client-v2: serializePrimitiveData rejects or silently corrupts String values for UInt64, UUID, Date and DateTime

Inizia in client-v2/src/main/java/com/clickhouse/client/api/data_formats/internal/SerializerUtils.java, in serializePrimitiveData, e segui i percorsi helper per UInt64, UUID, data e data-ora. Verifica il comportamento di conversione esistente e il controllo dell'intervallo di unsigned long correlato in BinaryStreamUtils. Il lavoro è completato quando i valori String vengono gestiti in modo coerente per questi tipi, i valori UInt64 non validi generano un errore invece di effettuare un wrapping silenzioso e i percorsi signed o numerici esistenti rimangono invariati.

Summary

SerializerUtils.serializePrimitiveData accepts String input for most scalar column types via the convertToInteger / convertToLong / convertToBigInteger / convertToBigDecimal / convertToString helpers, but a subset of types either reject a String outright or — for UInt64 — silently corrupt it. This makes String coercion inconsistent across the type switch.

Line references are against 0.9.5, file client-v2/src/main/java/com/clickhouse/client/api/data_formats/internal/SerializerUtils.java.

Behaviour today

The UInt64 case is a silent-corruption bug

convertToLong("-1")      // -> -1L        (Long.parseLong, signed)
writeUnsignedInt64(-1L)  // -> writeInt64 -> FF FF FF FF FF FF FF FF

ClickHouse reads that back as 18446744073709551615. So a negative String is written as the maximum UInt64 with no error.

This is already inconsistent within the library: BinaryStreamUtils.writeUnsignedInt64(OutputStream, BigInteger) range-checks via ClickHouseChecker.between(value, TYPE_BIG_INTEGER, BigInteger.ZERO, U_INT64_MAX), so a BigInteger of -1 is correctly rejected. Only the String path wraps.

Values in [2^63, 2^64-1] are legal UInt64 values and are encodable on the wire (writeUnsignedInt64(long) just emits the raw 8 bytes), but unreachable through the String path.

Proposed fix

1. UInt64 — use an unsigned-aware conversion at :522 rather than changing the shared convertToLong, which Int64 (:504) and UInt32 (:519) also use and which needs signed parsing:

/** UInt64 runs past Long.MAX_VALUE; the returned long carries the raw bit pattern. */
public static Long convertToUnsignedLong(Object value) {
    if (value instanceof String) {
        return Long.parseUnsignedLong((String) value);
    }
    return convertToLong(value);
}

2. UUID — add a String branch using UUID.fromString.
3. Date / Date32 / DateTime / DateTime64 — add a String branch in the writeDate / writeDate32 / writeDateTime / writeDateTime64 helpers, which already dispatch on LocalDate / ZonedDateTime / OffsetDateTime / Timestamp. Fixing them there also covers their other callers.

Backward compatibility

Items 2 and 3 are purely additive: those inputs throw today, so nothing that currently succeeds changes behaviour.

Item 1 has one deliberate behaviour change: "-1" into a UInt64 column currently writes 18446744073709551615 and would now throw NumberFormatException. That replaces silent corruption with a loud error and aligns the String path with the existing BigInteger overload. Suggest keeping it as its own commit so it can be reviewed separately from the additive branches.

The Long / Number path stays permissive and passes raw bits through unchanged — callers legitimately supply the bit pattern for values above 2^63.

Why it matters

Reported from the ClickHouse Flink connector. Its checkpoint state format serialises map keys as strings, so for a Map(K, V) column the key reaches serializeMapData (:478) → serializePrimitiveData as a String. The connector currently has to reject Map(UInt64, …), Map(UUID, …), Map(Date, …), Map(Date32, …) and Map(DateTime(64), …) at planning time and tell users to change their table DDL, even though all of these are legal ClickHouse map key types.

More generally, any caller passing a String value for one of these types — nested or top-level — hits the same behaviour.