Stringify numeric parameters bound to character columns (re: #163)

src/binary.cpp

@@ -29,6 +29,10 @@ char *DecimalChars::data() {
 	return characters.data();
 }
 
+SQLWCHAR *DecimalChars::wide_data() {
+	return wide_characters.data();
+}
+
 ScannerBlob::ScannerBlob() {
 }
 

src/include/binary.hpp

@@ -4,11 +4,17 @@
 #include <vector>
 
 #include "duckdb_extension_api.hpp"
+#include "odbc_api.hpp"
 
 namespace odbcscanner {
 
 struct DecimalChars {
 	std::vector<char> characters;
+	// Optional wide buffer populated by BindOdbcParam<DecimalChars> when the
+	// prepared parameter's expected SQL type is SQL_WCHAR / SQL_WVARCHAR /
+	// SQL_WLONGVARCHAR. Kept alongside `characters` so the binding's lifetime
+	// matches the ScannerValue.
+	std::vector<SQLWCHAR> wide_characters;
 
 	DecimalChars();
 
@@ -26,6 +32,8 @@ struct DecimalChars {
 	}
 
 	char *data();
+
+	SQLWCHAR *wide_data();
 };
 
 struct ScannerBlob {

src/include/scanner_value.hpp

@@ -163,6 +163,16 @@ class ScannerValue {
 
 	void TransformIntegralToDecimal();
 
+	// Stringifies an integral/float parameter in-place and re-tags it as
+	// TYPE_DECIMAL_AS_CHARS so the scanner binds via BindOdbcParam<DecimalChars>
+	// instead of a numeric C type. This avoids driver code paths that convert
+	// numeric-C → character-SQL, which are historically a common source of silent
+	// data corruption across ODBC drivers (e.g. Firebird ODBC ≤ 3.5.0, some
+	// MSSQL/MySQL releases). Wide character columns (SQL_WCHAR/SQL_WVARCHAR/
+	// SQL_WLONGVARCHAR) are handled by BindOdbcParam<DecimalChars> itself, which
+	// widens the buffer on demand.
+	void TransformNumericToChars();
+
 private:
 	void CheckType(param_type expected);
 

src/include/types.hpp

@@ -42,6 +42,10 @@ struct Types {
 
 	static const std::string UNKNOWN_DUCKDB_TYPE_NAME;
 
+	static bool IsCharacterSQLType(SQLSMALLINT t);
+
+	static bool IsWideCharacterSQLType(SQLSMALLINT t);
+
 	static const SQLSMALLINT SQL_SS_TIME2 = -154;
 	static const SQLSMALLINT SQL_SS_TIMESTAMPOFFSET = -155;
 

src/params.cpp

@@ -134,6 +134,35 @@ std::vector<SQLSMALLINT> Params::CollectTypes(QueryContext &ctx) {
 	return param_types;
 }
 
+// When the prepared-parameter's target is a character SQL type and the source is
+// numeric, stringify in the scanner. This short-circuits the driver's
+// numeric-C → character-SQL conversion path, which has produced silent data loss
+// in multiple ODBC drivers (FirebirdSQL/firebird-odbc-driver#292 and older
+// MSSQL/MySQL releases). Done here — before binding — so that post-SetExpectedTypes
+// the param's type_id is final. Wide targets (SQL_WCHAR/WVARCHAR/WLONGVARCHAR)
+// are handled by BindOdbcParam<DecimalChars>, which widens the buffer on demand.
+static void CoalesceNumericToCharsIfNeeded(ScannerValue &param) {
+	if (!Types::IsCharacterSQLType(param.ExpectedType())) {
+		return;
+	}
+	switch (param.ParamType()) {
+	case DUCKDB_TYPE_TINYINT:
+	case DUCKDB_TYPE_UTINYINT:
+	case DUCKDB_TYPE_SMALLINT:
+	case DUCKDB_TYPE_USMALLINT:
+	case DUCKDB_TYPE_INTEGER:
+	case DUCKDB_TYPE_UINTEGER:
+	case DUCKDB_TYPE_BIGINT:
+	case DUCKDB_TYPE_UBIGINT:
+	case DUCKDB_TYPE_FLOAT:
+	case DUCKDB_TYPE_DOUBLE:
+		param.TransformNumericToChars();
+		break;
+	default:
+		break;
+	}
+}
+
 void Params::SetExpectedTypes(QueryContext &ctx, const std::vector<SQLSMALLINT> &expected,
                               std::vector<ScannerValue> &actual) {
 	if (expected.size() != actual.size()) {
@@ -145,6 +174,7 @@ void Params::SetExpectedTypes(QueryContext &ctx, const std::vector<SQLSMALLINT>
 		ScannerValue &param = actual.at(i);
 		Types::CoalesceParameterType(ctx, param);
 		param.SetExpectedType(expected_type);
+		CoalesceNumericToCharsIfNeeded(param);
 	}
 }
 

src/scanner_value.cpp

@@ -1,5 +1,6 @@
 #include "scanner_value.hpp"
 
+#include <cstdio>
 #include <cstdlib>
 #include <cstring>
 #include <string>
@@ -522,4 +523,59 @@ void ScannerValue::TransformIntegralToDecimal() {
 	*this = ScannerValue(dec, false);
 }
 
+void ScannerValue::TransformNumericToChars() {
+	std::string str;
+	switch (type_id) {
+	case DUCKDB_TYPE_TINYINT:
+		str = std::to_string(static_cast<int32_t>(Value<int8_t>()));
+		break;
+	case DUCKDB_TYPE_UTINYINT:
+		str = std::to_string(static_cast<uint32_t>(Value<uint8_t>()));
+		break;
+	case DUCKDB_TYPE_SMALLINT:
+		str = std::to_string(Value<int16_t>());
+		break;
+	case DUCKDB_TYPE_USMALLINT:
+		str = std::to_string(Value<uint16_t>());
+		break;
+	case DUCKDB_TYPE_INTEGER:
+		str = std::to_string(Value<int32_t>());
+		break;
+	case DUCKDB_TYPE_UINTEGER:
+		str = std::to_string(Value<uint32_t>());
+		break;
+	case DUCKDB_TYPE_BIGINT:
+		str = std::to_string(Value<int64_t>());
+		break;
+	case DUCKDB_TYPE_UBIGINT:
+		str = std::to_string(Value<uint64_t>());
+		break;
+	case DUCKDB_TYPE_FLOAT: {
+		char buf[32];
+		std::snprintf(buf, sizeof(buf), "%.9g", static_cast<double>(Value<float>()));
+		str = buf;
+		break;
+	}
+	case DUCKDB_TYPE_DOUBLE: {
+		char buf[32];
+		std::snprintf(buf, sizeof(buf), "%.17g", Value<double>());
+		str = buf;
+		break;
+	}
+	default:
+		throw ScannerException("Invalid numeric param type for chars transform: " + std::to_string(type_id));
+	}
+
+	// Destroy the current (POD) value and repurpose the union as DecimalChars.
+	// Wide character targets are handled downstream by BindOdbcParam<DecimalChars>.
+	this->Destroy();
+	this->type_id = Params::TYPE_DECIMAL_AS_CHARS;
+	new (&this->val.decimal_chars) DecimalChars;
+	DecimalChars &dc = this->val.decimal_chars;
+	dc.characters.resize(str.size() + 1);
+	std::memcpy(dc.characters.data(), str.data(), str.size());
+	dc.characters[str.size()] = '\0';
+	this->len_bytes = static_cast<SQLLEN>(str.size());
+}
+
 } // namespace odbcscanner

src/types/decimal_type.cpp

@@ -9,6 +9,7 @@
 #include "connection.hpp"
 #include "diagnostics.hpp"
 #include "scanner_exception.hpp"
+#include "widechar.hpp"
 
 DUCKDB_EXTENSION_EXTERN
 
@@ -82,10 +83,37 @@ void TypeSpecific::BindOdbcParam<SQL_NUMERIC_STRUCT>(QueryContext &ctx, ScannerV
 
 template <>
 void TypeSpecific::BindOdbcParam<DecimalChars>(QueryContext &ctx, ScannerValue &param, SQLSMALLINT param_idx) {
-	SQLSMALLINT sqltype = param.ExpectedType() != SQL_PARAM_TYPE_UNKNOWN ? param.ExpectedType() : SQL_NUMERIC;
+	SQLSMALLINT expected = param.ExpectedType();
+	// Preserve the driver's column type when it is a CHAR/VARCHAR/WCHAR family:
+	// binding → the actual expected type avoids an extra driver-side coercion.
+	// For non-character expected types (e.g. SQL_NUMERIC), fall back to SQL_VARCHAR
+	// and let the driver parse the char buffer into the target type.
+	SQLSMALLINT sqltype = Types::IsCharacterSQLType(expected) ? expected : SQL_VARCHAR;
 	DecimalChars &dc = param.Value<DecimalChars>();
+
+	// Wide targets: widen the ASCII digits to SQLWCHAR and bind SQL_C_WCHAR.
+	// Some drivers do not reliably auto-convert SQL_C_CHAR → SQL_W*CHAR.
+	if (Types::IsWideCharacterSQLType(expected)) {
+		if (dc.wide_characters.empty()) {
+			WideString wstr = WideChar::Widen(dc.data(), static_cast<size_t>(dc.size<SQLLEN>()));
+			dc.wide_characters = std::move(wstr.vec);
+		}
+		SQLLEN length_chars = static_cast<SQLLEN>(dc.wide_characters.size() - 1);
+		param.LengthBytes() = length_chars * static_cast<SQLLEN>(sizeof(SQLWCHAR));
+		SQLRETURN ret = SQLBindParameter(
+		    ctx.hstmt(), param_idx, SQL_PARAM_INPUT, SQL_C_WCHAR, sqltype, static_cast<SQLULEN>(length_chars), 0,
+		    reinterpret_cast<SQLPOINTER>(dc.wide_data()), param.LengthBytes(), &param.LengthBytes());
+		if (!SQL_SUCCEEDED(ret)) {
+			std::string diag = Diagnostics::Read(ctx.hstmt(), SQL_HANDLE_STMT);
+			throw ScannerException("'SQLBindParameter' failed, type: " + std::to_string(sqltype) +
+			                       ", index: " + std::to_string(param_idx) + ", query: '" + ctx.query +
+			                       "', return: " + std::to_string(ret) + ", diagnostics: '" + diag + "'");
+		}
+		return;
+	}
+
 	SQLRETURN ret =
-	    SQLBindParameter(ctx.hstmt(), param_idx, SQL_PARAM_INPUT, SQL_C_CHAR, SQL_VARCHAR, param.LengthBytes(), 0,
+	    SQLBindParameter(ctx.hstmt(), param_idx, SQL_PARAM_INPUT, SQL_C_CHAR, sqltype, param.LengthBytes(), 0,
 	                     reinterpret_cast<SQLPOINTER>(dc.data()), param.LengthBytes(), &param.LengthBytes());
 	if (!SQL_SUCCEEDED(ret)) {
 		std::string diag = Diagnostics::Read(ctx.hstmt(), SQL_HANDLE_STMT);

src/types/float_types.cpp

@@ -54,6 +54,10 @@ static void BindOdbcParamInternal(QueryContext &ctx, SQLSMALLINT ctype, SQLSMALL
 	}
 }
 
+// If the expected SQL type is character, Params::SetExpectedTypes transforms the
+// ScannerValue to TYPE_DECIMAL_AS_CHARS before we get here — see the comment in
+// integer_types.cpp for the rationale.
+
 template <>
 void TypeSpecific::BindOdbcParam<float>(QueryContext &ctx, ScannerValue &param, SQLSMALLINT param_idx) {
 	SQLSMALLINT sqltype = param.ExpectedType() != SQL_PARAM_TYPE_UNKNOWN ? param.ExpectedType() : SQL_FLOAT;

src/types/integer_types.cpp

@@ -174,6 +174,15 @@ static SQLSMALLINT IntegralSQLType(ScannerValue &param, SQLSMALLINT def_sqltype)
 	return def_sqltype;
 }
 
+// Note: if the prepared-parameter's expected SQL type is a character type
+// (CHAR / VARCHAR / WCHAR family), the ScannerValue is expected to have been
+// transformed to TYPE_DECIMAL_AS_CHARS by Params::SetExpectedTypes before we get
+// here — dispatch will then route to BindOdbcParam<DecimalChars> instead of one of
+// the integer specializations below. Stringifying in the scanner avoids the
+// driver's numeric-C → character-SQL path, which has shipped silent-corruption bugs
+// across several ODBC drivers (Firebird ≤ 3.5.0 in
+// FirebirdSQL/firebird-odbc-driver#292; older MSSQL / MySQL releases too).
+
 template <>
 void TypeSpecific::BindOdbcParam<int8_t>(QueryContext &ctx, ScannerValue &param, SQLSMALLINT param_idx) {
 	SQLSMALLINT sqltype = IntegralSQLType(param, SQL_TINYINT);

src/types/types.cpp

@@ -32,6 +32,15 @@ const std::string Types::SQL_BLOB_TYPE_NAME = "BLOB";
 const std::string Types::SQL_CLOB_TYPE_NAME = "CLOB";
 const std::string Types::SQL_DB2_DBCLOB_TYPE_NAME = "DBCLOB";
 
+bool Types::IsCharacterSQLType(SQLSMALLINT t) {
+	return t == SQL_CHAR || t == SQL_VARCHAR || t == SQL_LONGVARCHAR || t == SQL_WCHAR || t == SQL_WVARCHAR ||
+	       t == SQL_WLONGVARCHAR;
+}
+
+bool Types::IsWideCharacterSQLType(SQLSMALLINT t) {
+	return t == SQL_WCHAR || t == SQL_WVARCHAR || t == SQL_WLONGVARCHAR;
+}
+
 ScannerValue Types::ExtractNotNullParam(DbmsQuirks &quirks, duckdb_type type_id, duckdb_vector vec, idx_t row_idx,
                                         idx_t param_idx) {
 	switch (type_id) {

test/sql/db2/db2_copy.test

@@ -400,5 +400,37 @@ NULL	NULL	11	s11	NULL	NULL	NULL	NULL
 statement ok
 SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE DUCKDB_TEST_COPY')
 
+# int → VARCHAR primary key round-trip (regression test for duckdb/odbc-scanner#161).
+# DB2 folds unquoted identifiers to upper case; column_quotes='' keeps the INSERT
+# column names unquoted so `id` resolves to the `ID` column created here.
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE INT_TO_VARCHAR_PK', ignore_exec_failure=TRUE)
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'),
+  'CREATE TABLE INT_TO_VARCHAR_PK (id VARCHAR(20) NOT NULL PRIMARY KEY)')
+
+query II
+SELECT completed, rows_processed FROM odbc_copy(getvariable('conn'),
+  dest_table='INT_TO_VARCHAR_PK',
+  batch_size=1,
+  column_quotes='',
+  source_query='SELECT i::INTEGER AS id FROM range(1, 501) t(i)')
+----
+1	500
+
+query I
+SELECT * FROM odbc_query(getvariable('conn'),
+  'SELECT id FROM INT_TO_VARCHAR_PK WHERE id IN (''1'', ''2'', ''42'', ''500'') ORDER BY id')
+----
+1
+2
+42
+500
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE INT_TO_VARCHAR_PK')
+
 statement ok
 SELECT odbc_close(getvariable('conn'))

test/sql/duckdb/duckdb_copy.test

@@ -88,5 +88,39 @@ _foo_col1_foo_
 statement ok
 SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE duckdb_test_copy')
 
+# int → VARCHAR primary key round-trip (regression test for duckdb/odbc-scanner#161).
+# The Firebird ODBC driver ≤ 3.5.0 silently dropped rows on this exact shape;
+# Params::SetExpectedTypes now stringifies the parameter in the scanner, which
+# sidesteps the driver's numeric-C → character-SQL path on every driver.
+# batch_size=1 exercises the per-row bind that was the most fragile shape.
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE IF EXISTS duckdb_int_to_varchar_pk')
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'),
+  'CREATE TABLE duckdb_int_to_varchar_pk (id VARCHAR(20) NOT NULL PRIMARY KEY)')
+
+query II
+SELECT completed, rows_processed FROM odbc_copy(getvariable('conn'),
+  dest_table='duckdb_int_to_varchar_pk',
+  batch_size=1,
+  column_quotes='',
+  source_query='SELECT i::INTEGER AS id FROM range(1, 501) t(i)')
+----
+1	500
+
+query I
+SELECT * FROM odbc_query(getvariable('conn'),
+  'SELECT id FROM duckdb_int_to_varchar_pk WHERE id IN (''1'', ''2'', ''42'', ''500'') ORDER BY id')
+----
+1
+2
+42
+500
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE duckdb_int_to_varchar_pk')
+
 statement ok
 SELECT odbc_close(getvariable('conn'))

test/sql/firebird/firebird_copy.test

@@ -397,5 +397,43 @@ NULL	NULL	11	s11	NULL	NULL	NULL	NULL
 statement ok
 SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE DUCKDB_TEST_COPY')
 
+# int → VARCHAR primary key round-trip (regression test for duckdb/odbc-scanner#161)
+#
+# Before the scanner stringified numeric parameters bound to a character column,
+# this exact shape on the Firebird ODBC driver ≤ 3.5.0 silently stored 11 rows
+# with NUL-byte-corrupted PK values. We exercise batch_size=1 because that was
+# the shape the original bug dropped rows on; column_quotes='' keeps the INSERT
+# column names unquoted so Firebird matches the `ID` column created below.
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE INT_TO_VARCHAR_PK', ignore_exec_failure=TRUE)
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'CREATE TABLE INT_TO_VARCHAR_PK (id VARCHAR(20) NOT NULL PRIMARY KEY)')
+
+query II
+SELECT completed, rows_processed FROM odbc_copy(getvariable('conn'),
+  dest_table='INT_TO_VARCHAR_PK',
+  batch_size=1,
+  column_quotes='',
+  source_query='SELECT i::INTEGER AS id FROM range(1, 501) t(i)')
+----
+1	500
+
+# Lexicographic sort: '1','2','42','500'. Checking sampled values guards against
+# the original bug, which stored distinct but corrupted strings (NUL bytes) while
+# returning SUCCESS from every driver call.
+query I
+SELECT * FROM odbc_query(getvariable('conn'),
+  'SELECT id FROM INT_TO_VARCHAR_PK WHERE id IN (''1'', ''2'', ''42'', ''500'') ORDER BY id')
+----
+1
+2
+42
+500
+
+statement ok
+SELECT * FROM odbc_query(getvariable('conn'), 'DROP TABLE INT_TO_VARCHAR_PK')
+
 statement ok
 SELECT odbc_close(getvariable('conn'))