Auto-entrace errors

[hadley]

No description provided.

[cderv]

I have looked into this traceback() change we identified. Here are some notes so that we can decide best way to adapt.

I found that new withRestarts() logic has a few implications

    continue <- withRestarts(
      with_handlers(
        {
          for (expr in tle$exprs) {
            ev <- withVisible(eval(expr, envir))
            watcher$capture_plot_and_output()
            watcher$print_value(ev$value, ev$visible, envir)
          }
          TRUE
        },
        handlers
      ),
      eval_continue = function() TRUE,
      eval_stop = function() FALSE,
      eval_error = signal_error
    )

In there, handlers needs to invoke restarts, otherwise it will skip the logic here. This is what happen with rlang::global_entrace() which adds a user_handler doing rlang::entrace(). This leads to continue being NULL, and it throws an error.

By commenting # rlang::global_entrace(), , the result will depends on rlang::abort() or stop() being used due to signal_error calling signalConditon():

• When using rlang::abort(), cnd is

   > str(cnd)
   List of 5
    $ message: chr "!"
    $ trace  :Classes 'rlang_trace', 'rlib_trace', 'tbl' and 'data.frame':	4 obs. of  6 variables:
     ..$ call       :List of 4
     .. ..$ : language f()
     .. ..$ : language g()
     .. ..$ : language h1()
     .. ..$ : language rlang::abort("!")
     ..$ parent     : int [1:4] 0 1 2 3
     ..$ visible    : logi [1:4] TRUE TRUE TRUE FALSE
     ..$ namespace  : chr [1:4] NA NA NA "rlang"
     ..$ scope      : chr [1:4] "global" "global" "global" "::"
     ..$ error_frame: logi [1:4] FALSE FALSE TRUE FALSE
     ..- attr(*, "version")= int 2
    $ parent : NULL
    $ rlang  :List of 1
     ..$ inherit: logi TRUE
    $ call   : language h1()
    - attr(*, "class")= chr [1:3] "rlang_error" "error" "condition"

  and signalCondition() will run and not stop

• When using stop(), cnd is

   > str(cnd)
   List of 2
    $ message: chr "!"
    $ call   : language h2()
    - attr(*, "class")= chr [1:3] "simpleError" "error" "condition"

  in that case, it will stop there - it seems signalCondition(cnd) wil l call stop(). Not sure how to find which handler it is calling exactly.

This happens in RStudio IDE context. To note that in this context, the special IDE setting comes also into the mix with

> getOption("error")
(function () 
{
    .rs.recordTraceback(TRUE, 5, .rs.enqueueError)
})()

when signalCondition() is called and no more handlers. From ?conditions

globalCallingHandlers establishes calling handlers globally. These handlers are only called as a last resort, after the other handlers dynamically registered with withCallingHandlers have been invoked. They are called before the error global option (which is the legacy interface for global handling of errors).

I'll try to see difference in other context.

More generally, with this continue <- withRestarts(), I believe handlers needs to return TRUE or FALSE, or really stop() or abort() effectively.

I'll keep looking.

[cderv]

From my digging into the situation, I believe rlang::entrace() won't throw an error in some situation (when this is already a rlang_error)
From rlang::entrace()

  if (!missing(cnd) && inherits(cnd, "rlang_error")) {
    poke_last_error(cnd)
    return()
  }

And so with the example of using global_entrace(), this will error because signal_error() returns NULL

This document with this PR version of evaluate

```{r}
rlang::global_entrace()
options(rlang_backtrace_on_error_report = "full")
```

```{r}
f <- function() g()
g <- function() h2()
h1 <- function() rlang::abort("!")
h2 <- function() stop("!")
f()
```

throws Error in `!continue`:

processing file: test.Rmd
  |....................................................................................................................................................| 100% [unnamed-chunk-3]Error in `!continue`:
! invalid argument type
Backtrace:
  1. xfun::in_dir("../rmarkdown", rmarkdown::render("test.Rmd"))
  2. rmarkdown::render("test.Rmd")
  3. knitr::knit(knit_input, knit_output, envir = envir, quiet = quiet)
  4. knitr:::process_file(text, output)
  7. knitr:::process_group(group)
  8. knitr:::call_block(x)
  9. knitr:::block_exec(params)
 10. knitr:::eng_r(options)
 13. knitr (local) evaluate(...)
 14. evaluate::evaluate(...)

Quitting from lines 17-22 [unnamed-chunk-3] (test.Rmd)

[cderv]

I wonder if this should directly by rlang::abort() instead of rlang::entrace() so that it really stops with traceback printed.

This is similar to what have been discussed in #226

[hadley]

You can't pass an existing condition object to abort() like you can with stop(). My read of the rlang docs does suggest that this is what entrace() does, so I think we'll need to get @lionel- to help.

[hadley]

I've updated with something I tried locally that should at least avoid the !continue problem.