sqlalchemy.ext.compiler

Synopsis

Usage involves the creation of one or more ~sqlalchemy.sql.expression.ClauseElement subclasses and one or more callables defining its compilation:

from sqlalchemy.ext.compiler import compiles
from sqlalchemy.sql.expression import ColumnClause

class MyColumn(ColumnClause):
    inherit_cache = True

@compiles(MyColumn)
def compile_mycolumn(element, compiler, **kw):
    return "[%s]" % element.name

Above, MyColumn extends ~sqlalchemy.sql.expression.ColumnClause, the base expression element for named column objects. The compiles decorator registers itself with the MyColumn class so that it is invoked when the object is compiled to a string:

from sqlalchemy import select

s = select(MyColumn('x'), MyColumn('y'))
print(str(s))

Produces:

SELECT [x], [y]

Dialect-specific compilation rules

Compilers can also be made dialect-specific. The appropriate compiler will be invoked for the dialect in use:

from sqlalchemy.schema import DDLElement

class AlterColumn(DDLElement):
    inherit_cache = False

    def __init__(self, column, cmd):
        self.column = column
        self.cmd = cmd

@compiles(AlterColumn)
def visit_alter_column(element, compiler, **kw):
    return "ALTER COLUMN %s ..." % element.column.name

@compiles(AlterColumn, 'postgresql')
def visit_alter_column(element, compiler, **kw):
    return "ALTER TABLE %s ALTER COLUMN %s ..." % (element.table.name,
                                                   element.column.name)

The second visit_alter_table will be invoked when any postgresql dialect is used.

Compiling sub-elements of a custom expression construct

The compiler argument is the ~sqlalchemy.engine.interfaces.Compiled object in use. This object can be inspected for any information about the in-progress compilation, including compiler.dialect, compiler.statement etc. The ~sqlalchemy.sql.compiler.SQLCompiler and ~sqlalchemy.sql.compiler.DDLCompiler both include a process() method which can be used for compilation of embedded attributes:

from sqlalchemy.sql.expression import Executable, ClauseElement

class InsertFromSelect(Executable, ClauseElement):
    inherit_cache = False

    def __init__(self, table, select):
        self.table = table
        self.select = select

@compiles(InsertFromSelect)
def visit_insert_from_select(element, compiler, **kw):
    return "INSERT INTO %s (%s)" % (
        compiler.process(element.table, asfrom=True, **kw),
        compiler.process(element.select, **kw)
    )

insert = InsertFromSelect(t1, select(t1).where(t1.c.x>5))
print(insert)

Produces:

"INSERT INTO mytable (SELECT mytable.x, mytable.y, mytable.z
                      FROM mytable WHERE mytable.x > :x_1)"

@compiles(MyConstraint)
def compile_my_constraint(constraint, ddlcompiler, **kw):
    kw['literal_binds'] = True
    return "CONSTRAINT %s CHECK (%s)" % (
        constraint.name,
        ddlcompiler.sql_compiler.process(
            constraint.expression, **kw)
    )

Enabling Autocommit on a Construct

Recall from the section :ref:`autocommit` that the _engine.Engine, when asked to execute a construct in the absence of a user-defined transaction, detects if the given construct represents DML or DDL, that is, a data modification or data definition statement, which requires (or may require, in the case of DDL) that the transaction generated by the DBAPI be committed (recall that DBAPI always has a transaction going on regardless of what SQLAlchemy does). Checking for this is actually accomplished by checking for the "autocommit" execution option on the construct. When building a construct like an INSERT derivation, a new DDL type, or perhaps a stored procedure that alters data, the "autocommit" option needs to be set in order for the statement to function with "connectionless" execution (as described in :ref:`dbengine_implicit`).

Currently a quick way to do this is to subclass .Executable, then add the "autocommit" flag to the _execution_options dictionary (note this is a "frozen" dictionary which supplies a generative union() method):

from sqlalchemy.sql.expression import Executable, ClauseElement

class MyInsertThing(Executable, ClauseElement):
    _execution_options = \
        Executable._execution_options.union({'autocommit': True})

More succinctly, if the construct is truly similar to an INSERT, UPDATE, or DELETE, .UpdateBase can be used, which already is a subclass of .Executable, _expression.ClauseElement and includes the autocommit flag:

from sqlalchemy.sql.expression import UpdateBase

class MyInsertThing(UpdateBase):
    def __init__(self, ...):
        ...

DDL elements that subclass .DDLElement already have the "autocommit" flag turned on.

Changing the default compilation of existing constructs

The compiler extension applies just as well to the existing constructs. When overriding the compilation of a built in SQL construct, the @compiles decorator is invoked upon the appropriate class (be sure to use the class, i.e. Insert or Select, instead of the creation function such as insert() or select()).

Within the new compilation function, to get at the "original" compilation routine, use the appropriate visit_XXX method - this because compiler.process() will call upon the overriding routine and cause an endless loop. Such as, to add "prefix" to all insert statements:

from sqlalchemy.sql.expression import Insert

@compiles(Insert)
def prefix_inserts(insert, compiler, **kw):
    return compiler.visit_insert(insert.prefix_with("some prefix"), **kw)

The above compiler will prefix all INSERT statements with "some prefix" when compiled.

Changing Compilation of Types

compiler works for types, too, such as below where we implement the MS-SQL specific 'max' keyword for String/VARCHAR:

@compiles(String, 'mssql')
@compiles(VARCHAR, 'mssql')
def compile_varchar(element, compiler, **kw):
    if element.length == 'max':
        return "VARCHAR('max')"
    else:
        return compiler.visit_VARCHAR(element, **kw)

foo = Table('foo', metadata,
    Column('data', VARCHAR('max'))
)

Subclassing Guidelines

A big part of using the compiler extension is subclassing SQLAlchemy expression constructs. To make this easier, the expression and schema packages feature a set of "bases" intended for common tasks. A synopsis is as follows:

• ~sqlalchemy.sql.expression.ClauseElement - This is the root expression class. Any SQL expression can be derived from this base, and is probably the best choice for longer constructs such as specialized INSERT statements.

• ~sqlalchemy.sql.expression.ColumnElement - The root of all "column-like" elements. Anything that you'd place in the "columns" clause of a SELECT statement (as well as order by and group by) can derive from this - the object will automatically have Python "comparison" behavior.

  ~sqlalchemy.sql.expression.ColumnElement classes want to have a type member which is expression's return type. This can be established at the instance level in the constructor, or at the class level if its generally constant:

  class timestamp(ColumnElement):
      type = TIMESTAMP()
      inherit_cache = True
  

• ~sqlalchemy.sql.functions.FunctionElement - This is a hybrid of a ColumnElement and a "from clause" like object, and represents a SQL function or stored procedure type of call. Since most databases support statements along the line of "SELECT FROM <some function>" FunctionElement adds in the ability to be used in the FROM clause of a select() construct:

  from sqlalchemy.sql.expression import FunctionElement
  
  class coalesce(FunctionElement):
      name = 'coalesce'
      inherit_cache = True
  
  @compiles(coalesce)
  def compile(element, compiler, **kw):
      return "coalesce(%s)" % compiler.process(element.clauses, **kw)
  
  @compiles(coalesce, 'oracle')
  def compile(element, compiler, **kw):
      if len(element.clauses) > 2:
          raise TypeError("coalesce only supports two arguments on Oracle")
      return "nvl(%s)" % compiler.process(element.clauses, **kw)
  

• ~sqlalchemy.schema.DDLElement - The root of all DDL expressions, like CREATE TABLE, ALTER TABLE, etc. Compilation of DDLElement subclasses is issued by a DDLCompiler instead of a SQLCompiler. DDLElement also features Table and MetaData event hooks via the execute_at() method, allowing the construct to be invoked during CREATE TABLE and DROP TABLE sequences.

• ~sqlalchemy.sql.expression.Executable - This is a mixin which should be used with any expression class that represents a "standalone" SQL statement that can be passed directly to an execute() method. It is already implicit within DDLElement and FunctionElement.

Most of the above constructs also respond to SQL statement caching. A subclassed construct will want to define the caching behavior for the object, which usually means setting the flag inherit_cache to the value of False or True. See the next section :ref:`compilerext_caching` for background.

Enabling Caching Support for Custom Constructs

SQLAlchemy as of version 1.4 includes a :ref:`SQL compilation caching facility <sql_caching>` which will allow equivalent SQL constructs to cache their stringified form, along with other structural information used to fetch results from the statement.

For reasons discussed at :ref:`caching_caveats`, the implementation of this caching system takes a conservative approach towards including custom SQL constructs and/or subclasses within the caching system. This includes that any user-defined SQL constructs, including all the examples for this extension, will not participate in caching by default unless they positively assert that they are able to do so. The .HasCacheKey.inherit_cache attribute when set to True at the class level of a specific subclass will indicate that instances of this class may be safely cached, using the cache key generation scheme of the immediate superclass. This applies for example to the "synopsis" example indicated previously:

class MyColumn(ColumnClause):
    inherit_cache = True

@compiles(MyColumn)
def compile_mycolumn(element, compiler, **kw):
    return "[%s]" % element.name

Above, the MyColumn class does not include any new state that affects its SQL compilation; the cache key of MyColumn instances will make use of that of the ColumnClause superclass, meaning it will take into account the class of the object (MyColumn), the string name and datatype of the object:

>>> MyColumn("some_name", String())._generate_cache_key()
CacheKey(
    key=('0', <class '__main__.MyColumn'>,
    'name', 'some_name',
    'type', (<class 'sqlalchemy.sql.sqltypes.String'>,
             ('length', None), ('collation', None))
), bindparams=[])

For objects that are likely to be used liberally as components within many larger statements, such as _schema.Column subclasses and custom SQL datatypes, it's important that caching be enabled as much as possible, as this may otherwise negatively affect performance.

An example of an object that does contain state which affects its SQL compilation is the one illustrated at :ref:`compilerext_compiling_subelements`; this is an "INSERT FROM SELECT" construct that combines together a _schema.Table as well as a _sql.Select construct, each of which independently affect the SQL string generation of the construct. For this class, the example illustrates that it simply does not participate in caching:

class InsertFromSelect(Executable, ClauseElement):
    inherit_cache = False

    def __init__(self, table, select):
        self.table = table
        self.select = select

@compiles(InsertFromSelect)
def visit_insert_from_select(element, compiler, **kw):
    return "INSERT INTO %s (%s)" % (
        compiler.process(element.table, asfrom=True, **kw),
        compiler.process(element.select, **kw)
    )

While it is also possible that the above InsertFromSelect could be made to produce a cache key that is composed of that of the _schema.Table and _sql.Select components together, the API for this is not at the moment fully public. However, for an "INSERT FROM SELECT" construct, which is only used by itself for specific operations, caching is not as critical as in the previous example.

For objects that are used in relative isolation and are generally standalone, such as custom :term:`DML` constructs like an "INSERT FROM SELECT", caching is generally less critical as the lack of caching for such a construct will have only localized implications for that specific operation.

Further Examples

from sqlalchemy.sql import expression
from sqlalchemy.ext.compiler import compiles
from sqlalchemy.types import DateTime

class utcnow(expression.FunctionElement):
    type = DateTime()
    inherit_cache = True

@compiles(utcnow, 'postgresql')
def pg_utcnow(element, compiler, **kw):
    return "TIMEZONE('utc', CURRENT_TIMESTAMP)"

@compiles(utcnow, 'mssql')
def ms_utcnow(element, compiler, **kw):
    return "GETUTCDATE()"

from sqlalchemy import (
            Table, Column, Integer, String, DateTime, MetaData
        )
metadata = MetaData()
event = Table("event", metadata,
    Column("id", Integer, primary_key=True),
    Column("description", String(50), nullable=False),
    Column("timestamp", DateTime, server_default=utcnow())
)

from sqlalchemy.sql import expression, case
from sqlalchemy.ext.compiler import compiles
from sqlalchemy.types import Numeric

class greatest(expression.FunctionElement):
    type = Numeric()
    name = 'greatest'
    inherit_cache = True

@compiles(greatest)
def default_greatest(element, compiler, **kw):
    return compiler.visit_function(element)

@compiles(greatest, 'sqlite')
@compiles(greatest, 'mssql')
@compiles(greatest, 'oracle')
def case_greatest(element, compiler, **kw):
    arg1, arg2 = list(element.clauses)
    return compiler.process(case([(arg1 > arg2, arg1)], else_=arg2), **kw)

Session.query(Account).\
        filter(
            greatest(
                Account.checking_balance,
                Account.savings_balance) > 10000
        )

from sqlalchemy.sql import expression
from sqlalchemy.ext.compiler import compiles

class sql_false(expression.ColumnElement):
    inherit_cache = True

@compiles(sql_false)
def default_false(element, compiler, **kw):
    return "false"

@compiles(sql_false, 'mssql')
@compiles(sql_false, 'mysql')
@compiles(sql_false, 'oracle')
def int_false(element, compiler, **kw):
    return "0"

from sqlalchemy import select, union_all

exp = union_all(
    select(users.c.name, sql_false().label("enrolled")),
    select(customers.c.name, customers.c.enrolled)
)