Tracking queries, object and Session Changes with Events

SQLAlchemy features an extensive Event Listening system used throughout the Core and ORM. Within the ORM, there are a wide variety of event listener hooks, which are documented at an API level at ORM Events. This collection of events has grown over the years to include lots of very useful new events as well as some older events that aren’t as relevant as they once were. This section will attempt to introduce the major event hooks and when they might be used.

Execute Events

New in version 1.4: The Session now features a single comprehensive hook designed to intercept all SELECT statements made on behalf of the ORM as well as bulk UPDATE and DELETE statements. This hook supersedes the previous QueryEvents.before_compile() event as well QueryEvents.before_compile_update() and QueryEvents.before_compile_delete().

Session features a comprehensive system by which all queries invoked via the Session.execute() method, which includes all SELECT statements emitted by Query as well as all SELECT statements emitted on behalf of column and relationship loaders, may be intercepted and modified. The system makes use of the SessionEvents.do_orm_execute() event hook as well as the ORMExecuteState object to represent the event state.

Basic Query Interception

SessionEvents.do_orm_execute() is firstly useful for any kind of interception of a query, which includes those emitted by Query with 1.x style as well as when an ORM-enabled 2.0 style select(), update() or delete() construct is delivered to Session.execute(). The ORMExecuteState construct provides accessors to allow modifications to statements, parameters, and options:

Session = sessionmaker(engine)

@event.listens_for(Session, "do_orm_execute")
def _do_orm_execute(orm_execute_state):
    if orm_execute_state.is_select:
        # add populate_existing for all SELECT statements


        # check if the SELECT is against a certain entity and add an
        # ORDER BY if so
        col_descriptions = orm_execute_state.statement.column_descriptions

        if col_descriptions[0]["entity"] is MyEntity:
            orm_execute_state.statement = statement.order_by(

The above example illustrates some simple modifications to SELECT statements. At this level, the SessionEvents.do_orm_execute() event hook intends to replace the previous use of the QueryEvents.before_compile() event, which was not fired off consistently for various kinds of loaders; additionally, the QueryEvents.before_compile() only applies to 1.x style use with Query and not with 2.0 style use of Session.execute().

Adding global WHERE / ON criteria

One of the most requested query-extension features is the ability to add WHERE criteria to all occurrences of an entity in all queries. This is achievable by making use of the with_loader_criteria() query option, which may be used on its own, or is ideally suited to be used within the SessionEvents.do_orm_execute() event:

from sqlalchemy.orm import with_loader_criteria

Session = sessionmaker(engine)

@event.listens_for(Session, "do_orm_execute")
def _do_orm_execute(orm_execute_state):
    if (
        and not orm_execute_state.is_column_load
        and not orm_execute_state.is_relationship_load
        orm_execute_state.statement = orm_execute_state.statement.options(
            with_loader_criteria(MyEntity.public == True)

Above, an option is added to all SELECT statements that will limit all queries against MyEntity to filter on public == True. The criteria will be applied to all loads of that class within the scope of the immediate query. The with_loader_criteria() option by default will automatically propagate to relationship loaders as well, which will apply to subsequent relationship loads, which includes lazy loads, selectinloads, etc.

For a series of classes that all feature some common column structure, if the classes are composed using a declarative mixin, the mixin class itself may be used in conjunction with the with_loader_criteria() option by making use of a Python lambda. The Python lambda will be invoked at query compilation time against the specific entities which match the criteria. Given a series of classes based on a mixin called HasTimestamp:

import datetime

class HasTimestamp:
    timestamp = mapped_column(DateTime,

class SomeEntity(HasTimestamp, Base):
    __tablename__ = "some_entity"
    id = mapped_column(Integer, primary_key=True)

class SomeOtherEntity(HasTimestamp, Base):
    __tablename__ = "some_entity"
    id = mapped_column(Integer, primary_key=True)

The above classes SomeEntity and SomeOtherEntity will each have a column timestamp that defaults to the current date and time. An event may be used to intercept all objects that extend from HasTimestamp and filter their timestamp column on a date that is no older than one month ago:

@event.listens_for(Session, "do_orm_execute")
def _do_orm_execute(orm_execute_state):
    if (
        and not orm_execute_state.is_column_load
        and not orm_execute_state.is_relationship_load
        one_month_ago = - datetime.timedelta(months=1)

        orm_execute_state.statement = orm_execute_state.statement.options(
                lambda cls: cls.timestamp >= one_month_ago,


The use of a lambda inside of the call to with_loader_criteria() is only invoked once per unique class. Custom functions should not be invoked within this lambda. See Using Lambdas to add significant speed gains to statement production for an overview of the “lambda SQL” feature, which is for advanced use only.

See also

ORM Query Events - includes working examples of the above with_loader_criteria() recipes.

Re-Executing Statements

Deep Alchemy

the statement re-execution feature involves a slightly intricate recursive sequence, and is intended to solve the fairly hard problem of being able to re-route the execution of a SQL statement into various non-SQL contexts. The twin examples of “dogpile caching” and “horizontal sharding”, linked below, should be used as a guide for when this rather advanced feature is appropriate to be used.

The ORMExecuteState is capable of controlling the execution of the given statement; this includes the ability to either not invoke the statement at all, allowing a pre-constructed result set retrieved from a cache to be returned instead, as well as the ability to invoke the same statement repeatedly with different state, such as invoking it against multiple database connections and then merging the results together in memory. Both of these advanced patterns are demonstrated in SQLAlchemy’s example suite as detailed below.

When inside the SessionEvents.do_orm_execute() event hook, the ORMExecuteState.invoke_statement() method may be used to invoke the statement using a new nested invocation of Session.execute(), which will then preempt the subsequent handling of the current execution in progress and instead return the Result returned by the inner execution. The event handlers thus far invoked for the SessionEvents.do_orm_execute() hook within this process will be skipped within this nested call as well.

The ORMExecuteState.invoke_statement() method returns a Result object; this object then features the ability for it to be “frozen” into a cacheable format and “unfrozen” into a new Result object, as well as for its data to be merged with that of other Result objects.

E.g., using SessionEvents.do_orm_execute() to implement a cache:

from sqlalchemy.orm import loading

cache = {}

@event.listens_for(Session, "do_orm_execute")
def _do_orm_execute(orm_execute_state):
    if "my_cache_key" in orm_execute_state.execution_options:
        cache_key = orm_execute_state.execution_options["my_cache_key"]

        if cache_key in cache:
            frozen_result = cache[cache_key]
            frozen_result = orm_execute_state.invoke_statement().freeze()
            cache[cache_key] = frozen_result

        return loading.merge_frozen_result(

With the above hook in place, an example of using the cache would look like:

stmt = (
    select(User).where( == "sandy").execution_options(my_cache_key="key_sandy")

result = session.execute(stmt)

Above, a custom execution option is passed to Select.execution_options() in order to establish a “cache key” that will then be intercepted by the SessionEvents.do_orm_execute() hook. This cache key is then matched to a FrozenResult object that may be present in the cache, and if present, the object is re-used. The recipe makes use of the Result.freeze() method to “freeze” a Result object, which above will contain ORM results, such that it can be stored in a cache and used multiple times. In order to return a live result from the “frozen” result, the merge_frozen_result() function is used to merge the “frozen” data from the result object into the current session.

The above example is implemented as a complete example in Dogpile Caching.

The ORMExecuteState.invoke_statement() method may also be called multiple times, passing along different information to the ORMExecuteState.invoke_statement.bind_arguments parameter such that the Session will make use of different Engine objects each time. This will return a different Result object each time; these results can be merged together using the Result.merge() method. This is the technique employed by the Horizontal Sharding extension; see the source code to familiarize.

Persistence Events

Probably the most widely used series of events are the “persistence” events, which correspond to the flush process. The flush is where all the decisions are made about pending changes to objects and are then emitted out to the database in the form of INSERT, UPDATE, and DELETE statements.


The SessionEvents.before_flush() hook is by far the most generally useful event to use when an application wants to ensure that additional persistence changes to the database are made when a flush proceeds. Use SessionEvents.before_flush() in order to operate upon objects to validate their state as well as to compose additional objects and references before they are persisted. Within this event, it is safe to manipulate the Session’s state, that is, new objects can be attached to it, objects can be deleted, and individual attributes on objects can be changed freely, and these changes will be pulled into the flush process when the event hook completes.

The typical SessionEvents.before_flush() hook will be tasked with scanning the collections, Session.dirty and Session.deleted in order to look for objects where something will be happening.

For illustrations of SessionEvents.before_flush(), see examples such as Versioning with a History Table and Versioning using Temporal Rows.


The SessionEvents.after_flush() hook is called after the SQL has been emitted for a flush process, but before the state of the objects that were flushed has been altered. That is, you can still inspect the, Session.dirty and Session.deleted collections to see what was just flushed, and you can also use history tracking features like the ones provided by AttributeState to see what changes were just persisted. In the SessionEvents.after_flush() event, additional SQL can be emitted to the database based on what’s observed to have changed.


SessionEvents.after_flush_postexec() is called soon after SessionEvents.after_flush(), but is invoked after the state of the objects has been modified to account for the flush that just took place. The, Session.dirty and Session.deleted collections are normally completely empty here. Use SessionEvents.after_flush_postexec() to inspect the identity map for finalized objects and possibly emit additional SQL. In this hook, there is the ability to make new changes on objects, which means the Session will again go into a “dirty” state; the mechanics of the Session here will cause it to flush again if new changes are detected in this hook if the flush were invoked in the context of Session.commit(); otherwise, the pending changes will be bundled as part of the next normal flush. When the hook detects new changes within a Session.commit(), a counter ensures that an endless loop in this regard is stopped after 100 iterations, in the case that an SessionEvents.after_flush_postexec() hook continually adds new state to be flushed each time it is called.

Mapper-level Flush Events

In addition to the flush-level hooks, there is also a suite of hooks that are more fine-grained, in that they are called on a per-object basis and are broken out based on INSERT, UPDATE or DELETE within the flush process. These are the mapper persistence hooks, and they too are very popular, however these events need to be approached more cautiously, as they proceed within the context of the flush process that is already ongoing; many operations are not safe to proceed here.

The events are:


It is important to note that these events apply only to the session flush operation , and not to the ORM-level INSERT/UPDATE/DELETE functionality described at ORM-Enabled INSERT, UPDATE, and DELETE statements. To intercept ORM-level DML, use the SessionEvents.do_orm_execute() event.

Each event is passed the Mapper, the mapped object itself, and the Connection which is being used to emit an INSERT, UPDATE or DELETE statement. The appeal of these events is clear, in that if an application wants to tie some activity to when a specific type of object is persisted with an INSERT, the hook is very specific; unlike the SessionEvents.before_flush() event, there’s no need to search through collections like in order to find targets. However, the flush plan which represents the full list of every single INSERT, UPDATE, DELETE statement to be emitted has already been decided when these events are called, and no changes may be made at this stage. Therefore the only changes that are even possible to the given objects are upon attributes local to the object’s row. Any other change to the object or other objects will impact the state of the Session, which will fail to function properly.

Operations that are not supported within these mapper-level persistence events include:

  • Session.add()

  • Session.delete()

  • Mapped collection append, add, remove, delete, discard, etc.

  • Mapped relationship attribute set/del events, i.e. someobject.related = someotherobject

The reason the Connection is passed is that it is encouraged that simple SQL operations take place here, directly on the Connection, such as incrementing counters or inserting extra rows within log tables.

There are also many per-object operations that don’t need to be handled within a flush event at all. The most common alternative is to simply establish additional state along with an object inside its __init__() method, such as creating additional objects that are to be associated with the new object. Using validators as described in Simple Validators is another approach; these functions can intercept changes to attributes and establish additional state changes on the target object in response to the attribute change. With both of these approaches, the object is in the correct state before it ever gets to the flush step.

Object Lifecycle Events

Another use case for events is to track the lifecycle of objects. This refers to the states first introduced at Quickie Intro to Object States.

All the states above can be tracked fully with events. Each event represents a distinct state transition, meaning, the starting state and the destination state are both part of what are tracked. With the exception of the initial transient event, all the events are in terms of the Session object or class, meaning they can be associated either with a specific Session object:

from sqlalchemy import event
from sqlalchemy.orm import Session

session = Session()

@event.listens_for(session, "transient_to_pending")
def object_is_pending(session, obj):
    print("new pending: %s" % obj)

Or with the Session class itself, as well as with a specific sessionmaker, which is likely the most useful form:

from sqlalchemy import event
from sqlalchemy.orm import sessionmaker

maker = sessionmaker()

@event.listens_for(maker, "transient_to_pending")
def object_is_pending(session, obj):
    print("new pending: %s" % obj)

The listeners can of course be stacked on top of one function, as is likely to be common. For example, to track all objects that are entering the persistent state:

@event.listens_for(maker, "pending_to_persistent")
@event.listens_for(maker, "deleted_to_persistent")
@event.listens_for(maker, "detached_to_persistent")
@event.listens_for(maker, "loaded_as_persistent")
def detect_all_persistent(session, instance):
    print("object is now persistent: %s" % instance)


All mapped objects when first constructed start out as transient. In this state, the object exists alone and doesn’t have an association with any Session. For this initial state, there’s no specific “transition” event since there is no Session, however if one wanted to intercept when any transient object is created, the InstanceEvents.init() method is probably the best event. This event is applied to a specific class or superclass. For example, to intercept all new objects for a particular declarative base:

from sqlalchemy.orm import DeclarativeBase
from sqlalchemy import event

class Base(DeclarativeBase):

@event.listens_for(Base, "init", propagate=True)
def intercept_init(instance, args, kwargs):
    print("new transient: %s" % instance)

Transient to Pending

The transient object becomes pending when it is first associated with a Session via the Session.add() or Session.add_all() method. An object may also become part of a Session as a result of a “cascade” from a referencing object that was explicitly added. The transient to pending transition is detectable using the SessionEvents.transient_to_pending() event:

@event.listens_for(sessionmaker, "transient_to_pending")
def intercept_transient_to_pending(session, object_):
    print("transient to pending: %s" % object_)

Pending to Persistent

The pending object becomes persistent when a flush proceeds and an INSERT statement takes place for the instance. The object now has an identity key. Track pending to persistent with the SessionEvents.pending_to_persistent() event:

@event.listens_for(sessionmaker, "pending_to_persistent")
def intercept_pending_to_persistent(session, object_):
    print("pending to persistent: %s" % object_)

Pending to Transient

The pending object can revert back to transient if the Session.rollback() method is called before the pending object has been flushed, or if the Session.expunge() method is called for the object before it is flushed. Track pending to transient with the SessionEvents.pending_to_transient() event:

@event.listens_for(sessionmaker, "pending_to_transient")
def intercept_pending_to_transient(session, object_):
    print("transient to pending: %s" % object_)

Loaded as Persistent

Objects can appear in the Session directly in the persistent state when they are loaded from the database. Tracking this state transition is synonymous with tracking objects as they are loaded, and is synonymous with using the InstanceEvents.load() instance-level event. However, the SessionEvents.loaded_as_persistent() event is provided as a session-centric hook for intercepting objects as they enter the persistent state via this particular avenue:

@event.listens_for(sessionmaker, "loaded_as_persistent")
def intercept_loaded_as_persistent(session, object_):
    print("object loaded into persistent state: %s" % object_)

Persistent to Transient

The persistent object can revert to the transient state if the Session.rollback() method is called for a transaction where the object was first added as pending. In the case of the ROLLBACK, the INSERT statement that made this object persistent is rolled back, and the object is evicted from the Session to again become transient. Track objects that were reverted to transient from persistent using the SessionEvents.persistent_to_transient() event hook:

@event.listens_for(sessionmaker, "persistent_to_transient")
def intercept_persistent_to_transient(session, object_):
    print("persistent to transient: %s" % object_)

Persistent to Deleted

The persistent object enters the deleted state when an object marked for deletion is deleted from the database within the flush process. Note that this is not the same as when the Session.delete() method is called for a target object. The Session.delete() method only marks the object for deletion; the actual DELETE statement is not emitted until the flush proceeds. It is subsequent to the flush that the “deleted” state is present for the target object.

Within the “deleted” state, the object is only marginally associated with the Session. It is not present in the identity map nor is it present in the Session.deleted collection that refers to when it was pending for deletion.

From the “deleted” state, the object can go either to the detached state when the transaction is committed, or back to the persistent state if the transaction is instead rolled back.

Track the persistent to deleted transition with SessionEvents.persistent_to_deleted():

@event.listens_for(sessionmaker, "persistent_to_deleted")
def intercept_persistent_to_deleted(session, object_):
    print("object was DELETEd, is now in deleted state: %s" % object_)

Deleted to Detached

The deleted object becomes detached when the session’s transaction is committed. After the Session.commit() method is called, the database transaction is final and the Session now fully discards the deleted object and removes all associations to it. Track the deleted to detached transition using SessionEvents.deleted_to_detached():

@event.listens_for(sessionmaker, "deleted_to_detached")
def intercept_deleted_to_detached(session, object_):
    print("deleted to detached: %s" % object_)


While the object is in the deleted state, the InstanceState.deleted attribute, accessible using inspect(object).deleted, returns True. However when the object is detached, InstanceState.deleted will again return False. To detect that an object was deleted, regardless of whether or not it is detached, use the InstanceState.was_deleted accessor.

Persistent to Detached

The persistent object becomes detached when the object is de-associated with the Session, via the Session.expunge(), Session.expunge_all(), or Session.close() methods.


An object may also become implicitly detached if its owning Session is dereferenced by the application and discarded due to garbage collection. In this case, no event is emitted.

Track objects as they move from persistent to detached using the SessionEvents.persistent_to_detached() event:

@event.listens_for(sessionmaker, "persistent_to_detached")
def intercept_persistent_to_detached(session, object_):
    print("object became detached: %s" % object_)

Detached to Persistent

The detached object becomes persistent when it is re-associated with a session using the Session.add() or equivalent method. Track objects moving back to persistent from detached using the SessionEvents.detached_to_persistent() event:

@event.listens_for(sessionmaker, "detached_to_persistent")
def intercept_detached_to_persistent(session, object_):
    print("object became persistent again: %s" % object_)

Deleted to Persistent

The deleted object can be reverted to the persistent state when the transaction in which it was DELETEd was rolled back using the Session.rollback() method. Track deleted objects moving back to the persistent state using the SessionEvents.deleted_to_persistent() event:

@event.listens_for(sessionmaker, "deleted_to_persistent")
def intercept_deleted_to_persistent(session, object_):
    print("deleted to persistent: %s" % object_)

Transaction Events

Transaction events allow an application to be notified when transaction boundaries occur at the Session level as well as when the Session changes the transactional state on Connection objects.

Attribute Change Events

The attribute change events allow interception of when specific attributes on an object are modified. These events include AttributeEvents.set(), AttributeEvents.append(), and AttributeEvents.remove(). These events are extremely useful, particularly for per-object validation operations; however, it is often much more convenient to use a “validator” hook, which uses these hooks behind the scenes; see Simple Validators for background on this. The attribute events are also behind the mechanics of backreferences. An example illustrating use of attribute events is in Attribute Instrumentation.