Reads one or more CSV files as structured records following the format described in RFC 4180.

  • Common

  • Advanced

# Common config fields, showing default values
  label: ""
    paths: [] # No default (required)
    parse_header_row: true
    delimiter: ','
    lazy_quotes: false
    auto_replay_nacks: true
# All config fields, showing default values
  label: ""
    paths: [] # No default (required)
    parse_header_row: true
    delimiter: ','
    lazy_quotes: false
    delete_on_finish: false
    batch_count: 1
    auto_replay_nacks: true

This input offers more control over CSV parsing than the file input.

When parsing with a header row each line of the file will be consumed as a structured object, where the key names are determined from the header now. For example, the following CSV file:

first foo,first bar,first baz
second foo,second bar,second baz

Would produce the following messages:

{"foo":"first foo","bar":"first bar","baz":"first baz"}
{"foo":"second foo","bar":"second bar","baz":"second baz"}

If, however, the field parse_header_row is set to false then arrays are produced instead, like follows:

["first foo","first bar","first baz"]
["second foo","second bar","second baz"]


This input adds the following metadata fields to each message:

- header
- path
- mod_time_unix
- mod_time (RFC3339)

You can access these metadata fields using function interpolation.

Note: The header field is only set when parse_header_row is true.

Output CSV column order

When creating CSV from Redpanda Connect messages, the columns must be sorted lexicographically to make the output deterministic. Alternatively, when using the csv input, one can leverage the header metadata field to retrieve the column order:

      - ./foo.csv
      - ./bar.csv
    parse_header_row: true

    - mapping: |
        map escape_csv {
          root = if this.re_match("[\"\n,]+") {
            "\"" + this.replace_all("\"", "\"\"") + "\""
          } else {

        let header = if count(@path) == 1 {
          @header.map_each(c -> c.apply("escape_csv")).join(",") + "\n"
        } else { "" }

        root = $header + @header.map_each(c -> this.get(c).string().apply("escape_csv")).join(",")

    path: ./output/${! @path.filepath_split().index(-1) }



A list of file paths to read from. Each file will be read sequentially until the list is exhausted, at which point the input will close. Glob patterns are supported, including super globs (double star).

Type: array

# Examples

  - /tmp/foo.csv
  - /tmp/bar/*.csv
  - /tmp/data/**/*.csv


Whether to reference the first row as a header row. If set to true the output structure for messages will be an object where field keys are determined by the header row. Otherwise, each message will consist of an array of values from the corresponding CSV row.

Type: bool

Default: true


The delimiter to use for splitting values in each record. It must be a single character.

Type: string

Default: ","


If set to true, a quote may appear in an unquoted field and a non-doubled quote may appear in a quoted field.

Type: bool

Default: false Requires version 4.1.0 or newer


Whether to delete input files from the disk once they are fully consumed.

Type: bool

Default: false


Optionally process records in batches. This can help to speed up the consumption of exceptionally large CSV files. When the end of the file is reached the remaining records are processed as a (potentially smaller) batch.

Type: int

Default: 1


Whether messages that are rejected (nacked) at the output level should be automatically replayed indefinitely, eventually resulting in back pressure if the cause of the rejections is persistent. If set to false these messages will instead be deleted. Disabling auto replays can greatly improve memory efficiency of high throughput streams as the original shape of the data can be discarded immediately upon consumption and mutation.

Type: bool

Default: true

This input is particularly useful when consuming CSV from files too large to parse entirely within memory. However, in cases where CSV is consumed from other input types it’s also possible to parse them using the Bloblang parse_csv method.