Sessions / Queries

I’m re-loading data with my Session but it isn’t seeing changes that I committed elsewhere

The main issue regarding this behavior is that the session acts as though the transaction is in the serializable isolation state, even if it’s not (and it usually is not). In practical terms, this means that the session does not alter any data that it’s already read within the scope of a transaction.

If the term “isolation level” is unfamiliar, then you first need to read this link:

Isolation Level

In short, serializable isolation level generally means that once you SELECT a series of rows in a transaction, you will get the identical data back each time you re-emit that SELECT. If you are in the next-lower isolation level, “repeatable read”, you’ll see newly added rows (and no longer see deleted rows), but for rows that you’ve already loaded, you won’t see any change. Only if you are in a lower isolation level, e.g. “read committed”, does it become possible to see a row of data change its value.

For information on controlling the isolation level when using the SQLAlchemy ORM, see Setting Transaction Isolation Levels / DBAPI AUTOCOMMIT.

To simplify things dramatically, the Session itself works in terms of a completely isolated transaction, and doesn’t overwrite any mapped attributes it’s already read unless you tell it to. The use case of trying to re-read data you’ve already loaded in an ongoing transaction is an uncommon use case that in many cases has no effect, so this is considered to be the exception, not the norm; to work within this exception, several methods are provided to allow specific data to be reloaded within the context of an ongoing transaction.

To understand what we mean by “the transaction” when we talk about the Session, your Session is intended to only work within a transaction. An overview of this is at Managing Transactions.

Once we’ve figured out what our isolation level is, and we think that our isolation level is set at a low enough level so that if we re-SELECT a row, we should see new data in our Session, how do we see it?

Three ways, from most common to least:

  1. We simply end our transaction and start a new one on next access with our Session by calling Session.commit() (note that if the Session is in the lesser-used “autocommit” mode, there would be a call to Session.begin() as well). The vast majority of applications and use cases do not have any issues with not being able to “see” data in other transactions because they stick to this pattern, which is at the core of the best practice of short lived transactions. See When do I construct a Session, when do I commit it, and when do I close it? for some thoughts on this.

  2. We tell our Session to re-read rows that it has already read, either when we next query for them using Session.expire_all() or Session.expire(), or immediately on an object using refresh. See Refreshing / Expiring for detail on this.

  3. We can run whole queries while setting them to definitely overwrite already-loaded objects as they read rows by using “populate existing”. This is an execution option described at Populate Existing.

But remember, the ORM cannot see changes in rows if our isolation level is repeatable read or higher, unless we start a new transaction.

“This Session’s transaction has been rolled back due to a previous exception during flush.” (or similar)

This is an error that occurs when a Session.flush() raises an exception, rolls back the transaction, but further commands upon the Session are called without an explicit call to Session.rollback() or Session.close().

It usually corresponds to an application that catches an exception upon Session.flush() or Session.commit() and does not properly handle the exception. For example:

from sqlalchemy import create_engine, Column, Integer
from sqlalchemy.orm import sessionmaker
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base(create_engine("sqlite://"))


class Foo(Base):
    __tablename__ = "foo"
    id = Column(Integer, primary_key=True)


Base.metadata.create_all()

session = sessionmaker()()

# constraint violation
session.add_all([Foo(id=1), Foo(id=1)])

try:
    session.commit()
except:
    # ignore error
    pass

# continue using session without rolling back
session.commit()

The usage of the Session should fit within a structure similar to this:

try:
    # <use session>
    session.commit()
except:
    session.rollback()
    raise
finally:
    session.close()  # optional, depends on use case

Many things can cause a failure within the try/except besides flushes. Applications should ensure some system of “framing” is applied to ORM-oriented processes so that connection and transaction resources have a definitive boundary, and so that transactions can be explicitly rolled back if any failure conditions occur.

This does not mean there should be try/except blocks throughout an application, which would not be a scalable architecture. Instead, a typical approach is that when ORM-oriented methods and functions are first called, the process that’s calling the functions from the very top would be within a block that commits transactions at the successful completion of a series of operations, as well as rolls transactions back if operations fail for any reason, including failed flushes. There are also approaches using function decorators or context managers to achieve similar results. The kind of approach taken depends very much on the kind of application being written.

For a detailed discussion on how to organize usage of the Session, please see When do I construct a Session, when do I commit it, and when do I close it?.

But why does flush() insist on issuing a ROLLBACK?

It would be great if Session.flush() could partially complete and then not roll back, however this is beyond its current capabilities since its internal bookkeeping would have to be modified such that it can be halted at any time and be exactly consistent with what’s been flushed to the database. While this is theoretically possible, the usefulness of the enhancement is greatly decreased by the fact that many database operations require a ROLLBACK in any case. Postgres in particular has operations which, once failed, the transaction is not allowed to continue:

test=> create table foo(id integer primary key);
NOTICE:  CREATE TABLE / PRIMARY KEY will create implicit index "foo_pkey" for table "foo"
CREATE TABLE
test=> begin;
BEGIN
test=> insert into foo values(1);
INSERT 0 1
test=> commit;
COMMIT
test=> begin;
BEGIN
test=> insert into foo values(1);
ERROR:  duplicate key value violates unique constraint "foo_pkey"
test=> insert into foo values(2);
ERROR:  current transaction is aborted, commands ignored until end of transaction block

What SQLAlchemy offers that solves both issues is support of SAVEPOINT, via Session.begin_nested(). Using Session.begin_nested(), you can frame an operation that may potentially fail within a transaction, and then “roll back” to the point before its failure while maintaining the enclosing transaction.

But why isn’t the one automatic call to ROLLBACK enough? Why must I ROLLBACK again?

The rollback that’s caused by the flush() is not the end of the complete transaction block; while it ends the database transaction in play, from the Session point of view there is still a transaction that is now in an inactive state.

Given a block such as:

sess = Session()  # begins a logical transaction
try:
    sess.flush()

    sess.commit()
except:
    sess.rollback()

Above, when a Session is first created, assuming “autocommit mode” isn’t used, a logical transaction is established within the Session. This transaction is “logical” in that it does not actually use any database resources until a SQL statement is invoked, at which point a connection-level and DBAPI-level transaction is started. However, whether or not database-level transactions are part of its state, the logical transaction will stay in place until it is ended using Session.commit(), Session.rollback(), or Session.close().

When the flush() above fails, the code is still within the transaction framed by the try/commit/except/rollback block. If flush() were to fully roll back the logical transaction, it would mean that when we then reach the except: block the Session would be in a clean state, ready to emit new SQL on an all new transaction, and the call to Session.rollback() would be out of sequence. In particular, the Session would have begun a new transaction by this point, which the Session.rollback() would be acting upon erroneously. Rather than allowing SQL operations to proceed on a new transaction in this place where normal usage dictates a rollback is about to take place, the Session instead refuses to continue until the explicit rollback actually occurs.

In other words, it is expected that the calling code will always call Session.commit(), Session.rollback(), or Session.close() to correspond to the current transaction block. flush() keeps the Session within this transaction block so that the behavior of the above code is predictable and consistent.

How do I make a Query that always adds a certain filter to every query?

See the recipe at FilteredQuery.

My Query does not return the same number of objects as query.count() tells me - why?

The Query object, when asked to return a list of ORM-mapped objects, will deduplicate the objects based on primary key. That is, if we for example use the User mapping described at Using ORM Declarative Forms to Define Table Metadata, and we had a SQL query like the following:

q = session.query(User).outerjoin(User.addresses).filter(User.name == "jack")

Above, the sample data used in the tutorial has two rows in the addresses table for the users row with the name 'jack', primary key value 5. If we ask the above query for a Query.count(), we will get the answer 2:

>>> q.count()
2

However, if we run Query.all() or iterate over the query, we get back one element:

>>> q.all()
[User(id=5, name='jack', ...)]

This is because when the Query object returns full entities, they are deduplicated. This does not occur if we instead request individual columns back:

>>> session.query(User.id, User.name).outerjoin(User.addresses).filter(
...     User.name == "jack"
... ).all()
[(5, 'jack'), (5, 'jack')]

There are two main reasons the Query will deduplicate:

  • To allow joined eager loading to work correctly - Joined Eager Loading works by querying rows using joins against related tables, where it then routes rows from those joins into collections upon the lead objects. In order to do this, it has to fetch rows where the lead object primary key is repeated for each sub-entry. This pattern can then continue into further sub-collections such that a multiple of rows may be processed for a single lead object, such as User(id=5). The dedpulication allows us to receive objects in the way they were queried, e.g. all the User() objects whose name is 'jack' which for us is one object, with the User.addresses collection eagerly loaded as was indicated either by lazy='joined' on the relationship() or via the joinedload() option. For consistency, the deduplication is still applied whether or not the joinedload is established, as the key philosophy behind eager loading is that these options never affect the result.

  • To eliminate confusion regarding the identity map - this is admittedly the less critical reason. As the Session makes use of an identity map, even though our SQL result set has two rows with primary key 5, there is only one User(id=5) object inside the Session which must be maintained uniquely on its identity, that is, its primary key / class combination. It doesn’t actually make much sense, if one is querying for User() objects, to get the same object multiple times in the list. An ordered set would potentially be a better representation of what Query seeks to return when it returns full objects.

The issue of Query deduplication remains problematic, mostly for the single reason that the Query.count() method is inconsistent, and the current status is that joined eager loading has in recent releases been superseded first by the “subquery eager loading” strategy and more recently the “select IN eager loading” strategy, both of which are generally more appropriate for collection eager loading. As this evolution continues, SQLAlchemy may alter this behavior on Query, which may also involve new APIs in order to more directly control this behavior, and may also alter the behavior of joined eager loading in order to create a more consistent usage pattern.

I’ve created a mapping against an Outer Join, and while the query returns rows, no objects are returned. Why not?

Rows returned by an outer join may contain NULL for part of the primary key, as the primary key is the composite of both tables. The Query object ignores incoming rows that don’t have an acceptable primary key. Based on the setting of the allow_partial_pks flag on Mapper, a primary key is accepted if the value has at least one non-NULL value, or alternatively if the value has no NULL values. See allow_partial_pks at Mapper.

I’m using joinedload() or lazy=False to create a JOIN/OUTER JOIN and SQLAlchemy is not constructing the correct query when I try to add a WHERE, ORDER BY, LIMIT, etc. (which relies upon the (OUTER) JOIN)

The joins generated by joined eager loading are only used to fully load related collections, and are designed to have no impact on the primary results of the query. Since they are anonymously aliased, they cannot be referenced directly.

For detail on this behavior, see The Zen of Joined Eager Loading.

Query has no __len__(), why not?

The Python __len__() magic method applied to an object allows the len() builtin to be used to determine the length of the collection. It’s intuitive that a SQL query object would link __len__() to the Query.count() method, which emits a SELECT COUNT. The reason this is not possible is because evaluating the query as a list would incur two SQL calls instead of one:

class Iterates:
    def __len__(self):
        print("LEN!")
        return 5

    def __iter__(self):
        print("ITER!")
        return iter([1, 2, 3, 4, 5])


list(Iterates())

output:

ITER!
LEN!

How Do I use Textual SQL with ORM Queries?

See:

I’m calling Session.delete(myobject) and it isn’t removed from the parent collection!

See Notes on Delete - Deleting Objects Referenced from Collections and Scalar Relationships for a description of this behavior.

why isn’t my __init__() called when I load objects?

See Constructors and Object Initialization for a description of this behavior.

how do I use ON DELETE CASCADE with SA’s ORM?

SQLAlchemy will always issue UPDATE or DELETE statements for dependent rows which are currently loaded in the Session. For rows which are not loaded, it will by default issue SELECT statements to load those rows and update/delete those as well; in other words it assumes there is no ON DELETE CASCADE configured. To configure SQLAlchemy to cooperate with ON DELETE CASCADE, see Using foreign key ON DELETE cascade with ORM relationships.

I set the “foo_id” attribute on my instance to “7”, but the “foo” attribute is still None - shouldn’t it have loaded Foo with id #7?

The ORM is not constructed in such a way as to support immediate population of relationships driven from foreign key attribute changes - instead, it is designed to work the other way around - foreign key attributes are handled by the ORM behind the scenes, the end user sets up object relationships naturally. Therefore, the recommended way to set o.foo is to do just that - set it!:

foo = session.get(Foo, 7)
o.foo = foo
Session.commit()

Manipulation of foreign key attributes is of course entirely legal. However, setting a foreign-key attribute to a new value currently does not trigger an “expire” event of the relationship() in which it’s involved. This means that for the following sequence:

o = session.scalars(select(SomeClass).limit(1)).first()

# assume the existing o.foo_id value is None;
# accessing o.foo will reconcile this as ``None``, but will effectively
# "load" the value of None
assert o.foo is None

# now set foo_id to something.  o.foo will not be immediately affected
o.foo_id = 7

o.foo is loaded with its effective database value of None when it is first accessed. Setting o.foo_id = 7 will have the value of “7” as a pending change, but no flush has occurred - so o.foo is still None:

# attribute is already "loaded" as None, has not been
# reconciled with o.foo_id = 7 yet
assert o.foo is None

For o.foo to load based on the foreign key mutation is usually achieved naturally after the commit, which both flushes the new foreign key value and expires all state:

session.commit()  # expires all attributes

foo_7 = session.get(Foo, 7)

# o.foo will lazyload again, this time getting the new object
assert o.foo is foo_7

A more minimal operation is to expire the attribute individually - this can be performed for any persistent object using Session.expire():

o = session.scalars(select(SomeClass).limit(1)).first()
o.foo_id = 7
Session.expire(o, ["foo"])  # object must be persistent for this

foo_7 = session.get(Foo, 7)

assert o.foo is foo_7  # o.foo lazyloads on access

Note that if the object is not persistent but present in the Session, it’s known as pending. This means the row for the object has not been INSERTed into the database yet. For such an object, setting foo_id does not have meaning until the row is inserted; otherwise there is no row yet:

new_obj = SomeClass()
new_obj.foo_id = 7

Session.add(new_obj)

# returns None but this is not a "lazyload", as the object is not
# persistent in the DB yet, and the None value is not part of the
# object's state
assert new_obj.foo is None

Session.flush()  # emits INSERT

assert new_obj.foo is foo_7  # now it loads

The recipe ExpireRelationshipOnFKChange features an example using SQLAlchemy events in order to coordinate the setting of foreign key attributes with many-to-one relationships.

Is there a way to automagically have only unique keywords (or other kinds of objects) without doing a query for the keyword and getting a reference to the row containing that keyword?

When people read the many-to-many example in the docs, they get hit with the fact that if you create the same Keyword twice, it gets put in the DB twice. Which is somewhat inconvenient.

This UniqueObject recipe was created to address this issue.

Why does post_update emit UPDATE in addition to the first UPDATE?

The post_update feature, documented at Rows that point to themselves / Mutually Dependent Rows, involves that an UPDATE statement is emitted in response to changes to a particular relationship-bound foreign key, in addition to the INSERT/UPDATE/DELETE that would normally be emitted for the target row. While the primary purpose of this UPDATE statement is that it pairs up with an INSERT or DELETE of that row, so that it can post-set or pre-unset a foreign key reference in order to break a cycle with a mutually dependent foreign key, it currently is also bundled as a second UPDATE that emits when the target row itself is subject to an UPDATE. In this case, the UPDATE emitted by post_update is usually unnecessary and will often appear wasteful.

However, some research into trying to remove this “UPDATE / UPDATE” behavior reveals that major changes to the unit of work process would need to occur not just throughout the post_update implementation, but also in areas that aren’t related to post_update for this to work, in that the order of operations would need to be reversed on the non-post_update side in some cases, which in turn can impact other cases, such as correctly handling an UPDATE of a referenced primary key value (see #1063 for a proof of concept).

The answer is that “post_update” is used to break a cycle between two mutually dependent foreign keys, and to have this cycle breaking be limited to just INSERT/DELETE of the target table implies that the ordering of UPDATE statements elsewhere would need to be liberalized, leading to breakage in other edge cases.