Bug (?) in vignette handling

Am 30.10.21 um 20:28 schrieb Duncan Murdoch:

On 29/10/2021 5:52 a.m., Martin Maechler wrote:

Duncan Murdoch
 ???? on Thu, 28 Oct 2021 13:18:54 -0400 writes:

 ???? > This StackOverflow post:
https://stackoverflow.com/q/69756236/2554330
 ???? > points out that objects created in one vignette are available
in a later
 ???? > vignette.? I don't think this should be happening:? vignettes
should be
 ???? > self-contained.

I strongly agree.

 ???? > The current answer there,
https://stackoverflow.com/a/69758025/2554330,
 ???? > suggests that "R CMD check" will detect this.? However,
sometimes one
 ???? > vignette can replace a standard function with a custom version,
and then
 ???? > both will work without generating an error, but the second
vignette
 ???? > won't do the same thing if run independently.

 ???? > For example, try these pure Sweave vignettes:

 ???? > -------------------------
 ???? > aaa3.Rnw:
 ???? > -------------------------
 ???? > \documentclass{article}
 ???? > %\VignetteIndexEntry{Sweave aaa3}
 ???? > \begin{document}

 ???? > <<>>=
 ???? > mean <- function(x) "I am the Sweave mean"
 ???? > @

 ???? > \end{document}

 ???? > ------------------------
 ???? > aaa4.Rnw:
 ???? > ------------------------

 ???? > \documentclass{article}
 ???? > %\VignetteIndexEntry{Sweave aaa4}
 ???? > \begin{document}

 ???? > <<>>=
 ???? > mean(1:5)
 ???? > @

 ???? > \end{document}

 ???? > Put these in a package, build and install the package, and
you'll see
 ???? > that the mean() function in aaa4.Rnw prints the result from the
 ???? > redefined mean in aaa3.Rnw.

Is it because R is *not* run with? --no-save --no-restore
accidentally?
Without looking, I would not expect that the vignettes are run
inside the same running R (even though that may speedup things)

I think for R CMD build they are run in one process, while for R CMD
check they are in separate processes.? R CMD build runs
tools::buildVignettes(), which runs code that's part of the vignette
build engine.

Thankfully R CMD check has been building the vignettes in separate R
processes already since R 3.6.0, so has hopefully identified most
problems until now. The corresponding env var is
_R_CHECK_BUILD_VIGNETTES_SEPARATELY_.

The standard (and exported!) buildVignettes() has been weaving all
vignettes in the same session ever since it was added back in 2002. This
approach is probably more efficient (avoiding repetitive package
loading), but carry-over effects seem both likely and undesirable
(thinking of vignettes as separate and independently reproducible
manuscripts about different aspects of a package). AFAICS, it is not
explicitly documented that buildVignettes() runs all vignettes in the
same R session, so at least this is no advertised feature.

The Sweave engine evaluates things in .GlobalEnv, so any leftover
objects will be visible there for the next vignette.? I think it's up to
the writer of each vignette engine whether there's any cleanup, but it
appears that neither Sweave nor knitr does any.

I think this is by design and also useful in interactive sessions to
investigate the environment after weaving.

One possible fix would be for buildVignettes() to make a snapshot of
what's in .GlobalEnv before processing any vignettes, and restoring it
after each one.? I've tried a weaker version of this:? it records the
names in .GlobalEnv at the start, and deletes anything new before
processing each vignette.? So vignettes could modify or delete what's
there, but not add anything.

I think you don't want to completely clear out .GlobalEnv, because
people might choose to run buildVignettes() in an R session and expect
the vignettes to see the contents there.

"make check" in R-devel doesn't complain about this change, but I'll let
R Core decide whether it's a good idea or not.? A patch is below.

Clearing the workspace would be an improvement, but I think it would be
even better for R CMD build to produce each vignette in a clean R
session, especially with regard to loaded packages. Changing
buildVignettes() to use clean R processes by default (I'd say even if
there is only one vignette) should be considered. I'd appreciate seeing
this report in Bugzilla to investigate further (and not forget).