SQLAlchemy: cascade delete

问题:

I must be missing something trivial with SQLAlchemy's cascade options because I cannot get a simple cascade delete to operate correctly -- if a parent element is a deleted, the children persist, with null foreign keys.

I've put a concise test case here:

from sqlalchemy import Column, Integer, ForeignKey
from sqlalchemy.orm import relationship

from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class Parent(Base):
    __tablename__ = "parent"
    id = Column(Integer, primary_key = True)

class Child(Base):
    __tablename__ = "child"
    id = Column(Integer, primary_key = True)
    parentid = Column(Integer, ForeignKey(Parent.id))
    parent = relationship(Parent, cascade = "all,delete", backref = "children")

engine = create_engine("sqlite:///:memory:")
Base.metadata.create_all(engine)
Session = sessionmaker(bind=engine)

session = Session()

parent = Parent()
parent.children.append(Child())
parent.children.append(Child())
parent.children.append(Child())

session.add(parent)
session.commit()

print "Before delete, children = {0}".format(session.query(Child).count())
print "Before delete, parent = {0}".format(session.query(Parent).count())

session.delete(parent)
session.commit()

print "After delete, children = {0}".format(session.query(Child).count())
print "After delete parent = {0}".format(session.query(Parent).count())

session.close()

Output:

Before delete, children = 3
Before delete, parent = 1
After delete, children = 3
After delete parent = 0

There is a simple, one-to-many relationship between Parent and Child. The script creates a parent, adds 3 children, then commits. Next, it deletes the parent, but the children persist. Why? How do I make the children cascade delete?

回答1:

The problem is that sqlalchemy considers Child as the parent, because that is where you defined your relationship (it doesn't care that you called it "Child" of course).

If you define the relationship on the Parent class instead, it will work:

children = relationship("Child", cascade="all,delete", backref="parent")

(note "Child" as a string: this is allowed when using the declarative style, so that you are able to refer to a class that is not yet defined)

You might want to add delete-orphan as well (delete causes children to be deleted when the parent gets deleted, delete-orphan also deletes any children that were "removed" from the parent, even if the parent is not deleted)

EDIT: just found out: if you really want to define the relationship on the Child class, you can do so, but you will have to define the cascade on the backref (by creating the backref explicitly), like this:

parent = relationship(Parent, backref=backref("children", cascade="all,delete"))

(implying from sqlalchemy.orm import backref)

回答2:

@Steven's asnwer is good when you are deleting through session.delete() which never happens in my case. I noticed that most of the time I delete through session.query().filter().delete() (which doesn't put elements in the memory and deletes directly from db). Using this method sqlalchemy's cascade='all, delete' doesn't work. There is a solution though: ON DELETE CASCADE through db (note: not all databases support it).

class Child(Base):
    __tablename__ = "children"

    id = Column(Integer, primary_key=True)
    parent_id = Column(Integer, ForeignKey("parents.id", ondelete='CASCADE'))

class Parent(Base):
    __tablename__ = "parents"

    id = Column(Integer, primary_key=True)
    child = relationship(Child, backref="parent", passive_deletes=True)

回答3:

Pretty old post, but I just spent an hour or two on this, so I wanted to share my finding, especially since some of the other comments listed aren't quite right.

TL;DR

Give the child table a foreign or modify the existing one, adding onedelete='CASCADE':

parent_id = db.Column(db.Integer, db.ForeignKey('parent.id', ondelete='CASCADE'))

And one of the following relationships:

a) This on the parent table:

children = db.relationship('Child', backref='parent', passive_deletes=True)

b) Or this on the child table:

parent = db.relationship('Parent', backref=backref('children', passive_deletes=True))

---

Details

First off, despite what the accepted answer says, the parent/child relationship is not established by using relationship, it's established by using ForeignKey. You can put the relationship on either the parent or child tables and it will work fine. Although, apparently on the child tables, you have to use the backref function in addition to the keyword argument.

Option 1 (preferred)

Second, SqlAlchemy supports two different kinds of cascading. The first, and the one I recommend, is built into your database and usually takes the form of a constraint on the foreign key declaration. In PostgreSQL it looks like this:

CONSTRAINT child_parent_id_fkey FOREIGN KEY (parent_id)
REFERENCES parent_table(id) MATCH SIMPLE
ON DELETE CASCADE

This means that when you delete a record from parent_table, then all the corresponding rows in child_table will be deleted for you by the database. It's fast and reliable and probably your best bet. You set this up in SqlAlchemy through ForeignKey like this (part of the child table definition):

parent_id = db.Column(db.Integer, db.ForeignKey('parent.id', ondelete='CASCADE'))
parent = db.relationship('Parent', backref=backref('children', passive_deletes=True))

The ondelete='CASCADE' is the part that creates the ON DELETE CASCADE on the table.

Gotcha!

There's an important caveat here. Notice how I have a relationship specified with passive_deletes=True? If you don't have that, the entire thing will not work. This is because by default when you delete a parent record SqlAlchemy does something really weird. It sets the foreign keys of all child rows to NULL. So if you delete a row from parent_table where id = 5, then it will basically execute

UPDATE child_table SET parent_id = NULL WHERE parent_id = 5

Why you would want this I have no idea. I'd be surprised if many database engines even allowed you to set a valid foreign key to NULL, creating an orphan. Seems like a bad idea, but maybe there's a use case. Anyway, if you let SqlAlchemy do this, you will prevent the database from being able to clean up the children using the ON DELETE CASCADE that you set up. This is because it relies on those foreign keys to know which child rows to delete. Once SqlAlchemy has set them all to NULL, the database can't delete them. Setting the passive_deletes=True prevents SqlAlchemy from NULLing out the foreign keys.

You can read more about passive deletes in the SqlAlchemy docs.

Option 2

The other way you can do it is to let SqlAlchemy do it for you. This is set up using the cascade argument of the relationship. If you have the relationship defined on the parent table, it looks like this:

children = relationship('Child', cascade='all,delete', backref='parent')

If the relationship is on the child, you do it like this:

parent = relationship('Parent', backref=backref('children', cascade='all,delete'))

Again, this is the child so you have to call a method called backref and putting the cascade data in there.

With this in place, when you delete a parent row, SqlAlchemy will actually run delete statements for you to clean up the child rows. This will likely not be as efficient as letting this database handle if for you so I don't recommend it.

Here are the SqlAlchemy docs on the cascading features it supports.

回答4:

Steven is correct in that you need to explicitly create the backref, this results in the cascade being applied on the parent (as opposed to it being applied to the child like in the test scenario).

However, defining the relationship on the Child does NOT make sqlalchemy consider Child the parent. It doesn't matter where the relationship is defined (child or parent), its the foreign key that links the two tables that determines which is the parent and which is the child.

It makes sense to stick to one convention though, and based on Steven's response, I'm defining all my child relationships on the parent.

回答5:

I struggled with the documentation as well, but found that the docstrings themselves tend to be easier than the manual. For example, if you import relationship from sqlalchemy.orm and do help(relationship), it will give you all the options you can specify for cascade. The bullet for "delete-orphan" says, "if an item of the child's type with no parent is detected, mark it for deletion. Note that this option prevents a pending item of the child's class from being persisted without a parent present."

I realize your issue was more with the way the documentation for defining parent-child relationships. But it seemed that you might also be having a problem with the cascade options, because "all" includes "delete." "delete orphan" is the only option that's not included in "all."

回答6:

Steven's answer is solid. I'd like to point out an additional implication.

By using relationship, you're making the app layer (Flask) responsible for referential integrity. That means other processes that access the database not through Flask, like a database utility or a person connecting to the database directly, will not experience those constraints and could change your data in a way that breaks the logical data model you worked so hard to design.

Whenever possible, use the ForeignKey approach described by d512 and Alex. The DB engine is very good at truly enforcing constraints (in an unavoidable way), so this is by far the best strategy for maintaining data integrity. The only time you need to rely on an app to handle data integrity is when the database can't handle them, e.g. versions of SQLite that don't support foreign keys.

If you need to create further linkage among entities to enable app behaviors like navigating parent-child object relationships, use backref in conjunction with ForeignKey.