target invalidated when referenced from another plan

[idavydov]

library(drake)

plan.a <- drake_plan(a=10)

make(plan.a)
#> target a
loadd(a)

plan.b <- drake_plan(b=a+10)

make(plan.b)
#> target b

make(plan.a)
#> Unloading targets from environment:
#>   a
#> target a

Note that target a was not changed, however it's invalidated.

This works as expected:

library(drake)

plan.a <- drake_plan(a=10)

make(plan.a)
#> target a
a.copy <- readd(a)

plan.b <- drake_plan(b=a.copy+10)

make(plan.b)
#> target b

make(plan.a)
#> All targets are already up to date.

[wlandau]

That's a great point. Please include all your targets in a single plan (with bind_plans(),rbind(), or dplyr::bind_rows()). drake thinks plan.a and plan.b are for different projects, and this is unlikely to change in future development. Sorry if this is inconvenient. The API I am developing in #674 should make it easier to keep things consolidated.

You could split up different plans into different projects with different sessions and caches. (See the cache argument to make() and https://ropenscilabs.github.io/drake-manual/store.html.)

[idavydov]

Thanks. Do you think this could detected with a warning? This behavior is not entirely obvious, and took a lot of time to debug for me.

Regarding my case, combining two plans is not a good option, since the second plan is created conditionally based on the results of the first plan. This is because not all branches of the graph can be computed (see second point of #528), so I'm trying to solve it with a preprocessing plan.

Having a separate cache is an option, I did it for a while. But I discovered that readd() and loadd() inside .Rmd are always using the default cache, and hard-coding cache for every readd() and loadd() call looks kinda ugly. I'm not sure this could be in principle solved by drake.

[wlandau]

Do you think this could detected with a warning?

I could be convinced. However, for large workflows, it seems like this would require a lot of cache operations, so I doubt it will scale well. I would be happy to review a pull request, though.