andygrove opened a new issue, #290:
URL: https://github.com/apache/datafusion-spark/issues/290
Part of #271 and apache/datafusion-comet#7300.
### What the function does
`from_xml(xmlStr, schema[, options])` parses one XML record into a `STRUCT`.
It was added in Spark 4.0.0.
- **Schema.** `schema` must be a foldable string: a DDL string (`'a INT, b
DOUBLE'` or `'STRUCT<...>'`) or
`schema_of_xml(...)`. A non-literal schema fails with
`INVALID_SCHEMA.NON_STRING_LITERAL`. A bad DDL string
fails with `PARSE_SYNTAX_ERROR`. The top level must be `STRUCT`, or
`VARIANT` from 4.1. Otherwise analysis
fails with `DATATYPE_MISMATCH.INVALID_XML_SCHEMA`. A `MAP` with a
non-`STRING` key anywhere in the schema
gives `DATATYPE_MISMATCH.INVALID_XML_MAP_KEY_TYPE`.
- **Options.** `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`.
- **Result.** The result type is the schema with nullability forced
everywhere (`asNullable`). It is nullable
and NULL-intolerant (NULL in, NULL out). The input is a string
(collation-aware), and the function is time
zone aware.
- **Root element.** The root element's name is ignored. Its attributes
become fields and its children are
matched against the schema, so `rowTag` has no effect: `XmlFunctionsSuite`
shows `<foo>`, `<ROW>` and `<p>`
all parsing.
Parsing uses `StaxXmlParser` over a StAX reader. The reader is not
namespace-aware and coalesces text. It does
not support DTDs or external entities, and skips comments, processing
instructions and DTD events. Records go
through `FailureSafeParser` with options from `XmlOptions`.
### Examples
```sql
SELECT from_xml('<p><a>1</a><b>0.8</b></p>', 'a INT, b DOUBLE');
-- {"a":1,"b":0.8}
SELECT from_xml('<p><time>26/08/2015</time></p>', 'time Timestamp',
map('timestampFormat', 'dd/MM/yyyy'));
-- {"time":2015-08-26 00:00:00}
SELECT
from_xml('<p><teacher>Alice</teacher><student><name>Bob</name><rank>1</rank></student><student><name>Charlie</name><rank>2</rank></student></p>',
'STRUCT<teacher: STRING, student: ARRAY<STRUCT<name: STRING, rank: INT>>>');
--
{"teacher":"Alice","student":[{"name":"Bob","rank":1},{"name":"Charlie","rank":2}]}
SELECT from_xml('<p><a>1</a></p>', 'struct<a:array<int>>');
-- {"a":[1]}
SELECT from_xml('<p><a>1</a><b>"2"</b></p>', 'struct<a:int,b:string>');
-- {"a":1,"b":""2""}
SELECT from_xml('<p><a>1</a><a>"2"</a></p>', 'struct<a:array<int>>');
-- {"a":[1]}
SELECT from_xml('<p><a>1</a><a></a></p>', 'struct<a:array<int>>');
-- {"a":[1,null]}
SELECT from_xml('<p><a attr="1"><b>2</b></a></p>',
'struct<a:map<string,int>>');
-- {"a":{"_attr":1,"b":2}}
SELECT from_xml('<p><d>12/15 2012</d><t>12/15 2012 15:15:15</t>}</p>', 'd
date, t timestamp', map('dateFormat', 'MM/dd yyyy', 'timestampFormat', 'MM/dd
yyyy HH:mm:ss'));
-- {"d":2012-12-15,"t":2012-12-15 15:15:15}
SELECT from_xml('<p><d>02-29</d></p>', 'd date', map('dateFormat', 'MM-dd'));
-- {"d":null}
SELECT from_xml('<p><a>1</a><a>2</a><a>3</a></p>',
schema_of_xml('<p><a>1</a><a>2</a></p>'));
-- {"a":[1,2,3]}
```
These come from the `ExpressionDescription` and `xml-functions.sql.out`
(4.2).
### Edge cases to test
- **Modes** (`mode`, case-insensitive; an unknown value falls back to
`PERMISSIVE` with a warning):
- `PERMISSIVE` (the default). A malformed record gives a struct of NULLs,
not a NULL struct. `<ROW><a>1</ROW>`
with `a INT` gives `{"a":null}`.
- A conversion failure inside an element keeps the partial row (the `[1]`
example above).
- If the schema contains the corrupt-record column
(`spark.sql.columnNameOfCorruptRecord`, default
`_corrupt_record`, or the `columnNameOfCorruptRecord` option), that
column gets the raw input string.
- `FAILFAST` raises `MALFORMED_RECORD_IN_PARSING.WITHOUT_SUGGESTION`.
- `DROPMALFORMED` raises `PARSE_MODE_UNSUPPORTED` ("doesn't support the
DROPMALFORMED mode").
- **Corrupt column in the middle.** `XmlFunctionsSuite` ("corrupt record
column in the middle") asserts that
schema `a INT, _unparsed STRING, b INT` on the valid
`<ROW><a>1</a><b>12</b></ROW>` gives `(1, null, null)`.
`FailureSafeParser` reads the parsed row by its index in the schema
without the corrupt column, so
every field after the corrupt column is shifted. Decide whether to copy
this bug and document the decision.
- **Type conversion** (`convertTo` / `castTo`, after trimming when
`ignoreSurroundingSpaces=true`, the default):
- `INT`, `BIGINT` and `FLOAT`/`DOUBLE` strip a leading `+` or `-`, parse
the rest with Java
`toInt`/`toLong`/`toDouble`, and then negate. By code reading this means
`-2147483648` and
`-9223372036854775808` overflow and become NULL.
- `FLOAT`/`DOUBLE` fall back to a locale `NumberFormat`.
- `TINYINT` and `SMALLINT` use `toByte`/`toShort` directly, and `DECIMAL`
uses the locale decimal parser.
- `BOOLEAN` accepts `true`, `false`, `1` and `0`, case-insensitive.
- Any failure nulls the field, or raises in `FAILFAST`.
- **Empty values.** An empty element (`<a></a>` or `<a/>`) is NULL for
non-string types. For `STRING` it is
`""` unless `nullValue` is `""`. A text value equal to `nullValue`
(default: none) becomes NULL.
- **`STRING` field with child elements.** The field gets the inner markup as
a string
(`currentStructureAsString`).
- **Attributes.** Attributes become fields named `attributePrefix + name`
(default `_`). With
`excludeAttribute=true` they are dropped. An element that has both
attributes and text maps the text to
`valueTag` (default `_VALUE`): `<a x="1">5</a>` with `a STRUCT<_VALUE:
INT, _x: INT>`. Option checks:
`valueTag` must be non-empty and different from `attributePrefix`.
- **Arrays.** Repeated elements fill an `ARRAY` field. A single occurrence
gives a one-element array.
`wildcardColName` (default `xs_any`) collects unmatched elements as
`STRING` or `ARRAY<STRING>`.
- **Field names.** Matching follows `spark.sql.caseSensitive` (default
false, so it is case-insensitive).
Unknown elements are skipped and missing ones are NULL.
`ignoreNamespace=true` strips the `prefix:`.
- **Datetimes.** `timestampFormat` has no default in read mode unless the
legacy parser policy is on. Also
cover `timestampNTZFormat`, `dateFormat` (an invalid date such as `02-29`
without a year gives NULL),
`timeFormat` (4.2, `TIME(p)`), `locale`, `timeZone`,
`spark.sql.session.timeZone`, DST transitions, and
`spark.sql.legacy.timeParserPolicy`.
- **Binary.** `BINARY` fields (4.1+) are parsed according to
`spark.sql.binaryOutputStyle`.
- **Bad options.** Validation errors from `require`
(`IllegalArgumentException`) cover an empty `rowTag`, a
`rowTag` with `<` or `>`, an empty `valueTag`, and a non-boolean value for
`excludeAttribute` and other
boolean options.
- **Inputs.** NULL input gives NULL. Also test the empty string,
whitespace-only input, multiple root elements,
trailing content after the root, and Unicode in names and text.
- **Round-trip.** `from_xml(to_xml(s), schema)` should round-trip structs
and arrays (`XmlFunctionsSuite`).
### Differences across Spark versions
- Not available in 3.4 or 3.5.
- 4.1: adds the `VARIANT` schema and the `singleVariantColumn` option,
`BINARY` parsing, and the
`useLegacyXMLParser` option and `spark.sql.legacy.useLegacyXMLParser`
config. These only affect the file
reader.
- 4.2: adds the `TIME` type and the `timeFormat` option.
### Implementation notes
- Spark: `XmlToStructs` in
`sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/xmlExpressions.scala`
and `XmlToStructsEvaluator` in
`.../expressions/xml/XmlExpressionEvalUtils.scala`. The parser is
`sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/xml/StaxXmlParser.scala`,
with
`StaxXmlParserUtils.scala` and `XmlOptions.scala`. Modes are in
`catalyst/util/FailureSafeParser.scala`.
- DataFusion core has no XML functions.
- Add a streaming XML parser dependency. `quick-xml` maps closely to the
StAX event model; it doesn't process
DTDs, and namespaces are opt-in. This crate has no XML dependency today.
The `xpath*` family needs a DOM and
an XPath engine instead (see the `xpath` issue), so the two may end up
with different crates.
- This crate has no Spark DDL schema parser. Share the approach chosen for
`from_json`: either parse the
literal at planning time, or let callers such as Comet pass the Arrow type
directly.
- A first version could support `PERMISSIVE` and `FAILFAST`, primitive,
struct, array and string-keyed map
types, attributes and `valueTag`, and leave `VARIANT`, XSD validation
(`rowValidationXSDPath`) and datetime
patterns for follow-ups.
- Tests: add `crates/sqllogictest/test_files/xml/from_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._
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]