Skip to content

Choices to remove `srcref` (and its buddies) when serializing objects

10 messages · Dipterix Wang, Tomas Kalibera, Lionel Henry +2 more

Original
Dipterix Wang
# 
Dear R devs,

I was digging into a package issue today when I realized R serialize function not always generate the same results on equivalent objects when users choose to run differently. For example, the following code

serialize(with(new.env(), { function(){} }), NULL, TRUE)

generates different results when I copy-paste into console vs when I use ctrl+shift+enter to source the file in RStudio. 

With a deeper inspect into the cause, I found that function and language get source reference when getOption("keep.source") is TRUE. This means the source reference will make the functions different while in most cases, whether keeping function source might not impact how a function behaves.

While it's OK that function serialize generates different results, functions such as `rlang::hash` and `digest::digest`, which depend on `serialize` might eventually deliver false positives on same inputs. I've checked source code in digest package hoping to get around this issue (for example serialize(..., refhook = ...)). However, my workaround did not work. It seems that the markers to the objects are different even if I used `refhook` to force srcref to be the same. I also tried `removeSource` and `rlang::zap_srcref`. None of them works directly on nested environments with multiple functions. 

I wonder how hard it would be to have options to discard source when serializing R objects? 

Currently my analyses heavily depend on digest function to generate file caches and automatically schedule pipelines (to update cache) when changes are detected. The pipelines save the hashes of source code, inputs, and outputs together so other people can easily verify the calculation without accessing the original data (which could be sensitive), or running hour-long analyses, or having to buy servers. All of these require `serialize` to produce the same results regardless of how users choose to run the code.

It would be great if this feature could be in the future R. Other pipeline packages such as `targets` and `drake` can also benefit from it.

Thanks,

- Dipterix
Ivan Krylov
# 
? Fri, 12 Jan 2024 00:11:45 -0500
Dipterix Wang <dipterix.wang at gmail.com> ?????:
Source references may be the main problem here, but not the only one.
There are also string encodings and function bytecode (which may or may
not be present and probably changes between R versions). I've been
collecting the ways that the objects that are identical() to each other
can serialize() differently in my package 'depcache'; I'm sure I missed
a few.

Admittedly, string encodings are less important nowadays (except on
older Windows and weirdly set up Unix-like systems). Thankfully, the
digest package already knows to skip the serialization header (which
contains the current version of R).

serialize() only knows about basic types [*], and source references are
implemented on top of these as objects of class 'srcref'. Sometimes
they are attached as attributes to other objects, other times (e.g. in
quote(function(){}), [**]) just sitting there as arguments to a call.

Sometimes you can hash the output of deparse(x) instead of serialize(x)
[***]. Text representations aren't without their own problems (e.g.
IEEE floating-point numbers not being representable as decimal
fractions), but at least deparsing both ignores the source references
and punts the encoding problem to the abstraction layer above it:
deparse() is the same for both '\uff' and iconv('\uff', 'UTF-8',
'latin1'): just "?".

Unfortunately, this doesn't solve the environment problem. For these,
you really need a way to canonicalize the reference-semantics objects
before serializing them without changing the originals, even in cases
like a <- new.env(); b <- new.env(); a$x <- b; b$x <- a. I'm not sure
that reference hooks can help with that. In order to implement it
properly, the fixup process will have to rely on global state and keep
weak references to the environments it visits and creates shadow copies
of.

I think it's not impossible to implement
serialize_to_canonical_representation() for an R package, but it will
be a lot of work to decide which parts are canonical and which should
be discarded.
Tomas Kalibera
#
On 1/12/24 06:11, Dipterix Wang wrote:
I don't think such functionality would belong to serialize(). This 
function is not meant to produce stable results based on the input, the 
serialized representation may even differ based on properties not seen 
by users.

I think an option to ignore source code would belong to a function that 
computes the hash, as other options of identical().

Tomas
4 days later
Dipterix Wang
# 
Could you recommend any packages/functions that compute hash such that the source references and sexpinfo_struct are ignored? Basically a version of `serialize` that convert R objects to raw without storing the ancillary source reference and sexpinfo.

I think most people would think of `digest` but that package uses `serialize` (see discussion https://github.com/eddelbuettel/digest/issues/200#issuecomment-1894289875)

  
  


  


      

    

    

      
      

        


  

        

      Tomas Kalibera
    

    

      
      #
    

  


  

      
On 1/16/24 20:16, Dipterix Wang wrote:

            

      
      
      
    

        
I think one could implement hashing on the fly without any 
serialization, similarly to how identical works, but I am not aware of 
any existing implementation. Again, if that wasn't clear: I don't think 
trying to compute a hash of an object from its serialized representation 
is a good idea - it is of course convenient, but has problems like the 
one you have ran into.

In some applications it may still be good enough: if by various tweaks, 
such as ensuring source references are off in your case, you achieve a 
state when false alarms are rare (identical objects have different 
hashes), and hence say unnecessary re-computation is rare, maybe it is 
good enough.

Tomas

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Lionel Henry
    

    

      
      #
    

  


  

          

      
      
      
    

        
We have one in vctrs but it's not exported:
https://github.com/r-lib/vctrs/blob/main/src/hash.c

The main use is vectorised hashing:

```
# Non-vectorised
vctrs:::obj_hash(1:10)
#> [1] 1e 77 ce 48

# Vectorised
vctrs:::vec_hash(1L)
#> [1] 70 a2 85 ef
vctrs:::vec_hash(1:2)
#> [1] 70 a2 85 ef bf 3c 2c cf

# vctrs semantics so dfs are vectors of rows
length(vctrs:::vec_hash(mtcars)) / 4
#> [1] 32
nrow(mtcars)
#> [1] 32
```

Best,
Lionel

On Wed, Jan 17, 2024 at 10:32?AM Tomas Kalibera

        
<tomas.kalibera at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Dipterix Wang
    

    

      
      #
    

  


  

          

      
      
      
    

        
Thanks for showing me this function. I have read the source code. That's a great idea. 

However, I think I might have missed something. When I tried vctrs::obj_hash, I couldn't get identical outputs.


``` r
options(keep.source = TRUE)
a <- function(){}
vctrs:::obj_hash(a)
#> [1] 68 e8 5a 0c
a <- function(){}
vctrs:::obj_hash(a)
#> [1] b2 6a 55 9c
a <-   function(){}
vctrs:::obj_hash(a)
#> [1] 01 a9 bc 30
options(keep.source = FALSE)
a <- function(){}
vctrs:::obj_hash(a)
#> [1] 93 d7 f2 72
a <- function(){}
vctrs:::obj_hash(a)
#> [1] f3 1d d2 f4
```

Created on 2024-01-17 with [reprex v2.1.0](https://reprex.tidyverse.org)

            

      
      
      
    

        
I really appreciate you answer my questions and solve my puzzles. I went back and read the R internal code for `serialize` and totally agree on this, that serialization is not a good idea for digesting R objects, especially on environments, expressions, and functions. 

What I want is a function that can produce the same and stable hash for identical objects. However, there is no function (given our best knowledge) on the market that can do this. `digest::digest` and `rlang::hash` are the first functions that come into my mind. Both are widely used, but they use serialize. The author of `digest` said:
	> "As you know,  digest takes and (ahem) "digests" what serialize gives it, so you would have to look into what serialize lets you do."

vctrs:::obj_hash is probably the closest to the implementation of `identical`, but the above examples give different results for identical objects.

The existence of digest:: digest and rlang::hash shows that there is a huge demand for this "ideal" hash function. However, I bet most people are using digest/hash "incorrectly".

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Ivan Krylov
    

    

      
      #
    

  


  

      
? Tue, 16 Jan 2024 14:16:19 -0500
Dipterix Wang <dipterix.wang at gmail.com> ?????:

            

      
      
      
    

        
I can show how this can be done, but it's not currently on CRAN or even
a well-defined package API. I have adapted a copy of R's serialize()
[*] with the following changes:

 * Function bytecode and flags are ignored:

f <- function() invisible()
depcache:::hash(f, 2) # This is plain FNV1a-64 of serialize() output
# [1] "9b7a1af5468deba4"
.Call(depcache:::C_hash2, f) # This is the new hash
[1] 91 5f b8 a1 b0 6b cb 40
f() # called once: function gets the MAYBEJIT_MASK flag
depcache:::hash(f, 2)
# [1] "7d30e05546e7a230"
.Call(depcache:::C_hash2, f)
# [1] 91 5f b8 a1 b0 6b cb 40
f() # called twice: function now has bytecode
depcache:::hash(f, 2)
# [1] "2a2cba4150e722b8"
.Call(depcache:::C_hash2, f)
# [1] 91 5f b8 a1 b0 6b cb 40 # new hash stays the same

 * Source references are ignored:

.Call(depcache:::C_hash2, \( ) invisible( ))
# [1] 91 5f b8 a1 b0 6b cb 40 # compare vs. above

# For quoted function definitions, source references have to be handled
# differently 
.Call(depcache:::C_hash2, quote(function(){}))
[1] 58 0d 44 8e d4 fd 37 6f
.Call(depcache:::C_hash2, quote(\( ){      }))
[1] 58 0d 44 8e d4 fd 37 6f

 * ALTREP is ignored:

identical(1:10, 1:10+0L)
# [1] TRUE
identical(serialize(1:10, NULL), serialize(1:10+0L, NULL))
# [1] FALSE
identical(
 .Call(depcache:::C_hash2, 1:10),
 .Call(depcache:::C_hash2, 1:10+0L)
)
# [1] TRUE

 * Strings not marked as bytes are encoded into UTF-8:

identical('\uff', iconv('\uff', 'UTF-8', 'latin1'))
# [1] TRUE
identical(
 serialize('\uff', NULL),
 serialize(iconv('\uff', 'UTF-8', 'latin1'), NULL)
)
# [1] FALSE
identical(
 .Call(depcache:::C_hash2, '\uff'),
 .Call(depcache:::C_hash2, iconv('\uff', 'UTF-8', 'latin1'))
)
# [1] TRUE

 * NaNs with different payloads (except NA_numeric_) are replaced by
   R_NaN.

One of the many downsides to the current approach is that we rely on
the non-API entry point getPRIMNAME() in order to hash builtins.
Looking at the source code for identical() is no help here, because it
uses the private PRIMOFFSET macro.

The bitstream being hashed is also, unfortunately, not exactly
compatible with R serialization format version 2: I had to ignore the
LEVELS of the language objects being hashed both because identical()
seems to ignore those and because I was missing multiple private
definitions (e.g. the MAYBEJIT flag) to handle them properly.

Then there's also the problem of immediate bindings [**]: I've seen bits
of vctrs, rstudio, rlang blow up when calling CAR() on SEXP objects that
are not safe to handle this way, but R_expand_binding_value() (used by
serialize()) is again a private function that is not accessible from
packages. identical() won't help here, because it compares reference
objects (which may or may not contain such immediate bindings) by their
pointer values instead of digging down into them.

Dropping the (already violated) requirement to be compatible with R
serialization bitstream will make it possible to simplify the code
further.

Finally:

a <- new.env()
b <- new.env()
a$x <- b$x <- 42
identical(a, b)
# [1] FALSE
.Call(depcache:::C_hash2, a)
# [1] 44 21 f1 36 5d 92 03 1b
.Call(depcache:::C_hash2, b)
# [1] 44 21 f1 36 5d 92 03 1b

...but that's unavoidable when looking at frozen object contents
instead of their live memory layout.

If you're interested, here's the development version of the package:
install.packages('depcache',contriburl='https://aitap.github.io/Rpackages')

  
    
  


  


      

    

    

      
      

        


  

        

      iuke-tier@ey m@iii@g oii uiow@@edu
    

    

      
      #
    

  


  

      
On Thu, 18 Jan 2024, Ivan Krylov via R-devel wrote:

        

            

      
      
      
    

        
What does 'blow up' mean? If it is anything other than signal a "bad
binding access" error then it would be good to have more details.

Best,

luke

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Ivan Krylov
    

    

      
      #
    

  


  

      
On Thu, 18 Jan 2024 09:59:31 -0600 (CST)

        
luke-tierney at uiowa.edu wrote:

        

            

      
      
      
    

        
My apologies for not being precise enough. I meant the "bad binding
access" error in all such cases.