prisma-driver-adapter-implementation

Implement SqlDriverAdapterFactory with connect(), then SqlDriverAdapter with queryRaw, executeRaw, executeScript, startTransaction, and dispose. Map arguments using both value and ArgType, return column names, types, and rows in stable order, and wrap driver failures in DriverAdapterError.

startTransaction acquires one dedicated connection, issues BEGIN, applies the isolation level, and returns a Transaction bound to that connection. commit() and rollback() are lifecycle cleanup hooks only; Prisma issues the SQL COMMIT or ROLLBACK via executeRaw, and savepoints are optional methods on the Transaction object.

Yes, through optional createSavepoint, rollbackToSavepoint, and releaseSavepoint methods on the Transaction object, implemented only where the provider supports them. Savepoint state must live on the transaction connection, never as adapter-global depth, because parallel transactions make global counters incorrect.

P2039 surfaces when an unmapped driver error reaches Prisma. Preserve originalCode and originalMessage in your DriverAdapterError so P2039 carries useful details, map known conditions like unique violations to structured MappedError kinds, and rethrow genuinely unexpected non-driver errors.

A shadow database is an isolated database created by connectToShadowDb() on SqlMigrationAwareDriverAdapterFactory so Migrate can validate migrations safely. It must use cryptographically unique quoted names, never point at the primary database, and be dropped during disposal or failure cleanup.