fmelt: create the string var once per measure group instead of once per row

[m-muecke]

There is more potential in this file, but that would've needed some refactoring and hence larger diffs.

ibrary(data.table)
setDTthreads(1L)
k = 10
for (n in c(1e2, 1e3, 1e4, 1e5, 1e6)) {
 set.seed(1)
 DT = data.table(id = seq_len(n))
 for (i in seq_len(k)) {
   set(DT, j = paste0("a_", i), value = runif(n))
   set(DT, j = paste0("b_", i), value = runif(n))
 }
 avars = paste0("a_", seq_len(k))
 bvars = paste0("b_", seq_len(k))
 print(bench::mark(
   list_varchar = melt(DT, id.vars = "id", measure.vars = list(avars, bvars), variable.factor = FALSE),
   measure_int = melt(DT, id.vars = "id", measure.vars = measure(value.name, num = as.integer, sep = "_")),
   measure_dbl = melt(DT, id.vars = "id", measure.vars = measure(value.name, num = as.numeric, sep = "_")),
   measure_chr = melt(DT, id.vars = "id", measure.vars = measure(value.name, num, sep = "_")),
   min_iterations = 30,
   min_time = 1.5,
   check = FALSE
 ))
}

rows	measure.vars = list(...), variable.factor = FALSE	measure(num = as.integer)	measure(num = as.numeric)	measure(num) (character)
1e2	50µs → 41µs (1.22x)	1.4ms → 1.4ms (1.00x)	1.4ms → 1.4ms (1.01x)	1.4ms → 1.4ms (1.00x)
1e3	167µs → 77µs (2.16x)	1.4ms → 1.4ms (1.01x)	1.5ms → 1.4ms (1.03x)	1.4ms → 1.4ms (1.02x)
1e4	1.3ms → 391µs (3.30x)	1.8ms → 1.5ms (1.22x)	1.7ms → 1.5ms (1.17x)	1.9ms → 1.8ms (1.09x)
1e5	12.6ms → 3.6ms (3.52x)	5.2ms → 1.9ms (2.75x)	4.5ms → 2.0ms (2.30x)	6.4ms → 5.0ms (1.28x)
1e6	124.4ms → 35.5ms (3.50x)	38.7ms → 6.3ms (6.16x)	31.9ms → 6.8ms (4.70x)	51.1ms → 37.3ms (1.37x)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.02%. Comparing base (1d87418) to head (40b5db9).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #7917   +/-   ##
=======================================
  Coverage   99.02%   99.02%           
=======================================
  Files          88       88           
  Lines       17381    17387    +6     
=======================================
+ Hits        17211    17217    +6     
  Misses        170      170           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tdhock]

this looks interesting thanks!!
My intuition is that the speedups should be minimal in realistic (large) data though?
Can you please try converting your benchmark code to an atime_versions test case prototype, to show speed differences with current master?
Docs are here https://github.com/Rdatatable/data.table/wiki/Performance-testing and https://rjournal.github.io/articles/RJ-2026-013/
but basically you convert for loop over n in your code to N argument in atime_versions.

[m-muecke]

CI seems to be failing, but this was the local run, there seems to be cedta() issue with atime, but believe got a work around working.

[tdhock]

thanks. those results look encouraging.
it is normal that atime CI is failing, until we merge #7901.

[tdhock]

On my machine with this code I get ~2x speedups for one test, but smaller speedups for the other

tinfo <- atime::atime_pkg_test_info("~/R/data.table")
test.cases <- grep("7917", names(tinfo$test.call), value=TRUE)
atime.results <- list()
for(Test in test.cases){
  tcall <- tinfo$test.call[[Test]]
  atime.results[[Test]] <- eval(tcall)
}
library(ggplot2)
Test <- test.cases[2]
plot(predict(atime::references_best(atime.results[[Test]])))+
  ggtitle(Test)+
  theme(text=element_text(size=30))

overall this looks good to me, can you please add a NEWS item?

[tdhock]

Please add these comments so we remember to change these commits after merging this PR?