What is JDF.jl?

JDF is a DataFrames serialization format with the following goals

• Fast save and load times
• Compressed storage on disk
• Enable disk-based data manipulation (not yet achieved)
• Supports machine learning workloads, e.g. mini-batch, sampling (not yet achieved)

JDF.jl is the Julia pacakge for all things related to JDF.

JDF stores a DataFrame in a folder with each column stored as a separate file. There is also a metadata.jls file that stores metadata about the original DataFrame. Collectively, the column files, the metadata file, and the folder is called a JDF "file".

JDF.jl is a pure-Julia solution and there are a lot of ways to do nifty things like compression and encapsulating the underlying struture of the arrays that's hard to do in R and Python. E.g. Python's numpy arrays are C objects, but all the vector types used in JDF are Julia data types.

The next major version of JDF will contain breaking changes. But don't worry I am fully committed to providing an automatic upgrade path. This means that you can safely use JDF.jl to save your data and not have to worry about the impending breaking change breaking all your JDF files.

using RDatasets, JDF, DataFrames

a = dataset("datasets", "iris");

first(a, 2)

2×5 DataFrame
 Row │ SepalLength  SepalWidth  PetalLength  PetalWidth  Species
     │ Float64      Float64     Float64      Float64     Cat…
─────┼───────────────────────────────────────────────────────────
   1 │         5.1         3.5          1.4         0.2  setosa
   2 │         4.9         3.0          1.4         0.2  setosa

By default JDF loads and saves DataFrames using multiple threads starting from Julia 1.3. For Julia < 1.3, it saves and loads using one thread only.

@time jdffile = JDF.save("iris.jdf", a)
@time a2 = DataFrame(JDF.load("iris.jdf"))

0.091923 seconds (157.33 k allocations: 9.226 MiB, 98.89% compilation tim
e)
  0.165332 seconds (197.31 k allocations: 11.476 MiB, 98.49% compilation ti
me)
150×5 DataFrame
 Row │ SepalLength  SepalWidth  PetalLength  PetalWidth  Species
     │ Float64      Float64     Float64      Float64     Cat…
─────┼─────────────────────────────────────────────────────────────
   1 │         5.1         3.5          1.4         0.2  setosa
   2 │         4.9         3.0          1.4         0.2  setosa
   3 │         4.7         3.2          1.3         0.2  setosa
   4 │         4.6         3.1          1.5         0.2  setosa
   5 │         5.0         3.6          1.4         0.2  setosa
   6 │         5.4         3.9          1.7         0.4  setosa
   7 │         4.6         3.4          1.4         0.3  setosa
   8 │         5.0         3.4          1.5         0.2  setosa
  ⋮  │      ⋮           ⋮            ⋮           ⋮           ⋮
 144 │         6.8         3.2          5.9         2.3  virginica
 145 │         6.7         3.3          5.7         2.5  virginica
 146 │         6.7         3.0          5.2         2.3  virginica
 147 │         6.3         2.5          5.0         1.9  virginica
 148 │         6.5         3.0          5.2         2.0  virginica
 149 │         6.2         3.4          5.4         2.3  virginica
 150 │         5.9         3.0          5.1         1.8  virginica
                                                   135 rows omitted

Simple checks for correctness

all(names(a2) .== names(a)) # true
all(skipmissing([all(a2[!,name] .== Array(a[!,name])) for name in names(a2)])) #true

true

You can load only a few columns from the dataset by specifying cols = [:column1, :column2]. For example

a2_selected = DataFrame(JDF.load("iris.jdf", cols = [:Species, :SepalLength, :PetalWidth]))

150×3 DataFrame
 Row │ SepalLength  PetalWidth  Species
     │ Float64      Float64     Cat…
─────┼────────────────────────────────────
   1 │         5.1         0.2  setosa
   2 │         4.9         0.2  setosa
   3 │         4.7         0.2  setosa
   4 │         4.6         0.2  setosa
   5 │         5.0         0.2  setosa
   6 │         5.4         0.4  setosa
   7 │         4.6         0.3  setosa
   8 │         5.0         0.2  setosa
  ⋮  │      ⋮           ⋮           ⋮
 144 │         6.8         2.3  virginica
 145 │         6.7         2.5  virginica
 146 │         6.7         2.3  virginica
 147 │         6.3         1.9  virginica
 148 │         6.5         2.0  virginica
 149 │         6.2         2.3  virginica
 150 │         5.9         1.8  virginica
                          135 rows omitted

The difference with loading the whole datasets and then subsetting the columns is that it saves time as only the selected columns are loaded from disk.

To take advatnage of some these convenience functions, you need to create a variable of type JDFFile pointed to the JDF file on disk. For example

jdf"path/to/JDF.jdf"

JDFFile{String}("path/to/JDF.jdf")

or

path_to_JDF = "path/to/JDF.jdf"
JDFFile(path_to_JDF)

JDFFile{String}("path/to/JDF.jdf")

You can load arbitrary col using the df[col] syntax. However, some of these operations are not yet optimized and hence may not be efficient.

afile = JDFFile("iris.jdf")

afile[:Species] # load Species column

150-element CategoricalArrays.CategoricalArray{String,1,UInt8}:
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 ⋮
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"

using Tables
ajdf = JDFFile("iris.jdf")
Tables.columnaccess(ajdf)

true

Tables.columns(ajdf)

JDFFile{String}("iris.jdf")

Tables.schema(ajdf)

Tables.Schema:
 :SepalLength  Float64
 :SepalWidth   Float64
 :PetalLength  Float64
 :PetalWidth   Float64
 :Species      CategoricalVector (alias for CategoricalArrays.CategoricalAr
ray{T, 1} where T)

getproperty(Tables.columns(ajdf), :Species)

150-element CategoricalArrays.CategoricalArray{String,1,UInt8}:
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 "setosa"
 ⋮
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"
 "virginica"

You can load each column of a JDF file from disk using iterations

jdffile = jdf"iris.jdf"
for col in eachcol(jdffile)
  # do something to col
  # where `col` is the content of one column of iris.jdf
end

To iterate through the columns names and the col

jdffile = jdf"iris.jdf"
for (name, col) in zip(names(jdffile), eachcol(jdffile))
  # `name::Symbol` is the name of the column
  #  `col` is the content of one column of iris.jdf
end

You can obtain the column names and number of columns ncol of a JDF, for example:

using JDF, DataFrames
df = DataFrame(a = 1:3, b = 1:3)
JDF.save(df, "plsdel.jdf")

names(jdf"plsdel.jdf") # [:a, :b]

# clean up
rm("plsdel.jdf", force = true, recursive = true)

DataFrame sizes are out of control. A 2GB CSV file can easily take up 10GB in RAM. One can use the function type_compress!(df) to compress any df::DataFrame. E.g.

type_compress!(df)

3×2 DataFrame
 Row │ a     b
     │ Int8  Int8
─────┼────────────
   1 │    1     1
   2 │    2     2
   3 │    3     3

The function looks at Int* columns and see if it can be safely "downgraded" to another Int* type with a smaller bits size. It will convert Float64 to Float32 if compress_float = true. E.g.

type_compress!(df, compress_float = true)

3×2 DataFrame
 Row │ a     b
     │ Int8  Int8
─────┼────────────
   1 │    1     1
   2 │    2     2
   3 │    3     3

String compression is planned and will likely employ categorical encoding combined with RLE encoding.

Here are some benchmarks using the Fannie Mae Mortgage Data. Please note that a reading of zero means that the method has failed to read or write.

JDF is a decent performer on both read and write and can achieve comparable performance to R's {fst}, once compiled. The JDF format also results in much smaller file size vs Feather.jl in this particular example (probably due to Feather.jl's inefficient storage of Union{String, Missing}).

Please note that the benchmarks were obtained on Julia 1.3+. On earlier versions of Julia where multi-threading isn't available, JDF is roughly 2x slower than as shown in the benchmarks.

I believe that restricting the types that JDF supports is vital for simplicity and maintainability.

There is support for

• WeakRefStrings.StringVector
• Vector{T}, Vector{Union{Mising, T}}, Vector{Union{Nothing, T}}
• CategoricalArrays.CategoricalVetors{T} and PooledArrays.PooledVector

where T can be String, Bool, Symbol, Char, SubString{String}, TimeZones.ZonedDateTime (experimental) and isbits types i.e. UInt*, Int*, and Float* Date* types etc.

RLEVectors support will be considered in the future when missing support arrives for RLEVectors.jl.

@bkamins's excellent DataFrames.jl tutorial contains a section on using JDF.jl.

When saving a JDF, each vector is Blosc compressed (using the default settings) if possible; this includes all T and Unions{Missing, T} types where T is isbits. For String vectors, they are first converted to a Run Length Encoding (RLE) representation, and the lengths component in the RLE are Blosc compressed.

I fully intend to develop JDF.jl into a language neutral format by version v0.4. However, I have other OSS commitments including R's {disk.frame} and hence new features might be slow to come onboard. But I am fully committed to making JDF files created using JDF.jl v0.2 or higher loadable in all future JDF.jl versions.

• Parallel read and write support is only available from Julia 1.3.
• The design of JDF was inspired by fst in terms of using compressions and allowing random-access to columns