Performance tips#

using DataFrames
using BenchmarkTools
using CategoricalArrays
using PooledArrays
using Random

Access by column number is faster than by name#

x = DataFrame(rand(5, 1000), :auto)
@btime $x[!, 500]; ## Faster
  3.095 ns (0 allocations: 0 bytes)
@btime $x.x500;  ## Slower
  11.272 ns (0 allocations: 0 bytes)

When working with data DataFrame use barrier functions or type annotation#

function f_bad() ## this function will be slow
    Random.seed!(1)
    x = DataFrame(rand(1000000, 2), :auto)
    y, z = x[!, 1], x[!, 2]
    p = 0.0
    for i in 1:nrow(x)
        p += y[i] * z[i]
    end
    p
end

@btime f_bad();
# if you run @code_warntype f_bad() then you notice
# that Julia does not know column types of `DataFrame`
  107.068 ms (5999022 allocations: 122.06 MiB)

solution 1 is to use barrier function (it should be possible to use it in almost any code) for the calculation. You will notice much less memopry allocations and faster performance.

function f_inner(y, z)
    p = 0.0
    for i in eachindex(y, z)
        p += y[i] * z[i]
    end
    p
end

function f_barrier()
    Random.seed!(1)
    x = DataFrame(rand(1000000, 2), :auto)
    f_inner(x[!, 1], x[!, 2])
end

@btime f_barrier();
  3.802 ms (44 allocations: 30.52 MiB)

or use inbuilt function if possible

using LinearAlgebra

function f_inbuilt()
    Random.seed!(1)
    x = DataFrame(rand(1000000, 2), :auto)
    dot(x[!, 1], x[!, 2])
end

@btime f_inbuilt();
  3.197 ms (44 allocations: 30.52 MiB)

solution 2 is to provide the types of extracted columns. However, there are cases in which you will not know these types.

function f_typed()
    Random.seed!(1)
    x = DataFrame(rand(1000000, 2), :auto)
    y::Vector{Float64}, z::Vector{Float64} = x[!, 1], x[!, 2]
    p = 0.0
    for i in 1:nrow(x)
        p += y[i] * z[i]
    end
    p
end

@btime f_typed();
  3.853 ms (44 allocations: 30.52 MiB)

In general for tall and narrow tables it is often useful to use Tables.rowtable, Tables.columntable or Tables.namedtupleiterator for intermediate processing of data in a type-stable way.

Consider using delayed DataFrame creation technique#

also notice the difference in performance between copying vs non-copying data frame creation

function f1()
    x = DataFrame([Vector{Float64}(undef, 10^4) for i in 1:100], :auto, copycols=false) ## we work with a DataFrame directly
    for c in 1:ncol(x)
        d = x[!, c]
        for r in 1:nrow(x)
            d[r] = rand()
        end
    end
    x
end

function f1a()
    x = DataFrame([Vector{Float64}(undef, 10^4) for i in 1:100], :auto) ## we work with a DataFrame directly
    for c in 1:ncol(x)
        d = x[!, c]
        for r in 1:nrow(x)
            d[r] = rand()
        end
    end
    x
end

function f2()
    x = Vector{Any}(undef, 100)
    for c in 1:length(x)
        d = Vector{Float64}(undef, 10^4)
        for r in eachindex(d)
            d[r] = rand()
        end
        x[c] = d
    end
    DataFrame(x, :auto, copycols=false) ## we delay creation of DataFrame after we have our job done
end

function f2a()
    x = Vector{Any}(undef, 100)
    for c in eachindex(x)
        d = Vector{Float64}(undef, 10^4)
        for r in eachindex(d)
            d[r] = rand()
        end
        x[c] = d
    end
    DataFrame(x, :auto) ## we delay creation of DataFrame after we have our job done
end

@btime f1();
@btime f1a();
@btime f2();
@btime f2a();
  27.952 ms (1949728 allocations: 37.40 MiB)
  28.930 ms (1950028 allocations: 45.03 MiB)
  1.160 ms (728 allocations: 7.66 MiB)
  1.605 ms (1028 allocations: 15.29 MiB)

You can add rows to a DataFrame in place and it is fast#

x = DataFrame(rand(10^6, 5), :auto)
y = DataFrame(transpose(1.0:5.0), :auto)
z = [1.0:5.0;]

@btime vcat($x, $y); ## creates a new DataFrame - slow
@btime append!($x, $y); ## in place - fast

x = DataFrame(rand(10^6, 5), :auto) ## reset to the same starting point
@btime push!($x, $z); ## add a single row in place - fast
  2.517 ms (212 allocations: 38.16 MiB)
  1.168 μs (29 allocations: 1.50 KiB)
  431.357 ns (16 allocations: 256 bytes)

Allowing missing as well as categorical slows down computations#

using StatsBase

function test(data) ## uses countmap function to test performance
    println(eltype(data))
    x = rand(data, 10^6)
    y = categorical(x)
    println(" raw:")
    @btime countmap($x)
    println(" categorical:")
    @btime countmap($y)
    nothing
end

test(1:10)
test([randstring() for i in 1:10])
test(allowmissing(1:10))
test(allowmissing([randstring() for i in 1:10]))
Int64
 raw:
  1.710 ms (8 allocations: 7.63 MiB)
 categorical:
  15.485 ms (1000004 allocations: 30.52 MiB)
String
 raw:
  21.243 ms (4 allocations: 448 bytes)
 categorical:
  31.704 ms (1000004 allocations: 30.52 MiB)
Union{Missing, Int64}
 raw:
  6.009 ms (4 allocations: 464 bytes)
 categorical:
  15.617 ms (1000004 allocations: 30.52 MiB)
Union{Missing, String}
 raw:
  21.557 ms (4 allocations: 448 bytes)
 categorical:
  32.430 ms (1000004 allocations: 30.52 MiB)

When aggregating use column selector and prefer integer, categorical, or pooled array grouping variable#

df = DataFrame(x=rand('a':'d', 10^7), y=1);
gdf = groupby(df, :x)

GroupedDataFrame with 4 groups based on key: x

First Group (2499512 rows): x = 'a': ASCII/Unicode U+0061 (category Ll: Letter, lowercase)
2499487 rows omitted
Rowxy
CharInt64
1a1
2a1
3a1
4a1
5a1
6a1
7a1
8a1
9a1
10a1
11a1
12a1
13a1
2499501a1
2499502a1
2499503a1
2499504a1
2499505a1
2499506a1
2499507a1
2499508a1
2499509a1
2499510a1
2499511a1
2499512a1

Last Group (2500217 rows): x = 'c': ASCII/Unicode U+0063 (category Ll: Letter, lowercase)
2500192 rows omitted
Rowxy
CharInt64
1c1
2c1
3c1
4c1
5c1
6c1
7c1
8c1
9c1
10c1
11c1
12c1
13c1
2500206c1
2500207c1
2500208c1
2500209c1
2500210c1
2500211c1
2500212c1
2500213c1
2500214c1
2500215c1
2500216c1
2500217c1

traditional syntax, slow

@btime combine(v -> sum(v.y), $gdf)
  15.907 ms (332 allocations: 19.08 MiB)
4×2 DataFrame
Rowxx1
CharInt64
1a2499512
2d2500017
3b2500254
4c2500217

use column selector

@btime combine($gdf, :y => sum)
  6.867 ms (198 allocations: 10.14 KiB)
4×2 DataFrame
Rowxy_sum
CharInt64
1a2499512
2d2500017
3b2500254
4c2500217
transform!(df, :x => categorical => :x);
gdf = groupby(df, :x)

GroupedDataFrame with 4 groups based on key: x

First Group (2499512 rows): x = CategoricalArrays.CategoricalValue{Char, UInt32} 'a'
2499487 rows omitted
Rowxy
Cat…Int64
1a1
2a1
3a1
4a1
5a1
6a1
7a1
8a1
9a1
10a1
11a1
12a1
13a1
2499501a1
2499502a1
2499503a1
2499504a1
2499505a1
2499506a1
2499507a1
2499508a1
2499509a1
2499510a1
2499511a1
2499512a1

Last Group (2500017 rows): x = CategoricalArrays.CategoricalValue{Char, UInt32} 'd'
2499992 rows omitted
Rowxy
Cat…Int64
1d1
2d1
3d1
4d1
5d1
6d1
7d1
8d1
9d1
10d1
11d1
12d1
13d1
2500006d1
2500007d1
2500008d1
2500009d1
2500010d1
2500011d1
2500012d1
2500013d1
2500014d1
2500015d1
2500016d1
2500017d1
@btime combine($gdf, :y => sum)
  6.851 ms (206 allocations: 10.62 KiB)
4×2 DataFrame
Rowxy_sum
Cat…Int64
1a2499512
2b2500254
3c2500217
4d2500017
transform!(df, :x => PooledArray{Char} => :x)
10000000×2 DataFrame
9999975 rows omitted
Rowxy
CharInt64
1a1
2a1
3d1
4b1
5a1
6d1
7a1
8b1
9c1
10b1
11a1
12c1
13c1
9999989c1
9999990c1
9999991d1
9999992c1
9999993b1
9999994b1
9999995d1
9999996c1
9999997d1
9999998c1
9999999d1
10000000d1
gdf = groupby(df, :x)

GroupedDataFrame with 4 groups based on key: x

First Group (2499512 rows): x = 'a': ASCII/Unicode U+0061 (category Ll: Letter, lowercase)
2499487 rows omitted
Rowxy
CharInt64
1a1
2a1
3a1
4a1
5a1
6a1
7a1
8a1
9a1
10a1
11a1
12a1
13a1
2499501a1
2499502a1
2499503a1
2499504a1
2499505a1
2499506a1
2499507a1
2499508a1
2499509a1
2499510a1
2499511a1
2499512a1

Last Group (2500217 rows): x = 'c': ASCII/Unicode U+0063 (category Ll: Letter, lowercase)
2500192 rows omitted
Rowxy
CharInt64
1c1
2c1
3c1
4c1
5c1
6c1
7c1
8c1
9c1
10c1
11c1
12c1
13c1
2500206c1
2500207c1
2500208c1
2500209c1
2500210c1
2500211c1
2500212c1
2500213c1
2500214c1
2500215c1
2500216c1
2500217c1
@btime combine($gdf, :y => sum)
  6.876 ms (200 allocations: 10.20 KiB)
4×2 DataFrame
Rowxy_sum
CharInt64
1a2499512
2d2500017
3b2500254
4c2500217

Use views instead of materializing a new DataFrame#

x = DataFrame(rand(100, 1000), :auto)
@btime $x[1:1, :]
  176.158 μs (3993 allocations: 159.03 KiB)
1×1000 DataFrame
900 columns omitted
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20x21x22x23x24x25x26x27x28x29x30x31x32x33x34x35x36x37x38x39x40x41x42x43x44x45x46x47x48x49x50x51x52x53x54x55x56x57x58x59x60x61x62x63x64x65x66x67x68x69x70x71x72x73x74x75x76x77x78x79x80x81x82x83x84x85x86x87x88x89x90x91x92x93x94x95x96x97x98x99x100
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.379780.4111870.8562930.7990660.390610.4859910.7122360.4111440.3423150.9621440.6840080.8926290.5122540.5769160.4418840.2864930.1033420.1434810.5925310.7752730.4346350.147430.1359480.4062040.1281720.9486330.7271550.4131540.7326880.3865140.1207650.05216720.1785280.7081480.4574370.3124590.01577150.3991610.6258630.2207480.865960.5594260.7448710.6839430.9624640.9471220.05291240.9301450.246580.5599140.1356830.9835990.2243410.808230.9789090.5555340.8829120.8690870.4446280.5456850.2151840.5717980.9109010.3774520.568790.4922640.4537110.3605860.3862660.9499540.4168630.3573120.3292120.3925360.03533490.492630.5001320.5091370.2302680.1797860.378619
@btime $x[1, :]
  22.833 ns (0 allocations: 0 bytes)
DataFrameRow (1000 columns)
900 columns omitted
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20x21x22x23x24x25x26x27x28x29x30x31x32x33x34x35x36x37x38x39x40x41x42x43x44x45x46x47x48x49x50x51x52x53x54x55x56x57x58x59x60x61x62x63x64x65x66x67x68x69x70x71x72x73x74x75x76x77x78x79x80x81x82x83x84x85x86x87x88x89x90x91x92x93x94x95x96x97x98x99x100
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.379780.4111870.8562930.7990660.390610.4859910.7122360.4111440.3423150.9621440.6840080.8926290.5122540.5769160.4418840.2864930.1033420.1434810.5925310.7752730.4346350.147430.1359480.4062040.1281720.9486330.7271550.4131540.7326880.3865140.1207650.05216720.1785280.7081480.4574370.3124590.01577150.3991610.6258630.2207480.865960.5594260.7448710.6839430.9624640.9471220.05291240.9301450.246580.5599140.1356830.9835990.2243410.808230.9789090.5555340.8829120.8690870.4446280.5456850.2151840.5717980.9109010.3774520.568790.4922640.4537110.3605860.3862660.9499540.4168630.3573120.3292120.3925360.03533490.492630.5001320.5091370.2302680.1797860.378619
@btime view($x, 1:1, :)
  22.833 ns (0 allocations: 0 bytes)
1×1000 SubDataFrame
900 columns omitted
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20x21x22x23x24x25x26x27x28x29x30x31x32x33x34x35x36x37x38x39x40x41x42x43x44x45x46x47x48x49x50x51x52x53x54x55x56x57x58x59x60x61x62x63x64x65x66x67x68x69x70x71x72x73x74x75x76x77x78x79x80x81x82x83x84x85x86x87x88x89x90x91x92x93x94x95x96x97x98x99x100
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.379780.4111870.8562930.7990660.390610.4859910.7122360.4111440.3423150.9621440.6840080.8926290.5122540.5769160.4418840.2864930.1033420.1434810.5925310.7752730.4346350.147430.1359480.4062040.1281720.9486330.7271550.4131540.7326880.3865140.1207650.05216720.1785280.7081480.4574370.3124590.01577150.3991610.6258630.2207480.865960.5594260.7448710.6839430.9624640.9471220.05291240.9301450.246580.5599140.1356830.9835990.2243410.808230.9789090.5555340.8829120.8690870.4446280.5456850.2151840.5717980.9109010.3774520.568790.4922640.4537110.3605860.3862660.9499540.4168630.3573120.3292120.3925360.03533490.492630.5001320.5091370.2302680.1797860.378619
@btime $x[1:1, 1:20]
  3.737 μs (70 allocations: 3.09 KiB)
1×20 DataFrame
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.37978
@btime $x[1, 1:20]
  23.447 ns (0 allocations: 0 bytes)
DataFrameRow (20 columns)
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.37978
@btime view($x, 1:1, 1:20)
  28.233 ns (0 allocations: 0 bytes)
1×20 SubDataFrame
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20
Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64Float64
10.2266990.8637130.1856640.9912530.4367370.3215610.220170.9973910.8894870.5214430.9585760.7357210.6246970.5609340.7058010.8025230.7487780.03863740.8320560.37978

This notebook was generated using Literate.jl.