"type_sum not exported" error on Windows with R 3.6

[gadenbuie]

I'm seeing a repeat of r-lib/pillar#462 but very specifically just on Windows with R 3.6.3. I've traced it to rlang::abort(), which calls pillar::type_sum() inside rlang:::trace_back(). I understand rlang and pillar are quite entangled in this area, so please let me know if I should move this issue to or open a complementary issue in rlang.

https://github.com/rstudio/learnr/runs/5521901016?check_suite_focus=true#step:7:210

== Failed tests ================================================================
-- Error (test-exercise.R:109:1): render_exercise() returns envir_result up to error --
Error: Error: 'type_sum' is not an exported object from 'namespace:pillar'
-- Error (test-knitr-hooks.R:31:1): Detection of chained setup cycle works -----
Error: Error: 'type_sum' is not an exported object from 'namespace:pillar'
-- Error (test-options-reveal_solution.R:32:1): Solutions are revealed or hidden with tutorial_options() --
Error: Error: 'type_sum' is not an exported object from 'namespace:pillar'
-- Error (test-options-reveal_solution.R:52:1): Solutions are revealed or hidden with global option --
Error: Error: 'type_sum' is not an exported object from 'namespace:pillar'

Here's a concise summary of the session info in the runner.

  - Session info ---------------------------------------------------------------
   setting  value
   version  R version 3.6.3 (2020-02-29)
   os       Windows Server x64 (build 20348)

  - Packages -------------------------------------------------------------------
   package      * version    date (UTC) lib source
   pillar         1.7.0      2022-02-01 [1] RSPM
   rlang          1.0.2      2022-03-04 [1] RSPM

[krlmlr]

Thanks. Does this come from rlang_type_sum() ? Should we check that the function exists to avoid this behavior?

rlang/R/types.R

Lines 474 to 482 in 22fe9e9

	# Use different generic name to avoid import warnings when loading
	# packages that import all of rlang after it has been load_all'd
	rlang_type_sum <- function(x) {
	if (is_installed("pillar")) {
	pillar::type_sum(x)
	} else {
	UseMethod("rlang_type_sum")
	}
	}

[lionel-]

Maybe is_installed() should check that the package is fully loaded?

We ran into similar issues before when loadNamespace() is called recursively with user hooks involved. Not sure how that could come into play here though.

[lionel-]

Hopefully the attached commit fixes this. I'll also try to send a patch for base R when I get the time.