r_make() does not pick up changes in from knitr_in()

[pat-s]

Prework

• Read and abide by drake's code of conduct.
• Search for duplicates among the existing issues, both open and closed.
• Advanced users: verify that the bug still persists in the current development version (i.e. remotes::install_github("ropensci/drake")) and mention the SHA-1 hash of the Git commit you install.

Description

When changing the dependencies of an Rmarkdown report tracked with knitr_in(), r_make() (and the manual way sourcing the config) sometimes do not pick up the following changes to dependencies loaded with drake::loadd():

• changes to object names
• commented/uncommented objects

This sounds in the first place like a trigger related problem.
However, this happens in an inconsistent fashion, i.e. the behavior is not persistent across the above mentioned actions.
Neither it is persistent across R sessions.
I've tried to find a reproducible pattern but was not successful.

A reprex seems to be complicated here since I can't interactively change the object name and then source the config again.
It could be related to triggers?
But I would expect that changing object names of dependencies of a report file changes the hash of the drake target.

Reproducible example

Not available.

Environment

Tested on v7.2.0 but I see no possible related changes to knitr_in() in the latest release.

Expected output

Changes of dependency object names in Rmarkdown reports get picked up by drake.

[wlandau]

I think this is because the processing steps in drake_config() are memoized (see create_drake_layout()). What this means is that some dependency checks are skipped if drake thinks they do not need to be repeated. If I am right, all we need to do is memoize against the relevant knitr / R Markdown source files. I will take a look tonight.

[wlandau]

A reprex:

library(drake)
  plan <- drake_plan(
  a = TRUE,
  b = TRUE,
  report_step = knitr_in("report.Rmd")
)

# Two versions of the report.
lines_a <- c(
  "---",
  "title: abc",
  "---",
  "",
  "```{r}",
  "readd(a)",
  "```"
)
lines_b <- c(
  "---",
  "title: abc",
  "---",
  "",
  "```{r}",
  "readd(b)",
  "```"
)

# Compute the DAG on the first report.
writeLines(lines_a, "report.Rmd")
cache <- storr::storr_environment()
for (i in 1:2) {
  config <- drake_config(
    plan,
    cache = cache,
    session_info = FALSE
  )
}
vis_drake_graph(config)

# drake does not recompute the DAG
# when the report has different dependencies.
writeLines(lines_b, "report.Rmd")
config <- drake_config(
  plan,
  cache = cache,
  session_info = FALSE
)

# The report step should depend on `b`, not `a`.
vis_drake_graph(config)

^{Created on 2019-05-27 by the reprex package (v0.3.0)}

[wlandau]

Yup, it was the memoization. Patch coming soon.

[wlandau]

Now fixed.

[pat-s]

Thanks for the instant fix, works like a charm!
Impressive that you even came up with a reprex for this non-trivial case. 🚀

[wlandau]

Thanks.

I may have actually introduced a bug, so I am reopening this. I need to check what happens if the knitr files are missing.

[wlandau]

Reproduced, fixed, and tested in 8adaee9.