CSV.jl

CSV.jl reads and writes comma-separated and other delimited text data. Its readers and writer implement the Tables.jl interfaces, so they work with any Julia table package.

Installation

Install the registered release in the Julia REPL:

] add CSV

CSV.jl 1.0 requires Julia 1.10 or later. Read Migrating to 1.0 before you update an application from CSV.jl 0.10.

First read and write

using CSV

input = IOBuffer("name,score\nAda,9.5\nGrace,10.0\n")
file = CSV.File(input)

(names(file), String.(file.name), collect(file.score))

([:name, :score], ["Ada", "Grace"], [9.5, 10.0])

CSV.File is a Tables.jl table. A table sink can consume it directly. The writer accepts any Tables.jl table:

output = IOBuffer()
CSV.write(output, (name=["Ada", "Grace"], score=[9.5, 10.0]))
String(take!(output))

"name,score\nAda,9.5\nGrace,10.0\n"

Select a reader

API	Use it when	What it does
`CSV.File`	You need a normal in-memory table	Parses every selected column
`CSV.read`	You want the columns in another table type	Parses like `CSV.File`, then lets the sink take the columns without copying
`CSV.lazy`	You need a fast first look or sparse cell access	Indexes the input; parses each cell when it is accessed
`CSV.Rows`	You process rows once and do not need columns	Indexes the input; parses cells on access
`CSV.Chunks`	You process a large table in batches	Parses one batch at a time; every batch has the same column types

CSV.lazy, CSV.Rows, and CSV.Chunks keep the whole input in memory (or memory-mapped) while you use them, so none of them streams an unbounded input. CSV.lazy and CSV.Rows also keep an index of every row; CSV.Chunks keeps one batch's. See Input and memory behavior for details. Results do not depend on the number of threads.

Data model

Date-time columns are Timestamp{Nanosecond} from Durations.jl (the Dates.Timestamp type where the standard library provides it). Wider dates use microseconds if all values fit exactly; otherwise the column stays text. Request Dates.DateTime with types when a consumer needs it.

Text columns use DataStrings.DataString by default. Short values are stored in the value; longer values live in buffers the column owns, so an eager table never refers to its source. Convert with String(value) when a standalone String is required, or pass stringtype=String to a reader.

Parse problems are structured data. A read prints one summary warning and keeps the problems; call CSV.problems(file) to inspect them, set on_error=:collect to skip the warning, or on_error=:error when a parse problem must stop the read with a CSV.ParseError.