arrow backend

[JanMarvin]

Hi @jeroen ,

this is a draft for an arrow backend. I just tried to include it from what I tried yesterday in #6. I make it pass the tests, but not everything can be done with this arrow backend. Since it can be called without any PR this is just elaborated toying around. I have no real world use case for this.

ct <- V8::v8(backend = "arrow")
ct$source("https://unpkg.com/underscore@1.13.7/underscore-min.js")
#> [1] "true"
ct$assign("flights", nycflights13::flights)
js_code <- 'var filtered = _.filter(flights, x => x.arr_delay > 720);'
ct$eval(js_code)
ct$get("filtered")[1:5, 1:8]
#> # A tibble: 5 × 8
#>    year month   day dep_time sched_dep_time dep_delay arr_time sched_arr_time
#>   <dbl> <dbl> <dbl>    <dbl>          <dbl>     <dbl>    <dbl>          <dbl>
#> 1  2013     1     1      848           1835       853     1001           1950
#> 2  2013     1     9      641            900      1301     1242           1530
#> 3  2013     1    10     1121           1635      1126     1239           1810
#> 4  2013    11     3      603           1645       798      829           1913
#> 5  2013    12     5      756           1700       896     1058           2020

ct <- V8::v8()
ct$source("https://unpkg.com/underscore@1.13.7/underscore-min.js")
#> [1] "true"
ct$assign("flights", nycflights13::flights)
js_code <- 'var filtered = _.filter(flights, x => x.arr_delay > 720);'
ct$eval(js_code)
ct$get("filtered")[1:5, 1:8]
#>   year month day dep_time sched_dep_time dep_delay arr_time sched_arr_time
#> 1 2013     1   1      848           1835       853     1001           1950
#> 2 2013     1   9      641            900      1301     1242           1530
#> 3 2013     1  10     1121           1635      1126     1239           1810
#> 4 2013    11   3      603           1645       798      829           1913
#> 5 2013    12   5      756           1700       896     1058           2020

[jeroen]

Wow this is pretty cool, thanks! Going to tinker with this some more when I have time. I wonder if we can even go further and share the in-memory data object without copying.

[JanMarvin]

Hm, I tried a bit, but apparently this is not so simple due to the sandbox in v8. All the hints I could find, caused the sandbox to bark while terminating the session due to a potential security violation.

But I asked Gemini to check the existing functions, and this is what I see

tmp <- tempfile(fileext = ".js")
curl::curl_download(
  url = "https://unpkg.com/underscore@1.13.7/underscore-min.js",
  destfile = tmp)

test <- function(backend = NULL, tmp = NULL) {
  ct <- V8::v8(backend = backend)
  ct$source(tmp)
  ct$call("_.filter", nycflights13::flights, V8::JS("function(x){return x.arr_delay > 720}"))
}

res <- microbenchmark::microbenchmark(
  test(backend = "arrow", tmp = tmp),
  test(backend = "jsonlite", tmp = tmp),
  times = 25, unit = "ms"
); res
#> Warning in microbenchmark::microbenchmark(test(backend = "arrow", tmp = tmp), :
#> less accurate nanosecond times to avoid potential integer overflows
#> Unit: milliseconds
#>                                   expr       min        lq      mean   median
#>     test(backend = "arrow", tmp = tmp)  287.8694  303.9558  354.6052  316.310
#>  test(backend = "jsonlite", tmp = tmp) 2621.3725 2676.6829 2761.8196 2703.873
#>         uq      max neval
#>   322.2923 1267.931    25
#>  2750.2416 4039.942    25

But this is a better comparison, the call function above is tweaked to avoid the construction of the function with the entire data included in the function body. The following should be a fair one to one comparison. This creates identical objects in v8, so both are directly comparable. Therefore the arrow backend creates a table from the Arrow table sourced from the ipc stream and when importing it creates an Arrow table to return the ipc stream.

test <- function(backend = NULL) {
  ct <- V8::v8(backend = backend)
  ct$assign("flights", nycflights13::flights)
  ct$get("flights")
}

res <- microbenchmark::microbenchmark(
  test(backend = "arrow"),
  test(backend = "jsonlite"),
  times = 5, unit = "ms"
); res
#> Warning in microbenchmark::microbenchmark(test(backend = "arrow"), test(backend
#> = "jsonlite"), : less accurate nanosecond times to avoid potential integer
#> overflows
#> Unit: milliseconds
#>                        expr      min       lq     mean   median       uq
#>     test(backend = "arrow") 6986.343 7040.548 7451.106 7641.929 7677.567
#>  test(backend = "jsonlite") 7421.892 7603.722 7905.523 7674.219 7795.122
#>       max neval
#>  7909.143     5
#>  9032.658     5

[JanMarvin]

getting the data is the slow part (probably due to the fact that the conversion from javascript table to Arrow table is happening in js). Might be faster using a wasm Arrow.

test <- function(backend = NULL) {
  ct <- V8::v8(backend = backend)
  ct$assign("flights", nycflights13::flights)
  # ct$get("flights")
}

res <- microbenchmark::microbenchmark(
  test(backend = "arrow"),
  test(backend = "jsonlite"),
  times = 5, unit = "ms"
); res
#> Warning in microbenchmark::microbenchmark(test(backend = "arrow"), test(backend
#> = "jsonlite"), : less accurate nanosecond times to avoid potential integer
#> overflows
#> Unit: milliseconds
#>                        expr      min        lq      mean    median        uq
#>     test(backend = "arrow")  180.601  186.3381  384.6102  195.7693  219.9445
#>  test(backend = "jsonlite") 2322.650 2326.7051 2604.6440 2390.2020 2410.7080
#>       max neval
#>  1140.398     5
#>  3572.955     5

[JanMarvin]

avoiding json speeds things up

test <- function(backend = NULL) {
  ct <- V8::v8(backend = backend)
  ct$assign("flights", nycflights13::flights)
  ct$get("flights")
}

res <- microbenchmark::microbenchmark(
  test(backend = "arrow"),
  test(backend = "jsonlite"),
  times = 5, unit = "ms"
); res
#> Warning in microbenchmark::microbenchmark(test(backend = "arrow"), test(backend
#> = "jsonlite"), : less accurate nanosecond times to avoid potential integer
#> overflows
#> Unit: milliseconds
#>                        expr      min       lq     mean   median       uq
#>     test(backend = "arrow") 4645.056 4645.183 4741.127 4655.773 4667.055
#>  test(backend = "jsonlite") 7717.207 8181.782 8765.938 8965.483 9028.921
#>       max neval
#>  5092.569     5
#>  9936.294     5