<!–- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at


Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. –>

User Manual

The goal of this documentation is to provide a brief introduction to the arrow data format, then provide a walk-through of the functionality provided in the Arrow.jl Julia package, with an aim to expose a little of the machinery "under the hood" to help explain how things work and how that influences real-world use-cases for the arrow data format.

The best place to learn about the Apache arrow project is the website itself, specifically the data format specification. Put briefly, the arrow project provides a formal speficiation for how columnar, "table" data can be laid out efficiently in memory to standardize and maximize the ability to share data across languages/platforms. In the current apache/arrow GitHub repository, language implementations exist for C++, Java, Go, Javascript, Rust, to name a few. Other database vendors and data processing frameworks/applications have also built support for the arrow format, allowing for a wide breadth of possibility for applications to "speak the data language" of arrow.

The Arrow.jl Julia package is another implementation, allowing the ability to both read and write data in the arrow format. As a data format, arrow specifies an exact memory layout to be used for columnar table data, and as such, "reading" involves custom Julia objects (Arrow.Table and Arrow.Stream), which read the metadata of an "arrow memory blob", then wrap the array data contained therein, having learned the type and size, amongst other properties, from the metadata. Let's take a closer look at what this "reading" of arrow memory really means/looks like.

Support for generic path-like types

Arrow.jl attempts to support any path-like type wherever a function takes a path as an argument. The Arrow.jl API should generically work as long as the type supports:

  • Base.open(path, mode)::I where I <: IO

When a custom IO subtype is returned (I) then the following methods also need to be defined:

  • Base.read(io::I, ::Type{UInt8}) or Base.read(io::I)
  • Base.write(io::I, x)

Reading arrow data

After installing the Arrow.jl Julia package (via ] add Arrow), and if you have some arrow data, let's say a file named data.arrow generated from the pyarrow library (a Python library for interfacing with arrow data), you can then read that arrow data into a Julia session by doing:

using Arrow

table = Arrow.Table("data.arrow")


The type of table in this example will be an Arrow.Table. When "reading" the arrow data, Arrow.Table first "mmapped" the data.arrow file, which is an important technique for dealing with data larger than available RAM on a system. By "mmapping" a file, the OS doesn't actually load the entire file contents into RAM at the same time, but file contents are "swapped" into RAM as different regions of a file are requested. Once "mmapped", Arrow.Table then inspected the metadata in the file to determine the number of columns, their names and types, at which byte offset each column begins in the file data, and even how many "batches" are included in this file (arrow tables may be partitioned into one or more "record batches" each containing portions of the data). Armed with all the appropriate metadata, Arrow.Table then created custom array objects (Arrow.ArrowVector), which act as "views" into the raw arrow memory bytes. This is a significant point in that no extra memory is allocated for "data" when reading arrow data. This is in contrast to if we wanted to read data from a csv file as columns into Julia structures; we would need to allocate those array structures ourselves, then parse the file, "filling in" each element of the array with the data we parsed from the file. Arrow data, on the other hand, is already laid out in memory or on disk in a binary format, and as long as we have the metadata to interpret the raw bytes, we can figure out whether to treat those bytes as a Vector{Float64}, etc. A sample of the kinds of arrow array types you might see when deserializing arrow data, include:

  • Arrow.Primitive: the most common array type for simple, fixed-size elements like integers, floats, time types, and decimals
  • Arrow.List: an array type where its own elements are also arrays of some kind, like string columns, where each element can be thought of as an array of characters
  • Arrow.FixedSizeList: similar to the List type, but where each array element has a fixed number of elements itself; you can think of this like a Vector{NTuple{N, T}}, where N is the fixed-size width
  • Arrow.Map: an array type where each element is like a Julia Dict; a list of key value pairs like a Vector{Dict}
  • Arrow.Struct: an array type where each element is an instance of a custom struct, i.e. an ordered collection of named & typed fields, kind of like a Vector{NamedTuple}
  • Arrow.DenseUnion: an array type where elements may be of several different types, stored compactly; can be thought of like Vector{Union{A, B}}
  • Arrow.SparseUnion: another array type where elements may be of several different types, but stored as if made up of identically lengthed child arrays for each possible type (less memory efficient than DenseUnion)
  • Arrow.DictEncoded: a special array type where values are "dictionary encoded", meaning the list of unique, possible values for an array are stored internally in an "encoding pool", whereas each stored element of the array is just an integer "code" to index into the encoding pool for the actual value.

And while these custom array types do subtype AbstractArray, there is no current support for setindex!. Remember, these arrays are "views" into the raw arrow bytes, so for array types other than Arrow.Primitive, it gets pretty tricky to allow manipulating those raw arrow bytes. Nevetheless, it's as simple as calling copy(x) where x is any ArrowVector type, and a normal Julia Vector type will be fully materialized (which would then allow mutating/manipulating values).

So, what can you do with an Arrow.Table full of data? Quite a bit actually!

Because Arrow.Table implements the Tables.jl interface, it opens up a world of integrations for using arrow data. A few examples include:

  • df = DataFrame(Arrow.Table(file)): Build a DataFrame, using the arrow vectors themselves; this allows utilizing a host of DataFrames.jl functionality directly on arrow data; grouping, joining, selecting, etc.
  • df = copy(DataFrame(Arrow.Table(file))): Build a DataFrame, where the columns are regular in-memory vectors (specifically, Base.Vectors and/or PooledVectors). This requires that you have enough memory to load the entire DataFrame into memory.
  • Tables.datavaluerows(Arrow.Table(file)) |> @map(...) |> @filter(...) |> DataFrame: use Query.jl's row-processing utilities to map, group, filter, mutate, etc. directly over arrow data.
  • Arrow.Table(file) |> SQLite.load!(db, "arrow_table"): load arrow data directly into an sqlite database/table, where sql queries can be executed on the data
  • Arrow.Table(file) |> CSV.write("arrow.csv"): write arrow data out to a csv file

A full list of Julia packages leveraging the Tables.jl inteface can be found here.

Apart from letting other packages have all the fun, an Arrow.Table itself can be plenty useful. For example, with tbl = Arrow.Table(file):

  • tbl[1]: retrieve the first column via indexing; the number of columns can be queried via length(tbl)
  • tbl[:col1] or tbl.col1: retrieve the column named col1, either via indexing with the column name given as a Symbol, or via "dot-access"
  • for col in tbl: iterate through columns in the table
  • AbstractDict methods like haskey(tbl, :col1), get(tbl, :col1, nothing), keys(tbl), or values(tbl)

Arrow types

In the arrow data format, specific logical types are supported, a list of which can be found here. These include booleans, integers of various bit widths, floats, decimals, time types, and binary/string. While most of these map naturally to types builtin to Julia itself, there are a few cases where the definitions are slightly different, and in these cases, by default, they are converted to more "friendly" Julia types (this auto conversion can be avoided by passing convert=false to Arrow.Table, like Arrow.Table(file; convert=false)). Examples of arrow to julia type mappings include:

  • Date, Time, Timestamp, and Duration all have natural Julia defintions in Dates.Date, Dates.Time, TimeZones.ZonedDateTime, and Dates.Period subtypes, respectively.
  • Char and Symbol Julia types are mapped to arrow string types, with additional metadata of the original Julia type; this allows deserializing directly to Char and Symbol in Julia, while other language implementations will see these columns as just strings
  • Similarly to the above, the UUID Julia type is mapped to a 128-bit FixedSizeBinary arrow type.
  • Decimal128 and Decimal256 have no corresponding builtin Julia types, so they're deserialized using a compatible type definition in Arrow.jl itself: Arrow.Decimal

Note that when convert=false is passed, data will be returned in Arrow.jl-defined types that exactly match the arrow definitions of those types; the authoritative source for how each type represents its data can be found in the arrow Schema.fbs file.

One note on performance: when writing TimeZones.ZonedDateTime columns to the arrow format (via Arrow.write), it is preferrable to "wrap" the columns in Arrow.ToTimestamp(col), as long as the column has ZonedDateTime elements that all share a common timezone. This ensures the writing process can know "upfront" which timezone will be encoded and is thus much more efficient and performant.

Custom types

To support writing your custom Julia struct, Arrow.jl utilizes the format's mechanism for "extension types" by allowing the storing of Julia type name and metadata in the field metadata. To "hook in" to this machinery, custom types can utilize the interface methods defined in the Arrow.ArrowTypes submodule. For example:

using Arrow

struct Person

# overload interface method for custom type Person; return a symbol as the "name"
# this instructs Arrow.write what "label" to include with a column with this custom type
ArrowTypes.arrowname(::Type{Person}) = :Person
# overload JuliaType on `Val{:Person}`, which is like a dispatchable string
# return our custom *type* Person; this enables Arrow.Table to know how the "label"
# on a custom column should be mapped to a Julia type and deserialized
ArrowTypes.JuliaType(::Val{:Person}) = Person

table = (col1=[Person(1, "Bob"), Person(2, "Jane")],)
io = IOBuffer()
Arrow.write(io, table)
table2 = Arrow.Table(io)

In this example, we're writing our table, which is a NamedTuple with one column named col1, which has two elements which are instances of our custom Person struct. We overload Arrowtypes.arrowname so that Arrow.jl knows how to serialize our Person struct. We then overload ArrowTypes.JuliaType so the deserialization process knows how to map from our type label back to our Person struct type. We can then write our data in the arrow format to an in-memory IOBuffer, then read the table back in using Arrow.Table. The table we get back will be an Arrow.Table, with a single Arrow.Struct column with element type Person.

Note that without calling Arrowtypes.JuliaType, we may get into a weird limbo state where we've written our table with Person structs out as a table, but when reading back in, Arrow.jl doesn't know what a Person is; deserialization won't fail, but we'll just get a Namedtuple{(:id, :name), Tuple{Int, String}} back instead of Person.

While this example is very simple, it shows the basics to allow a custom type to be serialized/deserialized. But the ArrowTypes module offers even more powerful functionality for "hooking" non-native arrow types into the serialization/deserialization processes. Let's walk through a couple more examples; if you've had enough custom type shenanigans, feel free to skip to the next section.

Let's take a look at how Arrow.jl allows serializing the nothing value, which is often referred to as the "software engineer's NULL" in Julia. While Arrow.jl treats missing as the default arrow NULL value, nothing is pretty similar, but we'd still like to treat it separately if possible. Here's how we enable serialization/deserialization in the ArrowTypes module:

ArrowTypes.ArrowKind(::Type{Nothing}) = ArrowTypes.NullKind()
ArrowTypes.ArrowType(::Type{Nothing}) = Missing
ArrowTypes.toarrow(::Nothing) = missing
const NOTHING = Symbol("JuliaLang.Nothing")
ArrowTypes.arrowname(::Type{Nothing}) = NOTHING
ArrowTypes.JuliaType(::Val{NOTHING}) = Nothing
ArrowTypes.fromarrow(::Type{Nothing}, ::Missing) = nothing

Let's walk through what's going on here, line-by-line:

  • ArrowKind overload: ArrowKinds are generic "categories" of types supported by the arrow format, like PrimitiveKind, ListKind, etc. They each correspond to a different data layout strategy supported in the arrow format. Here, we define nothing's kind to be NullKind, which means no actual memory is needed for storage, it's strictly a "metadata" type where we store the type and # of elements. In our Person example, we didn't need to overload this since types declared like struct T or mutable struct T are defined as ArrowTypes.StructKind by default
  • ArrowType overload: here we're signaling that our type (Nothing) maps to the natively supported arrow type of Missing; this is important for the serializer so it knows which arrow type it will be serializing. Again, we didn't need to overload this for Person since the serializer knows how to serialize custom structs automatically by using reflection methods like fieldnames(T) and getfield(x, i).
  • ArrowTypes.toarrow overload: this is a sister method to ArrowType; we said our type will map to the Missing arrow type, so here we actually define ___how___ it converts to the arrow type; and in this case, it just returns missing. This is yet another method that didn't show up for Person; why? Well, as we noted in ArrowType, the serializer already knows how to serialize custom structs by using all their fields; if, for some reason, we wanted to omit some fields or otherwise transform things, then we could define corresponding ArrowType and toarrow methods
  • arrowname overload: similar to our Person example, we need to instruct the serializer how to label our custom type in the arrow type metadata; here we give it the symbol Symbol("JuliaLang.Nothing"). Note that while this will ultimately allow us to disambiguate nothing from missing when reading arrow data, if we pass this data to other language implementations, they will only treat the data as missing since they (probably) won't know how to "understand" the JuliaLang.Nothing type label
  • JuliaType overload: again, like our Person example, we instruct the deserializer that when it encounters the JuliaLang.Nothing type label, it should treat those values as Nothing type.
  • And finally, fromarrow overload: this allows specifying how the native-arrow data should be converted back to our custom type. fromarrow(T, x...) by default will call T(x...), which is why we didn't need this overload for Person, but in this example, Nothing(missing) won't work, so we define our own custom conversion.

Let's run through one more complex example, just for fun and to really see how far the system can be pushed:

using Intervals
table = (col = [
const NAME = Symbol("JuliaLang.Interval")
ArrowTypes.arrowname(::Type{Interval{T, L, R}}) where {T, L, R} = NAME
const LOOKUP = Dict(
    "Closed" => Closed,
    "Unbounded" => Unbounded
ArrowTypes.arrowmetadata(::Type{Interval{T, L, R}}) where {T, L, R} = string(L, ".", R)
function ArrowTypes.JuliaType(::Val{NAME}, ::Type{NamedTuple{names, types}}, meta) where {names, types}
    L, R = split(meta, ".")
    return Interval{fieldtype(types, 1), LOOKUP[L], LOOKUP[R]}
ArrowTypes.fromarrow(::Type{Interval{T, L, R}}, first, last) where {T, L, R} = Interval{L, R}(first, R == Unbounded ? nothing : last)
io = Arrow.tobuffer(table)
tbl = Arrow.Table(io)

Again, let's break down what's going on here:

  • Here we're trying to save an Interval type in the arrow format; this type is unique in that it has two type parameters (Closed and Unbounded) that are not inferred/based on fields, but are just "type tags" on the type itself
  • Note that we define a generic arrowname method on all Intervals, regardless of type parameters. We just want to let arrow know which general type we're dealing with here
  • Next we use a new method ArrowTypes.arrowmetadata to encode the two non-field-based type parameters as a string with a dot delimiter; we encode this information here because remember, we have to match our arrowname Symbol typename in our JuliaType(::Val(name)) definition in order to dispatch correctly; if we encoded the type parameters in arrowname, we would need separate arrowname definitions for each unique combination of those two type parameters, and corresponding JuliaType definitions for each as well; yuck. Instead, we let arrowname be generic to our type, and store the type parameters for this specific column using arrowmetadata
  • Now in JuliaType, note we're using the 3-argument overload; we want the NamedTuple type that is the native arrow type our Interval is being serialized as; we use this to retrieve the 1st type parameter for our Interval, which is simply the type of the two first and last fields. Then we use the 3rd argument, which is whatever string we returned from arrowmetadata. We call L, R = split(meta, ".") to parse the two type parameters (in this case Closed and Unbounded), then do a lookup on those strings from a predefined LOOKUP Dict that matches the type parameter name as string to the actual type. We then have all the information to recreate the full Interval type. Neat!
  • The one final wrinkle is in our fromarrow method; Intervals that are Unbounded, actually take nothing as the 2nd argument. So letting the default fromarrow definition call Interval{T, L, R}(first, last), where first and last are both integers isn't going to work. Instead, we check if the R type parameter is Unbounded and if so, pass nothing as the 2nd arg, otherwise we can pass last.

This stuff can definitely make your eyes glaze over if you stare at it long enough. As always, don't hesitate to reach out for quick questions on the #data slack channel, or open a new issue detailing what you're trying to do.


In addition to Arrow.Table, the Arrow.jl package also provides Arrow.Stream for processing arrow data. While Arrow.Table will iterate all record batches in an arrow file/stream, concatenating columns, Arrow.Stream provides a way to iterate through record batches, one at a time. Each iteration yields an Arrow.Table instance, with columns/data for a single record batch. This allows, if so desired, "batch processing" of arrow data, one record batch at a time, instead of creating a single long table via Arrow.Table.

Custom application metadata

The Arrow format allows data producers to attach custom metadata to various Arrow objects.

Arrow.jl provides a convenient accessor for this metadata via Arrow.getmetadata. Arrow.getmetadata(t::Arrow.Table) will return an immutable AbstractDict{String,String} that represents the custom_metadata of the table's associated Schema (or nothing if no such metadata exists), while Arrow.getmetadata(c::Arrow.ArrowVector) will return a similar representation of the column's associated Field custom_metadata (or nothing if no such metadata exists).

To attach custom schema/column metadata to Arrow tables at serialization time, see the metadata and colmetadata keyword arguments to Arrow.write.

Writing arrow data

Ok, so that's a pretty good rundown of reading arrow data, but how do you produce arrow data? Enter Arrow.write.


With Arrow.write, you provide either an io::IO argument or a file_path to write the arrow data to, as well as a Tables.jl-compatible source that contains the data to be written.

What are some examples of Tables.jl-compatible sources? A few examples include:

  • Arrow.write(io, df::DataFrame): A DataFrame is a collection of indexable columns
  • Arrow.write(io, CSV.File(file)): read data from a csv file and write out to arrow format
  • Arrow.write(io, DBInterface.execute(db, sql_query)): Execute an SQL query against a database via the DBInterface.jl interface, and write the query resultset out directly in the arrow format. Packages that implement DBInterface include SQLite.jl, MySQL.jl, and ODBC.jl.
  • df |> @map(...) |> Arrow.write(io): Write the results of a Query.jl chain of operations directly out as arrow data
  • jsontable(json) |> Arrow.write(io): Treat a json array of objects or object of arrays as a "table" and write it out as arrow data using the JSONTables.jl package
  • Arrow.write(io, (col1=data1, col2=data2, ...)): a NamedTuple of AbstractVectors or an AbstractVector of NamedTuples are both considered tables by default, so they can be quickly constructed for easy writing of arrow data if you already have columns of data

And these are just a few examples of the numerous integrations.

In addition to just writing out a single "table" of data as a single arrow record batch, Arrow.write also supports writing out multiple record batches when the input supports the Tables.partitions functionality. One immediate, though perhaps not incredibly useful example, is Arrow.Stream. Arrow.Stream implements Tables.partitions in that it iterates "tables" (specifically Arrow.Table), and as such, Arrow.write will iterate an Arrow.Stream, and write out each Arrow.Table as a separate record batch. Another important point for why this example works is because an Arrow.Stream iterates Arrow.Tables that all have the same schema. This is important because when writing arrow data, a "schema" message is always written first, with all subsequent record batches written with data matching the initial schema.

In addition to inputs that support Tables.partitions, note that the Tables.jl itself provides the Tables.partitioner function, which allows providing your own separate instances of similarly-schema-ed tables as "partitions", like:

# treat 2 separate NamedTuples of vectors with same schema as 1 table, 2 partitions
tbl_parts = Tables.partitioner([(col1=data1, col2=data2), (col1=data3, col2=data4)])
Arrow.write(io, tbl_parts)

# treat an array of csv files with same schema where each file is a partition
# in this form, a function `CSV.File` is applied to each element of 2nd argument
csv_parts = Tables.partitioner(CSV.File, csv_files)
Arrow.write(io, csv_parts)


With Arrow.Writer, you instantiate an Arrow.Writer object, write sources using it, and then close it. This allows for incrmental writes to the same sink. It is similar to Arrow.append without having to close and re-open the sink in between writes and without the limitation of only supporting the IPC stream format.

Multithreaded writing

By default, Arrow.write will use multiple threads to write multiple record batches simultaneously (e.g. if julia is started with julia -t 8 or the JULIA_NUM_THREADS environment variable is set). The number of concurrent tasks to use when writing can be controlled by passing the ntasks keyword argument to Arrow.write. Passing ntasks=1 avoids any multithreading when writing.


Compression is supported when writing via the compress keyword argument. Possible values include :lz4, :zstd, or your own initialized LZ4FrameCompressor or ZstdCompressor objects; will cause all buffers in each record batch to use the respective compression encoding or compressor.