cat() in system.time() ?

On Fri, 15 Jul 2016, Martin Maechler wrote:

Hi Ben (and everyone else),

as this did not attract attention yet, let me start

Ben Bolker <bbolker at gmail.com>
    on Mon, 4 Jul 2016 11:49:40 -0400 writes:

   > Does anyone know if there's a reason that proc.time() uses cat()
   > rather than message() to print the output when there has been an

error

   > in the process of timing?

This is really not about proc.time(), but  about  system.time()
[ and I have corrected the  'Subject' accordingly ] ..

   > line 31 of time.R,

   >

https://github.com/wch/r-source/blob/e5b21d0397c607883ff25cca379687b86933d730/src/library/base/R/time.R#L31

   > on.exit(cat("Timing stopped at:", ppt(proc.time() - time), "\n"))

   > This means that as far as I can tell the general way to make sure
   > there is no output from a timed statement is ...

   > tt1 <- capture.output(tt0 <- suppressMessages(suppressWarnings(
   >                    try(<stuff to try>, silent=TRUE))))

   > (I know I could/should be using tryCatch() instead of try(), but

I don't

   > think it really matters here ... ?)

   > What would people think of a request to change this to message()
   > rather than cat() in the future ... ?  (This would mess up code

that is

   > already using capture.output() to store this information ...)

[I think that (last issue) would be acceptable.]

One reason of the current cat() may just be historical:
message() did not exist yet when  system.time()  was created.
However, I agree that that is not good enough a reason to keep
it. Much more important is the fact that it is *nice* that the
message

  Timing stopped at: ...

is printed in many cases when a system.time()d call is stopped
early.  Quite often for me this is *not* when an error happens
as your suppress*() contortions (;-) suggest, but rather when I
interrupt the long lasting call.
And the current setup nicely gives

  > i <- 0; system.time( while(TRUE) i <- i+1 )

    C-c C-c
  Timing stopped at: 1.001 0.084 1.086

  >

However, at least this simple case, also works fine with
message() instead of cat() ... as I just tried now.

A harder case are the "bad errors", e.g.,  memory overflow and
similar bad things...
cat() seems to be pretty robust, where as message() does invoke
a handler (exactly *why* you want it, right?) and that may be
harder to keep working correctly after certain errors than a
simple cat().

I hope that some real experts (on "context switching", "long
jumps", etc) would chime in now.

I don't have a strong opinion on whether message would be better or
not.  But now that the on.exit code is executed after unwinding the C
call back to the point of the system.time call there is no longer such
a robustness issue.

Best,

luke

Martin