Flaky test: client-v2 testDynamicWithPrimitives intermittently fails with ArithmeticException: Rounding necessary

[polyglotAI-bot]

Describe the bug

DataTypeTests.testDynamicWithPrimitives (module client-v2, group integration) is flaky: it intermittently fails with

java.lang.ArithmeticException: Rounding necessary
	at java.base/java.math.BigDecimal.divideAndRound(...)
	at ...DataTypeTests.testDynamicWithPrimitives(DataTypeTests.java)

Observed in CI on the client-v2 + ClickHouse 25.8 matrix job (roughly ~1% of runs); passes on most runs and most server versions.

Root cause

The shared test fixture DataTypesTestingPOJO assigns its four Decimal fields from an unseeded Random:

decimal32  = BigDecimal.valueOf(random.nextDouble());
decimal64  = BigDecimal.valueOf(random.nextDouble());
decimal128 = BigDecimal.valueOf(random.nextDouble());
decimal256 = BigDecimal.valueOf(random.nextDouble());

testDynamicWithPrimitives inserts each value into a Dynamic column and, for the Decimal* cases, verifies the round-trip with

BigDecimal tmpDec = row.getBigDecimal("field").stripTrailingZeros();
if (tmpDec.divide((BigDecimal) value, RoundingMode.UNNECESSARY).equals(BigDecimal.ONE)) { continue; }

When a BigDecimal is written into a Dynamic column, the type is inferred from the value's precision() (SerializerUtils.valueToColumnForDynamicType): the width is chosen by precision, but the scale is forced to that width's maximum (Decimal32→9, Decimal64→18, Decimal128→38, Decimal256→76). A small-magnitude random.nextDouble() can produce a value whose scale() exceeds that fixed scale (e.g. 0.0123456789 → precision 9 → Decimal32(9), but scale 10 > 9). The value is then rounded on write, the read-back value no longer equals the original, and divide(value, RoundingMode.UNNECESSARY) throws Rounding necessary. Because the input is random, the failure is non-deterministic.

Expected behaviour

The test should be deterministic. Making the Decimal samples fixed values whose scale is within each inferred width's capacity keeps the round-trip exact and preserves the assertion's intent, while also letting the four samples genuinely exercise Decimal32/64/128/256 (the random doubles almost always landed in Decimal64).

Environment

• clickhouse-java: main (client-v2)
• ClickHouse server: reproduced deterministically with a crafted small-scale value on current main; originally observed on CI with server 25.8