Load and save DataFrames#

We do not cover all features of the packages. Please refer to their documentation to learn them.

Here we’ll load CSV.jl to read and write CSV files and Arrow.jl, JLSO.jl, and serialization, which allow us to work with a binary format and JSONTables.jl for JSON interaction. Finally we consider a custom JDF.jl format.

using DataFrames
using Arrow
using CSV
using Serialization
using JLSO
using JSONTables
using CodecZlib
using ZipFile
using JDF
using StatsPlots ## for charts
using Mmap ## for compression

Let’s create a simple DataFrame for testing purposes,

x = DataFrame(
    A=[true, false, true], B=[1, 2, missing],
    C=[missing, "b", "c"], D=['a', missing, 'c']
)
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc

and use eltypes to look at the columnwise types.

eltype.(eachcol(x))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, Char}

CSV.jl#

Let’s use CSV to save x to disk; make sure x1.csv does not conflict with some file in your working directory.

CSV.write("x1.csv", x)
"x1.csv"

Now we can see how it was saved by reading x.csv.

print(read("x1.csv", String))
A,B,C,D
true,1,,a
false,2,b,
true,,c,c

We can also load it back as a data frame

y = CSV.read("x1.csv", DataFrame)
3×4 DataFrame
RowABCD
BoolInt64?String1?String1?
1true1missinga
2false2bmissing
3truemissingcc

Note that when loading in a DataFrame from a CSV the column type for columns :C :D have changed to use special strings defined in the InlineStrings.jl package.

eltype.(eachcol(y))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, InlineStrings.String1}
 Union{Missing, InlineStrings.String1}

Serialization by JDF.jl and JLSO.jl#

Now we use serialization to save x.

There are two ways to perform serialization. The first way is to use the Serialization.serialize as below:

Note that in general, this process will not work if the reading and writing are done by different versions of Julia, or an instance of Julia with a different system image.

open("x.bin", "w") do io
    serialize(io, x)
end

Now we load back the saved file to y variable. Again y is identical to x. However, please beware that if you session does not have DataFrames.jl loaded, then it may not recognize the content as DataFrames.jl

y = open(deserialize, "x.bin")
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc
eltype.(eachcol(y))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, Char}

JDF.jl#

JDF.jl is a relatively new package designed to serialize DataFrames. You can save a DataFrame with the savejdf function. For more details about design assumptions and limitations of JDF.jl please check out xiaodaigh/JDF.jl.

JDF.save("x.jdf", x);

To load the saved JDF file, one can use the loadjdf function

x_loaded = JDF.load("x.jdf") |> DataFrame
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc

You can see that they are the same

isequal(x_loaded, x)
true

JDF.jl offers the ability to load only certain columns from disk to help with working with large files. set up a JDFFile which is a on disk representation of x backed by JDF.jl

x_ondisk = jdf"x.jdf"
JDF.JDFFile{String}("x.jdf")

We can see all the names of x without loading it into memory

names(x_ondisk)
4-element Vector{Symbol}:
 :A
 :B
 :C
 :D

The below is an example of how to load only columns :A and :D

xd = JDF.load(x_ondisk; cols=["A", "D"]) |> DataFrame
3×2 DataFrame
RowAD
BoolChar?
1truea
2falsemissing
3truec

JLSO.jl#

Another way to perform serialization is by using the JLSO.jl library:

JLSO.save("x.jlso", :data => x)

Now we can load back the file to y

y = JLSO.load("x.jlso")[:data]
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc
eltype.(eachcol(y))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, Char}

JSONTables.jl#

Often you might need to read and write data stored in JSON format. JSONTables.jl provides a way to process them in row-oriented or column-oriented layout. We present both options below.

open(io -> arraytable(io, x), "x1.json", "w")
106
open(io -> objecttable(io, x), "x2.json", "w")
76
print(read("x1.json", String))
[{"A":true,"B":1,"C":null,"D":"a"},{"A":false,"B":2,"C":"b","D":null},{"A":true,"B":null,"C":"c","D":"c"}]
print(read("x2.json", String))
{"A":[true,false,true],"B":[1,2,null],"C":[null,"b","c"],"D":["a",null,"c"]}
y1 = open(jsontable, "x1.json") |> DataFrame
3×4 DataFrame
RowABCD
BoolInt64?String?String?
1true1missinga
2false2bmissing
3truemissingcc
eltype.(eachcol(y1))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, String}
y2 = open(jsontable, "x2.json") |> DataFrame
3×4 DataFrame
RowABCD
BoolInt64?String?String?
1true1missinga
2false2bmissing
3truemissingcc
eltype.(eachcol(y2))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, String}

Arrow.jl#

Finally we use Apache Arrow format that allows, in particular, for data interchange with R or Python.

Arrow.write("x.arrow", x)
"x.arrow"
y = Arrow.Table("x.arrow") |> DataFrame
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc
eltype.(eachcol(y))
4-element Vector{Type}:
 Bool
 Union{Missing, Int64}
 Union{Missing, String}
 Union{Missing, Char}

Note that columns of y are immutable

try
    y.A[1] = false
catch e
    show(e)
end
ReadOnlyMemoryError()

This is because Arrow.Table uses memory mapping and thus uses a custom vector types:

y.A
3-element Arrow.BoolVector{Bool}:
 1
 0
 1
y.B
3-element Arrow.Primitive{Union{Missing, Int64}, Vector{Int64}}:
 1
 2
  missing

You can get standard Julia Base vectors by copying a data frame

y2 = copy(y)
3×4 DataFrame
RowABCD
BoolInt64?String?Char?
1true1missinga
2false2bmissing
3truemissingcc
y2.A
3-element Vector{Bool}:
 1
 0
 1
y2.B
3-element Vector{Union{Missing, Int64}}:
 1
 2
  missing

Basic benchmarking#

Next, we’ll create some files, so be careful that you don’t already have these files in your working directory! In particular, we’ll time how long it takes us to write a DataFrame with 1000 rows and 100000 columns.

bigdf = DataFrame(rand(Bool, 10^4, 1000), :auto)

bigdf[!, 1] = Int.(bigdf[!, 1])
bigdf[!, 2] = bigdf[!, 2] .+ 0.5
bigdf[!, 3] = string.(bigdf[!, 3], ", as string")

println("First run")
First run
println("CSV.jl")
csvwrite1 = @elapsed @time CSV.write("bigdf1.csv", bigdf)
println("Serialization")
serializewrite1 = @elapsed @time open(io -> serialize(io, bigdf), "bigdf.bin", "w")
println("JDF.jl")
jdfwrite1 = @elapsed @time JDF.save("bigdf.jdf", bigdf)
println("JLSO.jl")
jlsowrite1 = @elapsed @time JLSO.save("bigdf.jlso", :data => bigdf)
println("Arrow.jl")
arrowwrite1 = @elapsed @time Arrow.write("bigdf.arrow", bigdf)
println("JSONTables.jl arraytable")
jsontablesawrite1 = @elapsed @time open(io -> arraytable(io, bigdf), "bigdf1.json", "w")
println("JSONTables.jl objecttable")
jsontablesowrite1 = @elapsed @time open(io -> objecttable(io, bigdf), "bigdf2.json", "w")
println("Second run")
println("CSV.jl")
csvwrite2 = @elapsed @time CSV.write("bigdf1.csv", bigdf)
println("Serialization")
serializewrite2 = @elapsed @time open(io -> serialize(io, bigdf), "bigdf.bin", "w")
println("JDF.jl")
jdfwrite2 = @elapsed @time JDF.save("bigdf.jdf", bigdf)
println("JLSO.jl")
jlsowrite2 = @elapsed @time JLSO.save("bigdf.jlso", :data => bigdf)
println("Arrow.jl")
arrowwrite2 = @elapsed @time Arrow.write("bigdf.arrow", bigdf)
println("JSONTables.jl arraytable")
jsontablesawrite2 = @elapsed @time open(io -> arraytable(io, bigdf), "bigdf1.json", "w")
println("JSONTables.jl objecttable")
jsontablesowrite2 = @elapsed @time open(io -> objecttable(io, bigdf), "bigdf2.json", "w")
CSV.jl
  8.713840 seconds (44.75 M allocations: 1.129 GiB, 1.41% gc time, 83.27% compilation time)
Serialization
  0.188762 seconds (253.52 k allocations: 11.954 MiB, 30.98% compilation time)
JDF.jl
  0.143229 seconds (84.33 k allocations: 148.583 MiB, 16.20% gc time, 52.44% compilation time)
JLSO.jl
  1.061989 seconds (315.27 k allocations: 22.427 MiB, 7.30% compilation time)
Arrow.jl
  5.948886 seconds (6.01 M allocations: 294.596 MiB, 0.49% gc time, 98.41% compilation time)
JSONTables.jl arraytable
 11.182910 seconds (229.63 M allocations: 5.497 GiB, 14.37% gc time, 0.09% compilation time)
JSONTables.jl objecttable
  0.519673 seconds (103.97 k allocations: 309.370 MiB, 50.90% gc time, 14.75% compilation time)
Second run
CSV.jl
  1.452166 seconds (44.40 M allocations: 1.113 GiB, 9.13% gc time)
Serialization
  0.141294 seconds (15.13 k allocations: 396.984 KiB, 6.67% compilation time)
JDF.jl
  0.100376 seconds (29.14 k allocations: 145.796 MiB, 18.20% gc time)
JLSO.jl
  0.975589 seconds (36.92 k allocations: 8.933 MiB)
Arrow.jl
  0.093417 seconds (80.88 k allocations: 5.165 MiB)
JSONTables.jl arraytable
 11.492409 seconds (229.63 M allocations: 5.497 GiB, 14.29% gc time, 0.08% compilation time)
JSONTables.jl objecttable
  0.208298 seconds (17.83 k allocations: 305.113 MiB, 9.17% gc time, 4.53% compilation time)
0.208514974
groupedbar(
    repeat(["CSV.jl", "Serialization", "JDF.jl", "JLSO.jl", "Arrow.jl", "JSONTables.jl\nobjecttable"],
        inner=2),
    [csvwrite1, csvwrite2, serializewrite1, serializewrite1, jdfwrite1, jdfwrite2,
        jlsowrite1, jlsowrite2, arrowwrite1, arrowwrite2, jsontablesowrite2, jsontablesowrite2],
    group=repeat(["1st", "2nd"], outer=6),
    ylab="Second",
    title="Write Performance\nDataFrame: bigdf\nSize: $(size(bigdf))"
)
_images/942af183e1c21d14f58de1e66f627e8f42cb7f0e84b5af7eb26d808491359df1.png
data_files = ["bigdf1.csv", "bigdf.bin", "bigdf.arrow", "bigdf1.json", "bigdf2.json"]
df = DataFrame(file=data_files, size=getfield.(stat.(data_files), :size))
append!(df, DataFrame(file="bigdf.jdf", size=reduce((x, y) -> x + y.size,
    stat.(joinpath.("bigdf.jdf", readdir("bigdf.jdf"))),
    init=0)))
sort!(df, :size)
6×2 DataFrame
Rowfilesize
StringInt64
1bigdf.arrow1742946
2bigdf.bin5200078
3bigdf.jdf5219956
4bigdf1.csv55085511
5bigdf2.json55089512
6bigdf1.json124030619
@df df plot(:file, :size / 1024^2, seriestype=:bar, title="Format File Size (MB)", label="Size", ylab="MB")
_images/d344746b994aaf71f662dca15687e8a9658facf6b75e24ea389d630f5b222574.png
println("First run")
println("CSV.jl")
csvread1 = @elapsed @time CSV.read("bigdf1.csv", DataFrame)
println("Serialization")
serializeread1 = @elapsed @time open(deserialize, "bigdf.bin")
println("JDF.jl")
jdfread1 = @elapsed @time JDF.load("bigdf.jdf") |> DataFrame
println("JLSO.jl")
jlsoread1 = @elapsed @time JLSO.load("bigdf.jlso")
println("Arrow.jl")
arrowread1 = @elapsed @time df_tmp = Arrow.Table("bigdf.arrow") |> DataFrame
arrowread1copy = @elapsed @time copy(df_tmp)
println("JSONTables.jl arraytable")
jsontablesaread1 = @elapsed @time open(jsontable, "bigdf1.json")
println("JSONTables.jl objecttable")
jsontablesoread1 = @elapsed @time open(jsontable, "bigdf2.json")
println("Second run")
csvread2 = @elapsed @time CSV.read("bigdf1.csv", DataFrame)
println("Serialization")
serializeread2 = @elapsed @time open(deserialize, "bigdf.bin")
println("JDF.jl")
jdfread2 = @elapsed @time JDF.load("bigdf.jdf") |> DataFrame
println("JLSO.jl")
jlsoread2 = @elapsed @time JLSO.load("bigdf.jlso")
println("Arrow.jl")
arrowread2 = @elapsed @time df_tmp = Arrow.Table("bigdf.arrow") |> DataFrame
arrowread2copy = @elapsed @time copy(df_tmp)
println("JSONTables.jl arraytable")
jsontablesaread2 = @elapsed @time open(jsontable, "bigdf1.json")
println("JSONTables.jl objecttable")
jsontablesoread2 = @elapsed @time open(jsontable, "bigdf2.json");
First run
CSV.jl
  2.875686 seconds (4.50 M allocations: 228.116 MiB, 0.84% gc time, 96.27% compilation time)
Serialization
  0.427552 seconds (9.49 M allocations: 155.522 MiB, 9.72% gc time, 8.69% compilation time)
JDF.jl
  0.145266 seconds (152.49 k allocations: 157.504 MiB, 12.09% gc time, 106.92% compilation time)
JLSO.jl
  0.338389 seconds (9.51 M allocations: 158.192 MiB, 6.01% gc time, 8.65% compilation time)
Arrow.jl
  0.501657 seconds (574.86 k allocations: 27.032 MiB, 98.51% compilation time)
  0.057483 seconds (14.02 k allocations: 10.299 MiB)
JSONTables.jl arraytable
  6.506571 seconds (271.10 k allocations: 1.772 GiB, 9.70% gc time)
JSONTables.jl objecttable
  0.383006 seconds (7.39 k allocations: 566.958 MiB, 3.40% gc time, 0.03% compilation time)
Second run
  0.911447 seconds (630.07 k allocations: 43.419 MiB, 0.71% gc time)
Serialization
  0.378857 seconds (9.47 M allocations: 154.451 MiB, 14.31% gc time)
JDF.jl
  0.333889 seconds (64.81 k allocations: 153.252 MiB, 85.91% gc time)
JLSO.jl
  0.330859 seconds (9.49 M allocations: 157.185 MiB, 5.60% gc time)
Arrow.jl
  0.006640 seconds (84.09 k allocations: 3.594 MiB)
  0.051059 seconds (14.02 k allocations: 10.299 MiB)
JSONTables.jl arraytable
  6.577269 seconds (271.10 k allocations: 1.772 GiB, 10.31% gc time)
JSONTables.jl objecttable
  0.356987 seconds (7.08 k allocations: 566.937 MiB, 2.21% gc time)

Exclude JSONTables due to much longer timing

groupedbar(
    repeat(["CSV.jl", "Serialization", "JDF.jl", "JLSO.jl", "Arrow.jl", "Arrow.jl\ncopy", ##"JSON\narraytable",
            "JSON\nobjecttable"], inner=2),
    [csvread1, csvread2, serializeread1, serializeread2, jdfread1, jdfread2, jlsoread1, jlsoread2,
        arrowread1, arrowread2, arrowread1 + arrowread1copy, arrowread2 + arrowread2copy,
        # jsontablesaread1, jsontablesaread2,
        jsontablesoread1, jsontablesoread2],
    group=repeat(["1st", "2nd"], outer=7),
    ylab="Second",
    title="Read Performance\nDataFrame: bigdf\nSize: $(size(bigdf))"
)
_images/d48aaf9f443966c9fe10c4bd65414001450b4a6847b58c4addb3cf5cd610c96a.png

Using gzip compression#

A common user requirement is to be able to load and save CSV that are compressed using gzip. Below we show how this can be accomplished using CodecZlib.jl. The same pattern is applicable to JSONTables.jl compression/decompression. Again make sure that you do not have file named df_compress_test.csv.gz in your working directory. We first generate a random data frame.

df = DataFrame(rand(1:10, 10, 1000), :auto)
10×1000 DataFrame
900 columns omitted
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20x21x22x23x24x25x26x27x28x29x30x31x32x33x34x35x36x37x38x39x40x41x42x43x44x45x46x47x48x49x50x51x52x53x54x55x56x57x58x59x60x61x62x63x64x65x66x67x68x69x70x71x72x73x74x75x76x77x78x79x80x81x82x83x84x85x86x87x88x89x90x91x92x93x94x95x96x97x98x99x100
Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64
1107156781641823109107923539102691417272343657933553179829396104639818917102766868525998778710464729710642845
29578381039471068917410964938843132181971498787418851541426324791128841792831225525545418658115681288472
31010214771310951055173635593532107725419176881010101251061018110567433344952132112449241010594558759651064987732210
4119112108104812987411010158944978428855777633826411491672104312245184881681014365934319555893103893101277911
557578224818101513915839644559367955559825166574778343275736476981643537109722426693372105339310424958634
61075643717278419831921317197833210929997837615399211058103871025984265776973599101531436596108825108962692107
726279226710617845817351311041510310103107552331038756514510771864484967993107461264374785568337169510925444332
81015391371684991041010921258679368975797529781946972810979410617492917276326108931031035110710410144459756953156
910102105102931087326627748846183544531155149771011089418616648873749177326710821010931096686966832625971619101062
1072234107465461039233217197145622110548582101122361010334671651224573523491061710264389284886171565546181031025

GzipCompressorStream comes from CodecZlib

open("df_compress_test.csv.gz", "w") do io
    stream = GzipCompressorStream(io)
    CSV.write(stream, df)
    close(stream)
end
df2 = CSV.File(transcode(GzipDecompressor, Mmap.mmap("df_compress_test.csv.gz"))) |> DataFrame
10×1000 DataFrame
900 columns omitted
Rowx1x2x3x4x5x6x7x8x9x10x11x12x13x14x15x16x17x18x19x20x21x22x23x24x25x26x27x28x29x30x31x32x33x34x35x36x37x38x39x40x41x42x43x44x45x46x47x48x49x50x51x52x53x54x55x56x57x58x59x60x61x62x63x64x65x66x67x68x69x70x71x72x73x74x75x76x77x78x79x80x81x82x83x84x85x86x87x88x89x90x91x92x93x94x95x96x97x98x99x100
Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64Int64
1107156781641823109107923539102691417272343657933553179829396104639818917102766868525998778710464729710642845
29578381039471068917410964938843132181971498787418851541426324791128841792831225525545418658115681288472
31010214771310951055173635593532107725419176881010101251061018110567433344952132112449241010594558759651064987732210
4119112108104812987411010158944978428855777633826411491672104312245184881681014365934319555893103893101277911
557578224818101513915839644559367955559825166574778343275736476981643537109722426693372105339310424958634
61075643717278419831921317197833210929997837615399211058103871025984265776973599101531436596108825108962692107
726279226710617845817351311041510310103107552331038756514510771864484967993107461264374785568337169510925444332
81015391371684991041010921258679368975797529781946972810979410617492917276326108931031035110710410144459756953156
910102105102931087326627748846183544531155149771011089418616648873749177326710821010931096686966832625971619101062
1072234107465461039233217197145622110548582101122361010334671651224573523491061710264389284886171565546181031025
df == df2
true

Using zip files#

Sometimes you may have files compressed inside a zip file. In such a situation you may use ZipFile.jl in conjunction an an appropriate reader to read the files. Here we first create a ZIP file and then read back its contents into a DataFrame.

df1 = DataFrame(rand(1:10, 3, 4), :auto)
3×4 DataFrame
Rowx1x2x3x4
Int64Int64Int64Int64
111065
27469
32722
df2 = DataFrame(rand(1:10, 3, 4), :auto)
3×4 DataFrame
Rowx1x2x3x4
Int64Int64Int64Int64
178102
210369
38195

And we show yet another way to write a DataFrame into a CSV file: Writing a CSV file into the zip file

w = ZipFile.Writer("x.zip")

f1 = ZipFile.addfile(w, "x1.csv")
write(f1, sprint(show, "text/csv", df1))

# write a second CSV file into zip file
f2 = ZipFile.addfile(w, "x2.csv", method=ZipFile.Deflate)
write(f2, sprint(show, "text/csv", df2))

close(w)

Now we read the compressed CSV file we have written:

z = ZipFile.Reader("x.zip");
# find the index index of file called x1.csv
index_xcsv = findfirst(x -> x.name == "x1.csv", z.files)
# to read the x1.csv file in the zip file
df1_2 = CSV.read(read(z.files[index_xcsv]), DataFrame)
3×4 DataFrame
Rowx1x2x3x4
Int64Int64Int64Int64
111065
27469
32722
df1_2 == df1
true
# find the index index of file called x2.csv
index_xcsv = findfirst(x -> x.name == "x2.csv", z.files)
# to read the x2.csv file in the zip file
df2_2 = CSV.read(read(z.files[index_xcsv]), DataFrame)
3×4 DataFrame
Rowx1x2x3x4
Int64Int64Int64Int64
178102
210369
38195
df2_2 == df2
true

Note that once you read a given file from z object its stream is all used-up (reaching its end). Therefore to read it again you need to close the file object z and open it again. Also do not forget to close the zip file once you are done.

close(z)

Remove generated files

rm("x.arrow")
rm("x.bin")
rm("x.zip")
rm("x.jlso")
rm("x1.csv")
rm("x1.json")
rm("x2.json")
rm("x.jdf", recursive=true)
rm("bigdf.jdf", recursive=true)
rm("df_compress_test.csv.gz")
rm("bigdf1.json")
rm("bigdf1.csv")
rm("bigdf2.json")
rm("bigdf.jlso")
rm("bigdf.bin")
rm("bigdf.arrow")

This notebook was generated using Literate.jl.