Halting drake plan makes it rebuild targets it already had built previously

Question

I'm currently using drake to run a set of >1k simulations. I've estimated that it would take about two days to run the complete set, but I also expect my computer to crash at any point during that period because, well, it has.

Apparently stopping the plan discards any targets that were already built so essentially this means I can't use drake for its intended purpose.

I suppose I could make a function that actually edits the R file where the plan is specified in order to make drake sequentially add targets to its cache but that seems utterly hackish.

Any ideas on how to deal with this?

EDIT: The actual problem seems to come from using set.seed inside my data generating functions. I was aware that drake already does this for the user in a way that ensures reproducibility, but I figured that if I just left my functions the way they were it wouldn't change anything since drake would be ensuring that the random seed I chose always ends up being the same? Guess not, but since I removed that step things are caching fine so the issue is solved.

Answer 1

To bring onlookers up to speed, I will try to spell out the problem. @zipzapboing, please correct me if my description is off-target.

Let's say you have a script that generates a drake plan and executes it.

library(drake)

simulate_data <- function(seed){
  set.seed(seed)
  rnorm(100)
}

seed_grid <- data.frame(
  id = paste0("target_", 1:3),
  seed = sample.int(1e6, 3)
)

print(seed_grid)
#>         id   seed
#> 1 target_1 581687
#> 2 target_2 700363
#> 3 target_3 914982

plan <- map_plan(seed_grid, simulate_data)

print(plan)
#> # A tibble: 3 x 2
#>   target   command                      
#>   <chr>    <chr>                        
#> 1 target_1 simulate_data(seed = 581687L)
#> 2 target_2 simulate_data(seed = 700363L)
#> 3 target_3 simulate_data(seed = 914982L)

make(plan)
#> target target_1
#> target target_2
#> target target_3
make(plan)
#> All targets are already up to date.

^{Created on 2018-11-12 by the reprex package (v0.2.1)}

The second make() worked just fine, right? But if you were to run the same script in a different session, you would end up with a different plan. The randomly-generated seed arguments to simulate_data() would be different, so all your targets would build from scratch.

library(drake)

simulate_data <- function(seed){
  set.seed(seed)
  rnorm(100)
}

seed_grid <- data.frame(
  id = paste0("target_", 1:3),
  seed = sample.int(1e6, 3)
)

print(seed_grid)
#>         id   seed
#> 1 target_1 654304
#> 2 target_2 252208
#> 3 target_3 781158

plan <- map_plan(seed_grid, simulate_data)

print(plan)
#> # A tibble: 3 x 2
#>   target   command                      
#>   <chr>    <chr>                        
#> 1 target_1 simulate_data(seed = 654304L)
#> 2 target_2 simulate_data(seed = 252208L)
#> 3 target_3 simulate_data(seed = 781158L)

make(plan)
#> target target_1
#> target target_2
#> target target_3

^{Created on 2018-11-12 by the reprex package (v0.2.1)}

One solution is to be extra careful to hold onto the same plan. However, there is an even easier way: just let drake set the seeds for you. drake automatically gives each target its own reproducible random seed. These target-level seeds are deterministically generated by a root seed (the seed argument to make()) and the names of the targets.

library(digest)
library(drake)
library(magrittr) # defines %>%

simulate_data <- function(){
  mean(rnorm(100))
}

plan <- drake_plan(target = simulate_data()) %>%
  expand_plan(values = 1:3)

print(plan)
#> # A tibble: 3 x 2
#>   target   command        
#>   <chr>    <chr>          
#> 1 target_1 simulate_data()
#> 2 target_2 simulate_data()
#> 3 target_3 simulate_data()

tmp <- rnorm(1)
digest(.Random.seed) # Fingerprint of the current seed.
#> [1] "0bbddc33a4afe7cd1c1742223764661c"

make(plan)
#> target target_1
#> target target_2
#> target target_3
make(plan)
#> All targets are already up to date.

# The targets have different seeds and different values.
readd(target_1)
#> [1] -0.05530201
readd(target_2)
#> [1] 0.03698055
readd(target_3)
#> [1] 0.05990671

clean() # Destroy the targets.
tmp <- rnorm(1) # Change the global seed.
digest(.Random.seed) # The seed changed.
#> [1] "5993aa5cff4b72a0e14fa58dc5c5e3bf"

make(plan)
#> target target_1
#> target target_2
#> target target_3

# The targets were regenerated with the same values (same seeds).
readd(target_1)
#> [1] -0.05530201
readd(target_2)
#> [1] 0.03698055
readd(target_3)
#> [1] 0.05990671

# You can recover a target's seed from its metadata.
seed <- diagnose(target_1)$seed
print(seed)
#> [1] 1875584181

# And you can use that seed to reproduce
# the target's value outside make().
set.seed(seed)
mean(rnorm(100))
#> [1] -0.05530201

^{Created on 2018-11-12 by the reprex package (v0.2.1)}

I really should write more in the manual about how seeds work in drake and highlight the original pitfall raised in this thread. I doubt you are the only one who struggled with this issue.