API Races - Conditional Updates

Background

On Cinder API nodes we have to check that requested action can be performed by checking request arguments and involved resources, and only if everything matches required criteria we will proceed with the RPC call to any of the other nodes.

Checking the conditions must be done in a non racy way to ensure that already checked requirements don’t change while we check remaining conditions. This is of utter importance, as Cinder uses resource status as a lock to prevent concurrent operations on a resource.

An simple example of this would be extending a volume, where we first check the status:

if volume['status'] != 'available':

Then update the status:

self.update(context, volume, {'status': 'extending'})

And finally make the RPC call:

self.volume_rpcapi.extend_volume(context, volume, new_size,
                                 reservations)

The problem is that that this code would allow races, as other request could have already changed the volume status between us getting the value and updating the DB.

There are multiple ways to fix this, such as:

  • Using a Distributed Locking Mechanism
  • Using DB isolation level
  • Using SQL SELECT ... FOR UPDATE
  • USING compare and swap mechanism in SQL query

Our tests showed that the best alternative was compare and swap and we decided to call this mechanism “Conditional Update” as it seemed more appropriate.

Conditional Update

Conditional Update is the mechanism we use in Cinder to prevent races when updating the DB. In essence it is the SQL equivalent of an UPDATE ... FROM ... WHERE; clause

It is implemented as an abstraction layer on top of SQLAlchemy ORM engine in our DB api layer and exposed for consumption in Cinder’s Persistent Versioned Objects through the conditional_update method so it can be used from any Vesioned Object instance that has persistence (Volume, Snapshot, Backup...).

Method signature is:

def conditional_update(self, values, expected_values=None, filters=(),
                       save_all=False, session=None, reflect_changes=True):
values:

Dictionary of key-value pairs with changes that we want to make to the resource in the DB.

expected_values:
 

Dictionary with conditions that must be met for the update to be executed.

Condition field.id == resource.id is implicit and there is no need to add it to the conditions.

If no expected_values argument is provided update will only go through if no field in the DB has changed. Dirty fields from the Versioned Object are excluded as we don’t know their original value.

filters:

Additional SQLAlchemy filters can be provided for more complex conditions.

save_all:

By default we will only be updating the DB with values provided in the values argument, but we can explicitly say that we also want to save object’s current dirty fields.

session:

A SQLAlchemy session can be provided, although it is unlikely to be needed.

reflect_changes:
 

On a successful update we will also update Versioned Object instance to reflect these changes, but we can prevent this instance update passing False on this argument.

Return Value:

We’ll return the number of changed rows. So we’ll get a 0 value if the conditional update has not been successful instead of an exception.

Basic Usage

  • Simple match

    The most basic example is doing a simple match, for example for a volume variable that contains a Versioned Object Volume class instance we may want to change the status to “deleting” and update the terminated_at field with current UTC time only if current status is “available” and the volume is not in a consistency group.

    values={'status': 'deleting',
            'terminated_at': timeutils.utcnow()}
    expected_values = {'status': 'available',
                       'consistencygroup_id': None}
    
    volume.conditional_update(values, expected_values)
    
  • Iterable match

    Conditions can contain not only single values, but also iterables, and the conditional update mechanism will correctly handle the presence of None values in the range, unlike SQL IN clause that doesn’t support NULL values.

    values={'status': 'deleting',
            'terminated_at': timeutils.utcnow()}
    expected_values={
        'status': ('available', 'error', 'error_restoring' 'error_extending'),
        'migration_status': (None, 'deleting', 'error', 'success'),
        'consistencygroup_id': None
    }
    
    volume.conditional_update(values, expected_values)
    
  • Exclusion

    In some cases we’ll need to set conditions on what is not in the DB record instead of what is is, for that we will use the exclusion mechanism provided by the Not class in all persistent objects. This class accepts single values as well as iterables.

    values={'status': 'deleting',
            'terminated_at': timeutils.utcnow()}
    expected_values={
        'attach_status': volume.Not('attached'),
        'status': ('available', 'error', 'error_restoring' 'error_extending'),
        'migration_status': (None, 'deleting', 'error', 'success'),
        'consistencygroup_id': None
    }
    
    volume.conditional_update(values, expected_values)
    
  • Filters

    We can use complex filters in the conditions, but these must be SQLAlchemy queries/conditions and as the rest of the DB methods must be properly abstracted from the API.

    Therefore we will create the medhot in cinder/db/sqlalchemy/api.py:

    def volume_has_snapshots_filter():
        return sql.exists().where(
            and_(models.Volume.id == models.Snapshot.volume_id,
                 ~models.Snapshot.deleted))
    

    Then expose this filter through the cinder/db/api.py:

    def volume_has_snapshots_filter():
        return IMPL.volume_has_snapshots_filter()
    

    And finally used in the API (notice how we are negating the filter at the API):

    filters = [~db.volume_has_snapshots_filter()]
    values={'status': 'deleting',
            'terminated_at': timeutils.utcnow()}
    expected_values={
        'attach_status': volume.Not('attached'),
        'status': ('available', 'error', 'error_restoring' 'error_extending'),
        'migration_status': (None, 'deleting', 'error', 'success'),
        'consistencygroup_id': None
    }
    
    volume.conditional_update(values, expected_values, filters)
    

Returning Errors

The most important downside of using conditional updates to remove API races is the inherent uncertainty of the cause of failure resulting in more generic error messages.

When we use the conditional_update method we’ll use returned value to determine the success of the operation, as a value of 0 indicates that no rows have been updated and the conditions were not met. But we don’t know which one, or which ones, were the cause of the failure.

There are 2 approaches to this issue:

  • On failure we go one by one checking the conditions and return the first one that fails.
  • We return a generic error message indicating all conditions that must be met for the operation to succeed.

It was decided that we would go with the second approach, because even though the first approach was closer to what we already had and would give a better user experience, it had considerable implications such as:

  • More code was needed to do individual checks making operations considerable longer and less readable. This was greatly alleviated using helper methods to return the errors.
  • Higher number of DB queries required to determine failure cause.
  • Since there could be races because DB contents could be changed between the failed update and the follow up queries that checked the values for the specific error, a loop would be needed to make sure that either the conditional update succeeds or one of the condition checks fails.
  • Having such a loop means that a small error in the code could lead to an endless loop in a production environment. This coding error could be an incorrect conditional update filter that would always fail or a missing or incorrect condition that checked for the specific issue to return the error.

A simple example of a generic error can be found in begin_detaching code:

@wrap_check_policy
def begin_detaching(self, context, volume):
    # If we are in the middle of a volume migration, we don't want the
    # user to see that the volume is 'detaching'. Having
    # 'migration_status' set will have the same effect internally.
    expected = {'status': 'in-use',
                'attach_status': 'attached',
                'migration_status': self.AVAILABLE_MIGRATION_STATUS}

    result = volume.conditional_update({'status': 'detaching'}, expected)

    if not (result or self._is_volume_migrating(volume)):
        msg = _("Unable to detach volume. Volume status must be 'in-use' "
                "and attach_status must be 'attached' to detach.")
        LOG.error(msg)
        raise exception.InvalidVolume(reason=msg)

Building filters on the API

SQLAlchemy filters created as mentioned above can create very powerful and complex conditions, but sometimes we may require a condition that, while more complex than the basic match and not match on the resource fields, it’s still quite simple. For those cases we can create filters directly on the API using the model field provided in Versioned Objects.

This model field is a reference to the ORM model that allows us to reference ORM fields.

We’ll use as an example changing the status field of a backup to “restoring” if the backup status is “available” and the volume where we are going to restore the backup is also in “available” state.

Joining of tables is implicit when using a model different from the one used for the Versioned Object instance.

  • As expected_values

    Since this is a matching case we can use expected_values argument to make the condition:

    values = {'status': 'restoring'}
    expected_values={'status': 'available',
                     objects.Volume.model.id: volume.id,
                     objects.Volume.model.status: 'available'}
    
  • As filters

    We can also use the filters argument to achieve the same results:

    filters = [objects.Volume.model.id == volume.id,
               objects.Volume.model.status == 'available']
    
  • Other filters

    If we are not doing a match for the condition the only available option will be to use filters argument. For example if we want to do a check on the volume size against the backup size:

    filters = [objects.Volume.model.id == volume.id,
               objects.Volume.model.size >= backup.model.size]
    

Using DB fields for assignment

  • Using non modified fields

    Similar to the way we use the fields to specify conditions, we can also use them to set values in the DB.

    For example when we disable a service we want to keep existing updated_at field value:

    values = {'disabled': True,
              'updated_at': service.model.updated_at}
    
  • Using modified field

    In some cases we may need to use a DB field that we are also updating, for example when we are updating the status but we also want to keep the old value in the previous_status field.

    values = {'status': 'retyping',
              'previous_status': volume.model.status}
    

    Conditional update mechanism takes into account that MySQL does not follow SQL language specs and adjusts the query creation accordingly.

  • Together with filters

    Using DB fields for assignment together with using them for values can give us advanced functionality like for example increasing a quota value based on current value and making sure we don’t exceed our quota limits.

    values = {'in_use': quota.model.in_use + volume.size}
    filters = [quota.model.in_use <= max_usage - volume.size]
    

Conditional value setting

Under certain circumstances you may not know what value should be set in the DB because it depends on another field or on another condition. For those cases we can use the Case class present in our persistent Versioned Objects which implements the SQL CASE clause.

The idea is simple, using Case class we can say which values to set in a field based on conditions and also set a default value if none of the conditions are True.

Conditions must be SQLAlchemy conditions, so we’ll need to use fields from the
model attribute.

For example setting the status to “maintenance” during migration if current status is “available” and leaving it as it was if it’s not can be done using the following:

values = {
    'status': volume.Case(
        [
            (volume.model.status == 'available', 'maintenance')
        ],
        else_=volume.model.status)
}

reflect_changes considerations

As we’ve already mentioned conditional_update method will update Versioned Object instance with provided values if the row in the DB has been updated, and in most cases this is OK since we can set the values directly because we are using simple values, but there are cases where we don’t know what value we should set in the instance, and is in those cases where the default reflect_changes value of True has performance implications.

There are 2 cases where Versioned Object conditional_update method doesn’t know the value it has to set on the Versioned Object instance, and they are when we use a field for assignment and when we are using the Case class, since in both cases the DB is the one deciding the value that will be set.

In those cases conditional_update will have to retrieve the value from the DB using get_by_id method, and this has a performance impact and therefore should be avoided when possible.

So the recommendation is to set reflect_changes to False when using Case class or using fields in the values argument if we don’t care about the stored value.

Limitations

We can only use functionality that works on all supported DBs, and that’s why we don’t allow multi table updates and will raise ProgrammingError exception even when the code is running against a DB engine that supports this functionality.

This way we make sure that we don’t inadvertently add a multi table update that works on MySQL but will surely fail on PostgreSQL.

MySQL DB engine also has some limitations that we should be aware of when creating our filters.

One that is very common is when we are trying to check if there is a row that matches a specific criteria in the same table that we are updating. For example, when deleting a Consistency Group we want to check that it is not being used as the source for a Consistency Group that is in the process of being created.

The straightforward way of doing this is using the core exists expression and use an alias to differentiate general query fields and the exists subquery. Code would look like this:

def cg_creating_from_src(cg_id):
   model = aliased(models.ConsistencyGroup)
   return sql.exists().where(and_(
       ~model.deleted,
       model.status == 'creating',
       conditions.append(model.source_cgid == cg_id)))

While this will work in SQLite and PostgreSQL, it will not work on MySQL and an error will be raised when the query is executed: “You can’t specify target table ‘consistencygroups’ for update in FROM clause”.

To solve this we have 2 options:

  • Create a specific query for MySQL engines using an update with a left self join, which is a feature only available in MySQL.
  • Use a trick -using a select subquery- that will work on all DBs.

Considering that it’s always better to have only 1 way of doing things and that SQLAlchemy doesn’t support MySQL’s non standard behavior we should generate these filters using the select subquery method like this:

def cg_creating_from_src(cg_id):
   subq = sql.select([models.ConsistencyGroup]).where(and_(
       ~model.deleted,
       model.status == 'creating')).alias('cg2')

   return sql.exists([subq]).where(subq.c.source_cgid == cgid)

Considerations for new ORM & Versioned Objects

Conditional update mechanism works using generic methods for getting an object from the DB as well as determining the model for a specific Versioned Object instance for field binding.

These generic methods rely on some naming rules for Versioned Object classes, ORM classes, and get methods, so when we are creating a new ORM class and adding the matching Versioned Object and access methods we must be careful to follow these rules or at least specify exceptions if we have a good reason not to follow these conventions.

Rules:

  • Versioned Object class name must be the same as the ORM class
  • Get method name must be ORM class converted to snake format with postfix “_get”. For example, for Volume ORM class expected method is volume_get, and for an imaginary MyORMClass it would be my_orm_class_get.
  • Get method must receive the context as the first argument and the id as the second one, although it may accept more optional arguments.

We should avoid diverging from these rules whenever is possible, but there are cases where this is not possible, for example BackupImport Versioned Object that really uses Backup ORM class. For cases such as this we have a way to set exceptions both for the generic get method and the model for a Versioned Object.

To add exceptions for the get method we have to add a new entry to GET_EXCEPTIONS dictionary mapping in cinder.db.sqlalchemy.api._get_get_method.

And for determining the model for the Versioned Object we have to add a new entry to VO_TO_MODEL_EXCEPTIONS dictionary mapping in cinder.db.sqlalchemy.api.get_model_for_versioned_object.