prisma-driver-adapter-implementation

Prisma SQL Driver Adapter Implementation

Use this guide with the exact @prisma/driver-adapter-utils version installed by the target Prisma release. Driver adapters are a protocol boundary: type-compatible code can still corrupt values, leak connections, or break transactions.

When to Apply

• Implementing SqlDriverAdapterFactory, SqlMigrationAwareDriverAdapterFactory, SqlDriverAdapter, or Transaction
• Adding nested-transaction/savepoint support
• Mapping driver values, column metadata, bind arguments, or database errors
• Debugging P2039, transaction leaks, shadow-database failures, or adapter-specific query behavior

Contract snapshot

interface SqlDriverAdapterFactory extends AdapterInfo {
  connect(): Promise<SqlDriverAdapter>
}

interface SqlMigrationAwareDriverAdapterFactory extends SqlDriverAdapterFactory {
  connectToShadowDb(): Promise<SqlDriverAdapter>
}

interface SqlDriverAdapter extends AdapterInfo {
  queryRaw(query: SqlQuery): Promise<SqlResultSet>
  executeRaw(query: SqlQuery): Promise<number>
  executeScript(script: string): Promise<void>
  startTransaction(isolationLevel?: IsolationLevel): Promise<Transaction>
  getConnectionInfo?(): ConnectionInfo
  dispose(): Promise<void>
}

interface Transaction extends AdapterInfo {
  readonly options: { usePhantomQuery: boolean }
  queryRaw(query: SqlQuery): Promise<SqlResultSet>
  executeRaw(query: SqlQuery): Promise<number>
  commit(): Promise<void>
  rollback(): Promise<void>
  createSavepoint?(name: string): Promise<void>
  rollbackToSavepoint?(name: string): Promise<void>
  releaseSavepoint?(name: string): Promise<void>
}

IsolationLevel currently includes READ UNCOMMITTED, READ COMMITTED, REPEATABLE READ, SNAPSHOT, and SERIALIZABLE; validate what the concrete database supports.

Priority rules

Query implementation

SqlQuery contains sql, args, and parallel argTypes. Map each argument using both value and ArgType; do not discard type/arity information. Execute in the driver's array/tuple row mode so column order is stable.

class ExampleQueryable {
  readonly provider = 'postgres' as const
  readonly adapterName = '@acme/adapter-example'

  constructor(protected readonly connection: DriverConnection) {}

  async queryRaw(query: SqlQuery): Promise<SqlResultSet> {
    try {
      const result = await this.connection.query({
        text: query.sql,
        values: query.args.map((value, index) =>
          mapArg(value, query.argTypes[index]),
        ),
        rowMode: 'array',
      })

      return {
        columnNames: result.fields.map((field) => field.name),
        columnTypes: result.fields.map(mapColumnType),
        rows: result.rows,
      }
    } catch (error) {
      throwAdapterError(error)
    }
  }

  async executeRaw(query: SqlQuery): Promise<number> {
    try {
      const result = await this.connection.execute(
        query.sql,
        query.args.map((value, index) => mapArg(value, query.argTypes[index])),
      )
      return result.rowsAffected ?? 0
    } catch (error) {
      throwAdapterError(error)
    }
  }
}

Result mapping

Return columnNames, columnTypes, and rows with identical lengths/order. Map driver metadata to ColumnTypeEnum deliberately:

• signed integer widths to Int32/Int64; preserve 64-bit values without JS number truncation
• decimal/numeric to Numeric using the representation expected by Prisma
• binary to Uint8Array/Bytes
• date-only, time-only, and timestamp to Date, Time, and DateTime
• UUID, JSON, enum, arrays, and provider-specific unknown values to their explicit types
• unsupported native types to DriverAdapterError({ kind: 'UnsupportedNativeDataType', type })

Test null, empty arrays, array element types, big integers, decimals, byte arrays, JSON, dates, and user-defined/unknown native types.

Script execution

executeScript must execute a migration script as the provider expects. Prefer the driver's native multi-statement/script facility or a real SQL parser. Naively splitting on ; breaks functions, triggers, quoted strings, and dialect-specific blocks. Only pass trusted scripts generated by Prisma Migrate to this method, and keep user-controlled input out of it. Use least privilege credentials for migration and shadow connections.

Transaction protocol

startTransaction must acquire one dedicated connection, start the database transaction, apply the requested isolation level, and return a Transaction bound to that same connection. If setup fails, release it immediately. If setup fails after BEGIN succeeded, issue a best effort ROLLBACK or explicitly destroy the connection instead of relying on driver specific release behavior.

async startTransaction(level?: IsolationLevel): Promise<Transaction> {
  const connection = await this.pool.acquire()
  try {
    const tx = new ExampleTransaction(connection, () => connection.release())
    await tx.executeRaw({ sql: 'BEGIN', args: [], argTypes: [] })
    if (level) {
      await tx.executeRaw({
        sql: `SET TRANSACTION ISOLATION LEVEL ${validateLevel(level)}`,
        args: [],
        argTypes: [],
      })
    }
    return tx
  } catch (error) {
    connection.release(error)
    throwAdapterError(error)
  }
}

The isolation level helper must look the level up in a fixed map from each supported isolation level to its provider SQL keywords. Any other value must fail as InvalidIsolationLevel, so raw strings never reach the statement. Some providers require setting the isolation level before the transaction begins, so order the statements to match the provider.

Commit and rollback

Prisma coordinates the SQL COMMIT/ROLLBACK through executeRaw. The transaction object's commit() and rollback() methods are lifecycle hooks: detach listeners and release the dedicated connection exactly once. They must not issue a second SQL commit/rollback. If a control statement failed, or the driver reports the connection is still inside a transaction at release time, destroy the connection instead of returning it to the pool.

class ExampleTransaction extends ExampleQueryable implements Transaction {
  readonly options = { usePhantomQuery: false }
  #closed = false

  constructor(connection: DriverConnection, private readonly release: () => void) {
    super(connection)
  }

  async commit() { this.finish() }
  async rollback() { this.finish() }

  private finish() {
    if (this.#closed) return
    this.#closed = true
    this.release()
  }

  async createSavepoint(name: string) {
    await this.control(`SAVEPOINT ${safeSavepoint(name)}`)
  }

  async rollbackToSavepoint(name: string) {
    await this.control(`ROLLBACK TO SAVEPOINT ${safeSavepoint(name)}`)
  }

  async releaseSavepoint(name: string) {
    await this.control(`RELEASE SAVEPOINT ${safeSavepoint(name)}`)
  }

  private async control(sql: string) {
    await this.executeRaw({ sql, args: [], argTypes: [] })
  }
}

Implement the optional savepoint methods only where the provider supports them. Validate/quote savepoint identifiers. The savepoint helper must accept only names that start with a letter or underscore, contain only letters, digits, and underscores, and fit the provider identifier length limit; it must reject anything else with a DriverAdapterError and then apply provider identifier quoting. After commit() or rollback() releases the connection, the transaction must reject further queries and savepoint calls with the closed transaction error kind of the installed version. For providers whose savepoints are intentionally no-ops, document and test that limitation.

Never keep transaction depth on the shared adapter. Parallel transactions make adapter-global depth incorrect; nested state belongs to the returned transaction connection and Prisma's savepoint calls.

Error mapping

Wrap recognized driver failures in DriverAdapterError. Map known conditions to MappedError kinds such as constraint violations, authentication/reachability, missing table/column/database, timeouts, closed transactions, invalid input, value range, and write conflicts.

For database errors, preserve originalCode and originalMessage even when falling back to the provider-specific raw variant:

import {
  DriverAdapterError,
  type Error as DriverAdapterErrorObject,
  type MappedError,
} from '@prisma/driver-adapter-utils'

function convertDriverError(error: DatabaseError): DriverAdapterErrorObject {
  return {
    originalCode: String(error.code ?? 'N/A'),
    originalMessage: error.message,
    ...mapKnownOrRaw(error),
  }
}

function mapKnownOrRaw(error: DatabaseError): MappedError {
  if (error.code === '23505') {
    return { kind: 'UniqueConstraintViolation', constraint: parsedConstraint(error) }
  }
  return {
    kind: 'postgres',
    code: String(error.code ?? 'N/A'),
    severity: error.severity ?? 'N/A',
    message: error.message,
    detail: error.detail,
    column: error.column,
    hint: error.hint,
  }
}

function throwAdapterError(error: unknown): never {
  if (!isDatabaseError(error)) throw error
  throw new DriverAdapterError(convertDriverError(error))
}

Prisma uses preserved original details when an unmapped driver error becomes P2039. Do not replace every unknown exception with a fabricated GenericJs id; rethrow genuinely unexpected non-driver errors so programming bugs remain visible.

Database error messages and details, including P2039 output, may contain row values or schema names. Applications should keep them in server logs and return generic errors to end users. Keep connection strings and credentials out of adapter error messages.

Factory, ownership, and shadow database

• connect() returns a fresh usable adapter connection/pool wrapper.
• Track whether the factory created the pool. dispose() closes owned pools and only detaches listeners from caller-owned pools unless an explicit option transfers ownership.
• Implement SqlMigrationAwareDriverAdapterFactory only when connectToShadowDb() can create an isolated shadow database, connect to it, and drop it during disposal/failure cleanup.
• Never point the shadow adapter at the primary database. Quote generated identifiers and use cryptographically unique names.
• Drop only the exact name this factory instance generated and recorded, which carries a fixed shadow prefix. Before dropping, check that the name equals the recorded name, carries the prefix, and differs from the primary database name, then issue the drop with a quoted identifier over a separate maintenance connection.
• getConnectionInfo() should accurately report schemaName, maxBindValues when applicable, and supportsRelationJoins.

Verification checklist

• Typecheck against the exact target @prisma/driver-adapter-utils version
• queryRaw preserves column order, types, nulls, and precision
• executeRaw reports affected rows correctly
• executeScript handles provider-specific multi-statement syntax
• Concurrent interactive transactions use distinct dedicated connections
• Success commits and releases once; failure rolls back and releases once
• Queries and savepoint calls after commit or rollback fail as closed transactions
• A transaction aborted midway never leaves a pooled connection still inside a transaction
• Nested transaction tests exercise create/rollback/release savepoint hooks
• Savepoint names containing quotes, semicolons, or whitespace are rejected
• Unsupported isolation levels fail as InvalidIsolationLevel
• Known constraints map to structured errors
• Unmapped database errors retain original code/message and surface useful P2039
• Dispose ownership is tested for internal and external pools
• Shadow database creation, use, failure cleanup, and disposal are isolated
• Run Prisma Client integration/E2E tests, not only adapter unit tests

Source references

• Driver adapter interfaces
• PostgreSQL adapter transaction implementation
• PostgreSQL adapter error mapping

These links follow a moving branch. Compare them against the release tag matching the installed version, and treat the typecheck against the exact installed package as authoritative.