Cryptic performance bug with hashing about pointblank

[yjunechoe]

I realize that I've made this issue more complicated than needed, and performance wasn't the right framing for understanding the problem. Closing in favor of #509

[yjunechoe]

Actually, the performance problem scales with the size of the data frame too (so you can ignore the convoluted reprex above). This is because hashing the (in most cases, global) environment will often also hash the table (and everything else) that's also available in the environment.

With that in mind, a newer, more minimal reprex on a 1-million row dataframe (using CRAN version this time, but the problem also exists on dev):

library(pointblank)
big_table <- small_table[sample(nrow(small_table), 1e6, replace = TRUE),]
agent <- create_agent(big_table)

profvis::profvis({
  agent %>% 
    col_vals_gt(
      columns = c("c", "d", "e"),
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, c, d, e)
    )
})

The slowdown is specifically being driven by hashing the vars() quosure (sha1.formula()) and the preconditions function (sha1.function()). Both methods hash the attached environment by default:

And of course, the problem also scales with the number of validation steps added to the agent:

system.time({
  agent %>% 
    col_vals_gt(
      columns = c,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, c)
    ) %>% 
    col_vals_gt(
      columns = d,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, d)
    ) %>% 
    col_vals_gt(
      columns = e,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, e)
    )
})
#> user  system elapsed 
#> 2.07    0.39    6.72

Resolving value and preconditions to string before hashing resolves the performance issue (I'm testing this on a local branch):

devtools::install_github("yjunechoe/pointblank@no_env_hash")
library(pointblank)

big_table <- small_table[sample(nrow(small_table), 1e6, replace = TRUE),]
agent <- create_agent(big_table)

system.time({
  agent %>% 
    col_vals_gt(
      columns = c,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, c)
    ) %>% 
    col_vals_gt(
      columns = d,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, d)
    ) %>% 
    col_vals_gt(
      columns = e,
      value = vars(a),
      preconditions = . %>% 
        dplyr::select(a, e)
    )
})
#> user  system elapsed 
#> 0.04    0.02    0.14

---

The "resolve to string before hashing" solution currently fails the multiagent tests (but passes all others) and it will break backwards compatibility as is. But since the multiagent has all the info that was used to generate the original hash, I wonder if we can (optionally) have the hash for each validation step be re-generated by the multiagent for the purposes of alignment (which I believe would still guarantee correct alignment for older agents)?