r_session$new() afects the random seed generator in a non reproducible way on Windows

[dfalbel]

For example, on Windows, if you run a few times:

set.seed(1)
digest::sha1(.Random.seed)
sess <- callr::r_session$new()
digest::sha1(.Random.seed)

You'll see that it that the RNG state is different after every execution.

[gaborcsardi]

Indeed, to generate a unique pipe name:

processx/src/win/stdio.c

Lines 93 to 98 in 118704a

	static void processx__unique_pipe_name(char* ptr, char* name, size_t size) {
	int r;
	GetRNGstate();
	r = (int)(unif_rand() * 65000);
	snprintf(name, size, "\\\\?\\pipe\\px\\%p-%lu", ptr + r, GetCurrentProcessId());
	PutRNGstate();

R CMD check forces me to use R's RNGs, so that is not OK, either?

[dfalbel]

I believe it's OK to use R's RNG, but starting a session should modify the RNG state in a predictable manner. Such that, the RNG state is consistent accross multiple runs. With the current behavior, if you do set.seed() and later use r_session$new(), then future outputs will not depend on the starting seed.

[gaborcsardi]

So how is the code in the original post different than this?

set.seed(1)
digest::sha1(.Random.seed)
rnorm(5)
digest::sha1(.Random.seed)

I am sorry, my argument is not with you, really. But I am also not entirely sure what to do here.

[dfalbel]

Here are two runs from the above and another two runs using the r_session:

> set.seed(1)
+ digest::sha1(.Random.seed)
+ rnorm(1)
+ digest::sha1(.Random.seed)
[1] "26d295cbeca7a799eb5b849b285aacf83794392c"
[1] -0.6264538
[1] "20e54edc8504855ccc29c3f9abb77ac5651300e8"
> set.seed(1)
+ digest::sha1(.Random.seed)
+ rnorm(1)
+ digest::sha1(.Random.seed)
[1] "26d295cbeca7a799eb5b849b285aacf83794392c"
[1] -0.6264538
[1] "20e54edc8504855ccc29c3f9abb77ac5651300e8"
> set.seed(1)
+ digest::sha1(.Random.seed)
+ sess <- callr::r_session$new()
+ digest::sha1(.Random.seed)
[1] "26d295cbeca7a799eb5b849b285aacf83794392c"
[1] "2a53052f8dfedfe31831a157ee373c93c695eaeb"
> set.seed(1)
+ digest::sha1(.Random.seed)
+ sess <- callr::r_session$new()
+ digest::sha1(.Random.seed)
[1] "26d295cbeca7a799eb5b849b285aacf83794392c"
[1] "00b61021b84384fb2292842dc1347bc75a9529ce"

Note that r_session is modifying the state differently between runs.

[gaborcsardi]

Well, yes, because it keeps generating random file names until one does not exist, and you are creating the second session while the first is still around, so that pipe file already exists. If you re-run it in a different session you get the same:

> RS -q -e "set.seed(1); rs <- callr::r_session[['new']](); digest::digest(.Random.seed)"
> set.seed(1); rs <- callr::r_session[['new']](); digest::digest(.Random.seed)
[1] "34ef6514589f20e41c2f1da6bdce71b1"

> RS -q -e "set.seed(1); rs <- callr::r_session[['new']](); digest::digest(.Random.seed)"
> set.seed(1); rs <- callr::r_session[['new']](); digest::digest(.Random.seed)
[1] "34ef6514589f20e41c2f1da6bdce71b1"

But yeah, I guess the file name is not something that is supposed to be reproducible, so I should not use the R RNG. Sadly, it is kind of hard to run the equivalent of basename(tempfile()) from the Windows C API. I'll come up with something.

[gaborcsardi]

I can do a release soon if this is important for you.

[dfalbel]

Thanks! No hurry for a release, that's not urgent!