it is in a Docker EE container as part of a CI/CD pipeline.

[Nowosad]

Hi @samvaltenbergs I think the problem is still there. We (@Robinlovelace and I) haven't found a solution yet..

[Robinlovelace]

To confirm: no I didn't find a solution. I'm not 100% sure what led to the issue being closed. Any info on that @hadley would be appreciated. I think one thing that would help developers would be reproducible examples.

[jayhesselberth]

If you are running very large vignettes it is going to be very difficult to debug this, and I would also advocate for breaking these up anyway (or at least trimming your example data sets down a bit).

If this is a bug, it's probably reproducible by a very short vignette.

[Robinlovelace]

Thanks for the feedback. Will try to create a reproducible example. Anyone else finding issues with memory usage: please add reproducible examples of what happened.

[Nowosad]

Hi @jayhesselberth, I spent an hour a few weeks trying to create a reproducible short example and I was unable to do so...

[hadley]

It was closed because we don’t have a reprex and it seems unlikely that the root cause is pkgdown.

[Robinlovelace]

I think it's possible to create a reproducible example. May be in bash or have lots of system() calls though. Will aim to do this when on a decent computer in a session I can afford to crash. Steps to reproduce as far as I remember:

git clone git@github.com:geocompr/geocompkg
cd geocompkg
R
pkgdown::build_site()

[samvaltenbergs]

@hadley, @Robinlovelace and @Nowosad I understand that it's difficult to troubleshoot without a reproducible example, but unfortunately I can't share the vignettes due to corporate policy (The Man)... I can show a screenshot that demonstrates the failure signature: building the articles by themselves works okay but when pkgdown tries to put the site together after all the vignettes are turned into HTML it hits a wall. And then if I immediately run the same build_articles() command (at which point the vignettes are already turned into HTML) everything is AOK.

What has me scratching my head is that the vignettes are all very simple: the only piece of R code that is executed across the seven vignettes is a call to Sys.Date(). That is, there isn't any R code being executed that could possibly create something in memory that's 11Gb in size.

What is even more strange is that the "11Gb failure" only occurs the first time build_articles() or build_site() is run for a given R session. If I delete the docs\articles folder (which will cause pkgdown to rebuild all the HTML vignettes) it runs without an issue. But if I then restart the R session and run it fresh, fails again...

Any ideas?

[Robinlovelace]

Reproducible example verified:

Heads-up @hadley and @jayhesselberth please try to reproduce.

[Robinlovelace]

Further evidence that this is an issue with build_site() and not the contents of the vignettes: when you build them individually, e.g. with the following commands, it works:

# build site:
pkgdown::build_home()

# build vignettes one by one, e.g.
pkgdown::build_article("sea-level-rise")

pkgdown::build_articles()
# build all articles
articles = list.files(path = "vignettes/", pattern = ".Rmd")
articles = gsub(pattern = ".Rmd", replacement = "", articles)
for(i in articles) {
  pkgdown::build_article(i)
}

[hadley]

@Robinlovelace that video doesn't really add anything. All build_articles() does is call build_article() with purrr::walk(), so it's already basically equivalent to your code. Including the contents of traceback() would be more useful.

[hadley]

I have started the process of installing all the geocompr dependencies so I can reproduce locally.

[Robinlovelace]

Sure. 1st time I've used the function. Hope the output is useful (means nothing to me)!

traceback()
5: stop(err[[2]])
4: get_result(output = out, options)
3: callr::r(function(..., crayon_enabled, crayon_colors, pkgdown_internet) {
       options(crayon.enabled = crayon_enabled, crayon.colors = crayon_colors, 
           pkgdown.internet = pkgdown_internet)
       pkgdown::build_site(...)
   }, args = args, show = TRUE, )
2: build_site_external(pkg = pkg, examples = examples, document = document, 
       run_dont_run = run_dont_run, seed = seed, lazy = lazy, override = override, 
       preview = preview)
1: pkgdown::build_site()

[Robinlovelace]

The good news: with the work-around I've managed to re-build and update the site: https://github.com/geocompr/geocompkg

Heads-up @samvaltenbergs that work-around #783 (comment) may be of use if build_site() is still playing-up for you.

[Robinlovelace]

btw @hadley apologies for all the dependencies.

[hadley]

After running pkgbuild::build_site(new_process = FALSE) I see:

> traceback()
10: knitr::knit_meta_add(old_knit_meta, attr(old_knit_meta, "knit_meta_id"))
9: render(input = md_file, output_file = html_file, output_format = override_output_format, 
       output_options = list(self_contained = FALSE), quiet = TRUE, 
       encoding = "UTF-8")
8: discover_rmd_resources(input_file, encoding, discover_single_resource)
7: rmarkdown::find_external_resources(input_path, "UTF-8") at rmarkdown.R#48
6: render_rmarkdown(pkg, input = input, output = output_file, output_format = format, 
       output_options = options, quiet = quiet) at build-articles.R#216
5: .f(.x[[i]], ...)
4: purrr::walk(pkg$vignettes$name, build_article, pkg = pkg, quiet = quiet, 
       lazy = lazy) at build-articles.R#133
3: build_articles(pkg, lazy = lazy, override = override, preview = FALSE) at build.r#367
2: build_site_local(pkg = pkg, examples = examples, document = document, 
       run_dont_run = run_dont_run, seed = seed, lazy = lazy, override = override, 
       preview = preview) at build.r#287
1: pkgdown::build_site(new_process = F)

[hadley]

If I run it again, it fails later:

Reading 'vignettes/solutions06.Rmd'
Error in CPL_geos_is_empty(st_geometry(x)) : 
  Evaluation error: IllegalArgumentException: Points of LinearRing do not form a closed linestring.

[Robinlovelace]

Thanks for reporting. I've raised an issue here: geocompx/geocompkg#2