[FLINK-40338][table-runtime] ELT() throws ClassCastException when index is not INT

[hulincup]

Problem fixed & how

ELT(index, expr, exprs...) accepts any INTEGER_NUMERIC index (TINYINT/SMALLINT/INT/BIGINT). However, EltFunction#eval declares index as java.lang.Number and indexes the varargs array with exprs[(int) index - 1]. Per JLS 5.5, casting a Number reference to int compiles to a checkcast to Integer followed by unboxing, so a Byte, Short, or Long value throws ClassCastException on the success path (1 <= index <= exprs.length).

SELECT ELT(CAST(2 AS BIGINT), 'scala', 'java');
-- java.lang.ClassCastException: class java.lang.Long cannot be cast to class java.lang.Integer

Same for CAST(2 AS TINYINT) and CAST(2 AS SMALLINT).

The out-of-range guard above the cast uses index.longValue(), so out-of-range indices of any type still return NULL correctly; the exception only fires in the valid range. That is why the existing test ELT(9223372036854775807, 'ab', 'b') passes (returns NULL before reaching the cast) and every other existing test uses an INT literal.

Present since FLINK-35987 introduced ELT; confirmed absent from release-1.20 and present from release-2.0.

Fix: narrow the already-unboxed long idx (computed above for the range check) via primitive narrowing (int) idx (JLS 5.1.3, no checkcast) instead of casting the Number reference.

Behavior modified

• previous: ELT with a non-INT INTEGER_NUMERIC index (TINYINT/SMALLINT/BIGINT) in the valid range 1 <= index <= exprs.length threw ClassCastException.
• now: non-INT integer indices correctly return the corresponding expression.
• impact: only the success path for non-INT integer indices; INT indices, NULL, and out-of-range behavior are unchanged.

Code refactored

EltFunction.eval: exprs[(int) index - 1] → exprs[(int) idx - 1], with a 3-line comment explaining the JLS rationale.

Features added

N/A

Functions optimized

N/A

Test plan

• Added 3 regression cases to StringFunctionsITCase.eltTestCases() covering TINYINT, SMALLINT, and BIGINT indices (the three types that previously threw ClassCastException). Each expects the correct expression ("java" for index 2).
• The local environment runs Java 8 and cannot build Flink master (requires Java 11+), so verification relies on CI:

  mvn -pl flink-table/flink-table-planner -am test -Dtest=StringFunctionsITCase
  

[flinkbot]

CI report:

• a1d7732 Azure: FAILURE

Bot commands

The @flinkbot bot supports the following commands:

• @flinkbot run azure re-run the last Azure build

[SEPURI-SAI-KRISHNA]

Hi @hulincup — I'm the reporter of FLINK-40338 and had asked to be assigned before this PR was opened, with a patch and regression tests ready. Per the contribute-code guide, PRs on unassigned tickets aren't reviewed or merged, so could we wait for a committer to decide who takes it?

[raminqaf]

Changes look good. Left a nit

[raminqaf]

I would remove this

Suggested change

// FLINK-40338: non-INT INTEGER_NUMERIC index must not throw ClassCastException

[SEPURI-SAI-KRISHNA]

Thanks for picking this up, the fix itself looks right. One coverage gap worth closing before merge.

All three new cases pass constant arguments, so ExpressionReducer folds the entire call during optimization and the generated runtime code is never reached. The plans:

-- ELT(CAST(2 AS TINYINT), 'scala', 'java')
== Optimized Execution Plan ==
Calc(select=[CAST('java' AS VARCHAR(5)) AS EXPR$0])     <- ELT folded away

-- ELT(b, 'scala', 'java')   where b TINYINT
== Optimized Execution Plan ==
Calc(select=[ELT(b, 'scala', 'java') AS EXPR$0])        <- ELT reaches the operator

To be clear, these cases do fail without the fix, because the reducer executes the function at plan time. But they only cover the constant-folding path, not the codegen'd operator path that a real job hits, so a future regression in the runtime path wouldn't be caught here.

The case just below at line 201 already uses the field-reference pattern, so extending the existing fields covers it:

.onFieldsWithData(null, null, null, new byte[] {1, 2, 3}, (byte) 2, (short) 2, 2L)
.andDataTypes(
        DataTypes.INT(), DataTypes.STRING(), DataTypes.BYTES(), DataTypes.BYTES(),
        DataTypes.TINYINT(), DataTypes.SMALLINT(), DataTypes.BIGINT())

and then $("f4").elt("scala", "java"), $("f5"), $("f6"), keeping one of the constant cases so the reducer path stays covered too.

I'm the reporter of FLINK-40338 and already have these written and verified locally (they fail with ClassCastException on java.lang.Byte/Short/Long without the fix). Happy to hand them over for you to include here, or to open them as a follow-up if you'd rather keep this PR as is, whichever you prefer.

[raminqaf]

Just in case write a test with non-literals too please

[SEPURI-SAI-KRISHNA]

@raminqaf agreed. Here's the patch I have locally, already verified, it fails with ClassCastException on java.lang.Byte, Short and Long before the fix and passes after:

.onFieldsWithData(null, null, null, new byte[] {1, 2, 3}, (byte) 2, (short) 2, 2L)
.andDataTypes(
        DataTypes.INT(), DataTypes.STRING(), DataTypes.BYTES(), DataTypes.BYTES(),
        DataTypes.TINYINT(), DataTypes.SMALLINT(), DataTypes.BIGINT())

.testResult(
        $("f4").elt("scala", "java"),
        "ELT(f4, 'scala', 'java')",
        "java",
        DataTypes.VARCHAR(5))
.testResult(
        $("f5").elt("scala", "java"),
        "ELT(f5, 'scala', 'java')",
        "java",
        DataTypes.VARCHAR(5))
.testResult(
        $("f6").elt("scala", "java"),
        "ELT(f6, 'scala', 'java')",
        "java",
        DataTypes.VARCHAR(5))

@hulincup feel free to take these directly.