Baked Queries

baked provides an alternative creational pattern for Query objects, which allows for caching of the object’s construction and string-compilation steps. This means that for a particular Query building scenario that is used more than once, all of the Python function invocation involved in building the query from its initial construction up through generating a SQL string will only occur once, rather than for each time that query is built up and executed.

The rationale for this system is to greatly reduce Python interpreter overhead for everything that occurs before the SQL is emitted. The caching of the “baked” system does not in any way reduce SQL calls or cache the return results from the database. A technique that demonstrates the caching of the SQL calls and result sets themselves is available in Dogpile Caching.

New in version 1.0.0.


The sqlalchemy.ext.baked extension is not for beginners. Using it correctly requires a good high level understanding of how SQLAlchemy, the database driver, and the backend database interact with each other. This extension presents a very specific kind of optimization that is not ordinarily needed. As noted above, it does not cache queries, only the string formulation of the SQL itself.


Usage of the baked system starts by producing a so-called “bakery”, which represents storage for a particular series of query objects:

from sqlalchemy.ext import baked

bakery = baked.bakery()

The above “bakery” will store cached data in an LRU cache that defaults to 200 elements, noting that an ORM query will typically contain one entry for the ORM query as invoked, as well as one entry per database dialect for the SQL string.

The bakery allows us to build up a Query object by specifying its construction as a series of Python callables, which are typically lambdas. For succinct usage, it overrides the += operator so that a typical query build-up looks like the following:

from sqlalchemy import bindparam

def search_for_user(session, username, email=None):

    baked_query = bakery(lambda session: session.query(User))
    baked_query += lambda q: q.filter( == bindparam('username'))

    baked_query += lambda q: q.order_by(

    if email:
        baked_query += lambda q: q.filter( == bindparam('email'))

    result = baked_query(session).params(username=username, email=email).all()

    return result

Following are some observations about the above code:

  1. The baked_query object is an instance of BakedQuery. This object is essentially the “builder” for a real orm Query object, but it is not itself the actual Query object.

  2. The actual Query object is not built at all, until the very end of the function when Result.all() is called.

  3. The steps that are added to the baked_query object are all expressed as Python functions, typically lambdas. The first lambda given to the bakery() function receives a Session as its argument. The remaining lambdas each receive a Query as their argument.

  4. In the above code, even though our application may call upon search_for_user() many times, and even though within each invocation we build up an entirely new BakedQuery object, all of the lambdas are only called once. Each lambda is never called a second time for as long as this query is cached in the bakery.

  5. The caching is achieved by storing references to the lambda objects themselves in order to formulate a cache key; that is, the fact that the Python interpreter assigns an in-Python identity to these functions is what determines how to identify the query on successive runs. For those invocations of search_for_user() where the email parameter is specified, the callable lambda q: q.filter( == bindparam('email')) will be part of the cache key that’s retrieved; when email is None, this callable is not part of the cache key.

  6. Because the lambdas are all called only once, it is essential that no variables which may change across calls are referenced within the lambdas; instead, assuming these are values to be bound into the SQL string, we use bindparam() to construct named parameters, where we apply their actual values later using Result.params().


The baked query probably looks a little odd, a little bit awkward and a little bit verbose. However, the savings in Python performance for a query which is invoked lots of times in an application are very dramatic. The example suite short_selects demonstrated in Performance illustrates a comparison of queries which each return only one row, such as the following regular query:

session = Session(bind=engine)
for id_ in random.sample(ids, n):
    session.query(Customer).filter( == id_).one()

compared to the equivalent “baked” query:

bakery = baked.bakery()
s = Session(bind=engine)
for id_ in random.sample(ids, n):
    q = bakery(lambda s: s.query(Customer))
    q += lambda q: q.filter( == bindparam('id'))

The difference in Python function call count for an iteration of 10000 calls to each block are:

test_baked_query : test a baked query of the full entity.
                   (10000 iterations); total fn calls 1951294

test_orm_query :   test a straight ORM query of the full entity.
                   (10000 iterations); total fn calls 7900535

In terms of number of seconds on a powerful laptop, this comes out as:

test_baked_query : test a baked query of the full entity.
                   (10000 iterations); total time 2.174126 sec

test_orm_query :   test a straight ORM query of the full entity.
                   (10000 iterations); total time 7.958516 sec

Note that this test very intentionally features queries that only return one row. For queries that return many rows, the performance advantage of the baked query will have less and less of an impact, proportional to the time spent fetching rows. It is critical to keep in mind that the baked query feature only applies to building the query itself, not the fetching of results. Using the baked feature is by no means a guarantee to a much faster application; it is only a potentially useful feature for those applications that have been measured as being impacted by this particular form of overhead.

Measure twice, cut once

For background on how to profile a SQLAlchemy application, please see the section Performance. It is essential that performance measurement techniques are used when attempting to improve the performance of an application.


The “lambda” approach above is a superset of what would be a more traditional “parameterized” approach. Suppose we wished to build a simple system where we build a Query just once, then store it in a dictionary for re-use. This is possible right now by just building up the query, and removing its Session by calling my_cached_query = query.with_session(None):

my_simple_cache = {}

def lookup(session, id_argument):
    if "my_key" not in my_simple_cache:
        query = session.query(Model).filter( == bindparam('id'))
        my_simple_cache["my_key"] = query.with_session(None)
        query = my_simple_cache["my_key"].with_session(session)

    return query.params(id=id_argument).all()

The above approach gets us a very minimal performance benefit. By re-using a Query, we save on the Python work within the session.query(Model) constructor as well as calling upon filter( == bindparam('id')), which will skip for us the building up of the Core expression as well as sending it to Query.filter(). However, the approach still regenerates the full Select object every time when Query.all() is called and additionally this brand new Select is sent off to the string compilation step every time, which for a simple case like the above is probably about 70% of the overhead.

To reduce the additional overhead, we need some more specialized logic, some way to memoize the construction of the select object and the construction of the SQL. There is an example of this on the wiki in the section BakedQuery, a precursor to this feature, however in that system, we aren’t caching the construction of the query. In order to remove all the overhead, we need to cache both the construction of the query as well as the SQL compilation. Let’s assume we adapted the recipe in this way and made ourselves a method .bake() that pre-compiles the SQL for the query, producing a new object that can be invoked with minimal overhead. Our example becomes:

my_simple_cache = {}

def lookup(session, id_argument):

    if "my_key" not in my_simple_cache:
        query = session.query(Model).filter( == bindparam('id'))
        my_simple_cache["my_key"] = query.with_session(None).bake()
        query = my_simple_cache["my_key"].with_session(session)

    return query.params(id=id_argument).all()

Above, we’ve fixed the performance situation, but we still have this string cache key to deal with.

We can use the “bakery” approach to re-frame the above in a way that looks less unusual than the “building up lambdas” approach, and more like a simple improvement upon the simple “reuse a query” approach:

bakery = baked.bakery()

def lookup(session, id_argument):
    def create_model_query(session):
        return session.query(Model).filter( == bindparam('id'))

    parameterized_query = bakery.bake(create_model_query)
    return parameterized_query(session).params(id=id_argument).all()

Above, we use the “baked” system in a manner that is very similar to the simplistic “cache a query” system. However, it uses two fewer lines of code, does not need to manufacture a cache key of “my_key”, and also includes the same feature as our custom “bake” function that caches 100% of the Python invocation work from the constructor of the query, to the filter call, to the production of the Select object, to the string compilation step.

From the above, if we ask ourselves, “what if lookup needs to make conditional decisions as to the structure of the query?”, this is where hopefully it becomes apparent why “baked” is the way it is. Instead of a parameterized query building off from exactly one function (which is how we thought baked might work originally), we can build it from any number of functions. Consider our naive example, if we needed to have an additional clause in our query on a conditional basis:

my_simple_cache = {}

def lookup(session, id_argument, include_frobnizzle=False):
    if include_frobnizzle:
        cache_key = "my_key_with_frobnizzle"
        cache_key = "my_key_without_frobnizzle"

    if cache_key not in my_simple_cache:
        query = session.query(Model).filter( == bindparam('id'))
        if include_frobnizzle:
            query = query.filter(Model.frobnizzle == True)

        my_simple_cache[cache_key] = query.with_session(None).bake()
        query = my_simple_cache[cache_key].with_session(session)

    return query.params(id=id_argument).all()

Our “simple” parameterized system must now be tasked with generating cache keys which take into account whether or not the “include_frobnizzle” flag was passed, as the presence of this flag means that the generated SQL would be entirely different. It should be apparent that as the complexity of query building goes up, the task of caching these queries becomes burdensome very quickly. We can convert the above example into a direct use of “bakery” as follows:

bakery = baked.bakery()

def lookup(session, id_argument, include_frobnizzle=False):
    def create_model_query(session):
        return session.query(Model).filter( == bindparam('id'))

    parameterized_query = bakery.bake(create_model_query)

    if include_frobnizzle:
        def include_frobnizzle_in_query(query):
            return query.filter(Model.frobnizzle == True)

        parameterized_query = parameterized_query.with_criteria(

    return parameterized_query(session).params(id=id_argument).all()

Above, we again cache not just the query object but all the work it needs to do in order to generate SQL. We also no longer need to deal with making sure we generate a cache key that accurately takes into account all of the structural modifications we’ve made; this is now handled automatically and without the chance of mistakes.

This code sample is a few lines shorter than the naive example, removes the need to deal with cache keys, and has the vast performance benefits of the full so-called “baked” feature. But still a little verbose! Hence we take methods like BakedQuery.add_criteria() and BakedQuery.with_criteria() and shorten them into operators, and encourage (though certainly not require!) using simple lambdas, only as a means to reduce verbosity:

bakery = baked.bakery()

def lookup(session, id_argument, include_frobnizzle=False):
    parameterized_query = bakery.bake(
        lambda s: s.query(Model).filter( == bindparam('id'))

    if include_frobnizzle:
        parameterized_query += lambda q: q.filter(Model.frobnizzle == True)

    return parameterized_query(session).params(id=id_argument).all()

Where above, the approach is simpler to implement and much more similar in code flow to what a non-cached querying function would look like, hence making code easier to port.

The above description is essentially a summary of the design process used to arrive at the current “baked” approach. Starting from the “normal” approaches, the additional issues of cache key construction and management, removal of all redundant Python execution, and queries built up with conditionals needed to be addressed, leading to the final approach.

Special Query Techniques

This section will describe some techniques for specific query situations.

Using IN expressions

The ColumnOperators.in_() method in SQLAlchemy historically renders a variable set of bound parameters based on the list of items that’s passed to the method. This doesn’t work for baked queries as the length of that list can change on different calls. To solve this problem, the :paramref:`.bindparam.expanding` parameter supports a late-rendered IN expression that is safe to be cached inside of baked query. The actual list of elements is rendered at statement execution time, rather than at statement compilation time:

bakery = baked.bakery()

baked_query = bakery(lambda session: session.query(User))
baked_query += lambda q: q.filter('username', expanding=True)))

result = baked_query.with_session(session).params(
  username=['ed', 'fred']).all()

Using Subqueries

When using Query objects, it is often needed that one Query object is used to generate a subquery within another. In the case where the Query is currently in baked form, an interim method may be used to retrieve the Query object, using the BakedQuery.to_query() method. This method is passed the Session or Query that is the argument to the lambda callable used to generate a particular step of the baked query:

bakery = baked.bakery()

# a baked query that will end up being used as a subquery
my_subq = bakery(lambda s: s.query(
my_subq += lambda q: q.filter( == Address.user_id)

# select a correlated subquery in the top columns list,
# we have the "session" argument, pass that
my_q = bakery(
  lambda s: s.query(, my_subq.to_query(s).as_scalar()))

# use a correlated subquery in some of the criteria, we have
# the "query" argument, pass that.
my_q += lambda q: q.filter(my_subq.to_query(q).exists())

New in version 1.3.

Using the before_compile event

As of SQLAlchemy 1.3.11, the use of the QueryEvents.before_compile() event against a particular Query will disallow the baked query system from caching the query, if the event hook returns a new Query object that is different from the one passed in. This is so that the QueryEvents.before_compile() hook may be invoked against a particular Query every time it is used, to accommodate for hooks that alter the query differently each time. To allow a QueryEvents.before_compile() to alter a Query() object, but still to allow the result to be cached, the event can be registered passing the bake_ok=True flag:

    Query, "before_compile", retval=True, bake_ok=True)
def my_event(query):
    for desc in query.column_descriptions:
        if desc['type'] is User:
            entity = desc['entity']
            query = query.filter(entity.deleted == False)
    return query

The above strategy is appropriate for an event that will modify a given Query in exactly the same way every time, not dependent on specific parameters or external state that changes.

New in version 1.3.11: - added the “bake_ok” flag to the QueryEvents.before_compile() event and disallowed caching via the “baked” extension from occurring for event handlers that return a new Query object if this flag is not set.

Disabling Baked Queries Session-wide

The flag :paramref:`.Session.enable_baked_queries` may be set to False, causing all baked queries to not use the cache when used against that Session:

session = Session(engine, enable_baked_queries=False)

Like all session flags, it is also accepted by factory objects like sessionmaker and methods like sessionmaker.configure().

The immediate rationale for this flag is to reduce memory use in the case that the query baking used by relationship loaders and other loaders is not desirable. It also can be used in the case that an application which is seeing issues potentially due to cache key conflicts from user-defined baked queries or other baked query issues can turn the behavior off, in order to identify or eliminate baked queries as the cause of an issue.

New in version 1.2.

Lazy Loading Integration

The baked query system is integrated into SQLAlchemy’s lazy loader feature as used by relationship(), and will cache queries for most lazy load conditions. A small subset of “lazy loads” may not be cached; these involve query options in conjunction with ad-hoc aliased structures that cannot produce a repeatable cache key.

Changed in version 1.2: “baked” queries are now the foundation of the lazy-loader feature of relationship().

Opting out with the bake_queries flag

The relationship() construct includes a flag :paramref:`.relationship.bake_queries` which when set to False will cause that relationship to opt out of caching queries. Additionally, the :paramref:`.Session.enable_baked_queries` setting can be used to disable all “baked query” use. These flags can be useful to conserve memory, when memory conservation is more important than performance for a particular relationship or for the application overall.

API Documentation

sqlalchemy.ext.baked.bakery(size=200, _size_alert=None)

Construct a new bakery.


an instance of Bakery

class sqlalchemy.ext.baked.BakedQuery(bakery, initial_fn, args=())

A builder object for query.Query objects.

add_criteria(fn, *args)

Add a criteria function to this BakedQuery.

This is equivalent to using the += operator to modify a BakedQuery in-place.

classmethod bakery(size=200, _size_alert=None)

Construct a new bakery.


an instance of Bakery


Return a Result object for this BakedQuery.

This is equivalent to calling the BakedQuery as a Python callable, e.g. result = my_baked_query(session).


Cancel any query caching that will occur on this BakedQuery object.

The BakedQuery can continue to be used normally, however additional creational functions will not be cached; they will be called on every invocation.

This is to support the case where a particular step in constructing a baked query disqualifies the query from being cacheable, such as a variant that relies upon some uncacheable value.


full – if False, only functions added to this BakedQuery object subsequent to the spoil step will be non-cached; the state of the BakedQuery up until this point will be pulled from the cache. If True, then the entire Query object is built from scratch each time, with all creational functions being called on each invocation.


Return the Query object for use as a subquery.

This method should be used within the lambda callable being used to generate a step of an enclosing BakedQuery. The parameter should normally be the Query object that is passed to the lambda:

sub_bq = self.bakery(lambda s: s.query(
sub_bq += lambda q: q.filter( == Address.user_id).correlate(Address)

main_bq = self.bakery(lambda s: s.query(Address))
main_bq += lambda q: q.filter(

In the case where the subquery is used in the first callable against a Session, the Session is also accepted:

sub_bq = self.bakery(lambda s: s.query(
sub_bq += lambda q: q.filter( == Address.user_id).correlate(Address)

main_bq = self.bakery(
    lambda s: s.query(, sub_bq.to_query(q).scalar_subquery())


a Query object or a class Session object, that is assumed to be within the context of an enclosing BakedQuery callable.

New in version 1.3.

with_criteria(fn, *args)

Add a criteria function to a BakedQuery cloned from this one.

This is equivalent to using the + operator to produce a new BakedQuery with modifications.

class sqlalchemy.ext.baked.Bakery(cls_, cache)

Callable which returns a BakedQuery.

This object is returned by the class method BakedQuery.bakery(). It exists as an object so that the “cache” can be easily inspected.

New in version 1.2.

class sqlalchemy.ext.baked.Result(bq, session)

Invokes a BakedQuery against a Session.

The Result object is where the actual query.Query object gets created, or retrieved from the cache, against a target Session, and is then invoked for results.


Return all rows.

Equivalent to Query.all().


return the ‘count’.

Equivalent to Query.count().

Note this uses a subquery to ensure an accurate count regardless of the structure of the original statement.

New in version 1.1.6.


Return the first row.

Equivalent to Query.first().


Retrieve an object based on identity.

Equivalent to Query.get().


Return exactly one result or raise an exception.

Equivalent to


Return one or zero results, or raise an exception for multiple rows.

Equivalent to Query.one_or_none().

New in version 1.0.9.

params(*args, **kw)

Specify parameters to be replaced into the string SQL statement.


Return the first element of the first result or None if no rows present. If multiple rows are returned, raises MultipleResultsFound.

Equivalent to Query.scalar().

New in version 1.1.6.


Add a criteria function that will be applied post-cache.

This adds a function that will be run against the Query object after it is retrieved from the cache. Functions here can be used to alter the query in ways that do not affect the SQL output, such as execution options and shard identifiers (when using a shard-enabled query object)


Result.with_post_criteria() functions are applied to the Query object after the query’s SQL statement object has been retrieved from the cache. Any operations here which intend to modify the SQL should ensure that BakedQuery.spoil() was called first.

New in version 1.2.