Skip to content

save() saves extra stuff if object is not evaluated

3 messages · Henrik Bengtsson, Luke Tierney

Original
Henrik Bengtsson
# 
Hi,

it looks like save() is saving all contents of the calling
environments if the object to be saved is *not* evaluated, although it
is not that simple either.  After many hours of troubleshooting, I'm
still confused.  Here is a reproducible example (also attached) with
output.  I let the code and the output talk for itself:

peek <- function(file, from=1, to=500) {
  cat("--------------------------------------\n")
  cat(sprintf("%s: %d bytes\n", file, file.info(file)$size))
  bfr <- suppressWarnings(readBin(file, what="character", n=to))
  bfr <- gsub("(\001|\002|\003|\004|\005|\016|\020|\036|\a|\n|\t)", "", bfr);
  bfr <- bfr[nchar(bfr) > 0];
  cat(bfr, sep="", "\n");
}

saveCache <- function(file, y, sources=NULL, eval=FALSE) {
  if (eval)
    dummy <- is.null(sources)
  base::save(file=file, sources, compress=FALSE)
}

aVariableNotSaved <- double(1e6)

main <- function() {
  # This 'big' variable is saved in case 1 below!
  big <- rep(letters, length.out=1e5)
  identifier <- "This string will be saved too!"

  y <- 1

  file <- "a.RData"
  saveCache(y, file=file)
  peek(file)

  file <- "a-eval.RData"
  saveCache(y, file=file, eval=TRUE)
  peek(file)

  file <- "b-noy.RData"
  saveCache(file=file)
  peek(file)

  file <- "b-noy-eval.RData"
  saveCache(file=file, eval=TRUE)
  peek(file)
}


# 1. Call saveCache() outside main()
eval(body(main))
# --------------------------------------
# a.RData: 238 bytes
# RDX2Xsources?filea.RData y?n $  n?$eval???n?
# --------------------------------------
# a-eval.RData: 58 bytes
# RDX2Xsources??
# --------------------------------------
# b-noy.RData: 230 bytes
# RDX2Xsources?file?b-noy.RData ?yv$  n?$eval???n?
# --------------------------------------
# b-noy-eval.RData: 58 bytes
# RDX2Xsources??

# 2. Call saveCache() from within main()
main()
# --------------------------------------
# a.RData: 900412 bytes
# RDX2Xsources?filea.RData y? a.RData ?=identifierThis
# string will be saved too!big??abcdefghijklmnopqrstuv
# wxyzabcdefghijklmnopqrstuvwxyzabcdefg
# --------------------------------------
# a-eval.RData: 58 bytes
# RDX2Xsources??
# --------------------------------------
# b-noy.RData: 230 bytes
# RDX2Xsources?file?b-noy.RData ?yv$  n?$eval???n?
# --------------------------------------
# b-noy-eval.RData: 58 bytes
# RDX2Xsources??

What is going on?

I get this on both R v2.3.0 patched (2006-04-28 r37936) and R v2.3.1
beta (2006-05-23 r38179) on my WinXP (with Rterm --vanilla).
Luke Tierney
#
On Thu, 25 May 2006, Henrik Bengtsson wrote:

            

      
      
      
    

        
No, it's exactly that simple.  Serialization follows and writes out
all reachable environments.  Unevaluated promises contain the
environments in which their evaluations are to occur; evaluated ones
have this field set to R_NilValue to eliminate this no longer needed
reference.

There are two environments involved: the calling environment in which
saveCache is called and the callee environment of the call to
saveCache where the body of saveCache is evaluated.  Because of
lexical scope the enclosing environment of the callee environment is
the closure environment of saveCache, which is .GlobalEnv.

The call to saveCache creates a promise for evaluating the default
value for 'source' _in the callee environment_. In the case with y the
callee environment includes a value of y which is a promise
referencing the calling environment (either .GlobalENv or the
environment of the call to main).  In the calls without y the value of
y in the calling environment is the missing value indicator, not a
promise.  So only with y and no eval is there a reference to the
calling environment that serialization then has to write out.

Best,

luke

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Henrik Bengtsson
    

    

      
      #
    

  


  

      
On 5/25/06, Luke Tierney <luke at stat.uiowa.edu> wrote:

            

      
      
      
    

        
Thank you very much for this sharp explanation.  It is now much
clearer to me what is going on.  Would it make sense to make save()
evaluate all non-evaluated arguments, e.g. is.null(list(...))?  ...or,
add an argument making this optional/default?

Best wishes,

Henrik