Fuel should allow user to deploy OpenStack to multiple racks with a scalable underlay network design so that the user can meet his or her expanding business needs without having to re-deploy/migrate workloads.
Current implementation of multi-rack support lacks a number of features in demand (all services via one network, networks shared between node network groups, VIPs can be allocated in different node network groups, arbitrary VIP addresses), and has a number of usability issues (manual setup of dnsmasq on master node, lack of validation, nodes do not have connection with node network groups until they are added into environment but in fact they have IPs from different node network groups, new routes are not applied when node network group is added to the deployed environment).
Current proposal deals with the issues listed above.
1. Make additional setup of dnsmasq on master node when node network groups are configured. User should do that by hands now. It should be done automatically when parameters of any of Admin networks are changed or Admin network is deleted. This will save a user from manual error-prone operations of setting this up via a command line. See https://bugs.launchpad.net/fuel/+bug/1495593
2. When a new node network group is added to the deployed environment new routes should be applied to network configuration of all nodes. This will resolve an issue with adding node network groups to the deployed environment. Reconfiguration should also be done when any of netmasks for networks of deployed environment is changed. See https://bugs.launchpad.net/fuel/+bug/1502842
3. VIP auto-allocation is restricted to controller node group now in Nailgun. It should be allowed to allocate VIP in any node group to allow proper separation of HA services into different nodes. But other restriction remains the same: VIP can be allocated only if all nodes which conform to its node_roles section are in the same node group. See https://bugs.launchpad.net/fuel/+bug/1487021
4. Provide an ability to share network (L2/L3 parameters) between several node network groups. As for now, each particular node network group have its own L2/L3 parameters for every network. It is 1:1 mapping. It will be possible to share arbitrary networks (use the same L2/L3 parameters) between several node network groups. It will be possible to use completely arbitrary mapping. To have some network shared between two node network groups user will have to setup equal L3 parameters for those (CIDR an Gateway should be equal, IP ranges do not matter). Routes between networks having same L3 will not be generated. Only networks with the same name can be shared this way. VLAN IDs of networks can be equal or different. VLAN IDs for the same network interface should be different. E.g. this ability is required to use a dedicated storage. See https://bugs.launchpad.net/fuel/+bug/1473047
5. Ability to set the floating IP range. Now, Nailgun doesn’t allow to set floating IP range from non-default node network group. So, network and controller nodes can be deployed in default nodegroup only. It should be allowed to set floating range within any of Public networks. So, it will be possible to deploy controller nodes in any node group. IPs from the floating range need to be added to the allocated addresses pool. See https://bugs.launchpad.net/fuel/+bug/1502829
GUI tasks are to be in separate ticket/spec.
Change 1. a. The “update_dnsmasq” task will be added to serve auto setup of dnsmasq on master node. This Nailgun task will “upload_file”, “puppet” and “cobbler_sync” orchestrator tasks to update dnsmasq configuration. b. Nailgun will run that task when some of IP ranges of Admin networks are changed via API. c. Nailgun API validator will check new Admin IP ranges against IPs of nodes currently visible to Fuel (bootstrap and deployed ones). Also, Nailgun will check IPS of newly added nodes against Admin IP ranges and mark nodes with errors if they are out of known Admin IP ranges.
Change 2. a. New Nailgun task will be added to rerun networking configuration on deployed nodes (or deployment tasks graph will be rebuilt). b. Deployment data for deployed nodes will be regenerated when required. c. Controller nodes maybe triggered for redeploy when required.
Change 3. a. VIPs addresses assignment should be reworked to determine the node network group where VIP address should be allocated (using node roles and network role from VIP description). Now the node network group that contains controllers is always taken for assgnment of VIPs. b. Nailgun should check that each VIP address should belong to exactly one nodegroup.
Change 4. a. Validation of network parameters (NetworkCheck class methods) should allow intersection of CIDRs and gateways (simultaneously only) of several networks with the same name from different (not the same) node network groups. In case of such intersection those networks are considered as one shared network and no routes should be generated between segments. b. Validation of network parameters (NetworkCheck class methods) should allow equal VLAN IDs within environment. Exception is that every single network interface cannot handle coincident VLAN IDs.
Change 5. a. Change the verification of floating ranges to allow floating range to match any of Public networks in environment. Now it may only match Public network from default node network group. b. Change the serialization of floating network to find appropriate Public network for floating ranges defined.
Change 1. New task name and new node error type will be added.
Change 2. New parameter for modular tasks is added: ‘reexecute_on’. This parameter is a list of Nailgun events on which this task should be re-executed. For example:
- id: netconfig type: puppet groups: [primary-controller, controller, cinder, cinder-vmware, compute, ceph-osd, primary-mongo, mongo, virt, ironic] required_for: [deploy_end] requires: [tools] reexecute_on: [deploy_changes] parameters: puppet_manifest: /etc/puppet/modules/osnailyfacter/modular/netconfig/netconfig.pp puppet_modules: /etc/puppet/modules timeout: 3600
It means that ‘netconfig’ task will be re-executed on ‘deploy_changes’ event.
Data formats remain the same. Flow for network configuration is changed for change 5.
We introduce new Nailgun hook in Astute - “cobbler_sync”. It runs cobbler sync to reload dnsmasq configuration. It does not affect other aspects of deployment orchestration.
New “nailgun::dnsmasq::dhcp_range” class added. It creates separate dnsmasq configuration file in /etc/dnsmasq.d/ directory for every dhcp-range. We also configure dnsmasq to use /etc/dnsmasq.d/ as conf-dir. /etc/dnsmasq.d directory is stored on the host system to prevent data loss due to container restart/rebuild.
This feature can be treated as a composition of several smaller changes. Seems, all of them can be implemented separately. But some tasks have dependencies: 8 depends on 1 and 5.
When a node group or a cluster is deleted there can be some nodes in bootstrap which have IPs corresponding to those deleted node groups. They cannot be provisioned as dnsmasq configuration does not contain information about those admin networks anymore. Nailgun marks such nodes as ‘error’ and sends appropriate notification.
User no longer needs to update dnsmasq.template file manually and append EXTRA_ADMIN_NETWORKS. All the DHCP related changes are being applied by Fuel automatically.
It’s possible to change admin network even for default node group (node group where Fuel node is). DHCP/PXE related settings provided via fuel-menu during Fuel node bootstrap are used only for initial DHCP configuration. After that Nailgun controls all the network related settings even for DHCP/PXE network. Please note that if you change DHCP/PXE CIDR for the default node group and it requires to change Fuel node IP address, you should change it on admin interface of Fuel node manually and check/update docker containers.
We should remove documentation about manual update of dnsmasq.template file here: https://docs.mirantis.com/openstack/fuel/fuel-7.0/operations.html#configuring-multiple-cluster-networks
Primary assignee: Alex Didenko
Other contributors: Aleksey Kasatkin, Sergey Vasilenko
Mandatory design review: Andrew Woodward, Sergey Vasilenko
In order to verify the quality of new features, automatic system tests will be expanded by the cases listed below:
1. Environment is deployed using slaves from non-default nodegroup as controller nodes. See https://blueprints.launchpad.net/fuel/+spec/test-custom-nodegroup-controllers
2. New nodegroup is added to operational environment. See https://blueprints.launchpad.net/fuel/+spec/test-nodegroup-add
3. Environment is deployed using default gateway from non-public network. See https://blueprints.launchpad.net/fuel/+spec/test-custom-default-gw
4. Deploy environment with few nodegroups and shared network parameters between them. See https://blueprints.launchpad.net/fuel/+spec/test-nodegroups-share-networks
5. Default IP range is changed for admin/pxe network. See https://bugs.launchpad.net/fuel/+bug/1513154
6. Slave nodes are bootstrapped and successfully deployed using non-eth0 interface for admin/pxe network. See https://bugs.launchpad.net/fuel/+bug/1513159
Also there is a need to align existing tests for multiple cluster networks with new features. See https://blueprints.launchpad.net/fuel/+spec/align-nodegroups-tests