WORK IN PROGRESS
The Ubuntu OpenStack HA reference architecture is a current, best practice deployment of OpenStack on Ubuntu 12.04 using a combination of tools and HA techniques to deliver high availability across an OpenStack deployment.
Before you start
Juju + MAAS
Its worth reading up on how to setup MAAS and Juju for your physical server environment prior to trying to deploy the Ubuntu OpenStack HA reference architecture using Juju.
How many servers?
Well for HA (N+1 resilience where possible) with all OpenStack services, you will need 28 servers. If you are trying this out Virtual-MAAS might be a good option. This should be announced as a beta-preview soon.
All configuration options should be placed in a file named 'local.yaml'. A full version of this file can be found here
Although all of the charms to support deployment of OpenStack will be available from the Juju Charm Store, its worth branching the bzr branches that support them locally; this means that if you need to tweak a charm for your specific deployment, its much easier:
mkdir precise ( cd precise bzr branch lp:charms/ceph bzr branch lp:charms/ceph-osd bzr branch lp:~openstack-charmers/charms/precise/mysql/ha-support mysql bzr branch lp:~openstack-charmers/charms/precise/rabbitmq-server/ha-support rabbitmq-server bzr branch lp:~openstack-charmers/charms/precise/hacluster/trunk hacluster bzr branch lp:~openstack-charmers/charms/precise/keystone/ha-support keystone bzr branch lp:~openstack-charmers/charms/precise/nova-cloud-controller/ha-support nova-cloud-controller bzr branch lp:~openstack-charmers/charms/precise/cinder/ha-support cinder bzr branch lp:~openstack-charmers/charms/precise/glance/ha-support glance bzr branch lp:~openstack-charmers/charms/precise/quantum-gateway/ha-support quantum-gateway bzr branch lp:~openstack-charmers/charms/precise/swift-proxy/ha-support swift-proxy bzr branch lp:~openstack-charmers/charms/precise/swift-storage/ha-support swift-storage bzr branch lp:~openstack-charmers/charms/precise/nova-compute/ha-support nova-compute bzr branch lp:~openstack-charmers/charms/precise/openstack-dashboard/ha-support openstack-dashboard ) export JUJU_REPOSITORY=`pwd`
Ceph is a key infrastructure component of the Ubuntu OpenStack HA reference architecture; it provides network accessible, resilient, scalable block storage to MySQL and RabbitMQ to support HA, as well as providing an natively resilient and scalable back-end for block storage (through Cinder) and for image storage (through Glance).
A Ceph deployment will typically consist of both Ceph Monitor (MON) Nodes (responsible for mapping the topology of a Ceph storage cluster and telling clients about it) and Ceph Object Storage Device (OSD) Nodes (responsible for storage data on devices). Some basic configuration is required to support deployment of Ceph using the Juju Charms for Ceph:
ceph: fsid: '6547bd3e-1397-11e2-82e5-53567c8d32dc' monitor-count: 3 monitor-secret: 'AQCXrnZQwI7KGBAAiPofmKEXKxu5bUzoYLVkbQ==' osd-devices: '/dev/vdb' osd-reformat: 'yes' source: 'cloud:precise-updates/grizzly' ceph-osd: osd-devices: '/dev/vdb' osd-reformat: 'yes' source: 'cloud:precise-updates/grizzly'
In this example, Ceph is configured with the provided fsid and secret (these should be unique for your environment) and will use the '/dev/vdb' block device if found for object storage - note that this is a whitelist of block storage devices to detect and use so something like '/dev/sdb /dev/sdd /dev/sde /dev/sdf dev/sdg' is perfectly valid and probably more realistic for physical server deployments.
Ceph is being sourced ('source') from the Ubuntu Cloud Archive for Grizzly to ensure we get the latest features.
The Ceph MON function is provided by the 'ceph' charm; as the monitor-count is set to '3' Ceph will not bootstrap itself and start responding to requests from clients until at least 3 service units have joined the ceph service. In order to operate the Ceph MON service must remain quorate at all time; so you can take one service unit out and still keep running. You can add more service-units to the ceph charm if you want - but 3 should provide enough for most Ceph storage clusters.
Note that the ceph charm will also slurp up and run OSD's on any available storage; for large deployments you might not want todo this but for proof-of-concept work its OK to just run with storage provided directly via the ceph service.
Additional storage is provided by the 'ceph-osd' charm; this allows additional service units to be spun up which purely provide object storage. Recommended for larger deployments.
First, deploy the ceph charm with a unit count of 3 to build the Ceph MON cluster:
juju deploy --config local.yaml -n 3 local:ceph
and then deploy some additional object storage nodes using the ceph-osd charm and relate them to the cluster.
juju deploy --config local.yaml -n 3 local:ceph-osd juju add-relation ceph ceph-osd
All of the above commands can be run in series with no pauses; the charms are clever enough to figure things out in the correct order.
By default, the CRUSH map (which tells Ceph where blocks should be stored for resilience etc..) is OSD centric; if you run multiple OSD's on a single server, Ceph will be device failure resilient but not server failure resilient as the default 3 replicas may be mapped onto OSD's on a single host.
Read the upstream documentation on how to tune the CRUSH map for your deployment requirements; this might land as a feature into the charm later but for now this bit requires manual tuning.
MySQL provides persistent, relational data storage for all OpenStack services; to provide MySQL in a highly-available configuration its deployed with Pacemaker and Corosync (HA tools) in an Active/Passive configuration. Shared block storage for MySQL data is provided by Ceph.
NOTE: For 12.04, its worth running with the Quantal LTS kernel (3.5) to pickup improvements in the Ceph rbd kernel driver.
The only additional configuration required by the MySQL charm is a VIP and subnet mask which will be used as the access point for other services to access the MySQL cluster:
mysql: vip: '192.168.77.8' vip_cidr: 19
The MySQL charm is deployed in-conjunction with the HACluster subordinate charm:
juju deploy --config local.yaml -n 2 local:mysql juju deploy local:hacluster mysql-hacluster juju add-relation mysql ceph juju add-relation mysql mysql-hacluster
After a period of time (it takes a while for all the relations to settle and for the cluster to configure and start), you should have a MySQL cluster listening on 192.168.77.8.
Its also possible to configure the HA cluster to use STONITH (Shoot The Other Node In The Head) based on the power management settings that have been configured in MAAS:
juju set mysql-hacluster stonith_enabled='True \ maas_url=http://192.168.64.1/MAAS \ maas_credentials=<SSO token>
You should be able to pick the maas_* settings from your Juju environments file.
Various Active/Active MySQL derivatives exist which could be used in place of MySQL; however for the Raring/Grizzly release cycle, only MySQL is in Ubuntu and fully supported by Canonical. Future releases of this architecture may use alternative MySQL solutions such as MariaDB or Percona with synchoronous replication support provided by Galera.
RabbitMQ provides a centralized message broker which the majority of OpenStack components use to communicate control plane requests around an OpenStack deployment. RabbitMQ does provide a native Active/Active architecture; however this is not yet well supported so for the Raring/Grizzly cycle RabbitMQ is deployed in Active/Passive configuration using Pacemaker and Corosync with Ceph providing shared block storage.
NOTE: For 12.04, its worth running with the Quantal LTS kernel (3.5) to pickup improvements in the Ceph rbd kernel driver.
The only additional configuration required by the RabbitMQ charm is a VIP and subnet mask which will be used as the access point for other services to access the RabbitMQ cluster:
rabbitmq-server: vip: '192.168.77.11' vip_cidr: 19
The RabbitMQ charm is deployed in-conjunction with the HACluster subordinate charm:
juju deploy --config local.yaml -n 2 local:rabbitmq-server rabbitmq-server juju deploy local:hacluster rabbitmq-hacluster juju add-relation rabbitmq-server ceph juju add-relation rabbitmq-server rabbitmq-hacluster
RabbitMQ will be accessible using the vip provided during configuration.
Various other messaging options exist for OpenStack other than RabbitMQ (although this is current see as the reference choice). Future versions of the HA reference architecture may provide alternative options utilizing ZeroMQ (for brokerless, scalable messaging) or Apache QPid.
Identity Service (Keystone)
Keystone provides central authentication and authorization services for all OpenStack services. It also holds the service catalog for all services in an OpenStack deployment. Keystone is generally stateless; in the reference architecture it can be scaled horizontally - requests are load balanced across all available service units.
The keystone charm requires basic configuration to be deployed in HA mode:
keystone: openstack-origin: 'cloud:precise-grizzly' admin-user: 'admin' admin-password: 'openstack' admin-token: 'ubuntutesting' vip: '192.168.77.1' vip_cidr: 19
The user/password/token should be specific to your deployment and can be used as an initial bootstrap admin account to start seeding your OpenStack cloud with actual user accounts.
The VIP and subnet mask are used to form the HA API endpoint for keystone requests. Keystone API requests will be load balanced across all available service units.
The Keystone charm is deployed in-conjunction with the HACluster subordinate charm:
juju deploy --config local.yaml -n 2 local:keystone juju deploy local:hacluster keystone-hacluster juju add-relation keystone keystone-hacluster
Keystone used MySQL for persistent storage of data:
juju add-relation keystone mysql
The keystone charm uses the Stateless API HA model (see below). Some state is stored on local disk (specifically service usernames and passwords). These are synced between services units during hook execution using SSH + unison.
The Cloud Controller provides the API endpoints for Nova (Compute) and Quantum (Networking) services; The API's are stateless; in the reference architecture this service can be scaled horizontally with API requests load balanced across all available service units.
The nova-cloud-controller charm has a large number of configuration options; in-line with other HA services, a VIP and subnet mask must be provided to host the HA API endpoints. In addition, configuration options for Quantum networking are also provided.
nova-cloud-controller: openstack-origin: 'cloud:precise-grizzly' vip: '192.168.77.2' vip_cidr: 19 network-manager: 'Quantum' conf-ext-net: 'no' ext-net-cidr: '192.168.64.0/19' ext-net-gateway: '192.168.64.1' pool-floating-start: '192.168.90.1' pool-floating-end: '192.168.95.254'
Note that the conf-ext-net option is disabled during service build; unfortunately configuring this during service build proved a bit racey but the external (public) network can be configured post deployment of the charm:
juju set nova-cloud-controller conf-ext-net=yes
The nova-cloud-controller charm is deployed in-conjunction with the HACluster subordinate charm:
juju deploy --config local.yaml -n 2 local:nova-cloud-controller juju deploy local:hacluster ncc-hacluster juju add-relation nova-cloud-controller ncc-hacluster juju add-relation nova-cloud-controller mysql juju add-relation nova-cloud-controller keystone juju add-relation nova-cloud-controller rabbitmq-server
The nova-cloud-controller charm uses the Stateless API HA model (see below).
Image Storage (Glance)
Glance provides multi-tenant image storage services for an OpenStack deployment; By default, Glance uses local storage to store uploaded images. The HA reference architecture uses Ceph in-conjunction with Glance to provide highly-available object storage; the design relegates Glance to being a stateless API and image registry service.
In-line with other OpenStack charms, Glance simply requires a VIP and subnet mask to host the Glance HA API endpoint:
glance: openstack-origin: 'cloud:precise-grizzly' vip: '192.168.77.4' vip_cidr: 19
juju deploy --config local.yaml -n 2 local:glance juju deploy local:hacluster glance-hacluster juju add-relation glance glance-hacluster juju add-relation glance mysql juju add-relation glance nova-cloud-controller juju add-relation glance ceph juju add-relation glance keystone
The glance charm uses the Stateless API HA model (see below).
Block Storage (Cinder)
Cinder provides block storage to tenant instances running with an OpenStack cloud. By default, Cinder uses local storage exposed via iSCSI which is inherently not highly-available. The HA reference architecture uses Ceph in conjunction with Cinder to provide highly-available, massively scalable block storage for tenant instances. Ceph block devices are accessed directly from compute nodes; this design relegates Cinder to being a stateless API and storage allocation service.
In-line with other OpenStack charms, Cinder requires a VIP and subnet mask to host the HA API endpoint. In addition, Cinder itself is explicitly configured not to use local block storage:
cinder: openstack-origin: 'cloud:precise-grizzly' block-device: 'None' vip: '192.168.77.3' vip_cidr: 19
juju deploy --config local.yaml -n 2 local:cinder juju deploy local:hacluster cinder-hacluster juju add-relation cinder cinder-hacluster juju add-relation cinder mysql juju add-relation cinder keystone juju add-relation cinder nova-cloud-controller juju add-relation cinder rabbitmq-server juju add-relation cinder ceph juju add-relation cinder glance
The cinder charm uses the stateless API HA model (see below).
Quantum provides the virtualized network infrastructure within an OpenStack deployment. Currently its provided as an alternative to Nova networking as Quantum has not got feature parity yet. Quantum in a HA mode is only supported in Grizzly due to the provision of a agent/scheduler infrastructure in this release.
Some aspects of Quantum (the API server for example) are integrated into other OpenStack charms; to complete the networking topology a Quantum Gateway is required to provide Layer 3 network routing and DHCP services for Layer 2 networks.
The quantum-gateway charm only requires configuration for the external network port that will be used for Layer 3 routing connectivity; This must not be the primary network interface on the server otherwise you will lose connectivity to the gateway server units!
quantum-gateway: openstack-origin: 'cloud:precise-grizzly' ext-port: 'eth1'
juju deploy --config local.yaml -n 2 local:quantum-gateway juju add-relation quantum-gateway mysql juju add-relation quantum-gateway rabbitmq-server juju add-relation quantum-gateway nova-cloud-controller
Quantum was due to have native HA support for Grizzly; however this feature did not land in full. Currently HA is implemented by re-allocating network resources on a failed service unit to good service units; this is orchestrated using the cluster-relation-departed hook in the quantum-gateway charm. Fail-over of services can take between 10-30 seconds.
Full HA is not possible on Nova Compute service units; however the nova-compute charm can be configured to support secure live migration of running instances between compute service units, supporting a managed, minimal disruption approach to maintenance of the underlying operating system.
nova-compute: openstack-origin: 'cloud:precise-grizzly' enable-live-migration: 'True' migration-auth-type: 'ssh'
juju deploy --config local.yaml -n 3 local:nova-compute juju add-relation nova-compute nova-cloud-controller juju add-relation nova-compute mysql juju add-relation nova-compute rabbitmq-server juju add-relation nova-compute glance juju add-relation nova-compute ceph
Live migration is facilitated using libvirt and qemu over a SSH connection. This includes live block migration. A shared filesystem provided by Ceph was considered; however this approach is not truly scalable and CephFS does not have 'stable' status yet.
Object Storage (Swift)
The Swift service provides multi-tenant object storage within an OpenStack deployment. Its analogous with Amazons S3 service. Objects are distributed across underlying Swift storage nodes for both resilience and scalability.
Swift is actually split into two charms; swift-proxy and swift-storage. For the HA reference architecture we configure Swift with three storage zones:
swift-proxy: openstack-origin: 'cloud:precise-grizzly' zone-assignment: 'manual' replicas: 3 swift-hash: 'fdfef9d4-8b06-11e2-8ac0-531c923c8fae' vip: '192.168.77.12' vip_cidr: 19 swift-storage-z1: openstack-origin: 'cloud:precise-grizzly' zone: 1 block-device: 'vdb' swift-storage-z2: openstack-origin: 'cloud:precise-grizzly' zone: 2 block-device: 'vdb' swift-storage-z3: openstack-origin: 'cloud:precise-grizzly' zone: 3 block-device: 'vdb'
In-line with other OpenStack charms, a VIP and subnet mask is provided to host the Swift HA API endpoint.
juju deploy --config local.yaml -n 2 local:swift-proxy juju deploy local:hacluster swift-hacluster juju deploy --config local.yaml local:swift-storage swift-storage-z1 juju deploy --config local.yaml local:swift-storage swift-storage-z2 juju deploy --config local.yaml local:swift-storage swift-storage-z3 juju add-relation swift-proxy swift-hacluster juju add-relation swift-proxy keystone juju add-relation swift-proxy swift-storage-z1 juju add-relation swift-proxy swift-storage-z2 juju add-relation swift-proxy swift-storage-z3
Need some notes about ring rebalancing and how the swift-proxy charm builds the rings without replicating data between nodes (for adam_g).
The Horizon service provides an end-user and administrator web portal with an OpenStack deployment. This service is completely stateless and can be scaled horizontally, with requests being load-balanced across all available service units.
openstack-dashboard: openstack-origin: 'cloud:precise-grizzly' vip: '192.168.77.5' vip_cidr: 19
juju deploy --config local.yaml -n 2 local:openstack-dashboard juju deploy local:hacluster dashboard-hacluster juju add-relation openstack-dashboard dashboard-hacluster juju add-relation openstack-dashboard keystone
Although this service is not an API service, it uses the same model for HA.
Keystone will always be listening on its VIP; source the following:
cat > novarc << EOF export OS_USERNAME=admin export OS_PASSWORD=openstack export OS_TENANT_NAME=admin export OS_AUTH_URL=http://192.168.77.1:5000/v2.0 export OS_REGION_NAME=RegionOne alias nova="nova --no-cache" EOF
Assuming you have deployed all services, keystone should provide an endpoint listing as detailed below:
keystone endpoint-list +----------------------------------+-----------+--------------------------------------------------+--------------------------------------------------+---------------------------------------------+ | id | region | publicurl | internalurl | adminurl | +----------------------------------+-----------+--------------------------------------------------+--------------------------------------------------+---------------------------------------------+ | 1ac5142878a34d0cb9e2290f23c916c6 | RegionOne | http://192.168.77.2:8774/v1.1/$(tenant_id)s | http://192.168.77.2:8774/v1.1/$(tenant_id)s | http://192.168.77.2:8774/v1.1/$(tenant_id)s | | 3836f45f29bb46b0a6709338f9dfc720 | RegionOne | http://192.168.77.2:3333 | http://192.168.77.2:3333 | http://192.168.77.2:3333 | | 4526045cbada4a7fa388b5154c32a626 | RegionOne | http://192.168.77.3:8776/v1/$(tenant_id)s | http://192.168.77.3:8776/v1/$(tenant_id)s | http://192.168.77.3:8776/v1/$(tenant_id)s | | 4cdbfb34997646c9abb552f03221d5be | RegionOne | http://192.168.77.4:9292 | http://192.168.77.4:9292 | http://192.168.77.4:9292 | | 6fef2877df7d4bc3a25ad04629c37abc | RegionOne | http://192.168.77.1:5000/v2.0 | http://192.168.77.1:5000/v2.0 | http://192.168.77.1:35357/v2.0 | | 9a1bad74efee4e5abfb4bce76847defb | RegionOne | http://192.168.77.2:8773/services/Cloud | http://192.168.77.2:8773/services/Cloud | http://192.168.77.2:8773/services/Cloud | | b382813b93064c6796ba8d13e51d5902 | RegionOne | http://192.168.77.2:9696 | http://192.168.77.2:9696 | http://192.168.77.2:9696 | | f21918422c664a399a25483d67078c6a | RegionOne | https://192.168.77.12:8080/v1/AUTH_$(tenant_id)s | https://192.168.77.12:8080/v1/AUTH_$(tenant_id)s | https://192.168.77.12:8080 | +----------------------------------+-----------+-------------------------------------------------+--------------------------------------------------+---------------------------------------------+
Scaling out services
It's possible to scale out the majority of services in the OpenStack HA reference architecture by simply adding units (in this example glance):
juju add-unit glance
This makes alot of sense for stateless services where requests are load-balanced across all available service units.
Additional units can be added to stateful services; however right now this just provides additional resilience over and above N+1; this might be useful to retaining full HA when doing a rolling-upgrade of an OpenStack cloud using something like Landscape.
Stateless API Server
For stateless API services, the OpenStack service is reconfigured to listen on [default port - 10], haproxy is installed and configured to listen on the default service port and to load balancer across all service units with the service and a Virtual IP is floated onto of the primary service unit.
This ensures that the full capacity of all service units in the service is used to service incoming API requests - an Active/Active model.
For services where state must be stored, such as for MySQL or RabbitMQ, state is stored on a shared block device provided by Ceph; this is mapped onto one (and only one) server at a time using the Ceph rbd kernel module.
The device (and its associated filesystem and contents) are placed under the control of Corosync and Pacemaker using the hacluster charm; this ensures that the persistent data is only writable from one service unit within a service at any point in time.
Services of this type are described as Active/Passive.
The hacluster charm deals with installing and configuring Corosync and Pacemaker based on what relation data has been provided by the principle charm that its been related to. This will include services to control from the cluster, shared block devices from ceph, filesystems on those block devices and VIP's.
If you need to check the cluster status of any service that utilizes the hacluster charm (glance in this example):
juju ssh glance/0 sudo crm status
This will output the current status of resources controller by Corosync and Pacemaker.
Leaders are elected by selecting the older peer within a given service deployment. This service unit will undertake activities such as creating underlying databases, issuing username and passwords and configuring HA services prior to full clustering.
Once a set of service units have been clustered using Corosync and Pacemaker, leader election is determine by which service unit holder the VIP through which the service is accessed. This service unit will then take ownership of singleton activity within the cluster.