Replication/Cluster Locality¶
Nova has the ability to control whether new instances are created on the same hypervisor (affinity) or on different hypervisors (anti-affinity). This behaviour is useful when setting up replication networks or clusters, so it is proposed to add support for it into Trove.
Launchpad Blueprint: https://blueprints.launchpad.net/trove/+spec/replication-cluster-locality
Problem Description¶
At times it is desirable to control where replicas in a replication set are created. In the case of read slaves, there may be a requirement to ensure that all replicas are on the same hypervisor, whereas for high availability it may be desired that all replicas are created on different hypervisors.
By the same token, clusters would benefit from having instances on different hypervisors.
Replication/cluster locality seeks to address this issue. (Granted, anti-affinity seems more relevant - and desirable - however the two go hand-in-hand as will be seen in the next section.)
Proposed Change¶
Nova has the capability of placing new instances on the same hypervisor, or on different ones. Within Nova this is handled through the use of ‘server groups.’ 1 Using this feature of Nova will require sending a server_group ‘hint’ to the nova_client.servers.create call. This functionality will be used ‘under the covers’ similar to how secgroups are managed now.
Trove will perform the following actions to accomplish this:
The create_instance method in the Task Manager will be changed to include a ‘locality’ argument.
def create_instance(self, context, instance_id, name, flavor,
image_id, databases, users, datastore_manager,
packages, volume_size, backup_id, availability_zone,
root_password, nics, overrides, slave_of_id,
cluster_config, volume_type, locality):
The create method in cluster/models.py will likewise be modified to include a ‘locality’ argument.
@classmethod
def create(cls, context, name, datastore, datastore_version,
instances, extended_properties, locality):
The locality argument will be used to create a server group with the corresponding policy (affinity or anti-affinity).
The server group details will be converted to a Nova ‘hint.’
This hint will be passed into the Nova client in the create call in taskmanager/models.py.
Note: If affinity is chosen and the hypervisor does not have enough resources, then some of the instances will fail to create. The same goes for choosing anti-affinity and there are not enough available hypervisors. There may be other cases where Nova is unable to create the instance (affinity chosen, but different AZ’s); in these cases the new instance will likely fail to spawn as well.
The Trove show and cluster-show commands will also be modified to show the ‘locality’ value (i.e. the server_group policy), if it exists.
Configuration¶
No changes
Database¶
No changes. The server_group id will be obtained from Nova when it’s needed.
Public API¶
An attribute ‘locality’ will be added to the data payload sent during a ‘create’ command. The request/response will look like:
Request:
POST v1/{tenant_id}/instances
{
"instance":
{
"volume":
{
"type": null,
"size": 1
},
"flavorRef": 7,
"name": "myinst",
"replica_count": 2,
"locality": "affinity"
}
}
Response:
{
"instance":
{
"status": "BUILD",
"updated": "2015-12-13T12:36:59",
"name": "myinst",
"links":
[
{
"href": "https://10.240.64.151:8779/v1.0/<tenant>/instances/<id>",
"rel": "self"
},
{
"href": "https://10.240.64.151:8779/instances/<id>",
"rel": "bookmark"
}
],
"created": "2015-12-13T12:36:59",
"id": "<id>",
"volume":
{
"size": 1
},
"flavor":
{
"id": "7",
"links":
[
{
"href": "https://10.240.64.151:8779/v1.0/<tenant>/flavors/7",
"rel": "self"
},
{
"href": "https://10.240.64.151:8779/flavors/7",
"rel": "bookmark"
}
]
},
"datastore":
{
"version": "5.6",
"type": "mysql"
},
"locality": "affinity"
}
}
The ‘show’ and ‘cluster-show’ commands will also now return an attribute ‘locality’ that will look like the one returned from their respective create commands.
Public API Security¶
No impact
Python API¶
A new argument ‘locality’ will be added to the Trove create command (this will be passed through to the Nova client as a hint by the Task Manager, after creating the corresponding server group). The new Python API signature will be:
def create(self, name, flavor_id, volume=None, databases=None, users=None,
restorePoint=None, availability_zone=None, datastore=None,
datastore_version=None, nics=None, configuration=None,
replica_of=None, slave_of=None, replica_count=None,
locality=None):
The Python API signature for cluster-create will be:
def create(self, name, datastore, datastore_version, instances=None,
locality=None):
CLI (python-troveclient)¶
The create command will now accept a –locality flag that can be one of two values: affinity and anti-affinity. The command would look like:
trove create my_instance 7 --size 1 --locality affinity
For clusters it will be:
trove cluster-create my_cluster mysql 5.6 --locality affinity \
--instance flavor=10,volume=1 \
--instance flavor=10,volume=1 \
--instance flavor=10,volume=1 \
Replicas can then be created in the usual fashion, with all following the locality setting of the master node. If adding replicas to an existing set, an exception will be thrown if –locality is specified, as this flag cannot be changed once it has been associated with an instance (this is a Nova restriction in that servers can’t be added or removed from a group manually). For example, the following command would fail:
trove create my_replica 7 --size 1 --locality affinity --replica_of <id>
When growing a cluster, the same locality will be applied to any new instances created. Because of this, the locality flag will not be available for cluster-grow.
The show and cluster-show commands will also display the locality value. For example:
> trove show my_instance
+-------------------+-------------------------+
| Property | Value |
+-------------------+-------------------------+
| created | 2015-12-13T12:36:59 |
| datastore | mysql |
| datastore_version | 5.6 |
| flavor | 7 |
| id | <id> |
| ip | 10.64.151.6 |
| name | my_instance |
| status | ACTIVE |
| updated | 2015-12-13T12:37:03 |
| volume | 1 |
| volume_used | 0.1 |
| locality | affinity |
+-------------------+-------------------------+
If locality is not specified, then no hint will be sent to Nova and it will follow its default algorithm for deciding what hypervisor to use.
For operator assistance in debugging issues the server group id could be displayed through a management call, but will be kept hidden from end users (just as the Nova instance id is not displayed).
Internal API¶
The Task Manager is responsible for creating the replication and cluster instances, so it will need to be aware of the locality flag. The relevant methods will be changed to include this, as described above.
Once the flag is converted to a server group, a ‘hint’ will be created to pass to the Nova client. The converted hint data will be equivalent to the corresponding ReST API values:
"os:scheduler_hints":
{
"group": "<id>"
}
Guest Agent¶
Since the server group must be created before the Nova instance is created, there are no anticipated Guest Agent changes.
Alternatives¶
The server_group id could be stored in the Trove database. The trade-off to any speed improvements (from not having to request the information from Nova) would be that Trove would have to manage this field properly. If it got out of sync then Trove wouldn’t function as expected.
Dashboard Impact (UX)¶
For each of the instance create and cluster create screens, a drop-down for locality will need to be added. This drop-down will have ‘None’, ‘affinity’, and ‘anti-affinity’ in it. The value chosen will need to be passed down into the respective create calls. The show commands will need to be enhanced as well to recognize the new ‘locality’ element.
Implementation¶
Assignee(s)¶
- Primary assignee:
peterstac
Milestones¶
- Target Milestone for completion:
Newton-1
Work Items¶
The work will be undertaken with the following tasks:
Client (Python and CLI) changes for replication
Server (API and TaskManager) changes for replication
Client (Python and CLI) changes for clustering
Server (API and TaskManager) changes for clustering
Upgrade Implications¶
Since this change is completely backwards compatible, no upgrade issues are expected.
Dependencies¶
None
Testing¶
The replication scenario test will be modified to use –locality=affinity and the results verified (i.e. that the Trove ‘show’ command returns the right value for the ‘locality’ attribute). A negative test with anti-affinity will also be created (since devstack runs on one hypervisor, this test should always fail to create replicas). As far as it is feasible with respect to resources consumed, this same pattern will be used in the cluster scenario tests.
Documentation Impact¶
This is a net-new feature, and as such will require documentation.
Appendix¶
None