[Muskan Gupta (muskan124947)]

Summary

SQL Server now supports implicit and explicit conversion between FLOAT16 and FLOAT32 vector subtypes. This PR updates existing tests that expected SQLException for cross-type operations to instead expect success, and adds comprehensive new test coverage for all conversion scenarios.

Motivation

The server-side change enables direct conversion between vector(N, float16) and vector(N, float32) without requiring intermediate varchar conversion. This improves usability and performance for vector operations across different precision levels.

Previously, cross-type operations like INSERT INTO f32_table SELECT * FROM f16_table or CAST(f16_col AS VECTOR(3, float32)) would fail with:

"Conversion of vector from data type float16 to float32 is not allowed."

With the server update, these operations now succeed with implicit conversion.

Changes

Updated existing tests (expect success instead of failure)

• VectorFloat16Test.java: Cross-type INSERT...SELECT and setObject tests now verify data arrives correctly instead of catching SQLException.
• VectorFloat16BulkCopyTest.java: Bulk copy cross-type tests (testBulkCopyFloat16ToFloat32Table, testBulkCopyFloat32ToFloat16Table, testMixedBulkCopyColumnMappingSwapped) now verify successful transfer with implicit conversion.
• VectorTest.java: Updated expected error message for exceeding max vector dimensions.

New test coverage

Test	Scenario
testExplicitCastFloat16ToFloat32	CAST(v AS VECTOR(3, float32)) with data assertion
testExplicitCastFloat32ToFloat16	CAST(v AS VECTOR(3, float16)) with data assertion
testConvertFloat16ToFloat32	CONVERT(VECTOR(3, float32), v) with data assertion
testConvertFloat32ToFloat16	CONVERT(VECTOR(3, float16), v) with data assertion
testAlterColumnFloat16ToFloat32	ALTER TABLE ... ALTER COLUMN v VECTOR(3, float32) with data assertion
testAlterColumnFloat32ToFloat16	ALTER TABLE ... ALTER COLUMN v VECTOR(3, float16) with data assertion
testCrossTypeUpdateFloat32ColWithFloat16Value	UPDATE a FLOAT32 column with a FLOAT16 Vector object, assert values
testCrossTypeUpdateFloat16ColWithFloat32Value	UPDATE a FLOAT16 column with a FLOAT32 Vector object, assert values
testStoredProcedureCrossTypeParameterPassing	SP declares @vec VECTOR(3), caller passes FLOAT16, assert values
testUdfCrossTypeParameterPassing	UDF declares @vec VECTOR(3, float16), called with FLOAT32 source, assert values
testTvfCrossTypeReturn	Inline TVF returns CAST(v AS VECTOR(3, float32)) from FLOAT16 source, assert values
testCrossTypeRoundTripPrecision	F32→F16→F32 round-trip preserves values within FLOAT16 precision
testFloat16PrecisionLossWithNonRepresentableValues	F32→F16 with 0.1/0.3/0.7 verifies lossy rounding contract
testCrossTypeDimensionMismatchFails	VECTOR(3, float16)→VECTOR(4, float32) rejects dimension mismatch
testNullVectorCrossTypeInsertSelect	NULL vector via cross-type INSERT...SELECT remains NULL
testNullVectorCrossTypeSetObject	setObject(null) into FLOAT16 column stores NULL correctly
testGetObjectCrossReadFloat16AsFloat32RoundTrip	Read FLOAT16 → setObject into FLOAT32 col → verify values end-to-end
testBulkCopyCrossTypeCSVFloat16ToFloat32	CSV declared as FLOAT16 imported into FLOAT32, per-row value assertion
testBulkCopyCrossTypeCSVFloat32ToFloat16	CSV declared as FLOAT32 imported into FLOAT16, per-row value assertion

New test resource

• BulkCopyCSVTestInputCrossTypeVector.csv — 3-row CSV with VECTOR(3) data for cross-type bulk copy tests.

Server capability covered

Capability	Tests
Implicit conversion in assignments (INSERT...SELECT)	testInsertFloat32FromFloat16Table, testInsertFloat16FromFloat32Table
Implicit conversion in parameter passing (setObject)	testSetObjectFloat32VectorIntoFloat16Column, testSetObjectFloat16VectorIntoFloat32Column
Explicit CAST	testExplicitCastFloat16ToFloat32, testExplicitCastFloat32ToFloat16
Explicit CONVERT	testConvertFloat16ToFloat32, testConvertFloat32ToFloat16
DDL (ALTER COLUMN)	testAlterColumnFloat16ToFloat32, testAlterColumnFloat32ToFloat16
SP cross-type parameter passing	testStoredProcedureCrossTypeParameterPassing
UDF cross-type parameter passing	testUdfCrossTypeParameterPassing
TVF cross-type return	testTvfCrossTypeReturn
UPDATE with cross-type value	testCrossTypeUpdateFloat32ColWithFloat16Value, testCrossTypeUpdateFloat16ColWithFloat32Value
Round-trip precision fidelity	testCrossTypeRoundTripPrecision
Precision-loss (lossy F32→F16)	testFloat16PrecisionLossWithNonRepresentableValues
Dimension mismatch still rejected	testCrossTypeDimensionMismatchFails
NULL vector cross-type handling	testNullVectorCrossTypeInsertSelect, testNullVectorCrossTypeSetObject
JDBC-layer cross-read round-trip	testGetObjectCrossReadFloat16AsFloat32RoundTrip
Bulk copy cross-type (table-to-table)	testBulkCopyFloat16ToFloat32Table, testBulkCopyFloat32ToFloat16Table, testMixedBulkCopyColumnMappingSwapped
Bulk copy cross-type (CSV)	testBulkCopyCrossTypeCSVFloat16ToFloat32, testBulkCopyCrossTypeCSVFloat32ToFloat16

Testing

All tests are tagged with @Tag(Constants.vectorFloat16Test) and require a server with vector version 2 support. Tests are excluded from default CI runs via excludedGroups in pom.xml.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 60.50%. Comparing base (1c2302b) to head (b9e3106).

Additional details and impacted files

@@             Coverage Diff              @@
##               main    #3005      +/-   ##
============================================
+ Coverage     60.26%   60.50%   +0.23%     
- Complexity     5168     5250      +82     
============================================
  Files           153      153              
  Lines         36646    36646              
  Branches       6727     6727              
============================================
+ Hits          22086    22173      +87     
+ Misses        10752    10735      -17     
+ Partials       3808     3738      -70     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

🟡 Not ready to approve

Several updated/new tests don’t fully validate converted vector contents (and two CSV tests ORDER BY a VECTOR column, which may fail on servers that don’t support ordering on VECTOR).

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

This PR updates the JDBC driver’s vector datatype test suite to align with a SQL Server capability change: implicit/explicit conversions between vector(..., float16) and vector(..., float32) now succeed where they previously failed, and new scenarios (CAST/CONVERT, DDL, DML, and bulk copy) are covered.

Changes:

• Updates existing FLOAT16/FLOAT32 cross-type tests to expect successful execution (rather than SQLException) and validate results.
• Adds new FLOAT16/FLOAT32 conversion tests covering CAST/CONVERT, ALTER COLUMN, UPDATE, and routine (SP/UDF/TVF) scenarios.
• Adds a new CSV resource used for cross-type bulk copy CSV test cases.

File summaries

File	Description
src/test/resources/BulkCopyCSVTestInputCrossTypeVector.csv	Adds CSV input data for cross-type VECTOR bulk copy CSV tests.
src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorTest.java	Broadens the accepted error message for “exceeds max vector dimension” test expectations.
src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java	Converts cross-type tests from “expected failure” to “expected success” and adds many new conversion scenarios.
src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/bulkcopy/VectorFloat16BulkCopyTest.java	Updates cross-type bulk copy tests to succeed and adds CSV cross-type bulk copy coverage.

Review details

Suppressed comments (9)

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/bulkcopy/VectorFloat16BulkCopyTest.java:1173

• Ordering by a VECTOR column can fail on servers that don't support ordering/comparison for VECTOR types, and it isn't needed for validating row count here. Dropping the ORDER BY makes the test more robust.

                try (ResultSet rs = stmt.executeQuery("SELECT v FROM " + dstTable + " ORDER BY v")) {

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:363

• This test asserts the converted vector's type and dimension count, but doesn't verify the element values after the FLOAT32→FLOAT16 cross-type INSERT...SELECT. Adding a data assertion would better validate the conversion behavior (with FLOAT16 tolerance).

                assertTrue(rs.next(), "No result found in FLOAT16 table after cross-type insert.");
                Vector result = rs.getObject("f16_col", Vector.class);
                assertNotNull(result, "Retrieved FLOAT16 vector is null.");
                assertEquals(VectorDimensionType.FLOAT16, result.getVectorDimensionType());
                assertEquals(3, result.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:412

• After inserting a FLOAT32 Vector via setObject into a FLOAT16 column, the test only checks type and dimension count. It should also validate that the stored values match the source vector within FLOAT16 tolerance.

                assertTrue(rs.next(), "No result found in FLOAT16 table after cross-type insert.");
                Vector result = rs.getObject("v", Vector.class);
                assertNotNull(result, "Retrieved FLOAT16 vector is null.");
                assertEquals(VectorDimensionType.FLOAT16, result.getVectorDimensionType());
                assertEquals(3, result.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:461

• After inserting a FLOAT16 Vector via setObject into a FLOAT32 column, the test only checks type and dimension count. It should also validate that the stored values match the source vector (FLOAT16→FLOAT32 should be exact).

                assertTrue(rs.next(), "No result found in FLOAT32 table after cross-type insert.");
                Vector result = rs.getObject("v", Vector.class);
                assertNotNull(result, "Retrieved FLOAT32 vector is null.");
                assertEquals(VectorDimensionType.FLOAT32, result.getVectorDimensionType());
                assertEquals(3, result.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:606

• The Javadoc says the ALTER COLUMN conversion "preserves data", but the test currently only verifies type and dimension count. Add a value assertion to ensure the converted vector elements are preserved.

            try (ResultSet rs = stmt.executeQuery("SELECT v FROM " + table + " WHERE id = 1")) {
                assertTrue(rs.next());
                Vector result = rs.getObject("v", Vector.class);
                assertNotNull(result);
                assertEquals(VectorDimensionType.FLOAT32, result.getVectorDimensionType());
                assertEquals(3, result.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:635

• The Javadoc says the ALTER COLUMN conversion preserves data within FLOAT16 precision, but the test currently only verifies type and dimension count. Add a value assertion to validate the actual conversion result.

            try (ResultSet rs = stmt.executeQuery("SELECT v FROM " + table + " WHERE id = 1")) {
                assertTrue(rs.next());
                Vector result = rs.getObject("v", Vector.class);
                assertNotNull(result);
                assertEquals(VectorDimensionType.FLOAT16, result.getVectorDimensionType());
                assertEquals(3, result.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/VectorFloat16Test.java:2344

• This test says it verifies the data was converted and stored correctly, but it currently only checks vector types and dimension counts. Add assertions on the stored element values to actually validate the cross-type conversion.

                Vector f32 = rs.getObject("v_float32", Vector.class);
                Vector f16 = rs.getObject("v_float16", Vector.class);
                assertNotNull(f32, "FLOAT32 result is null.");
                assertNotNull(f16, "FLOAT16 result is null.");
                assertEquals(VectorDimensionType.FLOAT32, f32.getVectorDimensionType());
                assertEquals(VectorDimensionType.FLOAT16, f16.getVectorDimensionType());
                assertEquals(3, f32.getDimensionCount());
                assertEquals(3, f16.getDimensionCount());

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/bulkcopy/VectorFloat16BulkCopyTest.java:1048

• This cross-type bulk copy test checks only the destination vector type and dimension count. Adding a value assertion would better validate that the implicit FLOAT16→FLOAT32 conversion preserved the expected element values.

                    Vector result = rs.getObject("v", Vector.class);
                    assertNotNull(result, "Retrieved FLOAT32 vector is null.");
                    assertEquals(VectorDimensionType.FLOAT32, result.getVectorDimensionType());
                    assertEquals(3, result.getDimensionCount());
                }

src/test/java/com/microsoft/sqlserver/jdbc/datatypes/vector/bulkcopy/VectorFloat16BulkCopyTest.java:1092

• This cross-type bulk copy test checks only the destination vector type and dimension count. Adding a value assertion would better validate that the implicit FLOAT32→FLOAT16 conversion produced the expected values (within FLOAT16 tolerance).

                    Vector result = rs.getObject("v", Vector.class);
                    assertNotNull(result, "Retrieved FLOAT16 vector is null.");
                    assertEquals(VectorDimensionType.FLOAT16, result.getVectorDimensionType());
                    assertEquals(3, result.getDimensionCount());
                }

• Files reviewed: 4/4 changed files
• Comments generated: 3
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.