Django ORM and locking table

Question

My problem is as follows:

I have a car dealer A, and a db table named sold_cars. When a car is being sold I create entry in this table.

Table ha

Answer 1

from contextlib import contextmanager
from django.db import transaction
from django.db.transaction import get_connection


@contextmanager
def lock_table(model):
    with transaction.atomic():
        cursor = get_connection().cursor()
        cursor.execute(f'LOCK TABLE {model._meta.db_table}')
        try:
            yield
        finally:
            cursor.close()

This is very similar to @jdepoix solution, but a but more dense.

You can use it like this:

with lock_table(MyModel):
    MyModel.do_something()

Note that this only works with PostgreSQL and uses python 3.6's f-strings a.k.a. literal string interpolation.

Answer 2

I think this code snippet meets your need, assuming you are using MySQL. If not, you may need to tweak the syntax a little, but the idea should still work.

Source: Locking tables

class LockingManager(models.Manager):
    """ Add lock/unlock functionality to manager.

    Example::

        class Job(models.Model):

            manager = LockingManager()

            counter = models.IntegerField(null=True, default=0)

            @staticmethod
            def do_atomic_update(job_id)
                ''' Updates job integer, keeping it below 5 '''
                try:
                    # Ensure only one HTTP request can do this update at once.
                    Job.objects.lock()

                    job = Job.object.get(id=job_id)
                    # If we don't lock the tables two simultanous
                    # requests might both increase the counter
                    # going over 5
                    if job.counter < 5:
                        job.counter += 1                                        
                        job.save()

                finally:
                    Job.objects.unlock()


    """    

    def lock(self):
        """ Lock table. 

        Locks the object model table so that atomic update is possible.
        Simulatenous database access request pend until the lock is unlock()'ed.

        Note: If you need to lock multiple tables, you need to do lock them
        all in one SQL clause and this function is not enough. To avoid
        dead lock, all tables must be locked in the same order.

        See http://dev.mysql.com/doc/refman/5.0/en/lock-tables.html
        """
        cursor = connection.cursor()
        table = self.model._meta.db_table
        logger.debug("Locking table %s" % table)
        cursor.execute("LOCK TABLES %s WRITE" % table)
        row = cursor.fetchone()
        return row

    def unlock(self):
        """ Unlock the table. """
        cursor = connection.cursor()
        table = self.model._meta.db_table
        cursor.execute("UNLOCK TABLES")
        row = cursor.fetchone()
        return row  

Answer 3

I would recommend using the F() expression instead of locking the entire table. If your app is being heavily used, locking the table will have significant performance impact.

The exact scenario you described is mentioned in Django documentation here. Based on your scenario, here's the code you can use:

from django.db.models import F


# Populate sold_cars as you normally do..
# Before saving, use the "F" expression
sold_cars.order_num =F('order_num') + 1
sold_cars.save()

# You must do this before referring to order_num:
sold_cars.refresh_from_db()
# Now you have the database-assigned order number in sold_cars.order_num

Note that if you set order_num during an update operation, use the following instead:

sold_cars.update(order_num=F('order_num')+1)
sold_cars.refresh_from_db()

Since database is in charge of updating the field, there won't be any race conditions or duplicated order_num values. Plus, this approach is much faster than one with locked tables.

Answer 4

I know this question is a bit older, but I just had the same issue and wanted to share my learnings.

I wasn't quite satisfied with st0nes answer, since (at least for postgres) a LOCK TABLE statement can only be issued within a transaction. And although in Django usually almost everything happens within a transaction, this LockingManager does not make sure, that you actually are within a transaction, at least to my understanding. Also I didn't want to completely change the Models Manager just to be able to lock it at one spot and therefore I was more looking for something that works kinda like the with transaction.atomic():, but also locks a given Model.

So I came up with this:

from django.conf import settings
from django.db import DEFAULT_DB_ALIAS
from django.db.transaction import Atomic, get_connection


class LockedAtomicTransaction(Atomic):
    """
    Does a atomic transaction, but also locks the entire table for any transactions, for the duration of this
    transaction. Although this is the only way to avoid concurrency issues in certain situations, it should be used with
    caution, since it has impacts on performance, for obvious reasons...
    """
    def __init__(self, model, using=None, savepoint=None):
        if using is None:
            using = DEFAULT_DB_ALIAS
        super().__init__(using, savepoint)
        self.model = model

    def __enter__(self):
        super(LockedAtomicTransaction, self).__enter__()

        # Make sure not to lock, when sqlite is used, or you'll run into problems while running tests!!!
        if settings.DATABASES[self.using]['ENGINE'] != 'django.db.backends.sqlite3':
            cursor = None
            try:
                cursor = get_connection(self.using).cursor()
                cursor.execute(
                    'LOCK TABLE {db_table_name}'.format(db_table_name=self.model._meta.db_table)
                )
            finally:
                if cursor and not cursor.closed:
                    cursor.close()

So if I now want to lock the model ModelToLock, this can be used like this:

with LockedAtomicTransaction(ModelToLock):
    # do whatever you want to do
    ModelToLock.objects.create()

EDIT: Note that I have only tested this using postgres. But to my understanding, it should also work on mysql just like that.