Choose the caching method for each target?

[wlandau]

Prework

• Read and abide by drake's code of conduct.
• Search for duplicates among the existing issues, both open and closed.

Description

A Keras model can only be saved using the same R session that fit the model. So if you have a plan with target(keras_model, format = "keras") and you make() it with HPC, you must set caching = "worker". Worker caching is almost always slower than master caching, especially for a large number of targets, so maybe it would be nice to set different caching methods for different targets. As with other custom columns such as "trigger", a "caching" column would override the corresponding argument to make().

library(drake)
drake_plan(
    x = target(fit_model(1), caching = "master"),
    y = target(fit_model(1), caching = "worker")
)
#> # A tibble: 2 x 3
#>   target command      caching
#>   <chr>  <expr>       <chr>  
#> 1 x      fit_model(1) master 
#> 2 y      fit_model(1) worker

^{Created on 2019-08-18 by the reprex package (v0.3.0)}

Hesitation

I am not sure about this feature because

1. Worker caching is almost always slower, so it might be nice to find workarounds instead of developing it seriously as a feature.
2. For Keras models specifically, the workaround is easy: instead of using target(format = "keras"), simply call serialize_model() before you return the fitted model. Then, the model will still be valid even if you send it to a different R process.