Skip to content

Implement Spark-compatible to_xml #291

Open

Description

@andygrove

Part of #271 and apache/datafusion-comet#7300.

What the function does

to_xml(expr[, options]) serializes a STRUCT (or a VARIANT, from 4.1) to an XML string. It was added in
Spark 4.0.0.

  • Arguments. Any other input type fails analysis with DATATYPE_MISMATCH.UNEXPECTED_INPUT_TYPE
    (requiredType "STRUCT"). options must be a map(...) of string literals. Otherwise the error is
    INVALID_OPTIONS.NON_MAP_FUNCTION or INVALID_OPTIONS.NON_STRING_TYPE. Wrong arity gives
    WRONG_NUM_ARGS.WITHOUT_SUGGESTION (expectedNum [1, 2]).
  • Result. The result is a nullable STRING (the session default string type). The function is
    NULL-intolerant and time zone aware.

Output comes from StaxXmlGenerator, which wraps a Woodstox XMLStreamWriter with structure validation off.

  • Row element. The struct is written as one element named rowTag (default ROW). No XML declaration and
    no rootTag wrapper are written; those are file-writer only.
  • Indentation. The default indent is four spaces, which gives newline-separated, indented output.
    indent set to '' gives compact output.
  • Fields. A field whose name starts with attributePrefix (default _), and is not valueTag (default
    _VALUE), becomes an attribute without the prefix. Attributes are written before child elements. The
    valueTag field is written as the element's text.
  • NULL values. When nullValue is not set (the default), NULL fields are omitted. Otherwise the
    nullValue text is written.
  • Arrays. An array field becomes repeated elements with the field's name. An array nested directly in an
    array uses arrayElementName (default item).
  • Maps. A map becomes child elements keyed by the map key. Keys with the attribute prefix become
    attributes.

Examples

SELECT to_xml(named_struct('a', 1, 'b', 2));
-- <ROW>
--     <a>1</a>
--     <b>2</b>
-- </ROW>
SELECT to_xml(named_struct('a', 1, 'b', 2), map('indent', ''));
-- <ROW><a>1</a><b>2</b></ROW>
SELECT to_xml(named_struct('time', to_timestamp('2015-08-26', 'yyyy-MM-dd')), map('timestampFormat', 'dd/MM/yyyy', 'indent', ''));
-- <ROW><time>26/08/2015</time></ROW>
SELECT to_xml(named_struct('time', TIME'14:30:45.123456'), map('rowTag', 'record', 'indent', ''));
-- <record><time>14:30:45.123456</time></record>
SELECT to_xml(array(named_struct('a', 1, 'b', 2)));
-- error: DATATYPE_MISMATCH.UNEXPECTED_INPUT_TYPE
SELECT to_xml(map('a', 1));
-- error: DATATYPE_MISMATCH.UNEXPECTED_INPUT_TYPE

These come from the ExpressionDescription and xml-functions.sql.out (4.2). XmlFunctionsSuite also has
these cases:

  • with session time zone America/Los_Angeles, the timestamp 1800-01-01 00:00:00 is written as
    1800-01-01T00:00:00.000-07:52:58;
  • with timestampFormat yyyy-MM-dd HH:mm:ss.SSSSSS, it writes 2019-11-18 11:56:00.123456.

Edge cases to test

  • Value formatting.
    • Numbers use Java toString, so DOUBLE and FLOAT print as 1.0E10, NaN and Infinity, and
      DECIMAL uses Decimal.toString. Booleans print as true or false.
    • DATE uses dateFormat (default yyyy-MM-dd).
    • TIMESTAMP uses timestampFormat. The default is yyyy-MM-dd'T'HH:mm:ss[.SSS][XXX];
      spark.sql.legacy.timeParserPolicy=LEGACY changes it to ...ss.SSSXXX, and the internal
      supportSecondOffsetFormat setting changes the offset to [XXXXX]. The value is rendered in the session
      time zone or the timeZone option. Test DST and historical (LMT) offsets, as in the 1800 case above.
    • TIMESTAMP_NTZ uses timestampNTZFormat (default yyyy-MM-dd'T'HH:mm:ss[.SSS]), and TIME uses
      timeFormat (4.2).
    • BINARY follows spark.sql.binaryOutputStyle.
  • Escaping and names.
    • <, & and > in text, and quotes in attributes, are escaped by the writer.
    • Field names that aren't valid XML names ('1a', a name with a space) fail when validateName=true
      (default). Check the exact error against Spark.
  • Empty struct values. Nested NULL structs are omitted. A struct whose fields are all NULL writes an empty
    element. Compare the exact output (<ROW></ROW> or <ROW/>) against Spark.
  • Attributes and valueTag. For named_struct('_id', 1, '_VALUE', 'x'), attribute fields are written
    first, whatever the field order.
  • Arrays. Arrays of primitives, arrays of structs, and NULL array elements (omitted unless nullValue is
    set). Nested arrays use arrayElementName.
  • Option validation. attributePrefix must be non-empty (require fails otherwise), and rowTag must
    not contain angle brackets.
  • Round-trips. to_xml followed by from_xml with the same schema should round-trip.
  • NULL input gives NULL.

Differences across Spark versions

  • Not available in 3.4 or 3.5.
  • 4.1: accepts VARIANT input, and adds the singleVariantColumn handling in StaxXmlGenerator.write.
  • 4.2: adds TIME and the timeFormat option.

Implementation notes

  • Spark: StructsToXml in sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/xmlExpressions.scala,
    StructsToXmlEvaluator in .../expressions/xml/XmlExpressionEvalUtils.scala, and the writer in
    sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlGenerator.scala with
    IndentingXMLStreamWriter.scala. Options are in XmlOptions.scala.
  • DataFusion core has no XML functions.
  • A writer can be hand-rolled, or built on quick-xml's Writer. The crate has no XML dependency today.
    Match the indentation rules of IndentingXMLStreamWriter exactly. Reuse the datetime formatting helpers
    that to_json and date_format use, if they exist by then.
  • Tests: add crates/sqllogictest/test_files/xml/to_xml.slt and document the function with #[user_doc].

This issue was generated by an LLM (Claude) from the Apache Spark source code and tests. Examples come from Spark's ExpressionDescription examples, test suites and .sql.out golden files, but please verify them against Spark before relying on them.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions