Custom Types

A variety of methods exist to redefine the behavior of existing types as well as to provide new ones.

Overriding Type Compilation

A frequent need is to force the “string” version of a type, that is the one rendered in a CREATE TABLE statement or other SQL function like CAST, to be changed. For example, an application may want to force the rendering of BINARY for all platforms except for one, in which is wants BLOB to be rendered. Usage of an existing generic type, in this case LargeBinary, is preferred for most use cases. But to control types more accurately, a compilation directive that is per-dialect can be associated with any type:

  1. from sqlalchemy.ext.compiler import compiles
  2. from sqlalchemy.types import BINARY
  4. @compiles(BINARY, "sqlite")
  5. def compile_binary_sqlite(type_, compiler, **kw):
  6.     return "BLOB"

The above code allows the usage of BINARY, which will produce the string BINARY against all backends except SQLite, in which case it will produce BLOB.

See the section Changing Compilation of Types, a subsection of Custom SQL Constructs and Compilation Extension, for additional examples.

Augmenting Existing Types

The TypeDecorator allows the creation of custom types which add bind-parameter and result-processing behavior to an existing type object. It is used when additional in-Python marshalling of data to and/or from the database is required.

Note

The bind- and result-processing of TypeDecorator is in addition to the processing already performed by the hosted type, which is customized by SQLAlchemy on a per-DBAPI basis to perform processing specific to that DBAPI. While it is possible to replace this handling for a given type through direct subclassing, it is never needed in practice and SQLAlchemy no longer supports this as a public use case.

ORM Tip

The TypeDecorator can be used to provide a consistent means of converting some type of value as it is passed into and out of the database. When using the ORM, a similar technique exists for converting user data from arbitrary formats which is to use the validates() decorator. This technique may be more appropriate when data coming into an ORM model needs to be normalized in some way that is specific to the business case and isn’t as generic as a datatype.

Object NameDescription

TypeDecorator

Allows the creation of types which add additional functionality to an existing type.

class sqlalchemy.types.``TypeDecorator(\args, **kwargs*)

Allows the creation of types which add additional functionality to an existing type.

This method is preferred to direct subclassing of SQLAlchemy’s built-in types as it ensures that all required functionality of the underlying type is kept in place.

Typical usage:

  1. import sqlalchemy.types as types
  3. class MyType(types.TypeDecorator):
  4.     '''Prefixes Unicode values with "PREFIX:" on the way in and
  5.     strips it off on the way out.
  6.     '''
  8.     impl = types.Unicode
  10.     def process_bind_param(self, value, dialect):
  11.         return "PREFIX:" + value
  13.     def process_result_value(self, value, dialect):
  14.         return value[7:]
  16.     def copy(self, **kw):
  17.         return MyType(self.impl.length)

The class-level impl attribute is required, and can reference any TypeEngine class. Alternatively, the load_dialect_impl() method can be used to provide different type classes based on the dialect given; in this case, the impl variable can reference TypeEngine as a placeholder.

Types that receive a Python type that isn’t similar to the ultimate type used may want to define the TypeDecorator.coerce_compared_value() method. This is used to give the expression system a hint when coercing Python objects into bind parameters within expressions. Consider this expression:

  1. mytable.c.somecol + datetime.date(2009, 5, 15)

Above, if “somecol” is an Integer variant, it makes sense that we’re doing date arithmetic, where above is usually interpreted by databases as adding a number of days to the given date. The expression system does the right thing by not attempting to coerce the “date()” value into an integer-oriented bind parameter.

However, in the case of TypeDecorator, we are usually changing an incoming Python type to something new - TypeDecorator by default will “coerce” the non-typed side to be the same type as itself. Such as below, we define an “epoch” type that stores a date value as an integer:

  1. class MyEpochType(types.TypeDecorator):
  2.     impl = types.Integer
  4.     epoch = datetime.date(1970, 1, 1)
  6.     def process_bind_param(self, value, dialect):
  7.         return (value - self.epoch).days
  9.     def process_result_value(self, value, dialect):
  10.         return self.epoch + timedelta(days=value)

Our expression of somecol + date with the above type will coerce the “date” on the right side to also be treated as MyEpochType.

This behavior can be overridden via the TypeDecorator.coerce_compared_value() method, which returns a type that should be used for the value of the expression. Below we set it such that an integer value will be treated as an Integer, and any other value is assumed to be a date and will be treated as a MyEpochType:

  1. def coerce_compared_value(self, op, value):
  2.     if isinstance(value, int):
  3.         return Integer()
  4.     else:
  5.         return self

Warning

Note that the behavior of coerce_compared_value is not inherited by default from that of the base type. If the TypeDecorator is augmenting a type that requires special logic for certain types of operators, this method must be overridden. A key example is when decorating the JSON and JSONB types; the default rules of TypeEngine.coerce_compared_value() should be used in order to deal with operators like index operations:

  1. class MyJsonType(TypeDecorator):
  2.     impl = postgresql.JSON
  4.     def coerce_compared_value(self, op, value):
  5.         return self.impl.coerce_compared_value(op, value)

Without the above step, index operations such as mycol['foo'] will cause the index value 'foo' to be JSON encoded.

Class signature

class sqlalchemy.types.TypeDecorator (sqlalchemy.sql.expression.SchemaEventTarget, sqlalchemy.types.TypeEngine)

  • class Comparator(expr)

    A Comparator that is specific to TypeDecorator.

    User-defined TypeDecorator classes should not typically need to modify this.

    Class signature

    class sqlalchemy.types.TypeDecorator.Comparator (sqlalchemy.types.Comparator)

    • method sqlalchemy.types.TypeDecorator.Comparator.operate(op, \other, **kwargs*)

      Operate on an argument.

      This is the lowest level of operation, raises NotImplementedError by default.

      Overriding this on a subclass can allow common behavior to be applied to all operations. For example, overriding ColumnOperators to apply func.lower() to the left and right side:

      1. class MyComparator(ColumnOperators):
      2.     def operate(self, op, other):
      3.         return op(func.lower(self), func.lower(other))

      Parameters

      • op – Operator callable.

      • *other – the ‘other’ side of the operation. Will be a single scalar for most operations.

      • **kwargs – modifiers. These may be passed by special operators such as ColumnOperators.contains().

    • method sqlalchemy.types.TypeDecorator.Comparator.reverse_operate(op, other, \*kwargs*)

      Reverse operate on an argument.

      Usage is the same as operate().

  • method sqlalchemy.types.TypeDecorator.__init__(\args, **kwargs*)

    Construct a TypeDecorator.

    Arguments sent here are passed to the constructor of the class assigned to the impl class level attribute, assuming the impl is a callable, and the resulting object is assigned to the self.impl instance attribute (thus overriding the class attribute of the same name).

    If the class level impl is not a callable (the unusual case), it will be assigned to the same instance attribute ‘as-is’, ignoring those arguments passed to the constructor.

    Subclasses can override this to customize the generation of self.impl entirely.

  • method sqlalchemy.types.TypeDecorator.bind_expression(bindparam)

    Given a bind value (i.e. a BindParameter instance), return a SQL expression in its place.

    This is typically a SQL function that wraps the existing bound parameter within the statement. It is used for special data types that require literals being wrapped in some special database function in order to coerce an application-level value into a database-specific format. It is the SQL analogue of the TypeEngine.bind_processor() method.

    The method is evaluated at statement compile time, as opposed to statement construction time.

    Note that this method, when implemented, should always return the exact same structure, without any conditional logic, as it may be used in an executemany() call against an arbitrary number of bound parameter sets.

    See also

    Applying SQL-level Bind/Result Processing

  • method sqlalchemy.types.TypeDecorator.bind_processor(dialect)

    Provide a bound value processing function for the given Dialect.

    This is the method that fulfills the TypeEngine contract for bound value conversion. TypeDecorator will wrap a user-defined implementation of process_bind_param() here.

    User-defined code can override this method directly, though its likely best to use process_bind_param() so that the processing provided by self.impl is maintained.

    • Parameters

      dialect – Dialect instance in use.

    This method is the reverse counterpart to the result_processor() method of this class.

  • method sqlalchemy.types.TypeDecorator.coerce_compared_value(op, value)

    Suggest a type for a ‘coerced’ Python value in an expression.

    By default, returns self. This method is called by the expression system when an object using this type is on the left or right side of an expression against a plain Python object which does not yet have a SQLAlchemy type assigned:

    1. expr = table.c.somecolumn + 35

    Where above, if somecolumn uses this type, this method will be called with the value operator.add and 35. The return value is whatever SQLAlchemy type should be used for 35 for this particular operation.

  • attribute sqlalchemy.types.TypeDecorator.coerce_to_is_types = (<class ‘NoneType’>,)

    Specify those Python types which should be coerced at the expression level to “IS <constant>” when compared using == (and same for IS NOT in conjunction with !=).

    For most SQLAlchemy types, this includes NoneType, as well as bool.

    TypeDecorator modifies this list to only include NoneType, as typedecorator implementations that deal with boolean types are common.

    Custom TypeDecorator classes can override this attribute to return an empty tuple, in which case no values will be coerced to constants.

  • method sqlalchemy.types.TypeDecorator.column_expression(column)

    Given a SELECT column expression, return a wrapping SQL expression.

    This is typically a SQL function that wraps a column expression as rendered in the columns clause of a SELECT statement. It is used for special data types that require columns to be wrapped in some special database function in order to coerce the value before being sent back to the application. It is the SQL analogue of the TypeEngine.result_processor() method.

    The method is evaluated at statement compile time, as opposed to statement construction time.

    See also

    Applying SQL-level Bind/Result Processing

  • attribute sqlalchemy.types.TypeDecorator.comparator_factory

    A Comparator class which will apply to operations performed by owning ColumnElement objects.

    The comparator_factory attribute is a hook consulted by the core expression system when column and SQL expression operations are performed. When a Comparator class is associated with this attribute, it allows custom re-definition of all existing operators, as well as definition of new operators. Existing operators include those provided by Python operator overloading such as ColumnOperators.__add__() and ColumnOperators.__eq__(), those provided as standard attributes of ColumnOperators such as ColumnOperators.like() and ColumnOperators.in_().

    Rudimentary usage of this hook is allowed through simple subclassing of existing types, or alternatively by using TypeDecorator. See the documentation section Redefining and Creating New Operators for examples.

  • method sqlalchemy.types.TypeDecorator.compare_values(x, y)

    Given two values, compare them for equality.

    By default this calls upon TypeEngine.compare_values() of the underlying “impl”, which in turn usually uses the Python equals operator ==.

    This function is used by the ORM to compare an original-loaded value with an intercepted “changed” value, to determine if a net change has occurred.

  • method sqlalchemy.types.TypeDecorator.copy(\*kw*)

    Produce a copy of this TypeDecorator instance.

    This is a shallow copy and is provided to fulfill part of the TypeEngine contract. It usually does not need to be overridden unless the user-defined TypeDecorator has local state that should be deep-copied.

  • method sqlalchemy.types.TypeDecorator.get_dbapi_type(dbapi)

    Return the DBAPI type object represented by this TypeDecorator.

    By default this calls upon TypeEngine.get_dbapi_type() of the underlying “impl”.

  • method sqlalchemy.types.TypeDecorator.literal_processor(dialect)

    Provide a literal processing function for the given Dialect.

    Subclasses here will typically override TypeDecorator.process_literal_param() instead of this method directly.

    By default, this method makes use of TypeDecorator.process_bind_param() if that method is implemented, where TypeDecorator.process_literal_param() is not. The rationale here is that TypeDecorator typically deals with Python conversions of data that are above the layer of database presentation. With the value converted by TypeDecorator.process_bind_param(), the underlying type will then handle whether it needs to be presented to the DBAPI as a bound parameter or to the database as an inline SQL value.

    New in version 0.9.0.

  • method sqlalchemy.types.TypeDecorator.load_dialect_impl(dialect)

    Return a TypeEngine object corresponding to a dialect.

    This is an end-user override hook that can be used to provide differing types depending on the given dialect. It is used by the TypeDecorator implementation of type_engine() to help determine what type should ultimately be returned for a given TypeDecorator.

    By default returns self.impl.

  • method sqlalchemy.types.TypeDecorator.process_bind_param(value, dialect)

    Receive a bound parameter value to be converted.

    Subclasses override this method to return the value that should be passed along to the underlying TypeEngine object, and from there to the DBAPI execute() method.

    The operation could be anything desired to perform custom behavior, such as transforming or serializing data. This could also be used as a hook for validating logic.

    This operation should be designed with the reverse operation in mind, which would be the process_result_value method of this class.

    • Parameters

      • value – Data to operate upon, of any type expected by this method in the subclass. Can be None.

      • dialect – the Dialect in use.

  • method sqlalchemy.types.TypeDecorator.process_literal_param(value, dialect)

    Receive a literal parameter value to be rendered inline within a statement.

    This method is used when the compiler renders a literal value without using binds, typically within DDL such as in the “server default” of a column or an expression within a CHECK constraint.

    The returned string will be rendered into the output string.

    New in version 0.9.0.

  • method sqlalchemy.types.TypeDecorator.process_result_value(value, dialect)

    Receive a result-row column value to be converted.

    Subclasses should implement this method to operate on data fetched from the database.

    Subclasses override this method to return the value that should be passed back to the application, given a value that is already processed by the underlying TypeEngine object, originally from the DBAPI cursor method fetchone() or similar.

    The operation could be anything desired to perform custom behavior, such as transforming or serializing data. This could also be used as a hook for validating logic.

    • Parameters

      • value – Data to operate upon, of any type expected by this method in the subclass. Can be None.

      • dialect – the Dialect in use.

  1. This operation should be designed to be reversible by the “process\_bind\_param” method of this class.

  • method sqlalchemy.types.TypeDecorator.result_processor(dialect, coltype)

    Provide a result value processing function for the given Dialect.

    This is the method that fulfills the TypeEngine contract for result value conversion. TypeDecorator will wrap a user-defined implementation of process_result_value() here.

    User-defined code can override this method directly, though its likely best to use process_result_value() so that the processing provided by self.impl is maintained.

    • Parameters

      • dialect – Dialect instance in use.

      • coltype – A SQLAlchemy data type

  1. This method is the reverse counterpart to the [`bind_processor()`](#sqlalchemy.types.TypeDecorator.bind_processor "sqlalchemy.types.TypeDecorator.bind_processor") method of this class.

TypeDecorator Recipes

A few key TypeDecorator recipes follow.

Coercing Encoded Strings to Unicode

A common source of confusion regarding the Unicode type is that it is intended to deal only with Python unicode objects on the Python side, meaning values passed to it as bind parameters must be of the form u'some string' if using Python 2 and not 3. The encoding/decoding functions it performs are only to suit what the DBAPI in use requires, and are primarily a private implementation detail.

The use case of a type that can safely receive Python bytestrings, that is strings that contain non-ASCII characters and are not u'' objects in Python 2, can be achieved using a TypeDecorator which coerces as needed:

  1. from sqlalchemy.types import TypeDecorator, Unicode
  3. class CoerceUTF8(TypeDecorator):
  4.     """Safely coerce Python bytestrings to Unicode
  5.     before passing off to the database."""
  7.     impl = Unicode
  9.     def process_bind_param(self, value, dialect):
  10.         if isinstance(value, str):
  11.             value = value.decode('utf-8')
  12.         return value

Rounding Numerics

Some database connectors like those of SQL Server choke if a Decimal is passed with too many decimal places. Here’s a recipe that rounds them down:

  1. from sqlalchemy.types import TypeDecorator, Numeric
  2. from decimal import Decimal
  4. class SafeNumeric(TypeDecorator):
  5.     """Adds quantization to Numeric."""
  7.     impl = Numeric
  9.     def __init__(self, *arg, **kw):
  10.         TypeDecorator.__init__(self, *arg, **kw)
  11.         self.quantize_int = - self.impl.scale
  12.         self.quantize = Decimal(10) ** self.quantize_int
  14.     def process_bind_param(self, value, dialect):
  15.         if isinstance(value, Decimal) and \
  16.             value.as_tuple()[2] < self.quantize_int:
  17.             value = value.quantize(self.quantize)
  18.         return value

Store Timezone Aware Timestamps as Timezone Naive UTC

Timestamps in databases should always be stored in a timezone-agnostic way. For most databases, this means ensuring a timestamp is first in the UTC timezone before it is stored, then storing it as timezone-naive (that is, without any timezone associated with it; UTC is assumed to be the “implicit” timezone). Alternatively, database-specific types like PostgreSQLs “TIMESTAMP WITH TIMEZONE” are often preferred for their richer functionality; however, storing as plain UTC will work on all databases and drivers. When a timezone-intelligent database type is not an option or is not preferred, the TypeDecorator can be used to create a datatype that convert timezone aware timestamps into timezone naive and back again. Below, Python’s built-in datetime.timezone.utc timezone is used to normalize and denormalize:

  1. import datetime
  3. class TZDateTime(TypeDecorator):
  4.     impl = DateTime
  6.     def process_bind_param(self, value, dialect):
  7.         if value is not None:
  8.             if not value.tzinfo:
  9.                 raise TypeError("tzinfo is required")
  10.             value = value.astimezone(datetime.timezone.utc).replace(
  11.                 tzinfo=None
  12.             )
  13.         return value
  15.     def process_result_value(self, value, dialect):
  16.         if value is not None:
  17.             value = value.replace(tzinfo=datetime.timezone.utc)
  18.         return value

Backend-agnostic GUID Type

Receives and returns Python uuid() objects. Uses the PG UUID type when using PostgreSQL, CHAR(32) on other backends, storing them in stringified hex format. Can be modified to store binary in CHAR(16) if desired:

  1. from sqlalchemy.types import TypeDecorator, CHAR
  2. from sqlalchemy.dialects.postgresql import UUID
  3. import uuid
  5. class GUID(TypeDecorator):
  6.     """Platform-independent GUID type.
  8.     Uses PostgreSQL's UUID type, otherwise uses
  9.     CHAR(32), storing as stringified hex values.
  11.     """
  12.     impl = CHAR
  14.     def load_dialect_impl(self, dialect):
  15.         if dialect.name == 'postgresql':
  16.             return dialect.type_descriptor(UUID())
  17.         else:
  18.             return dialect.type_descriptor(CHAR(32))
  20.     def process_bind_param(self, value, dialect):
  21.         if value is None:
  22.             return value
  23.         elif dialect.name == 'postgresql':
  24.             return str(value)
  25.         else:
  26.             if not isinstance(value, uuid.UUID):
  27.                 return "%.32x" % uuid.UUID(value).int
  28.             else:
  29.                 # hexstring
  30.                 return "%.32x" % value.int
  32.     def process_result_value(self, value, dialect):
  33.         if value is None:
  34.             return value
  35.         else:
  36.             if not isinstance(value, uuid.UUID):
  37.                 value = uuid.UUID(value)
  38.             return value

Marshal JSON Strings

This type uses simplejson to marshal Python data structures to/from JSON. Can be modified to use Python’s builtin json encoder:

  1. from sqlalchemy.types import TypeDecorator, VARCHAR
  2. import json
  4. class JSONEncodedDict(TypeDecorator):
  5.     """Represents an immutable structure as a json-encoded string.
  7.     Usage::
  9.         JSONEncodedDict(255)
  11.     """
  13.     impl = VARCHAR
  15.     def process_bind_param(self, value, dialect):
  16.         if value is not None:
  17.             value = json.dumps(value)
  19.         return value
  21.     def process_result_value(self, value, dialect):
  22.         if value is not None:
  23.             value = json.loads(value)
  24.         return value

Adding Mutability

The ORM by default will not detect “mutability” on such a type as above - meaning, in-place changes to values will not be detected and will not be flushed. Without further steps, you instead would need to replace the existing value with a new one on each parent object to detect changes:

  1. obj.json_value["key"] = "value"  # will *not* be detected by the ORM
  3. obj.json_value = {"key": "value"}  # *will* be detected by the ORM

The above limitation may be fine, as many applications may not require that the values are ever mutated once created. For those which do have this requirement, support for mutability is best applied using the sqlalchemy.ext.mutable extension. For a dictionary-oriented JSON structure, we can apply this as:

  1. json_type = MutableDict.as_mutable(JSONEncodedDict)
  3. class MyClass(Base):
  4.     #  ...
  6.     json_data = Column(json_type)

See also

Mutation Tracking

Dealing with Comparison Operations

The default behavior of TypeDecorator is to coerce the “right hand side” of any expression into the same type. For a type like JSON, this means that any operator used must make sense in terms of JSON. For some cases, users may wish for the type to behave like JSON in some circumstances, and as plain text in others. One example is if one wanted to handle the LIKE operator for the JSON type. LIKE makes no sense against a JSON structure, but it does make sense against the underlying textual representation. To get at this with a type like JSONEncodedDict, we need to coerce the column to a textual form using cast() or type_coerce() before attempting to use this operator:

  1. from sqlalchemy import type_coerce, String
  3. stmt = select(my_table).where(
  4.     type_coerce(my_table.c.json_data, String).like('%foo%'))

TypeDecorator provides a built-in system for working up type translations like these based on operators. If we wanted to frequently use the LIKE operator with our JSON object interpreted as a string, we can build it into the type by overriding the TypeDecorator.coerce_compared_value() method:

  1. from sqlalchemy.sql import operators
  2. from sqlalchemy import String
  4. class JSONEncodedDict(TypeDecorator):
  6.     impl = VARCHAR
  8.     def coerce_compared_value(self, op, value):
  9.         if op in (operators.like_op, operators.not_like_op):
  10.             return String()
  11.         else:
  12.             return self
  14.     def process_bind_param(self, value, dialect):
  15.         if value is not None:
  16.             value = json.dumps(value)
  18.         return value
  20.     def process_result_value(self, value, dialect):
  21.         if value is not None:
  22.             value = json.loads(value)
  23.         return value

Above is just one approach to handling an operator like “LIKE”. Other applications may wish to raise NotImplementedError for operators that have no meaning with a JSON object such as “LIKE”, rather than automatically coercing to text.

Applying SQL-level Bind/Result Processing

As seen in the section Augmenting Existing Types, SQLAlchemy allows Python functions to be invoked both when parameters are sent to a statement, as well as when result rows are loaded from the database, to apply transformations to the values as they are sent to or from the database. It is also possible to define SQL-level transformations as well. The rationale here is when only the relational database contains a particular series of functions that are necessary to coerce incoming and outgoing data between an application and persistence format. Examples include using database-defined encryption/decryption functions, as well as stored procedures that handle geographic data. The PostGIS extension to PostgreSQL includes an extensive array of SQL functions that are necessary for coercing data into particular formats.

Any TypeEngine, UserDefinedType or TypeDecorator subclass can include implementations of TypeEngine.bind_expression() and/or TypeEngine.column_expression(), which when defined to return a non-None value should return a ColumnElement expression to be injected into the SQL statement, either surrounding bound parameters or a column expression. For example, to build a Geometry type which will apply the PostGIS function ST_GeomFromText to all outgoing values and the function ST_AsText to all incoming data, we can create our own subclass of UserDefinedType which provides these methods in conjunction with func:

  1. from sqlalchemy import func
  2. from sqlalchemy.types import UserDefinedType
  4. class Geometry(UserDefinedType):
  5.     def get_col_spec(self):
  6.         return "GEOMETRY"
  8.     def bind_expression(self, bindvalue):
  9.         return func.ST_GeomFromText(bindvalue, type_=self)
  11.     def column_expression(self, col):
  12.         return func.ST_AsText(col, type_=self)

We can apply the Geometry type into Table metadata and use it in a select() construct:

  1. geometry = Table('geometry', metadata,
  2.               Column('geom_id', Integer, primary_key=True),
  3.               Column('geom_data', Geometry)
  4.             )
  6. print(select(geometry).where(
  7.   geometry.c.geom_data == 'LINESTRING(189412 252431,189631 259122)'))

The resulting SQL embeds both functions as appropriate. ST_AsText is applied to the columns clause so that the return value is run through the function before passing into a result set, and ST_GeomFromText is run on the bound parameter so that the passed-in value is converted:

  1. SELECT geometry.geom_id, ST_AsText(geometry.geom_data) AS geom_data_1
  2. FROM geometry
  3. WHERE geometry.geom_data = ST_GeomFromText(:geom_data_2)

The TypeEngine.column_expression() method interacts with the mechanics of the compiler such that the SQL expression does not interfere with the labeling of the wrapped expression. Such as, if we rendered a select() against a label() of our expression, the string label is moved to the outside of the wrapped expression:

  1. print(select(geometry.c.geom_data.label('my_data')))

Output:

  1. SELECT ST_AsText(geometry.geom_data) AS my_data
  2. FROM geometry

Another example is we decorate BYTEA to provide a PGPString, which will make use of the PostgreSQL pgcrypto extension to encrypt/decrypt values transparently:

  1. from sqlalchemy import create_engine, String, select, func, \
  2.         MetaData, Table, Column, type_coerce, TypeDecorator
  4. from sqlalchemy.dialects.postgresql import BYTEA
  6. class PGPString(TypeDecorator):
  7.     impl = BYTEA
  9.     def __init__(self, passphrase):
  10.         super(PGPString, self).__init__()
  11.         self.passphrase = passphrase
  13.     def bind_expression(self, bindvalue):
  14.         # convert the bind's type from PGPString to
  15.         # String, so that it's passed to psycopg2 as is without
  16.         # a dbapi.Binary wrapper
  17.         bindvalue = type_coerce(bindvalue, String)
  18.         return func.pgp_sym_encrypt(bindvalue, self.passphrase)
  20.     def column_expression(self, col):
  21.         return func.pgp_sym_decrypt(col, self.passphrase)
  23. metadata = MetaData()
  24. message = Table('message', metadata,
  25.                 Column('username', String(50)),
  26.                 Column('message',
  27.                     PGPString("this is my passphrase")),
  28.             )
  30. engine = create_engine("postgresql://scott:tiger@localhost/test", echo=True)
  31. with engine.begin() as conn:
  32.     metadata.create_all(conn)
  34.     conn.execute(message.insert(), username="some user",
  35.                                 message="this is my message")
  37.     print(conn.scalar(
  38.             select(message.c.message).\
  39.                 where(message.c.username == "some user")
  40.         ))

The pgp_sym_encrypt and pgp_sym_decrypt functions are applied to the INSERT and SELECT statements:

  1. INSERT INTO message (username, message)
  2.   VALUES (%(username)s, pgp_sym_encrypt(%(message)s, %(pgp_sym_encrypt_1)s))
  3.   {'username': 'some user', 'message': 'this is my message',
  4.     'pgp_sym_encrypt_1': 'this is my passphrase'}
  6. SELECT pgp_sym_decrypt(message.message, %(pgp_sym_decrypt_1)s) AS message_1
  7.   FROM message
  8.   WHERE message.username = %(username_1)s
  9.   {'pgp_sym_decrypt_1': 'this is my passphrase', 'username_1': 'some user'}

See also

PostGIS Integration

Redefining and Creating New Operators

SQLAlchemy Core defines a fixed set of expression operators available to all column expressions. Some of these operations have the effect of overloading Python’s built-in operators; examples of such operators include ColumnOperators.__eq__() (table.c.somecolumn == 'foo'), ColumnOperators.__invert__() (~table.c.flag), and ColumnOperators.__add__() (table.c.x + table.c.y). Other operators are exposed as explicit methods on column expressions, such as ColumnOperators.in_() (table.c.value.in_(['x', 'y'])) and ColumnOperators.like() (table.c.value.like('%ed%')).

The Core expression constructs in all cases consult the type of the expression in order to determine the behavior of existing operators, as well as to locate additional operators that aren’t part of the built-in set. The TypeEngine base class defines a root “comparison” implementation Comparator, and many specific types provide their own sub-implementations of this class. User-defined Comparator implementations can be built directly into a simple subclass of a particular type in order to override or define new operations. Below, we create a Integer subclass which overrides the ColumnOperators.__add__() operator:

  1. from sqlalchemy import Integer
  3. class MyInt(Integer):
  4.     class comparator_factory(Integer.Comparator):
  5.         def __add__(self, other):
  6.             return self.op("goofy")(other)

The above configuration creates a new class MyInt, which establishes the TypeEngine.comparator_factory attribute as referring to a new class, subclassing the Comparator class associated with the Integer type.

Usage:

  1. >>> sometable = Table("sometable", metadata, Column("data", MyInt))
  2. >>> print(sometable.c.data + 5)
  3. sometable.data goofy :data_1

The implementation for ColumnOperators.__add__() is consulted by an owning SQL expression, by instantiating the Comparator with itself as the expr attribute. The mechanics of the expression system are such that operations continue recursively until an expression object produces a new SQL expression construct. Above, we could just as well have said self.expr.op("goofy")(other) instead of self.op("goofy")(other).

When using Operators.op() for comparison operations that return a boolean result, the Operators.op.is_comparison flag should be set to True:

  1. class MyInt(Integer):
  2.     class comparator_factory(Integer.Comparator):
  3.         def is_frobnozzled(self, other):
  4.             return self.op("--is_frobnozzled->", is_comparison=True)(other)

New methods added to a Comparator are exposed on an owning SQL expression using a __getattr__ scheme, which exposes methods added to Comparator onto the owning ColumnElement. For example, to add a log() function to integers:

  1. from sqlalchemy import Integer, func
  3. class MyInt(Integer):
  4.     class comparator_factory(Integer.Comparator):
  5.         def log(self, other):
  6.             return func.log(self.expr, other)

Using the above type:

  1. >>> print(sometable.c.data.log(5))
  2. log(:log_1, :log_2)

Unary operations are also possible. For example, to add an implementation of the PostgreSQL factorial operator, we combine the UnaryExpression construct along with a custom_op to produce the factorial expression:

  1. from sqlalchemy import Integer
  2. from sqlalchemy.sql.expression import UnaryExpression
  3. from sqlalchemy.sql import operators
  5. class MyInteger(Integer):
  6.     class comparator_factory(Integer.Comparator):
  7.         def factorial(self):
  8.             return UnaryExpression(self.expr,
  9.                         modifier=operators.custom_op("!"),
  10.                         type_=MyInteger)

Using the above type:

  1. >>> from sqlalchemy.sql import column
  2. >>> print(column('x', MyInteger).factorial())
  3. x !

See also

Operators.op()

TypeEngine.comparator_factory

Creating New Types

The UserDefinedType class is provided as a simple base class for defining entirely new database types. Use this to represent native database types not known by SQLAlchemy. If only Python translation behavior is needed, use TypeDecorator instead.

Object NameDescription

UserDefinedType

Base for user defined types.

class sqlalchemy.types.``UserDefinedType

Base for user defined types.

This should be the base of new types. Note that for most cases, TypeDecorator is probably more appropriate:

  1. import sqlalchemy.types as types
  3. class MyType(types.UserDefinedType):
  4.     def __init__(self, precision = 8):
  5.         self.precision = precision
  7.     def get_col_spec(self, **kw):
  8.         return "MYTYPE(%s)" % self.precision
  10.     def bind_processor(self, dialect):
  11.         def process(value):
  12.             return value
  13.         return process
  15.     def result_processor(self, dialect, coltype):
  16.         def process(value):
  17.             return value
  18.         return process

Once the type is made, it’s immediately usable:

  1. table = Table('foo', meta,
  2.     Column('id', Integer, primary_key=True),
  3.     Column('data', MyType(16))
  4.     )

The get_col_spec() method will in most cases receive a keyword argument type_expression which refers to the owning expression of the type as being compiled, such as a Column or cast() construct. This keyword is only sent if the method accepts keyword arguments (e.g. **kw) in its argument signature; introspection is used to check for this in order to support legacy forms of this function.

New in version 1.0.0: the owning expression is passed to the get_col_spec() method via the keyword argument type_expression, if it receives **kw in its signature.

Class signature

class sqlalchemy.types.UserDefinedType (sqlalchemy.types.TypeEngine)

Working with Custom Types and Reflection

It is important to note that database types which are modified to have additional in-Python behaviors, including types based on TypeDecorator as well as other user-defined subclasses of datatypes, do not have any representation within a database schema. When using database the introspection features described at Reflecting Database Objects, SQLAlchemy makes use of a fixed mapping which links the datatype information reported by a database server to a SQLAlchemy datatype object. For example, if we look inside of a PostgreSQL schema at the definition for a particular database column, we might receive back the string "VARCHAR". SQLAlchemy’s PostgreSQL dialect has a hardcoded mapping which links the string name "VARCHAR" to the SQLAlchemy VARCHAR class, and that’s how when we emit a statement like Table('my_table', m, autoload_with=engine), the Column object within it would have an instance of VARCHAR present inside of it.

The implication of this is that if a Table object makes use of type objects that don’t correspond directly to the database-native type name, if we create a new Table object against a new MetaData collection for this database table elsewhere using reflection, it will not have this datatype. For example:

  1. >>> from sqlalchemy import Table, Column, MetaData, create_engine, PickleType, Integer
  2. >>> metadata = MetaData()
  3. >>> my_table = Table("my_table", metadata, Column('id', Integer), Column("data", PickleType))
  4. >>> engine = create_engine("sqlite://", echo='debug')
  5. >>> my_table.create(engine)
  6. INFO sqlalchemy.engine.base.Engine
  7. CREATE TABLE my_table (
  8.     id INTEGER,
  9.     data BLOB
  10. )

Above, we made use of PickleType, which is a TypeDecorator that works on top of the LargeBinary datatype, which on SQLite corresponds to the database type BLOB. In the CREATE TABLE, we see that the BLOB datatype is used. The SQLite database knows nothing about the PickleType we’ve used.

If we look at the datatype of my_table.c.data.type, as this is a Python object that was created by us directly, it is PickleType:

  1. >>> my_table.c.data.type
  2. PickleType()

However, if we create another instance of Table using reflection, the use of PickleType is not represented in the SQLite database we’ve created; we instead get back BLOB:

  1. >>> metadata_two = MetaData()
  2. >>> my_reflected_table = Table("my_table", metadata_two, autoload_with=engine)
  3. INFO sqlalchemy.engine.base.Engine PRAGMA main.table_info("my_table")
  4. INFO sqlalchemy.engine.base.Engine ()
  5. DEBUG sqlalchemy.engine.base.Engine Col ('cid', 'name', 'type', 'notnull', 'dflt_value', 'pk')
  6. DEBUG sqlalchemy.engine.base.Engine Row (0, 'id', 'INTEGER', 0, None, 0)
  7. DEBUG sqlalchemy.engine.base.Engine Row (1, 'data', 'BLOB', 0, None, 0)
  9. >>> my_reflected_table.c.data.type
  10. BLOB()

Typically, when an application defines explicit Table metadata with custom types, there is no need to use table reflection because the necessary Table metadata is already present. However, for the case where an application, or a combination of them, need to make use of both explicit Table metadata which includes custom, Python-level datatypes, as well as Table objects which set up their Column objects as reflected from the database, which nevertheless still need to exhibit the additional Python behaviors of the custom datatypes, additional steps must be taken to allow this.

The most straightforward is to override specific columns as described at Overriding Reflected Columns. In this technique, we simply use reflection in combination with explicit Column objects for those columns for which we want to use a custom or decorated datatype:

  1. >>> metadata_three = MetaData()
  2. >>> my_reflected_table = Table("my_table", metadata_three, Column("data", PickleType), autoload_with=engine)

The my_reflected_table object above is reflected, and will load the definition of the “id” column from the SQLite database. But for the “data” column, we’ve overridden the reflected object with an explicit Column definition that includes our desired in-Python datatype, the PickleType. The reflection process will leave this Column object intact:

  1. >>> my_reflected_table.c.data.type
  2. PickleType()

A more elaborate way to convert from database-native type objects to custom datatypes is to use the DDLEvents.column_reflect() event handler. If for example we knew that we wanted all BLOB datatypes to in fact be PickleType, we could set up a rule across the board:

  1. from sqlalchemy import BLOB
  2. from sqlalchemy import event
  3. from sqlalchemy import PickleType
  4. from sqlalchemy import Table
  6. @event.listens_for(Table, "column_reflect")
  7. def _setup_pickletype(inspector, table, column_info):
  8.     if isinstance(column_info["type"], BLOB):
  9.         column_info["type"] = PickleType()

When the above code is invoked before any table reflection occurs (note also it should be invoked only once in the application, as it is a global rule), upon reflecting any Table that includes a column with a BLOB datatype, the resulting datatype will be stored in the Column object as PickleType.

In practice, the above event-based approach would likely have additional rules in order to affect only those columns where the datatype is important, such as a lookup table of table names and possibly column names, or other heuristics in order to accurately determine which columns should be established with an in Python datatype.