Roll-back tracking full language objects for internal calls

[yjunechoe]

This is a small amendment of #543, which introduced a problem in serialization size due to validation_set$capture_stack$pb_call storing language objects with environments attached.

Now, we simply track the name of the internal function in string (ex: "tbl_val_comparison") as opposed to the full call (ex: tbl_val_comparison(...)). No user-visible changes.

What used to be 13kB -> 208kB is now 1Kb -> 133b. Reprex from #567:

# Setup
agent <- 
  create_agent(
    tbl = pointblank::small_table,
    tbl_name = "small_table",
    label = "`create_agent()` example.",
    actions = action_levels(
      warn_at = 0.10,
      stop_at = 0.25,
      notify_at = 0.35
    )
  )

agent <-
  agent %>% 
  col_exists(columns = c(date, date_time)) %>%
  col_vals_regex(
    columns = b,
    regex = "[0-9]-[a-z]{3}-[0-9]{3}"
  ) %>%
  rows_distinct() %>%
  col_vals_gt(columns = d, value = 100) %>%
  col_vals_lte(columns = c, value = 5) %>%
  col_vals_between(
    columns = c,
    left = vars(a), right = vars(d),
    na_pass = TRUE
  ) %>%
  interrogate()
  
  
# In memory
pb_call <- lapply(agent$validation_set$capture_stack, `[[`, "pb_call")
scales::label_bytes()(
  as.integer(object.size(pb_call))
)
#> [1] "1 kB"

# Serialized
f <- tempfile()
saveRDS(pb_call, f)
scales::label_bytes()(
  file.size(f)
)
#> [1] "133 B"

# Equivalence
identical(pb_call, readRDS(f))
#> [1] TRUE

pb_call
#> [[1]]
#> [1] "tbl_col_exists"
#> 
#> [[2]]
#> [1] "tbl_col_exists"
#> 
#> [[3]]
#> [1] "tbl_val_regex"
#> 
#> [[4]]
#> [1] "tbl_rows_distinct"
#> 
#> [[5]]
#> [1] "tbl_val_comparison"
#> 
#> [[6]]
#> [1] "tbl_val_comparison"
#> 
#> [[7]]
#> [1] "tbl_vals_between"

Rider: fixes a typo in examples of ?create_agent

[yjunechoe]

@rich-iannone I think this fixes the original issue and should be ready to go!

One Q for this PR while I'm on it - do we need a test for memory size at serialization? I suppose it'd only exist to guard against regressions like this, but could be useful since it can easily go unnoticed - though I'm unsure how it's best implemented (some upper bound size?) or whether it's stable enough to be useful as a test at all (across OS, R setups, etc.).

[rich-iannone]

@rich-iannone I think this fixes the original issue and should be ready to go!

This is amazing!! Thank you so much for working with the user to track this down (and of course for the PR).

One Q for this PR while I'm on it - do we need a test for memory size at serialization? I suppose it'd only exist to guard against regressions like this, but could be useful since it can easily go unnoticed - though I'm unsure how it's best implemented (some upper bound size?) or whether it's stable enough to be useful as a test at all (across OS, R setups, etc.).

This seems somewhat brittle as a test. There might be a convenient way to monitor the size of serialized agents through a separate workflow that just reports the sizes as an artifact (which we could check every now and then).

[yjunechoe]

This seems somewhat brittle as a test. There might be a convenient way to monitor the size of serialized agents through a separate workflow that just reports the sizes as an artifact (which we could check every now and then).

Thanks - this makes sense! For now I just added a manual test manual_tests/tests_agent_serialization_size.R that at least checks for the worst-case scenario of the entire user environment being snapshotted and written out as part of the agent, when using saveRDS().

[rich-iannone]

LGTM!

[rich-iannone]

Thanks again!