feat(conditions): improved error handling (#1365)

sebffischer · web-flow · commit a3356c13226d · 2025-08-16T10:43:45.000+02:00
diff --git a/R/Learner.R b/R/Learner.R
@@ -572,6 +572,9 @@ Learner = R6Class("Learner",
     #' If the training step fails, the `$model` field of the original learner is `NULL`.
     #' The results are reproducible across the different encapsulation methods.
     #'
+    #' Note that for errors of class `Mlr3ErrorConfig`, the function always errs and no fallback learner
+    #' is trained.
+    #'
     #' Also see the section on error handling the mlr3book:
     #' \url{https://mlr3book.mlr-org.com/chapters/chapter10/advanced_technical_aspects_of_mlr3.html#sec-error-handling}
     #'
@@ -580,15 +583,20 @@ Learner = R6Class("Learner",
     #'  See the description for details.
     #' @param fallback [Learner]\cr
     #'  The fallback learner for failed predictions.
+    #' @param when (`function(condition)`)\cr
+    #'  Function that takes in the condition and returns `logical(1)` indicating whether to run the fallback learner.
+    #'  If `NULL` (default), the fallback is always trained, except for errors of class `Mlr3ErrorConfig`.
     #'
     #' @return `self` (invisibly).
     #' @examples
     #' learner = lrn("classif.rpart")
     #' fallback = lrn("classif.featureless")
     #' learner$encapsulate("try", fallback = fallback)
-    encapsulate = function(method, fallback = NULL) {
+    encapsulate = function(method, fallback = NULL, when = NULL) {
       assert_choice(method, c("none", "try", "evaluate", "callr", "mirai"))
 
+      private$.when = assert_function(when, null.ok = TRUE)
+
       if (method != "none") {
         assert_learner(fallback, task_type = self$task_type)
 
@@ -702,7 +710,6 @@ Learner = R6Class("Learner",
       private$.use_weights
     },
 
-
     #' @field model (any)\cr
     #' The fitted model. Only available after `$train()` has been called.
     model = function(rhs) {
@@ -750,7 +757,6 @@ Learner = R6Class("Learner",
       get_log_condition(self$state, "error")
     },
 
-
     #' @field hash (`character(1)`)\cr
     #' Hash (unique identifier) for this object.
     #' The hash is calculated based on the learner id, the parameter settings, the predict type, the fallback hash, the parallel predict setting, the validate setting, and the predict sets.
@@ -780,7 +786,6 @@ Learner = R6Class("Learner",
 
       assert_string(rhs, .var.name = "predict_type")
       if (rhs %nin% self$predict_types) {
-
         stopf("Learner '%s' does not support predict type '%s'", self$id, rhs)
       }
       private$.predict_type = rhs
@@ -840,6 +845,7 @@ Learner = R6Class("Learner",
   ),
 
   private = list(
+    .when = NULL,
     .use_weights = NULL,
     .encapsulation = c(train = "none", predict = "none"),
     .fallback = NULL,
diff --git a/R/LearnerClassifDebug.R b/R/LearnerClassifDebug.R
@@ -80,7 +80,8 @@ LearnerClassifDebug = R6Class("LearnerClassifDebug", inherit = LearnerClassif,
         iter                 = p_iter,
         early_stopping       = p_lgl(default = FALSE, tags = "train"),
         count_marshaling     = p_lgl(default = FALSE, tags = "train"),
-        check_pid            = p_lgl(default = TRUE, tags = "train")
+        check_pid            = p_lgl(default = TRUE, tags = "train"),
+        config_error         = p_lgl(default = FALSE, tags = "train")
       )
       super$initialize(
         id = "classif.debug",
@@ -161,6 +162,10 @@ LearnerClassifDebug = R6Class("LearnerClassifDebug", inherit = LearnerClassif,
     .validate = NULL,
     .train = function(task) {
       pv = self$param_set$get_values(tags = "train")
+      if (isTRUE(pv$config_error)) {
+        error_config("You misconfigured the learner")
+      }
+
       pv$count_marshaling = pv$count_marshaling %??% FALSE
       roll = function(name) {
         name %chin% names(pv) && pv[[name]] > runif(1L)
diff --git a/R/benchmark_grid.R b/R/benchmark_grid.R
@@ -15,7 +15,7 @@
 #'   The grid will be generated based on the Cartesian product of learners and pairs.
 #'
 #' @section Errors and Warnings:
-#' * `varying_predict_types`: This warning will be thrown if the learners have different `predict_type`s.
+#' * `Mlr3WarningVaryingPredictTypes`: This warning will be thrown if the learners have different `predict_type`s.
 #'
 #' @param tasks (list of [Task]).
 #' @param learners (list of [Learner]).
@@ -79,7 +79,10 @@ benchmark_grid = function(tasks, learners, resamplings, param_values = NULL, pai
     assert_param_values(param_values, n_learners = length(learners))
   }
   if (length(unique(map_chr(unique(learners), "predict_type"))) > 1) {
-    warningf("Multiple predict types detected, this will mean that you cannot evaluate the same measures on all learners.", class = "varying_predict_types") # nolint
+    warning_config(
+      "Multiple predict types detected, this will mean that you cannot evaluate the same measures on all learners.",
+      class = "Mlr3WarningVaryingPredictTypes"
+    )
   }
 
   if (assert_flag(paired)) {
diff --git a/R/worker.R b/R/worker.R
@@ -108,9 +108,22 @@ learner_train = function(learner, task, train_row_ids = NULL, test_row_ids = NUL
     .compute = getOption("mlr3.mirai_encapsulation", "mlr3_encapsulation")
   )
 
-  log = append_log(NULL, "train", result$log$class, result$log$msg)
+  cond = result$log[class == "error", "condition"][[1L]]
+
+  cond = if (length(cond)) {
+    cond = cond[[1L]]
+  }
+
+  when = get_private(learner)$.when
+  catch_error = (!is.null(cond)) && (!inherits(cond, "Mlr3ErrorConfig")) && (is.null(when) || when(cond))
+
+  log = append_log(NULL, "train", result$log$class, result$log$msg, log_error = catch_error)
   train_time = result$elapsed
 
+  if (!is.null(cond) && !catch_error) {
+    stop(cond)
+  }
+
   learner$state = set_class(insert_named(learner$state, list(
     model = result$result$model,
     log = log,
@@ -495,7 +508,7 @@ process_model_after_predict = function(learner, store_models, is_sequential, unm
   }
 }
 
-append_log = function(log = NULL, stage = NA_character_, class = NA_character_, msg = character()) {
+append_log = function(log = NULL, stage = NA_character_, class = NA_character_, msg = character(), log_error = TRUE) {
   if (is.null(log)) {
     log = data.table(
       stage = factor(levels = c("train", "predict")),
@@ -506,7 +519,7 @@ append_log = function(log = NULL, stage = NA_character_, class = NA_character_,
 
   if (length(msg)) {
     pwalk(list(stage, class, msg), function(s, c, m) {
-      if (c == "error") lg$error("%s: %s", s, m)
+      if (c == "error" && log_error) lg$error("%s: %s", s, m)
       if (c == "warning") lg$warn("%s: %s", s, m)
     })
     log = rbindlist(list(log, data.table(stage = stage, class = class, msg = msg)), use.names = TRUE)
diff --git a/man/Learner.Rd b/man/Learner.Rd
diff --git a/man/benchmark_grid.Rd b/man/benchmark_grid.Rd
diff --git a/man/mlr_learners_classif.debug.Rd b/man/mlr_learners_classif.debug.Rd
diff --git a/tests/testthat/test_Learner.R b/tests/testthat/test_Learner.R
@@ -839,6 +839,86 @@ test_that("Learner printer for encapsulation", {
   expect_output(print(lrn("classif.rpart")$encapsulate("none")), "Encapsulation: none \\(fallback: -\\)")
 })
 
+test_that("error conditions are working: callr", {
+  l = lrn("classif.debug",
+    timeout = c(train = 0.01),
+    # Sys.sleep does not get interrupted reliably
+    sleep_train = function() while (TRUE) NULL
+  )
+
+  l$encapsulate(
+    "callr",
+    lrn("classif.featureless"),
+    when = function(cond) {
+      !inherits(cond, "Mlr3ErrorTimeout")
+    }
+  )
+
+  expect_error(l$train(tsk("iris")), regexp = "reached elapsed time limit")
+  l$configure(error_train = 1, sleep_train = NULL, timeout = c(train = Inf, predict = Inf))
+  expect_error(l$train(tsk("iris")), regexp = NA)
+})
+
+test_that("error conditions are working: evaluate", {
+  l = lrn("classif.debug",
+    timeout = c(train = 0.2),
+    # Sys.sleep does not get interrupted reliably
+    sleep_train = function() while (TRUE) NULL
+  )
+
+  l$encapsulate(
+    "evaluate",
+    lrn("classif.featureless"),
+    function(x) {
+      !inherits(x, "Mlr3ErrorTimeout")
+    }
+  )
+
+  expect_error(l$train(tsk("iris")), regexp = "reached elapsed time limit")
+  l$configure(error_train = 1, sleep_train = NULL, timeout = c(train = Inf, predict = Inf))
+  expect_error(l$train(tsk("iris")), regexp = NA)
+})
+
+test_that("error conditions are working: try", {
+  l = lrn("classif.debug",
+    timeout = c(train = 0.01),
+    # Sys.sleep does not get interrupted reliably
+    sleep_train = function() while (TRUE) NULL
+  )
+
+  l$encapsulate(
+    "try",
+    lrn("classif.featureless"),
+    function(x) {
+      !inherits(x, "Mlr3ErrorTimeout")
+    }
+  )
+
+  expect_error(l$train(tsk("iris")), regexp = "reached elapsed time limit")
+  l$configure(error_train = 1, sleep_train = NULL, timeout = c(train = Inf, predict = Inf))
+  expect_error(l$train(tsk("iris")), regexp = NA)
+})
+
+test_that("error conditions are working: mirai", {
+  l = lrn("classif.debug",
+    timeout = c(train = 0.01),
+    # Sys.sleep does not get interrupted reliably
+    sleep_train = function() while (TRUE) NULL
+  )
+
+  l$encapsulate(
+    "mirai",
+    lrn("classif.featureless"),
+    function(x) {
+      !inherits(x, "Mlr3ErrorTimeout")
+    }
+  )
+
+  expect_error(l$train(tsk("iris")), regexp = "reached elapsed time limit")
+  l$configure(error_train = 1, sleep_train = NULL, timeout = c(train = Inf, predict = Inf))
+  expect_error(l$train(tsk("iris")), regexp = NA)
+})
+
 test_that("oob_error is available without storing models via $.extract_oob_error()", {
   LearnerDummyOOB = R6::R6Class("LearnerDummyOOB", inherit = LearnerClassif,
     public = list(
@@ -872,3 +952,11 @@ test_that("oob_error is available without storing models via $.extract_oob_error
 
   expect_equal(rr$aggregate(msr("oob_error")), c(oob_error = 0.123))
 })
+
+test_that("config error does not trigger callback", {
+  l = lrn("classif.debug", config_error = TRUE)
+  l$encapsulate("evaluate", lrn("classif.featureless"), function(x) TRUE)
+  expect_error(l$train(tsk("iris")), regexp = "You misconfigured the learner")
+  l$encapsulate("evaluate", lrn("classif.featureless"))
+  expect_error(l$train(tsk("iris")), regexp = "You misconfigured the learner")
+})
diff --git a/tests/testthat/test_errorhandling.R b/tests/testthat/test_errorhandling.R
@@ -88,4 +88,3 @@ test_that("encapsulation / benchmark", {
   expect_equal(aggr$warnings, 3L)
   expect_equal(aggr$errors, 3L)
 })
-
