Fast SQLAlchemy counting (avoid query.count() subquery)

gistfile1.py

def get_count(q):
    count_q = q.statement.with_only_columns([func.count()]).order_by(None)
    count = q.session.execute(count_q).scalar()
    return count

q = session.query(TestModel).filter(...).order_by(...)

# Slow: SELECT COUNT(*) FROM (SELECT ... FROM TestModel WHERE ...) ...
print q.count()

# Fast: SELECT COUNT(*) FROM TestModel WHERE ...
print get_count(q)

nice jj

This is amazing.

👍

Thanks. I used it like this:

db.session.execute(
                    db.session
                        .query(TableName)
                        .filter_by(x_id=y.id)
                        .statement.with_only_columns([func.count()]).order_by(None)
                    ).scalar()

In case of 'joined' relations it can make bad queries. So I cancel all joined to be lazy:

count_q = q.options(lazyload('*')).statement.with_only_columns([func.count()]).order_by(None)
count = q.session.execute(count_q).scalar()

@hagai26 I am running into the same bad queries with joins.

.options(lazyload('*')) doesn't change the result or the resulting sql

count = get_count(
        User
        .join(Partnership)
        .filter(
            Customer.partnership_id == Partnership.id,
            Partnership.user_id == user_id
        )
    )

which results with 1

If I use .count() it is returning 2 (the correct answer)

Any ideas?

ended up using with_entities instead

def get_count(q):
    return q.with_entities(func.count()).scalar()

I added support for more stuff. Updated code:

from sqlalchemy import func, distinct
from sqlalchemy.orm import lazyload

def get_count(q):
	disable_group_by = False
	if len(q._entities) > 1:
		# currently support only one entity
		raise Exception('only one entity is supported for get_count, got: %s' % q)
	entity = q._entities[0]
	if hasattr(entity, 'column'):
		# _ColumnEntity has column attr - on case: query(Model.column)...
		col = entity.column
		if q._group_by and q._distinct:
			# which query can have both?
			raise NotImplementedError
		if q._group_by or q._distinct:
			col = distinct(col)
		if q._group_by:
			# need to disable group_by and enable distinct - we can do this because we have only 1 entity
			disable_group_by = True
		count_func = func.count(col)
	else:
		# _MapperEntity doesn't have column attr - on case: query(Model)...
		count_func = func.count()
	if q._group_by and not disable_group_by:
		count_func = count_func.over(None)
	count_q = q.options(lazyload('*')).statement.with_only_columns([count_func]).order_by(None)
	if disable_group_by:
		count_q = count_q.group_by(None)
	return q.session.execute(count_q).scalar()

Very helpful, thanks!

This is pretty cool, but I can't seem to get it working against a view. The underlying query I'm testing is the equivalent of select * from myview, no query params (yet) but the query it generates is select count(*) as count_1 which always returns 1. It seems the from clause just gets dropped. Do you have any thoughts about this?

UPDATE: So if there are no filters in the query the solution above results in select count(*) as count_1 and adding a filter like '1=1' changes the query to select count(*) as count_1 where 1=1, so there still isn't a table to count from.

SOLUTION: I added the following code to force the existance of a filter in the case where there are no filters;

    if hasattr(model_class, 'columns'):
        q = q.filter(model_class.columns.get('id') == model_class.columns.get('id'))
    else:
        q = q.filter(getattr(model_class, 'id') == getattr(model_class, 'id'))

It's not a great solution, but it seems to be working. I'll let you know if I can refine this (which I'm sure I can)

Thank you!

The solution in the gist was working, but I was getting warnings like SAWarning: Column 'id' on table being replaced by Column('id', Integer(), table=, primary_key=True, nullable=False), which has the same key. Consider use_labels for select() statements..

Using with_entities as suggested in @nchudleigh's comment got rid of those warnings.

Not All Heroes wear capes

I was into this problem for hours, finally got the solution. Really appreciate the solution.

If sharding is at play, make sure to pass in shard_id=YOUR_SHARD_ID as a param to execute as follows: q.session.execute(count_q, shard_id=YOUR_SHARD_ID).scalar()

amazing work.
it's so odd that query.count() makes a subquery, what's the reason behind that?

Nice function! FYI it does not consider the query limit.

So i extended the function for my usage.

count = q.session.execute(count_q).scalar()
return min(q._limit, count) if q._limit else count

I've implemented this in a subclass of Session with count() overridden, so now I could use query.count() and it does not make a sub-query.
you should implement this method in a subclass of Session, I also sub-classed the Column class, but its not really needed if you don't want.
you must pass this sub-classed Session class to your session maker.

THIS IS INSIDE SUB-CLASSED SESSION:

def count(self, **options):
    """
    returns the count of rows the sql formed by this `Query` would return.
    this method is overridden to prevent inefficient count() of sqlalchemy `Query`
    which produces a subquery.

    this method generates a single sql query like below:
    select count(column, ...)
    from table
    where ...

    :keyword bool distinct: specifies that count should
                            be executed on distinct select.
                            defaults to False if not provided.

    :keyword bool fallback: specifies that count should
                            be executed using original sqlalchemy
                            function which produces a subquery.
                            defaults to False if not provided.

    :rtype: int
    """

    fallback = options.get('fallback', False)
    needs_fallback = False
    columns = []
    for single_column in self.selectable.columns:
        if not isinstance(single_column, CoreColumn):
            if fallback is False:
                raise UnsupportedQueryStyleError('Current query does not have pure columns '
                                                 'in its expression. if you need to apply a '
                                                 'keyword like "DISTINCT", you should apply '
                                                 'it by passing "distinct=True" keyword to '
                                                 'query method and do not apply it in query '
                                                 'structure itself. for example instead of '
                                                 'writing '
                                                 '"store.query(distinct(Entity.id)).count()" '
                                                 'you should write this in the following form '
                                                 '"store.query(Entity.id).count()" and then '
                                                 'pass "distinct=True" in options of query '
                                                 'method. if you want the sqlalchemy original '
                                                 'style of count() which produces a subquery, '
                                                 'it is also possible to fallback to that '
                                                 'default sqlalchemy count() but keep in '
                                                 'mind that, that method is not efficient. '
                                                 'you could pass "fallback=True" in options '
                                                 'to fallback to default mode if CoreSession '
                                                 'failed to execute count().')
            else:
                needs_fallback = True
                break

        fullname = single_column.get_table_fullname()
        if fullname not in (None, ''):
            columns.append(fullname)

    if needs_fallback is True:
        return super().count()

    func_count = func.count()
    if len(columns) > 0:
        distinct = options.get('distinct', False)
        column_clause = ', '.join(columns)
        if distinct is True:
            column_clause = 'distinct {clause}'.format(clause=column_clause)
        func_count = func.count(column_clause)

    statement = self.options(lazyload('*')).statement.with_only_columns(
        [func_count]).order_by(None)

    store = get_current_store()
    return store.execute(statement).scalar()

THIS IS INSIDE SUB-CLASSED COLUMN:

def get_table_fullname(self):
    """
    gets the current column's table fullname, if the
    column has no table, it returns None.

    :rtype: str
    """

    if self.table is None:
        return None

    for single_column in self.table.columns:
        if single_column.name == self.name:
            for base_column in single_column.base_columns:
                return '{table}.{column}'.format(table=base_column.table.fullname,
                                                 column=self._get_real_name())

    return None

def _get_real_name(self):
    """
    gets column's real name.

    :rtype: str
    """

    for column in self.base_columns:
        return column.name

Then you could use it as normal select:
count = session.query(SomeEntity.Field1, SomeEntity.Field2).count()
and this will produce:
select count(field1, field2) from some_table

the full project is also available on GitHub.
https://github.com/mononobi/pyrin

I am not an expert on this, but it doesnt give me the right result in a query which has a join, filter and contains_eager option

@mononobi

I've implemented this in a subclass of Session with count() overridden, so now I could use query.count() and it does not make a sub-query ...

why not just override default query's count?

original 'count' in default query:

    def count(self):
        r"""Return a count of rows this the SQL formed by this :class:`Query`
        would return.

        This generates the SQL for this Query as follows::

            SELECT count(1) AS count_1 FROM (
                SELECT <rest of query follows...>
            ) AS anon_1

        The above SQL returns a single row, which is the aggregate value
        of the count function; the :meth:`_query.Query.count`
        method then returns
        that single integer value.

        .. warning::

            It is important to note that the value returned by
            count() is **not the same as the number of ORM objects that this
            Query would return from a method such as the .all() method**.
            The :class:`_query.Query` object,
            when asked to return full entities,
            will **deduplicate entries based on primary key**, meaning if the
            same primary key value would appear in the results more than once,
            only one object of that primary key would be present.  This does
            not apply to a query that is against individual columns.

            .. seealso::

                :ref:`faq_query_deduplicating`

                :ref:`orm_tutorial_query_returning`

        For fine grained control over specific columns to count, to skip the
        usage of a subquery or otherwise control of the FROM clause, or to use
        other aggregate functions, use :attr:`~sqlalchemy.sql.expression.func`
        expressions in conjunction with :meth:`~.Session.query`, i.e.::

            from sqlalchemy import func

            # count User records, without
            # using a subquery.
            session.query(func.count(User.id))

            # return count of user "id" grouped
            # by "name"
            session.query(func.count(User.id)).\
                    group_by(User.name)

            from sqlalchemy import distinct

            # count distinct "name" values
            session.query(func.count(distinct(User.name)))

        """
        col = sql.func.count(sql.literal_column("*"))
        return self.from_self(col).scalar()

overridden 'count' in custom query which is subclass of BaseQuery

    def count(self):
		disable_group_by = False
		if len(self._entities) > 1:
			# currently support only one entity
			raise Exception('only one entity is supported for get_count, got: %s' % q)
		entity = self._entities[0]
		if hasattr(entity, 'column'):
			# _ColumnEntity has column attr - on case: query(Model.column)...
			col = entity.column
			if self._group_by and self._distinct:
				# which query can have both?
				raise NotImplementedError
			if self._group_by or self._distinct:
				col = distinct(col)
			if self._group_by:
				# need to disable group_by and enable distinct - we can do this because we have only 1 entity
				disable_group_by = True
			count_func = func.count(col)
		else:
			# _MapperEntity doesn't have column attr - on case: query(Model)...
			count_func = func.count()
		if self._group_by and not disable_group_by:
			count_func = count_func.over(None)
		count_q = self.options(lazyload('*')).statement.with_only_columns([count_func]).order_by(None)
		if disable_group_by:
			count_q = count_q.group_by(None)
		return self.session.execute(count_q).scalar()

had to say, original count with self.from_self(col).scalar() is clear enough, and no troubles caused by group by, distinct, or may be even more advanced query, what ever.

Hi all!

If you using ORM querys with model as first argument — the comments below are good to use.

But, in my project i use many querys like:
db.session.query(User.id, UserInfo.name).outerjoin(UserInfo, UserInfo.user_id == User.id)

And prev examples are not able to use.

So i write class that process the query and removes unusable joins for count query.
This class checks joins of query and where conditions then removes unnecessary joins.

Usage:

q = db.session.query(
    User.id, User.nname, UserInfo.theme_id
).outerjoin(
    UserInfo, UserInfo.user_id == User.id
).filter(
    User.id > 0,
    UserInfo.name.like('123')
)

# where=True — take a look on "where"; no drop joins if it exists in "filter" statements
# fields=False — not take a look on fields; drop joins if it exists in fields
count_q = DBRemoveJoin(q, fields=False, where=True, debug=False).process()
count_q = count_q.statement.with_only_columns([func.count('*')]).order_by(None)

print(count_q)

Output:

SELECT count(:count_2) AS count_1
FROM "User" LEFT OUTER JOIN "UserInfo" ON "UserInfo".user_id = "User".id
WHERE "User".id > :id_1 AND "UserInfo".name LIKE :name_1

If you comment UserInfo where statement:

q = db.session.query(
    User.id, User.nname, UserInfo.theme_id
).outerjoin(
    UserInfo, UserInfo.user_id == User.id
).filter(
    User.id > 0,
    # UserInfo.name.like('123')
)

Result will be:

SELECT count(:count_2) AS count_1
FROM "User"
WHERE "User".id > :id_1

Class code:

import collections
import six
from sqlalchemy.orm.util import _ORMJoin, outerjoin
from sqlalchemy import Table, Column

from collections import OrderedDict


def iterable(arg):
    return (
        isinstance(arg, collections.Iterable)
        and not isinstance(arg, six.string_types)
    )


class DBRemoveJoin:
    q = None

    fields = True
    where = True
    debug = False

    table_main = None
    table_need = []
    table_joins = {}

    column_items = []
    where_items = []

    def __init__(self, q, fields=True, where=True, debug=False):
        self.fields = fields
        self.where = where
        self.debug = debug
        self.q = q.filter()
        self.table_need = []
        self.column_items = []
        self.where_items = []
        self.table_joins = OrderedDict()

    @staticmethod
    def clause_columns_process(clause):
        ret = []
        column_objs = []
        if hasattr(clause, 'left'):
            column_objs.append(clause.left)
        if hasattr(clause, 'right'):
            column_objs.append(clause.right)
        if column_objs:
            for column_obj in column_objs:
                if isinstance(column_obj, Column):
                    where_item = {
                        'name': column_obj.name,
                        'column': column_obj,
                        'table_name': str(column_obj.table.name)
                    }
                    ret.append(where_item)
        else:
            for tbl in clause._from_objects:
                where_item = {
                    'name': None,
                    'column': None,
                    'table_name': str(tbl.name)
                }
                ret.append(where_item)
        return ret

    @staticmethod
    def clause_columns(clauses):
        ret = []
        if not iterable(clauses):
            clauses = [clauses]
        for clause in clauses:
            ret += DBRemoveJoin.clause_columns_process(clause)
        return ret
            
    def table_joins_process_join(self, table):
        if isinstance(table.left, Table):
            self.table_main = table.left
        else:
            self.table_joins_process_join(table.left)

        right_str = str(table.right)
        if not right_str:
            right_str = str(table.right.selectable.name)

        right_el = {
            'name': right_str,
            'table': table.right,
            'onclause': table.onclause.expression,
            'table_need': [],
            'need': False
        }

        clause_columns = self.clause_columns(table.onclause)
        clause_columns_skip_table_name = [right_el['name'], self.table_main]
        for clause_column in clause_columns:
            if clause_column['table_name'] not in clause_columns_skip_table_name:
                right_el['table_need'].append(clause_column['table_name'])

        self.table_joins[right_str] = right_el

    def table_joins_process(self):
        for table in self.q._from_obj:
            if isinstance(table, _ORMJoin):
                self.table_joins_process_join(table)
            else:
                pass

    def table_need_add(self, table_name):
        if table_name in self.table_need:
            return
        self.table_need.append(table_name)
        table_join = self.table_joins.get(table_name)
        if table_join:
            table_join['need'] = True
            for sub_table_name in table_join['table_need']:
                self.table_need_add(sub_table_name)

    def fields_process(self):
        if not self.fields:
            return
        for column in self.q.statement.columns:
            column_obj = list(column.base_columns)[0]
            column_el = {
                'name': column.name,
                'column': column_obj,
                'table_name': str(column_obj.table.name)
            }
            self.column_items.append(column_el)
            self.table_need_add(column_el['table_name'])

    def where_process_item(self, clause):
        if clause is None:
            return
        if iterable(clause):
            for el in clause:
                self.where_process_item(el)
            return
        items = self.clause_columns(clause)
        for item in items:
            self.table_need_add(item['table_name'])

    def where_process(self):
        if not self.where:
            return
        if not iterable(self.q.whereclause):
            clauses = self.q.whereclause
        else:
            clauses = self.q.whereclause.clauses
        self.where_process_item(clauses)

    def query_process(self):
        if not self.table_joins:
            return

        self.q._from_obj = (self.table_main, )
        for table in self.table_joins.values():
            if not table['need']:
                continue
            self.q._from_obj = (outerjoin(self.q._from_obj[0], table['table'], table['onclause']), )

    def process(self):
        self.table_joins_process()
        self.fields_process()
        self.where_process()
        self.query_process()
        return self.q

Keep in mind that is home project code.

Sorry for my english (=

If anyone is having issues of the ORM "dropping" the FROM statement all together. In my case it was solved by referencing a column in the func.count() call. Instead of using a literal like func.count(1) just reference a column (func.count(Table.id)) & the ORM won't delete the from clause in the resulting query.

Nice function! FYI it does not consider the query limit.

So i extended the function for my usage.

count = q.session.execute(count_q).scalar()
return min(q._limit, count) if q._limit else count

THIS METHOD DOES NOT APPLY TO QUERIES WITH LIMIT AND OFFSET

What if I want to know the total count? Normally in an API's pagination, (page, page_size, total_count) is needed.

Say

SELECT  count(*) AS total_count
FROM user
WHERE user.is_deleted = false
 LIMIT 20 OFFSET 5;

If you execute this in a DB, it'll return nothing, which is None in Python. In fact, the count_q cannot include limit and offset:

Although you can work around like this, it still isn't ideal due to overheads.

Solution

To avoid using limit and offset in the query.

Hi,
somebody an idea how that looks like in sqlalchemy 1.4?
I am getting:

sqlalchemy.exc.ArgumentError: Query has only expression-based entities - can't find property named "template". where template is something from my db I guess

⚠️ YOU DO NOT NEED THIS ⚠️

https://dba.stackexchange.com/questions/168022/performance-of-count-in-subquery

⚠️ IF YOUR QUERY IS SLOW, YOU SHOULD REMOVE ORDER BY FIRST ⚠️

count_stmt = select(func.count()).select_from(statement.order_by(None).subquery())
await session.scalar(count_stmt)

This worked perfectly for me after I upgraded from sa 1.3 to 1.4 and a count query went from 80ms to 800ms.

by switching form the built in .count() to the suggested first gist the query went back to 80ms. I believe the problem was .count() loading all columns into python which is not required and very slow for thousands or rows, using with_only_columns and removing the sub query took it back to 80ms. No idea what broke it in 1.4.

    count_q = q.statement.with_only_columns([func.count()]).order_by(None)
    count = q.session.execute(count_q).scalar()

I'm not sure why people are putting warnings on this gist or writing very long replies expecting help. If you are stuck ask on Stackoverflow!

My thanks to @hest

This worked perfectly for me after I upgraded from sa 1.3 to 1.4 and a count query went from 80ms to 800ms.

by switching form the built in .count() to the suggested first gist the query went back to 80ms. I believe the problem was .count() loading all columns into python which is not required and very slow for thousands or rows, using with_only_columns and removing the sub query took it back to 80ms. No idea what broke it in 1.4.

    count_q = q.statement.with_only_columns([func.count()]).order_by(None)
    count = q.session.execute(count_q).scalar()

I'm not sure why people are putting warnings on this gist or writing very long replies expecting help. If you are stuck ask on Stackoverflow!

My thanks to @hest

This thread is generally a kind of misinformation. If you check the stackoverflow link above, these sqls are exactly the same:

-- on postgres
EXPLAIN ANALYZE SELECT COUNT(*) FROM some_big_table WHERE some_col = 'some_val'

EXPLAIN ANALYZE SELECT COUNT(*) FROM ( SELECT col1, col2, col3, col4 FROM some_big_table WHERE some_col = 'some_val' )

If you find your query executing slow, the first is to try removing order by.

If you find your query executing slow, the first is to try removing order by.

I tried this, and the methods outlined in the stackoverflow link, and it did not work. Please don't say I'm providing misinformation by outlining what worked for me it is rude.

@davidjb99

Sorry for letting you misunderstand. I meant the gist (which was posted in 2015) is likely misinformation now, not your comment.

This is the method, I'm using

def get_count(self, model_fields, filter_clause):
        """ Note: filter_clause should not be 'None' or 'Null' for this method to work """
        query = self.session.query().with_entities(*model_fields)
        query = query.filter(filter_clause)
        count_query = query.statement \
                .with_only_columns([func.count()]) \
                .order_by(None)
        result = query.session.execute(count_query).scalar()
        return result