fix deleteat! and subset! performance

[bkamins]

Explanation in https://discourse.julialang.org/t/learning-to-benchmark-and-find-the-best-function-to-select-a-subset-of-a-dataframe/91704/12.

Benchmarks

This PR

julia> using Random

julia> Random.seed!(1234)
TaskLocalRNG()

julia> df = DataFrame(rand(10^6, 100), :auto);

julia> df.id = rand(1:100, 10^6);

julia> inds = rand(Bool, 10^6);

julia> x = copy(df); @time deleteat!(x, inds);
  0.084989 seconds (29.05 k allocations: 1.471 MiB, 11.83% compilation time)

julia> x = copy(df); @time deleteat!(x, inds);
  0.074356 seconds (305 allocations: 4.766 KiB)

julia> x = copy(df); @time subset!(x, :x1 => Returns(inds));
  0.091337 seconds (485 allocations: 262.422 KiB)

julia> x = copy(df); @time subset!(x, :x1 => Returns(inds));
  0.148983 seconds (485 allocations: 262.422 KiB)

julia> x = groupby(copy(df), :id); @time subset!(x, :x1 => x -> rand(Bool, length(x)));     
  0.207012 seconds (1.15 M allocations: 63.651 MiB, 31.11% compilation time)

julia> x = groupby(copy(df), :id); @time subset!(x, :x1 => x -> rand(Bool, length(x)));     
  0.216079 seconds (1.15 M allocations: 63.632 MiB, 30.41% compilation time)

1.4.4 release

julia> using Random

julia> Random.seed!(1234)
TaskLocalRNG()

julia> df = DataFrame(rand(10^6, 100), :auto);

julia> df.id = rand(1:100, 10^6);

julia> inds = rand(Bool, 10^6);

julia> x = copy(df); @time deleteat!(x, inds);
  0.300327 seconds (307 allocations: 3.817 MiB)

julia> x = copy(df); @time deleteat!(x, inds);
  0.303013 seconds (307 allocations: 3.817 MiB)

julia> x = copy(df); @time subset!(x, :x1 => Returns(inds));
  0.307057 seconds (483 allocations: 4.074 MiB)

julia> x = copy(df); @time subset!(x, :x1 => Returns(inds));
  0.300665 seconds (483 allocations: 4.074 MiB)

julia> x = groupby(copy(df), :id); @time subset!(x, :x1 => x -> rand(Bool, length(x)));
  0.437468 seconds (1.15 M allocations: 67.427 MiB, 15.85% compilation time)

julia> x = groupby(copy(df), :id); @time subset!(x, :x1 => x -> rand(Bool, length(x)));
  0.417350 seconds (1.15 M allocations: 67.425 MiB, 15.70% compilation time)

[nalimilan]

Interesting. Have you checked with more columns and with a lower percentage of dropped rows? I would expect the findall approach to be less slow (and maybe faster) in these cases.

[bkamins]

I would expect the findall approach to be less slow (and maybe faster) in these cases.

You are right! 🧠

The threshold value I assessed empirically is less than 5% observations when it is better (it probably also depends on number of columns, but I wanted to have something relatively simple). I have proposed an adaptive algorithm switching between two approaches as needed.

[nalimilan]

Cool. Can you check when there are many columns? That's a use case that we care about.

[bkamins]

Here is an example: 100 columns, 10^6 rows. Tested with 5.5% rows to drop (so a bit above 5% threshold).

Setup:

df = DataFrame(rand(10^6, 100), :auto)

This PR:

julia> t = 0.055;

julia> Random.seed!(1234);

julia> idx = rand(10^6) .< t;

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.095975 seconds (303 allocations: 4.734 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.096654 seconds (303 allocations: 4.734 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.094744 seconds (303 allocations: 4.734 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.094374 seconds (303 allocations: 4.734 KiB)

Current release:

julia> t = 0.055;

julia> Random.seed!(1234);

julia> idx = rand(10^6) .< t;

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.092951 seconds (304 allocations: 432.203 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.087335 seconds (304 allocations: 432.203 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.091471 seconds (304 allocations: 432.203 KiB)

julia> dfc = copy(df); @time deleteat!(dfc, idx);
  0.092145 seconds (304 allocations: 432.203 KiB)

I did some more tests on even wider tables and it seems that a more precise threshold is 6% on my laptop, so I changed it to that value.

[nalimilan]

OK, great!

[bkamins]

Thank you!