Skip to content

Implement Spark-compatible is_variant_null #277

Open

Description

@andygrove

Part of #273 and apache/datafusion-comet#7300.

What the function does

is_variant_null(v) returns true only when v is a variant null, meaning a JSON null stored inside a VARIANT.
In every other case it returns false, and that includes a SQL NULL input. The result is a non-nullable
BOOLEAN. The only accepted argument type is VARIANT, and there are no implicit casts. Added in Spark 4.0.0.

The check is a single byte: the first value byte must be 0x00 (basic type PRIMITIVE, type id NULL). This depends
on the VARIANT Arrow representation proposed in the parse_json issue, which uses the canonical
arrow.parquet.variant extension with metadata and value binary children.

Examples

SELECT is_variant_null(parse_json('null'));
-- true
SELECT is_variant_null(parse_json('"null"'));
-- false
SELECT is_variant_null(parse_json('13'));
-- false
SELECT is_variant_null(parse_json(null));
-- false
SELECT is_variant_null(variant_get(parse_json('{"a":null, "b":"spark"}'), '$.c'));
-- false
SELECT is_variant_null(variant_get(parse_json('{"a":null, "b":"spark"}'), '$.a'));
-- true
SELECT is_variant_null(variant_get(parse_json('{ "a": null, "b": {"c": null, "d": [13, null]} }'), '$.b.d[1]'));
-- true
SELECT is_variant_null(variant_get(parse_json('{ "a": null, "b": {"c": null, "d": [13, null]} }'), '$.b.d'));
-- false

These come from the ExpressionDescription and VariantEndToEndSuite.

Edge cases to test

  • SQL NULL returns false, not NULL. StaticInvoke uses propagateNull = false, returnNullable = false.
    A missing path in variant_get gives a SQL NULL, so it also returns false.
  • Cases that return false:
    • scalars: the string "null", "", 0, true, false, large ints such as 2147483648, and decimals;
    • a long string ("spark" repeated 100 times);
    • arrays and objects, including [], [null, null] and [{"a" : 13, "b" : null}].
      These all come from VariantExpressionEvalUtilsSuite."isVariantNull".
  • A variant null nested in an object or array reached through variant_get without a target type is still a
    variant null, so the result is true. With a concrete target type variant_get returns SQL NULL instead, so
    is_variant_null(variant_get(v, '$.a', 'int')) is false.
  • A malformed value with an empty value binary raises MALFORMED_VARIANT ("Variant binary is malformed. Please
    check the data source is valid."). Spark does not validate anything beyond the first byte.
  • Check that metadata is ignored: any valid header, including sorted_strings set or 2-4 byte offsets, gives the
    same answer.
  • Test both a column and a literal input. Also test a variant produced by to_variant_object or by
    cast(x as variant) from a NULL array element.
  • The result does not depend on ANSI mode.

Differences across Spark versions

  • 3.4 and 3.5 have no VARIANT type.
  • No differences between 4.0.4, 4.1.3 and 4.2.0.

Implementation notes

  • Spark: IsVariantNull in
    sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/variant/variantExpressions.scala.
    The logic is in VariantExpressionEvalUtils.isVariantNull, in the same package.
  • DataFusion core has no VARIANT support. arrow-rs parquet_variant::Variant::Null can be used for a decoded
    check, but reading value[0] == 0 from the value child of the struct array is enough and avoids decoding.
  • When the struct row itself is null, return false. A row with an empty value binary should return
    MALFORMED_VARIANT.
  • [EPIC] Native Variant Support datafusion-comet#5438 tracks native Variant support in Comet.
  • This is blocked on the VARIANT type representation from the parse_json issue. Put it in a new
    crates/spark/src/function/variant/ module. Add crates/sqllogictest/test_files/variant/is_variant_null.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