perf(bootstrap): stop serializing the task list with every mirai batch

DESCRIPTION

@@ -37,6 +37,6 @@ Suggests:
     pkgload,
     rmarkdown,
     Rmpfr,
-    testthat (>= 3.0.0)
+    testthat (>= 3.1.7)
 VignetteBuilder: knitr
 Config/testthat/edition: 3

R/bootstrap.R

@@ -215,14 +215,7 @@ start_mirai_bootstrap <- function(inputs, object) {
   mapped <- tryCatch(
     mirai::mirai_map(
       batches,
-      function(batch, original_fit, dev_path) {
-        if (!is.null(dev_path) && !pkgload::is_dev_package("singlehit")) {
-          pkgload::load_all(dev_path, quiet = TRUE)
-        }
-        dplyr::bind_rows(lapply(batch, function(task) {
-          singlehit:::bootstrap_refit(original_fit, task$sample_data, task$replicate)
-        }))
-      },
+      bootstrap_worker(),
       .args = list(original_fit = object, dev_path = dev_path),
       .compute = compute
     ),
@@ -245,6 +238,25 @@ start_mirai_bootstrap <- function(inputs, object) {
   )
 }
 
+# The function mirai sends to each daemon with every batch. Its environment is
+# globalenv() so that mirai does not serialize the frame that created it, which
+# holds the full task list. The body uses only namespaced calls, so it needs no
+# free variables from that frame. Source references are removed because under
+# pkgload::load_all() they would carry the whole source file with each batch.
+bootstrap_worker <- function() {
+  worker <- function(batch, original_fit, dev_path) {
+    if (!is.null(dev_path) && !pkgload::is_dev_package("singlehit")) {
+      pkgload::load_all(dev_path, quiet = TRUE)
+    }
+    dplyr::bind_rows(lapply(batch, function(task) {
+      singlehit:::bootstrap_refit(original_fit, task$sample_data, task$replicate)
+    }))
+  }
+  worker <- utils::removeSource(worker)
+  environment(worker) <- globalenv()
+  worker
+}
+
 bind_bootstrap_results <- function(results) {
   failed <- vapply(results, function(result) inherits(result, c("miraiError", "errorValue")), logical(1))
   if (any(failed)) {

tests/testthat/test-bootstrap-dispatch.R

@@ -0,0 +1,39 @@
+test_that("the mirai worker does not capture the frame that creates it", {
+  worker <- bootstrap_worker()
+
+  # A closure built inside start_mirai_bootstrap() would carry that frame, which
+  # holds every task, and mirai would serialize it with each batch.
+  expect_identical(environment(worker), globalenv())
+  expect_lt(length(serialize(worker, NULL)), 10000L)
+})
+
+test_that("the mirai worker refits a batch exactly like the sequential path", {
+  fit <- fit_dose_response(ward_fixture(), "beta_poisson")
+  tasks <- generate_bootstrap_tasks(fit, fit$data$response, 3L, 42L)
+
+  batched <- bootstrap_worker()(tasks, fit, NULL)
+  sequential <- run_sequential_bootstrap(tasks, fit)
+
+  expect_equal(tibble::as_tibble(batched), tibble::as_tibble(sequential))
+})
+
+test_that("start_mirai_bootstrap sends mirai the detached worker", {
+  skip_if_not_installed("mirai")
+  fit <- fit_dose_response(ward_fixture(), "beta_poisson")
+  inputs <- prepare_bootstrap(fit, 3L, "observed", 42L, "mirai", NULL, NULL)
+  captured <- NULL
+  local_mocked_bindings(
+    daemons_set = function(...) TRUE,
+    mirai_map = function(.x, .f, ...) {
+      captured <<- .f
+      list()
+    },
+    .package = "mirai"
+  )
+
+  job <- start_mirai_bootstrap(inputs, fit)
+
+  expect_s3_class(job, "qdr_bootstrap_job")
+  expect_identical(environment(captured), globalenv())
+  expect_lt(length(serialize(captured, NULL)), 10000L)
+})