django-pg-bulk-update

Django extension to update multiple table records with similar (but not equal) conditions in efficient way on PostgreSQL

Requirements

• Python 3.5+ Previous versions may also work, but are not tested in CI
• django >= 1.8
  Previous versions may also work, but are not tested in CI.
  django.postgres.contrib fields are also supported (available since django 1.8) django.postgres.contrib.JSONField is supported since django 1.9
• pytz for python before 3.3
• typing for python before 3.5
• psycopg2-binary
• PostgreSQL 9.4+
  Previous versions may also work, but haven't been tested.
  JSONB operations are available for PostgreSQL 9.4+. INSERT .. ON CONFLICT is used for PostgreSQL 9.5+.

Installation

Install via pip:
pip install django-pg-bulk-update
or via setup.py:
python setup.py install

Usage

You can make queries in 2 ways:

• Declaring a custom manager for your model
• Calling query functions directly

Query functions

There are 4 query helpers in this library. There parameters are unified and described in the section below.

• bulk_update(model, values, key_fields='id', using=None, set_functions=None, key_fields_ops=(), where=None, returning=None, batch_size=None, batch_delay=0)
  This function updates multiple records of given model in single database query.
  Functions forms raw sql query for PostgreSQL. It's work is not guaranteed on other databases.
  Function returns number of updated records.

• bulk_update_or_create(model, values, key_fields='id', using=None, set_functions=None, update=True, key_is_unique=True, returning=None, batch_size=None, batch_delay=0)
  This function finds records by key_fields. It creates not existing records with data, given in values.
  If update flag is set, it updates existing records with data, given in values.

  There are two ways, this function may work:

  1. Use INSERT ... ON CONFLICT statement. It is safe, but requires PostgreSQL 9.5+ and unique index on key fields. This behavior is used by default.
  2. 3-query transaction:
     + Search for existing records
     + Create not existing records (if values have any)
     + Update existing records (if values have any and update flag is set)
     This behavior is used by default on PostgreSQL before 9.5 and if key_is_unique parameter is set to False. Note that transactional update has a known race condition issue that can't be fixed.

  Function returns number of records inserted or updated by query.

• bulk_create(model, values, using=None, set_functions=None, returning=None, batch_size=None, batch_delay=0)
  This function creates multiple records of given model in single database query.
  It's functionality is the same as django's QuerySet.bulk_create, but it is implemented on this library bases and can be more effective in some cases (for instance, for wide models).

• pdnf_clause(key_fields, field_values, key_fields_ops=())
  Pure django implementation of principal disjunctive normal form. It is base on combining Q() objects.
  Condition will look like:

    SELECT ... WHERE (a = x AND b = y AND ...) OR (a = x1 AND b = y1  AND ...) OR ...

  Function returns a django.db.models.Q instance

Function parameters

• model: Type[Model] A subclass of django.db.models.Model to update

• values: Union[Union[TUpdateValuesValid, Dict[Any, Dict[str, Any]]], Iterable[Dict[str, Any]]]
  Data to update. All items must update same fields!!!
  Parameter can have one of 2 forms:

  • Iterable of dicts. Each dict contains both key and update data. Each dict must contain all key_fields as keys. You can't update key_fields with this format.
  • Dict of key_values: update_fields_dict
    Can not be used for bulk_create function. You can use this format to update key_fields
    • key_values can be tuple or single object. If tuple, key_values length must be equal to key_fields length. If single object, key_fields is expected to have 1 element
    • update_fields_dict is a dictionary {field_name: update_value} to update

• key_fields: Union[str, Iterable[str]] Optional. Field names, which are used as update conditions. Parameter can have one of 2 forms:

  • String for single key field. Primary key is used by default.
  • Iterable of strings for multiple key fields.

• using: Optional[str]
  Optional. Database alias to query. If not set, 'default' database is used.

• set_functions: Optional[Dict[str, Union[str, AbstractSetFunction]]]
  Optional. Functions which will be used to set values.
  If given, it should be a dictionary:

  • Key is a field name, function is applied to

  • Value is a function alias name or AbstractSetFunction instance.
    Available function aliases:

    • 'eq', '='
      Simple assign operator. It used by default for fields that are not mentioned in the dict.
    • 'incr', '+'
      Adds field value to previous one. It can be used for all numeric database types.
    • 'concat', '||'
      Concatenates field value to previous one. It can be used for string types, JSONField, HStoreField, ArrayField.
    • 'eq_not_null'
      This function can be used, if you want to update value only if it is not None.
    • 'union'
      This function combines ArrayField value with previous one, removing duplicates.
    • 'array_remove'
      This function deletes value from ArrayField field using array_remove PSQL Function.
    • 'now', 'NOW'
      This function sets field value to NOW() database function. Doesn't require any value in values parameter.
    • You can define your own set function. See section below.

    Increment, union and concatenate functions concern NULL as default value. You can see default values in sections below.

• key_field_ops: Union[Dict[str, Union[str, AbstractClauseOperator]], Iterable[Union[str, AbstractClauseOperator]]] Optional. Operators, which are used to fined records for update. Operators are applied to key_fields.
  If some fields are not given, equality operator is used. bulk_update_or_create function always uses equality operator Parameter can have one of 2 forms:

  • Iterable of operator alias names or AbstractClauseOperator instances. Order of iterable must be the same as key_fields.
  • Dictionary:
    • Key is a field name, function is applied to
    • Value is a function alias name of set function or AbstractSetFunction instance.
      Available name aliases:
  • 'eq', '=', '==' Simple equality condition. It is used by default.
  • '!eq', '!=', '<>' Not equal operator
  • 'in' Searches for records, which have field from values list. Value should be an iterable of correct field values.
  • '!in' Searches for records, which have field not from values list. Value should be an iterable of correct field values.
  • 'lt', '<'
  • 'lte', '<='
  • 'gt', '>'
  • 'gte', '>='
  • 'between' Searches for records, which have field between a and b. Value should be iterable with 2 items.
  • 'is_null', 'isnull' Checks field value for been NULL. Value should be boolean (true for IS NULL, false for IS NOT NULL)
  • You can define your own clause operator. See section below.

• where: Optional[WhereNode]
  This parameter is used to filter data before doing bulk update, using QuerySet filter and exclude methods. Generated condition should not contain annotations and other table references.
  NOTE: parameter is not supported in bulk_update_or_create

• returning: Optional[Union[str, Iterable[str]]]
  If this parameter is set, it can be:

  1. A field name string
  2. An iterable of field names
  3. '*' string to return all model fields

  Query returns django_pg_returning.ReturningQuerySet instead of rows count.
  Using this feature requires django-pg-returning library installed (it is not in requirements, though).

• batch_size: Optional[int]
  If this parameter is set, values are split into batches of given size. Each batch is processed separately. Note that batch_size != number of records processed if you use key_field_ops other than 'eq'

• batch_delay: float
  If batch_size is set, this parameter sets time to sleep in seconds between batches execution

• update: bool
  If flag is not set, bulk_update_or_create function will not update existing records, only creating not existing.

• key_is_unique: bool Defaults to True. Settings this flag to False forces library to use 3-query transactional update_or_create.

• field_values: Iterable[Union[Iterable[Any], dict]]
  Field values to use in pdnf_clause function. They have simpler format than update functions. It can come in 2 formats:

  • An iterable of tuples in key_fields order ( (x, y), (x1, y1), ...)
  • An iterable of dicts with field name as key ({'a': x, 'b': y}, ...)

Examples

from django.db import models
from django_pg_bulk_update import bulk_update, bulk_update_or_create, pdnf_clause

# Test model
class TestModel(models.Model):
    name = models.CharField(max_length=50)
    int_field = models.IntegerField()

# Create test data
created = TestModel.objects.pg_bulk_create([
    {'id': i, 'name': "item%d" % i, 'int_field': 1} for i in range(1, 4)
])
print(created)
# Outputs 3

# Create test data returning
created = TestModel.objects.pg_bulk_create([
    {'id': i, 'name': "item%d" % i, 'int_field': 1} for i in range(4, 6)
], returning='*')
print(created)
print(type(res), list(res.values_list('id', 'name', 'int_field')))
# Outputs: 
# <class 'django_pg_returning.queryset.ReturningQuerySet'>
# [
#    (4, "item4", 1),
#    (5, "item5", 1)
# ]

# Update by id field
updated = bulk_update(TestModel, [{
    "id": 1,
    "name": "updated1",
}, {
    "id": 2,
    "name": "updated2"
}])

print(updated)
# Outputs: 2

# Update returning
res = bulk_update(TestModel, [{
    "id": 1,
    "name": "updated1",
}, {
    "id": 2,
    "name": "updated2"
}], returning=('id', 'name', 'int_field'))

print(type(res), list(res.values_list('id', 'name', 'int_field')))
# Outputs: 
# <class 'django_pg_returning.queryset.ReturningQuerySet'>
# [
#    (1, "updated1", 1),
#    (2, "updated2", 1)
# ]

# Call update by name field
updated = bulk_update(TestModel, {
    "updated1": {
        "int_field": 2
    },
    "updated2": {
        "int_field": 3
    }
}, key_fields="name")

print(updated)
# Outputs: 2

print(list(TestModel.objects.all().order_by("id").values("id", "name", "int_field")))
# Outputs: [
#     {"id": 1, "name": "updated1", "int_field": 2},
#     {"id": 2, "name": "updated2", "int_field": 3},
#     {"id": 3, "name": "item3", "int_field": 1}
# ]

# Increment int_field by 3 and set name to 'incr' for records where id >= 2 and int_field < 3
updated = bulk_update(TestModel, {
    (2, 3): {
        "int_field": 3,
        "name": "incr"
    }
}, key_fields=['id', 'int_field'], key_fields_ops={'int_field': '<', 'id': 'gte'}, set_functions={'int_field': '+'})

print(updated)
# Outputs: 1

print(list(TestModel.objects.all().order_by("id").values("id", "name", "int_field")))
# Outputs: [
#     {"id": 1, "name": "updated1", "int_field": 2},
#     {"id": 2, "name": "updated2", "int_field": 3},
#     {"id": 3, "name": "incr", "int_field": 4}
# ]
 
 
res = bulk_update_or_create(TestModel, [{
    "id": 3,
    "name": "_concat1",
    "int_field": 4
}, {
    "id": 4,
    "name": "concat2",
    "int_field": 5
}], set_functions={'name': '||'})

print(res)
# Outputs: 2

print(list(TestModel.objects.all().order_by("id").values("id", "name", "int_field")))
# Outputs: [
#     {"id": 1, "name": "updated1", "int_field": 2},
#     {"id": 2, "name": "updated2", "int_field": 3},
#     {"id": 3, "name": "incr_concat1", "int_field": 4},
#     {"id": 4, "name": "concat2", "int_field": 5},
# ]

# Find records where 
# id IN [1, 2, 3] AND name = 'updated2' OR id IN [3, 4, 5] AND name = 'concat2' OR id IN [2, 3, 4] AND name = 'updated1'
        cond = pdnf_clause(['id', 'name'], [([1, 2, 3], 'updated2'),
                                            ([3, 4, 5], 'concat2'),
                                            ([2, 3, 4], 'updated1')], key_fields_ops={'id': 'in'})
data = TestModel.objects.filter(cond).order_by('int_field').values_list('int_field', flat=True)
print(list(data))
# Outputs: [3, 5]

Using custom manager and query set

In order to simplify using bulk_create, bulk_update and bulk_update_or_create functions, you can use a custom manager.
It automatically fills:

• model parameter
• using parameter (extracts queryset write database)
• where parameter (applies queryset filters, if called as QuerySet method). Not supported in bulk_update_or_create.
  You can change database to use with Manager.db_manager() or QuerySet.using() methods.
  The rest parameters are the same as above.

Note: As django 2.2 introduced bulk_update method, library methods were renamed to pg_bulk_create, pg_bulk_update and pg_bulk_update_or_create respectively.

Example:

from django.db import models
from django_pg_bulk_update.manager import BulkUpdateManager

# Test model
class TestModel(models.Model):
    objects = BulkUpdateManager()
    
    name = models.CharField(max_length=50)
    int_field = models.IntegerField()
    
# Now you can use functions like:
TestModel.objects.pg_bulk_create([
    # Any data here
], set_functions=None)

TestModel.objects.pg_bulk_update([
    # Any data here
], key_fields='id', set_functions=None, key_fields_ops=())

# Update only records with id gtreater than 5 
TestModel.objects.filter(id__gte=5).pg_bulk_update([
    # Any data here
], key_fields='id', set_functions=None, key_fields_ops=())

TestModel.objects.pg_bulk_update_or_create([
    # Any data here
], key_fields='id', set_functions=None, update=True)           

If you already have a custom manager, you can replace QuerySet to BulkUpdateQuerySet:

from django.db import models
from django.db.models.manager import BaseManager
from django_pg_bulk_update.manager import BulkUpdateQuerySet


class CustomManager(BaseManager.from_queryset(BulkUpdateQuerySet)):
    pass
    
    
# Test model
class TestModel(models.Model):
    objects = CustomManager()
    
    name = models.CharField(max_length=50)
    int_field = models.IntegerField()

If you already have a custom QuerySet, you can inherit it from BulkUpdateMixin:

from django.db import models
from django.db.models.manager import BaseManager
from django_pg_bulk_update.manager import BulkUpdateMixin


class CustomQuerySet(BulkUpdateMixin, models.QuerySet):
    pass
    
    
class CustomManager(BaseManager.from_queryset(CustomQuerySet)):
    pass
    
    
# Test model
class TestModel(models.Model):
    objects = CustomManager()
    
    name = models.CharField(max_length=50)
    int_field = models.IntegerField()

Custom clause operator

You can define your own clause operator, creating AbstractClauseOperator subclass and implementing:

• names attribute
• def get_django_filter(self, name) method
• One of def get_sql_operator(self) or def get_sql(self, table_field, value) When clause is formed, it calls get_sql() method. In order to simplify method usage of simple field <op> value operators, by default get_sql() forms this condition, calling get_sql_operator() method, which returns .

Optionally, you can change def format_field_value(self, field, val, connection, cast_type=True, **kwargs) method, which formats value according to field rules

Example:

from django_pg_bulk_update import bulk_update
from django_pg_bulk_update.clause_operators import AbstractClauseOperator

class LTClauseOperator(AbstractClauseOperator):
    names = {'lt', '<'}

    def get_django_filter(self, name):  # type: (str) -> str
        """
        This method should return parameter name to use in django QuerySet.fillter() kwargs
        :param name: Name of parameter
        :return: String with filter
        """
        return '%s__lt' % name

    def get_sql_operator(self):  # type: () -> str
        """
        If get_sql operator is simple binary operator like "field <op> val", this functions returns operator
        :return: str
        """
        return '<'
        

# Usage examples
# import you function here before calling an update
bulk_update(TestModel, [], key_field_ops={'int_field': 'lt'})
bulk_update(TestModel, [], key_field_ops={'int_field': LTClauseOperator()})

You can use class instance directly in key_field_ops parameter or use its aliases from names attribute.
When update function is called, it searches for all imported AbstractClauseOperator subclasses and takes first class which contains alias in names attribute.

Custom set function

You can define your own set function, creating AbstractSetFunction subclass and implementing:

• names attribute
• supported_field_classes attribute
• One of:
  • def get_sql_value(self, field, val, connection, val_as_param=True, with_table=False, for_update=True, **kwargs) method This method defines new value to set for parameter. It is called from get_sql(...) method by default.
  • def get_sql(self, field, val, connection, val_as_param=True, with_table=False, for_update=True, **kwargs) method This method sets full sql and it params to use in set section of update query.
    By default it returns: "%s" = self.get_sql_value(...), params

Optionally, you can change:

• def format_field_value(self, field, val, connection, cast_type=False, **kwargs) method, if input data needs special formatting.
• def modify_create_params(self, model, key, kwargs) method, to change data before passing them to model constructor in bulk_update_or_create(). This method is used in 3-query transactional update only. INSERT ... ON CONFLICT uses for_update flag of get_sql() and get_sql_value() functions

Example:

from django_pg_bulk_update import bulk_update
from django_pg_bulk_update.set_functions import AbstractSetFunction

class CustomSetFunction(AbstractSetFunction):
    # Set function alias names
    names = {'func_alias_name'}

    # Names of django field classes, this function supports. You can set None (default) to support any field.
    supported_field_classes = {'IntegerField', 'FloatField', 'AutoField', 'BigAutoField'}

    def get_sql_value(self, field, val, connection, val_as_param=True, with_table=False, for_update=True, **kwargs):
        """
        Returns value sql to set into field and parameters for query execution
        This method is called from get_sql() by default.
        :param field: Django field to take format from
        :param val: Value to format
        :param connection: Connection used to update data
        :param val_as_param: If flag is not set, value should be converted to string and inserted into query directly.
            Otherwise a placeholder and query parameter will be used
        :param with_table: If flag is set, column name in sql is prefixed by table name
        :param for_update: If flag is set, returns update sql. Otherwise - insert SQL
        :param kwargs: Additional arguments, if needed
        :return: A tuple: sql, replacing value in update and a tuple of parameters to pass to cursor
        """
        # If operation is incremental, it should be ready to get NULL in database
        null_default, null_default_params = self._parse_null_default(field, connection, **kwargs)
        
        # Your function/operator should be defined here
        tpl = 'COALESCE("%s", %s) + %s'

        if val_as_param:
            sql, params = self.format_field_value(field, val, connection)
            return tpl % (field.column, null_default, sql), null_default_params + params
        else:
            return tpl % (field.column, null_default, str(val)), null_default_params
            
            
# Usage examples
# import you function here before calling an update
bulk_update(TestModel, [], set_functions={'int_field': 'func_alias_name'})
bulk_update(TestModel, [], set_functions={'int_field': CustomSetFunction()})

You can use class instance directly in set_functions parameter or use its aliases from names attribute.
When update function is called, it searches for all imported AbstractSetFunction subclasses and takes first class which contains alias in names attribute.

Compatibility

Library supports django.contrib.postgres.fields:

• ArrayField
• JSONField
• HStoreField
• RangeField (IntegerRangeField, BigIntegerRangeField, FloatRangeField, DateTimeRangeField, DateRangeField)

Note that ArrayField and HStoreField are available since django 1.8, JSONField - since django 1.9.
RangeField supports are available since PostgreSQL 9.2, psycopg2 since 2.5 and django since 1.8.
PostgreSQL before 9.4 doesn't support jsonb, and so - JSONField.
PostgreSQL 9.4 supports JSONB, but doesn't support concatenation operator (||). In order to support this set function a special function for postgres 9.4 was written. Add a migration to create it:

from django.db import migrations,
from django_pg_bulk_update.compatibility import Postgres94MergeJSONBMigration

class Migration(migrations.Migration):
    dependencies = []

    operations = [
        Postgres94MergeJSONBMigration()
    ]

PostgreSQL before 9.5 doesn't support INSERT ... ON CONFLICT statement. So 3-query transactional update will be used.

Performance

Test background:

• Django 2.0.2
• PostgreSQL 10.2
• Python 3.6.3
• 1000 pre-created records
  Updating records one by one took 51,68 seconds.
  Updating records with bulk_update took 0.13 seconds.
  You can write your own tests, based on test.test_performance and running it.

Development

This is an Open source project developed by M1ha-Shvn under BSD 3 license. Feel free to create issues and make pull requests.
Library test system is based on django.test. You can find them in tests directory.

Tests requirements

• PostgreSQL
• Pypi libraries listed in requirements-test.txt file

Running tests

Running in docker

1. Install docker and docker-compose
2. Run docker-compose run run_tests in project directory

Running in virtual environment

1. Install all requirements listed above
2. Create virtual environment
3. Create a superuser named 'test' on your local Postgres instance:

CREATE ROLE test;
ALTER ROLE test WITH SUPERUSER;
ALTER ROLE test WITH LOGIN;
ALTER ROLE test PASSWORD 'test';
CREATE DATABASE test OWNER test;
CREATE DATABASE test2 OWNER test;

3. Install requirements
   pip3 install -U -r requirements-test.txt
4. Start tests
   python3 runtests.py

Alternatives

django 2.2+ bulk_update difference

Pros:

• bulk_update_or_create() method
• Ability to use complex set functions
• Ability to use complex conditions
• Ability to update primary key
• pdnf_clause helper
• Django 1.7+ support
• Ability to make delay between batches
• Ability to return affected rows instead of rowcount (using Postgres RETURNING feature)

Cons:

• PostgreSQL only
• Django method supports Func objects (in library's backlog)
• Ability to update parents/children (using extra queries)

django-bulk-update difference

Pros:

• bulk_update_or_create() method
• Ability to use complex set functions
• Ability to use complex conditions
• pdnf_clause helper
• Django 1.7 support
• Ability to make delay between batches
• Ability to return affected rows instead of rowcount (using Postgres RETURNING feature)

Cons:

• PostgreSQL only