CellsV2 - Move quota tables to API database

https://blueprints.launchpad.net/nova/+spec/cells-quota-api-db

As part of the CellsV2 work we are in the process of splitting the current cell database. Quotas in nova are global and should apply across cells. Because of this their data needs to reside in the API database.

Problem description

Quotas for projects and users need to be enforced across cell boundaries. Quotas are also exposed in the API. If quotas remain in the cell database then the API would have to me modified to expose cell information. Alternatively quota data would have to be replicated across all cells which would make it difficult to enforce a global quota.

Use Cases

Operators and users wish to enforce a quota that is applicable across cells.

Proposed change

We propose to move the quota related tables that currently reside in the cell database to the API database. These tables are:

quotas
quota_classes
quota_usages
project_user_quotas
reservations

Database models will be created in the API database. These will closely match the existing models in the cell database. Some modifications may be made to these tables to clean up the existing data model.

The Quota object in nova makes use of functions in quota.py. It is mostly an RPC facade to the quota.py API. The DbQuotaDriver in quota.py will be modified to access the API database.

Wrapper methods will be created for the existing database access. Wrapper methods that perform any get type operation on the quotas, quota_classes, project_user_quotas tables will be modified to load initially from the API database. If the item is not found then it will be searched for in the cell database.

Because there will often be attempt to access entries in project_user_quotas and quotas that are empty we will need to cache a value that indicates the migration is complete. Without this reads of empty entries in these tables will be attempted twice. Instead the flow will be:

use_api_database_only = False
value = get_from_api_database()

if use_database_only == False:
    if not value and is_cell_database_empty():
        use_api_database_only = True
else:
    value = get_from_cell_database()

Migration functions will be created to migrate data from the cell to API database. These methods will be added to the nova manage online_data_migrations command.

Migration of flavors has already been completed and for the above tables we will generally follow this example. [3]

The quota_usages and reservations table will not read from both databases. No data migration will be made for usages and reservations. New database access methods will be created that access these tables in the API datbase. The DbQuotaDriver code will be modified to use these methods. When usages or reservations are found to be missing in the API database a sync call will be made on the quota to write the current usage to the API database.

Alternatives

Without changing the nature of quotas there is likely no alternative for the DbQuotaDriver but to move its tables to the API database. Alternatives for quotas in general include ‘Quotas Reimagined’ [1] and the ‘Delimiter’ [2] project. The former specification would likely still make use of some of the existing tables that would require migration. A separate quotas library or service would create an entirely new quota data store.

Data model impact

The data model impact will be large as many new tables will be created in the API database. The models will not be detailed here as they are essentially unchanged from the cell database.

REST API impact

None - Cells implementation should not be exposed in the API.

Security impact

None, moving the Quota data to the API databse keeps the implementation as close as possible to the current model. It should be no more or less difficult than now to break quota enforcement.

Notifications impact

None

Other end user impact

None

Performance Impact

The performance impact of the changes will be negligible. However, the performance of the DbQuotaDriver will also not be improved for the large-scale deployments that CellsV2 is targeting.

During migration there will be a larger performance impact to quota operations and therefore build requests. This comes as database accesses for the quota, project_user_quota, and quota_classes tables may require two database requests. Also after upgrade the quota_usages and reservations tables will not be migrated. This will require a sync call until these tables are populated.

This performance impact will be short term and will last until data has been migrated.

Other deployer impact

Deployers will have to be aware of the data migration when upgrading. They will have to know about nova-manage commands to migrate data to the API database.

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

<mjdoffma@us.ibm.com>

Other contributors:

None

Work Items

  • Create database models and migrations for quota tables in the API database.

  • Create databases models and migrations for quota usages and reservations.

  • Create database access and wrapper methods for API datbase.

  • Modify the DbQuotaDriver to use the new database access methods.

  • Ensure that the DbQuotaDriver re-syncs quotas when usages are not found in the API datbase.

  • Create migration methods to move data to the API database and add this method to online_data_migrations.

Dependencies

None

Testing

  • Add unit and functional tests for new database models.

  • Add new unit tests for database access wrapper methods.

  • Add new functional tests for data migration.

  • Modify unit tests for the quota driver.

  • Enhance existing functional tests for quotas.

Documentation Impact

Operator documentation may need to me modified to include details of upgrading and migrating data using nova-manage command.

References

History

Revisions

Release Name

Description

Newton

Introduced