Writing data

CSV.write(sink, table; keywords...) writes any Tables.jl table to a file path or IO and returns the sink.

using CSV

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

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

Column tables render in parallel on up to ntasks tasks (Threads.nthreads() by default); the output is the same for every ntasks value. Row tables are written sequentially without being collected. The writer works in bounded blocks of rows, so its memory use does not grow with the output size.

Headers, appending, and sinks

The default writes the Tables.jl column names as the first row. Use writeheader=false to omit them, or pass header=[...] with one name per column to replace them. header=true and header=false also turn the header on and off.

append=true opens a path for append instead of replacement. It omits the header unless writeheader=true is explicit. An IO sink is written at its current position, like Base.write; for an IO, append only decides whether the header is written.

CSV.write(sink, CSV.Chunks(source)) streams every batch of a chunked read under one header without collecting the batches.

bom=true writes a UTF-8 byte-order mark before new output. It is not written again during append.

using CSV

output = IOBuffer()
CSV.write(output, (a=[1, 2], b=[3, 4]); header=["left", "right"])
String(take!(output))

"left,right\n1,3\n2,4\n"

Delimiters, quotes, and newlines

The writer dialect options are:

:minimal quotes text that contains a delimiter, quote, carriage return, newline, or leading or trailing space. It also quotes an empty string. :all quotes every text field. :none throws an ArgumentError for text that contains a structural byte and for an empty string, which would read back as missing.

Quote-related characters and decimal must be single ASCII characters, given as a Char or a one-character string. A multi-byte delimiter is written as-is, and a value containing its first byte is quoted. A newline can be a character or string such as "\r\n". An unknown keyword is an ArgumentError.

using CSV

output = IOBuffer()
CSV.write(output, (value=["plain", "with,comma", ""]); quotestyle=:minimal)
String(take!(output))

""

Missing and empty text

missingstring controls the text written for missing and defaults to an unquoted empty field. A present empty string is always quoted when quoting is enabled. This matches how CSV.jl reads empty fields:

• an unquoted empty field is missing; and
• a quoted empty field is a present empty string.

Choose a non-empty missingstring if another system does not keep that distinction.

Numbers and dates

floatformat accepts a Printf-style format such as "%.3f". Its default uses Julia's shortest round-trip floating-point representation. decimal replaces the decimal point in floating-point output.

Numbers are quoted when their text contains a delimiter, quote character, or line break; quotestyle=:none rejects such values. When reading decimal-comma output, configure the reader to use decimal=',' as well.

dateformat accepts a Dates.jl format for all Date, DateTime, Timestamp, and Time values. Other values use their ordinary text representation.

DataDecimals values keep their decimal spelling, and decimal applies to them too: decimal=',' writes 1.20 as 1,20, quoted when the delimiter is also a comma.

using CSV, Dates

output = IOBuffer()
CSV.write(output, (x=[1.23456], day=[Date(2026, 8, 19)]);
          floatformat="%.2f", dateformat="dd/mm/yyyy", delim=';')
String(take!(output))

"x;day\n1.23;19/08/2026\n"

Gzip output

compress=:auto writes gzip when a path ends in .gz (in any letter case). Use compress=:gzip (or true) to force gzip for any sink, or compress=:none (or false) to disable it. Compression is streamed to the sink. With more than one task, a column table's row blocks compress in parallel, one gzip member each, so the file is a multi-member gzip stream (RFC 1952). Readers such as gunzip and CSV.File concatenate the members. A consumer that reads only one gzip member needs output written with ntasks=1.

CSV.write("output.csv.gz", table)                  # detected from the suffix
CSV.write(io, table; compress=:gzip)               # explicit for an IO

Partitioned output

partition=true writes each value from Tables.partitions(table) to a separate sink, at most ntasks at a time. A vector of sinks must contain one sink per partition. A base path generates paths by appending _1, _2, and so on, and CSV.write returns them.

CSV.write(["part-1.csv", "part-2.csv"], partitioned_table; partition=true)
CSV.write("part.csv", partitioned_table; partition=true) # ["part.csv_1", "part.csv_2"]

Other options

quotestrings=true is an alias for quotestyle=:all.

transform=(column, value) -> new_value transforms each cell before writing. The callback uses a 1-based column index and runs in row-major order. A custom transform disables parallel rendering so callback order stays stable.

bufsize is the largest allowed rendered row size in bytes and defaults to 4 MiB. Increase it when one output row is larger. It is a safety bound, not a whole-file staging buffer.

Writing is streaming. If a later row fails, the sink contains the rows written before the error and nothing of the failing row.

RowWriter

CSV.RowWriter(table; keywords...) is an iterator of complete CSV-formatted row strings. It yields the header first unless writeheader=false. It accepts the same dialect, value-format, and BOM options as CSV.write.

using CSV

rows = collect(CSV.RowWriter((id=[1, 2], label=["a", "b"])))
join(rows)

"id,label\n1,a\n2,b\n"

Rows render on demand. join(CSV.RowWriter(table)) is byte-identical to CSV.write(io, table) with the same options.