Tabular data is routinely transferred on the web as "CSV", but the definition of "CSV" in practice is very loose. This document outlines a basic data model or infoset for tabular data and metadata about that tabular data. It also contains some non-normative information about a best practice syntax for tabular data, for mapping into that data model, to contribute to the standardisation of CSV syntax by IETF. Various methods of locating metadata are also provided.
The CSV on the Web Working Group was chartered to produce a Recommendation "Access methods for CSV Metadata" as well as Recommendations for "Metadata vocabulary for CSV data" and "Mapping mechanism to transforming CSV into various Formats (e.g., RDF, JSON, or XML)". This document aims to primarily satisfy the first of those Recommendations (see ), though it also specifies an underlying model and therefore starting point for the other chartered Recommendations.
This document is based on IETF's [[!RFC4180]] which is an Informational RFC. The working group's expectation is that future suggestions to refine RFC 4180 will be relayed to the IETF (e.g. around I18N and multi-part packaging) and contribute to its discussions about moving CSV to the Standards track.
Many "CSV" files embed metadata, for example in lines before the header row of the CSV document. This specification does not define any formats for embedding metadata within CSV files, aside from the names of columns in the header row.
Tabular data is data that is structured into rows, each of which contains information about some thing. Each row contains the same number of cells (although some of these cells may be empty), which provide values of properties of the thing described by the row. In tabular data, cells within the same column provide values for the same property of the thing described by the particular row. This is what differentiates tabular data from other line-oriented formats.
Tabular data is routinely transferred on the web in a textual format called "CSV", but the definition of "CSV" in practice is very loose. Some people use the term to mean any delimited text file. Others stick more closely to the most standard definition of CSV that there is, [[!RFC4180]]. Appendix A describes the various ways in which CSV is defined.
There are different levels of data models for tabular data:
The core tabular data model can be used to describe a table that lacks any annotations, whether those annotations are embedded within a CSV file or arise from a separate metadata document.
Data is held in a table. The properties of a table are:
A column represents a vertical arrangement of cells within a table. The properties of a column are:
A row represents a horizontal arrangement of cells within a table. The properties of a row are:
A cell represents a cell at the intersection of a row and a column within a table. The properties of a cell are:
null, if the string value is an empty string
Should an empty string value always count as a
null semantic value within the core data model (ie when there is no other metadata telling the processor how to interpret that value), or should the semantic value be an empty string? GitHub >
There presence or absence of quotes around a value within a CSV file is a syntactic detail that is not reflected in the tabular data model. In other words, there is no distinction in the model between the second value in
a,,z and the second value in
An annotated table is a table that is annotated with additional metadata. The table MAY have any number of properties in addition to those provided in the core tabular data model described in which provide information about the table as a whole. The values of these properties may be lists, structured objects, or atomic values. Annotations on a table may include:
The columns within an annotated table are all annotated columns: columns which MAY have any number of properties in addition to those provided in the core tabular data model described in . The annotations on a column might provide information about how to interpret the cells in the column or information about the column as a whole. Examples of annotations might be:
The rows within an annotated table are all annotated rows: rows which MAY have any number of properties in addition to those provided in the core tabular data model described in . The annotations on a row provide additional metadata about the information held in the row, such as:
The cells within an annotated row are all annotated cells: cells which MAY have any number of properties in addition to those provided in the core tabular data model described in . The annotations on a cell provide metadata about the value held in the cell, particularly when this overrides the information provided for the annotated column and annotated row that the cell falls within. Annotations on a cell might be:
The value of an annotated cell MAY be of a datatype other than a string. For example, annotations might enable a processor to understand the string value of the cell as representing a number or a date.
The permitted types of names and values of annotations need to be spelled out here. GitHub >
A group of tables comprises a set of tables (which may be annotated tables) and a set of annotations (properties and values) that relate to the set.
Tables can be loosely related to each other simply through annotations; not all tables that are related to each other need to grouped together. Groups of tables are useful because they can be annotated with metadata that applies to all the tables in the group.
As described in , tabular data may have a number of annotations associated with it. Here we describe the different methods that can be used to locate those annotations given a link to a CSV file.
In most methods of locating metadata described here, metadata is provided within a separate document. The syntax of this document is defined in the Metadata Vocabulary for Tabular Data specification. These documents can include things like:
When creating a set of annotations from metadata, if the same property is specified in two locations then information "closer" to the end user or interpreter of the document should override other information, and information "closer" to the document itself should override information "further" from the document. Explicitly, the order of preference is:
Linkheader associated with the tabular data file, see
Processors MUST attempt to locate a metadata document based on each of these locations in order, and use first metadata document that is successfully located in this way.
What should processors do if they have been passed a metadata file and located a CSV file from that metadata file. Should they still check for metadata files related specifically to the CSV file itself? For example, say that an application has been pointed at a metadata file at
http://example.org/metadata.json which references
http://example.org/toilets.csv but there is also a metadata file at
http://example.org/toilets.csv-metadata.json. If the processor had been originally pointed to
http://example.org/toilets.csv then it would have located the file-specific metadata at
http://example.org/toilets.csv-metadata.json, but coming via
http://example.org/metadata.json means that file-specific metadata is skipped. GitHub >
Processors SHOULD provide users with the facility to provide their own metadata for tabular data files that they process.
Most syntaxes for tabular data provide a facility for embedding some metadata within the tabular data file itself. The definition of a syntax for tabular data SHOULD include a description of how the syntax maps to an annotated data model. See for a description of this for CSV files.
When retrieving a tabular data file via HTTP, the response MAY include a
Link header with
type=application/csv-metadata+json that points to a metadata file that describes the tabular data file. If, by inspection, the referenced file is not a valid metadata file then it MUST be ignored. If there is more than one valid metadata file linked to through multiple
Link headers, then the metadata referenced by
Link headers that appear later in the response are used in preference to those referenced by earlier
Suggest that we drop the requirement to ignore invalid metadata files, given that the correct metadata file should be identified through the type of the link. GitHub >
If no metadata document has been located through higher-priority mechanisms, processors MUST attempt to locate a file-specific metadata document. This method is only used if the URL of the tabular data file does not contain a query component.
In this case, the file-specific metadata document should be located at
location-metadata.json. For example, if the tabular data file is at
http://example.org/south-west/devon.csv then processors should attempt to locate a metadata document at
http://example.org/south-west/devon.csv-metadata.json. If the metadata file found at this location does not explicitly point to the relevant tabular data file then it MUST be ignored.
If no metadata document has been located through higher-priority mechanisms, processors MUST attempt to locate a directory-level metadata document. This method is only used if the URL of the tabular data file does not contain a query component.
The common metadata document for a directory can be located by resolving the relative URL
metadata.json against the base URL of the tabular data file. For example, if the tabular data file is at
http://example.org/south-west/devon.csv then processors should attempt to locate a metadata file at
http://example.org/south-west/metadata.json. If the metadata file found at this location does not explicitly point to the relevant tabular data file then it MUST be ignored.
There is no standard for CSV, and there are many variants of CSV used on the web today. This section defines a method for outputting tabular data adhering to the core tabular data model described in into a standard, CSV-based, syntax. Compliant applications that output this format must meet each of the constraints.
We are actively working with the IETF to develop a standard for CSV, which is outside the scope of the Working Group. The details here aim to help shape that standard based on our requirements.
This section does not seek to describe how applications that input textual tabular data should interpret it, except that any data that is in the format defined here should be understood as defined here.
This syntax is not compliant with
text/csv as defined in [[!RFC4180]]: it permits characters other than ASCII, and it permits line endings other than
CRLF. Supporting the full set of Unicode characters by using UTF-8 and supporting
LF line endings are important characteristics for data formats that are used internationally and on non-Windows platforms. However, all files that adhere to [[!RFC4180]]'s definition of CSV are compliant CSV+ files.
The appropriate content type for a CSV+ file is
text/csv. For example, when a CSV+ file is transmitted via HTTP, the HTTP response MUST include a
Content-Type header with the value
See below for issues relating to whether we should instead define a different content type.
CSV+ files SHOULD be encoded using UTF-8. If a CSV+ file is not encoded using UTF-8, the encoding MUST be specified through the
charset parameter in the
RFC4180 defines the default charset as US-ASCII because that was (at the time RFC4180 was written) the default charset for all
text/* media types. This has been superseded with RFC6657.
Section 3 of RFC6657 states "new subtypes of the "text" media type SHOULD NOT define a default "charset" value. If there is a strong reason to do so despite this advice, they SHOULD use the "UTF-8" [RFC3629] charset as the default."
Do we have a strong reason to specify a default charset? Should we be defining
application/csv instead, to avoid doing unrecommended things with a
text/* media type.
The ends of rows in a CSV+ file MUST be either
U+000D U+000A) or
U+000A). Line endings within escaped cells are not normalised.
Section 4.1.1 of RFC2046 specifies that "The canonical form of any MIME "text" subtype MUST always represent a line break as a CRLF sequence. Similarly, any occurrence of CRLF in MIME "text" MUST represent a line break. Use of CR and LF outside of line break sequences is also forbidden."
Should we be defining
application/csv instead, to prevent having to adhere to this rule, or should we stick to the
Each line of a CSV+ file MUST contain the same number of comma-separated values.
Values that contain commas, line endings or double quotes MUST be escaped by having the entire value wrapped in double quotes. There MUST NOT be whitespace before or after the double quotes. Within these escaped cells, any double quotes MUST be escaped with two double quotes (
The first line of a CSV+ file SHOULD contain a comma-separated list of names of columns. This is known as the header line and provides names for the columns. There are no constraints on these names.
If a CSV+ file does not include a header line, this MUST be specified using the
header parameter of the media type:
Bidirectional content does not alter the definition of rows or the assignment of cells to columns. Whether or not a CSV+ file contains right-to-left characters, the first column's content is the first cell of each row, which is the text prior to the first occurrence of a comma within that row.
For example, Egyptian Referendum results are available as a CSV file at https://egelections-2011.appspot.com/Referendum2012/results/csv/EG.csv. Over the wire and in non-Unicode-aware text editors, the CSV looks like:
المحافظة,نسبة موافق,نسبة غير موافق,عدد الناخبين,الأصوات الصحيحة,الأصوات الباطلة,نسبة المشاركة,موافق,غير موافق القليوبية,60.0,40.0,"2,639,808","853,125","15,224",32.9,"512,055","341,070" الجيزة,66.7,33.3,"4,383,701","1,493,092","24,105",34.6,"995,417","497,675" القاهرة,43.2,56.8,"6,580,478","2,254,698","36,342",34.8,"974,371","1,280,327" قنا,84.5,15.5,"1,629,713","364,509","6,743",22.8,"307,839","56,670" ...
Within this CSV file, the first column appears as the content of each line before the first comma and is named
المحافظة (appearing at the start of each row as
المحافظة in the example, which is displaying the relevant characters from left to right in the order they appear "on the wire").
The CSV translates to a table model that looks like:
|Column / Row||column 1||column 2||column 3||column 4||column 5||column 6||column 7||column 8||column 9|
|row 1 (header)||المحافظة||نسبة موافق||نسبة غير موافق||عدد الناخبين||الأصوات الصحيحة||الأصوات الباطلة||نسبة المشاركة||موافق||غير موافق|
The fragment identifier
#col=3 identifies the third of the columns, named
نسبة غير موافق (appearing as
نسبة غير موافق in the example).
[[!tabular-metadata]] defines how this table model should be displayed by compliant applications, and how metadata can affect the display. The default is for the display to be determined by the content of the table. For example, if this CSV were turned into an HTML table for display into a web page, it should be displayed with the first column on the right and the last on the left, as follows:
|غير موافق||موافق||نسبة المشاركة||الأصوات الباطلة||الأصوات الصحيحة||عدد الناخبين||نسبة غير موافق||نسبة موافق||المحافظة|
The fragment identifier
#col=3 still identifies the third of the columns, named
نسبة غير موافق, which appears in the HTML display as the third column from the right and is what those who read right-to-left would think of as the third column.
Note that this display matches that shown on the original website.
An alternative approach is for the CSV to be parsed into a table model in which the columns are numbered in the reverse, for tables which are either marked as or detected to be right-to-left tables. For example, we could introduce a
bidi=rtl or similar media type parameter, and use this to determine whether the first column in table generated from the CSV is the text before the first comma in each line or the text after the last comma in the line.
In the example above, if the CSV were served with
bidi=rtl, or the table was detected as being a right-to-left table, then the column numbering in the model would be reversed:
|Column / Row||column 9||column 8||column 7||column 6||column 5||column 4||column 3||column 2||column 1|
|row 1 (header)||المحافظة||نسبة موافق||نسبة غير موافق||عدد الناخبين||الأصوات الصحيحة||الأصوات الباطلة||نسبة المشاركة||موافق||غير موافق|
This would require a change to [[RFC7111]] but that might be required by updates to the definition of
text/csv in any case. With the change, the fragment identifier
#col=3 would then refer to the third column from the right, named
If the model were defined in this way, there would be no need to determine the order of the columns when displayed using a metadata property. Columns would always be displayed with the first column (numbered 1 in the model) on the left. The final display in HTML, for example, would be exactly as above. The only difference would be that
#col=3 would refer to the third column from the left.
We note that using media type parameters is problematic because publishers might not have the ability to set them on their servers, and because they can easily get lost as a file is republished or emailed between people.
We invite comment on the best way to approach bidirectionality in CSV files.
This grammar is a generalization of that defined in [[!RFC4180]] and is included for reference only.
The EBNF used here is defined in XML 1.0 [[EBNF-NOTATION]].
We should probably place further restrictions on QCHAR and SCHAR to avoid control characters. If header weren’t optional, it would be better defined as in RFC4180, but if the syntax allows it to be optional, this would make it not an LL(1) grammar, which isn’t too much of an issue.
As described in , there may be many formats which an application might interpret into the tabular data model described in , including using different separators or fixed format tables, multiple tables within a single file, or ones that have metadata lines before a table header.
Standardising the parsing of CSV is outside the chartered scope of the Working Group. This non-normative section is intended to help the creators of parsers handle the wide variety of CSV-based formats that they may encounter due to the current lack of standardisation of the format.
This section describes an algorithm for parsing formats other than the plain CSV+ format specified in . It is impossible to do this in a fully automated manner, so this algorithm depends on the following flags being set externally (eg through user input):
""is used to escape
"within an escaped cell).
end. The default is
When parsing, should we:
The algorithm for parsing a document containing tabular data is as follows:
Split the header and data rows into cells using the delimiter. Values that are enclosed within the quote character may contain the delimiter. The quote character may be escaped using the escape character where it appears within cells. If the escape character is not the same as the quote character then the escape character is also used to escape the character that immediately follows it.
If trim is
start then whitespace from the start of values that are not enclosed must be removed from the value. If trim is
end then whitespace from the end of values that are not enclosed must be removed from the value.
truethen ignore any rows in which all the cell values are empty strings.
This appendix outlines various ways in which CSV is defined.
[[!RFC4180]] defines CSV with the following ABNF grammar:
file = [header CRLF] record *(CRLF record) [CRLF] header = name *(COMMA name) record = field *(COMMA field) name = field field = (escaped / non-escaped) escaped = DQUOTE *(TEXTDATA / COMMA / CR / LF / 2DQUOTE) DQUOTE non-escaped = *TEXTDATA COMMA = %x2C CR = %x0D DQUOTE = %x22 LF = %x0A CRLF = CR LF TEXTDATA = %x20-21 / %x23-2B / %x2D-7E
Of particular note here are:
TEXTDATAindicates that only non-control ASCII characters are permitted within a CSV file. This restriction is routinely ignored in practice, and is impractical on the international web.
CRLF. This makes it harder to produce CSV files on Unix-based systems where the usual line ending is
headerparameter on the media type indicates whether the header is present or not.
Excel is a common tool for both creating and reading CSV documents, and therefore the CSV that it produces is a de facto standard.
The following describes the behaviour of Microsoft Excel for Mac 2011 with an English locale. Further testing is needed to see the behaviour of Excel in other situations.
Excel generates CSV files encoded using Windows-1252 with
LF line endings. Characters that cannot be represented within Windows-1252 are replaced by underscores. Only those cells that need escaping (eg because they contain commas or double quotes) are escaped, and double quotes are escaped with two double quotes.
Dates and numbers are formatted as displayed, which means that formatting can lead to information being lost or becoming inconsistent.
When opening CSV files, Excel interprets CSV files saved in UTF-8 as being encoded as Windows-1252 (whether or not a BOM is present). It correctly deals with double quoted cells, except that it converts line breaks within cells into spaces. It understands
CRLF as a line break. It detects dates (formatted as
YYYY-MM-DD) and formats them in the default date formatting for files.
Excel provides more control when importing CSV files into Excel. However, it does not properly understand UTF-8 (with or without BOM). It does however properly understand UTF-16 and can read non-ASCII characters from a UTF-16-encoded file.
A particular quirk in the importing of CSV is that if a cell contains a line break, the final double quote that escapes the cell will be included within it.
When tabular data is copied from Excel, it is copied in a tab-delimited format, with
LF line breaks.
Downloaded CSV files are encoded in UTF-8, without a BOM, and with
LF line endings. Dates and numbers are formatted as they appear within the spreadsheet.
CSV files can be imported as UTF-8 (with or without BOM).
CRLF line endings are correctly recognised. Dates are reformatted to the default date format on load.
Tabular Data Packages place the following restrictions on CSV files:
As a starting point, CSV files included in a Tabular Data Package package must conform to the RFC for CSV (4180 - Common Format and MIME Type for Comma-Separated Values (CSV) Files). In addition:
- File names MUST end with
- Files MUST be encoded as UTF-8
Files MUST have a single header row. This row MUST be the first row in the file.
- Terminology: each column in the CSV file is termed a field and its
nameis the string in that column in the header row.
nameMUST be unique amongst fields and MUST contain at least one character
- There are no further restrictions on the form of the
namebut it is RECOMMENDED that it contain only alphanumeric characters together with “ .-_”
- Rows in the file MUST NOT contain more fields than are in the header row (though they may contain less)
- Each file MUST have an entry in the
resourcesarray in the
- The resource metadata MUST include a
schemaattribute whose value MUST conform to the JSON Table Schema
- All fields in the CSV files MUST be described in the
CSV files generated by different applications often vary in their syntax, e.g. use of quoting characters, delimiters, etc. To encourage conformance, CSV files in a CSV files in a Tabular Data Package SHOULD
- Use “,” as field delimiters
- Use “rn” or “n” as line terminators
If a CSV file does not follow these rules then its specific CSV dialect MUST be documented. The resource hash for the resource in the
- Include a
dialectkey that conforms to that described in the CSV Dialect Description Format
Applications processing the CSV file SHOULD read use the
dialectof the CSV file to guide parsing.
More details of behaviour of other tools should go here. This should include the most popular CSV parsing/generating libraries in common programming languages. Test files which include non-ASCII characters, double quotes and line breaks within cells are: