`doFuture` warning about PRNG use in parallel projections

[fweber144]

When running a projection in parallel (across draws) using the doFuture package, the doFuture package claims that the pseudo-random number generator (PRNG) had been used. This is unexpected because the projection is deterministic, which can be checked by inspecting the source code of fit_glm_ridge_callback() (including all functions used by it, and functions used by these, and so on).

Reprex:

data("df_gaussian", package = "projpred")
df_gaussian <- df_gaussian[1:29, ]
dat <- data.frame(y = df_gaussian$y, df_gaussian$x)
rfit <- rstanarm::stan_glm(y ~ X1 + X2 + X3 + X4 + X5,
                           data = dat,
                           chains = 1,
                           iter = 500,
                           seed = 1140350788,
                           refresh = 0)

library(projpred)

doFuture::registerDoFuture()
future::plan(future::multisession, workers = 4)
progressr::handlers(global = TRUE)
trigger_default <- options(projpred.prll_prj_trigger = 0L)

prj <- project(rfit,
               predictor_terms = c("X1", "X2"),
               seed = 46782345)

The last line throws 4 warnings (one per parallel worker), the last one (here for worker 4) being:

Warning: UNRELIABLE VALUE: One of the foreach() iterations (‘doFuture-4’)
unexpectedly generated random numbers without declaring so. There is a risk
that those random numbers are not statistically sound and the overall results
might be invalid. To fix this, use '%dorng%' from the 'doRNG' package instead
of '%dopar%'. This ensures that proper, parallel-safe random numbers are
produced via the L'Ecuyer-CMRG method. To disable this check, set option
'doFuture.rng.onMisuse' to "ignore".

[fweber144]

I think I found the reason for this: The Rcpp package introduces

projpred/src/RcppExports.cpp

Line 43 in 0362770

Rcpp::RNGScope rcpp_rngScope_gen;

(same for glm_elnet_c() in

projpred/src/RcppExports.cpp

Line 19 in 0362770

Rcpp::RNGScope rcpp_rngScope_gen;

), which probably alters the PRNG state. So the doFuture warning seems to be a false alarm. This is confirmed by changing the GLM submodel fitter:

data("df_gaussian", package = "projpred")
dat <- data.frame(y = df_gaussian$y, df_gaussian$x)
rfit <- rstanarm::stan_glm(y ~ X1 + X2 + X3 + X4 + X5,
                           data = dat,
                           chains = 1,
                           iter = 500,
                           seed = 1140350788,
                           refresh = 0)

library(projpred)

doFuture::registerDoFuture()
future::plan(future::multisession, workers = 4)
progressr::handlers(global = TRUE)
trigger_default <- options(projpred.prll_prj_trigger = 0L)

options(projpred.glm_fitter = "fit_glm_callback")
### Default would be:
# options(projpred.glm_fitter = "fit_glm_ridge_callback")
###

prj <- project(rfit,
               predictor_terms = c("X1", "X2"),
               seed = 46782345)

That way, the warnings are not thrown anymore. So if possible, we should make Rcpp omit those Rcpp::RNGScope lines or make fit_glm_callback() the default GLM submodel fitter (instead of fit_glm_ridge_callback()). The latter suggestion might be a significant change.