library(rush)
branin = function(x1, x2) {
(x2 - 5.1 / (4 * pi^2) * x1^2 + 5 / pi * x1 - 6)^2 + 10 * (1 - 1 / (8 * pi)) * cos(x1) + 10
}
wl_random_search = function(rush, branin) {
while (rush$n_finished_tasks < 100) {
xs = list(x1 = runif(1, -5, 10), x2 = runif(1, 0, 15))
key = rush$push_running_tasks(xss = list(xs))
tryCatch({
if (runif(1) < 0.5) stop("Random Error")
ys = list(y = branin(xs$x1, xs$x2))
rush$finish_tasks(key, yss = list(ys))
}, error = function(e) {
condition = list(message = e$message)
rush$fail_tasks(key, conditions = list(condition))
})
}
}rush provides error-handling mechanisms for two failure modes: standard R errors during task evaluation and unexpected worker failures such as crashes or lost connections. If errors cannot be caught automatically, the worker loop can be debugged locally.
Simple R Errors
We use the random search example from the tutorial and introduce a random error with 50% probability. Within the worker loop, users are responsible for catching errors and marking the corresponding task as "failed" using the $fail_tasks() method.
We initialize the network and start the workers.
rush = rsh(
network = "test-simple-error",
config = redux::redis_config())
mirai::daemons(4)
rush$start_workers(
worker_loop = wl_random_search,
n_workers = 4,
branin = branin)When an error occurs, the task is marked as "failed" and the error message is stored in the "message" column. This ensures that errors do not interrupt the overall execution and allows subsequent inspection and reevaluation of failed tasks.
rush$fetch_failed_tasks() x1 x2 worker_id condition keys
<num> <num> <char> <list> <char>
1: 2.9187981 3.026696 unisex_gal... <list[1]> 78eac550-1...
2: 7.7340875 7.979802 unisex_gal... <list[1]> 6c369591-8...
3: 0.7677242 11.097560 unisex_gal... <list[1]> 609417a9-6...
4: 4.2320750 2.624436 unisex_gal... <list[1]> 32ae15ef-7...
5: -1.0000879 11.713371 scotophobi... <list[1]> 4d16bc5d-c...
---
119: 1.3812328 4.249511 semiration... <list[1]> fa991a7a-9...
120: -0.7862603 8.588903 unisex_gal... <list[1]> ad19175e-9...
121: 5.8514690 3.056105 scotophobi... <list[1]> 290be219-5...
122: -2.6919508 8.754054 semiration... <list[1]> ac504349-a...
123: 6.0918077 12.277266 scotophobi... <list[1]> b9b54acd-d...Handling Failing Workers
When a worker fails due to a crash or lost connection, its tasks may remain in the "running" state indefinitely. We simulate a segmentation fault by terminating the worker process.
wl_failed_worker = function(rush) {
xs = list(x1 = runif(1, -5, 10), x2 = runif(1, 0, 15))
key = rush$push_running_tasks(xss = list(xs))
tools::pskill(Sys.getpid(), tools::SIGKILL)
}
rush = rsh(
network = "test-failed-workers",
config = redux::redis_config())
mirai::daemons(2)
worker_ids = rush$start_workers(
worker_loop = wl_failed_worker,
n_workers = 2)The $detect_lost_workers() method identifies such workers and updates their state to "terminated". For workers started with $start_local_workers() or $start_workers(), lost worker detection works automatically by checking process status. Workers started via $worker_script() require an additional heartbeat mechanism (see the manager vignette).
rush$detect_lost_workers()[1] "hated_hyrax_54f39e63" "confiscable_shelduck_35826399"When a worker fails, the state of any task it was evaluating is set to "failed".
rush$fetch_failed_tasks() x1 x2 worker_id condition keys
<num> <num> <char> <list> <char>
1: 0.214842 9.790674 hated_hyra... <list[1]> 4c1ceee5-8...
2: 5.606061 4.941651 confiscabl... <list[1]> 14e1f4ad-4...Debugging
When the worker loop fails due to an uncaught error, the loop can be executed locally to reproduce and inspect the failure. Consider the following worker loop, which generates an error for large values of x1.
To debug the worker loop locally, a RushWorker instance is instantiated manually and passed as argument to the worker loop.
rush_worker = RushWorker$new(network_id = "test-error")
wl_error(rush_worker)Error in `wl_error()`:
! Unexpected errorWhen an error is raised in the main process, traceback() can be called to examine the stack trace and breakpoints can be set within the worker loop to inspect the program state. Note that certain errors such as missing packages may not be reproducible locally but can be identified by running the worker loop in a separate process and using $detect_lost_workers().
rush = rsh(
network = "test-error",
config = redux::redis_config())
mirai::daemons(1)
rush$start_workers(
worker_loop = wl_error,
n_workers = 1)
rush$detect_lost_workers()[1] "aeronautic_africanpiedkingfisher_fef242b1"Output and message logs can be written to files via the message_log and output_log arguments.
rush = rsh(
network = "test-error",
config = redux::redis_config())
message_log = tempdir()
output_log = tempdir()
mirai::daemons(1)
worker_ids = rush$start_workers(
worker_loop = wl_error,
n_workers = 1,
message_log = message_log,
output_log = output_log)
Sys.sleep(5)
readLines(file.path(message_log, sprintf("message_%s.log", worker_ids[1])))[1] "Debug message logging on worker weaponless_kakarikis_1fb90a7e started"[1] "[1] \"Debug output logging on worker weaponless_kakarikis_1fb90a7e started\""