Merge "Add notes for DCN operational availability modes"

doc/source/install/advanced_deployment/distributed_compute_node.rst

@@ -28,6 +28,101 @@ to be co-located at the same physical location or datacenter. See
 Such an architecture is referred to as "Distributed Compute Node" or "DCN" for
 short.
 
+Supported failure modes and High Availability recommendations
+-------------------------------------------------------------
+
+Handling negative scenarios for DCN starts from the deployment planning, like
+choosing some particular SDN solution over provider networks to meet the
+expected SLA.
+
+Loss of control plane connectivity
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+A failure of the central control plane affects all DCN edge sites. There is no
+autonomous control planes at the edge. No OpenStack control plane API or CLI
+operations can be executed locally in that case. For example, you cannot create
+a snapshot of a Nova VM, or issue an auth token, nor can you delete an image or
+a VM.
+
+.. note:: A single Controller service failure normally induces
+   no downtime for edge sites and should be handled as for usual HA deployments.
+
+Loss of an edge site
+^^^^^^^^^^^^^^^^^^^^
+
+Running Nova VM instances will keep running. If stopped running, you need the
+control plane back to recover the stopped or crashed workloads.
+
+.. note:: A single Compute service failure normally affects only its edge site
+   without additional downtime induced for neighbor edge sites or the central
+   control plane.
+
+OpenStack infrastructure services, like Nova Compute, will automatically
+reconnect to MariaDB database cluster and RabbitMQ broker when the control
+plane's uplink is back. No timed out operations can be resumed though and need
+to be retried manually.
+
+It is recommended to maintain each DCN edge site as a separate Availability Zone
+(AZ) for Nova/Neutron and Cinder services.
+
+Improving resiliency for N/S and E/W traffic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Reliability of the central control plane may be enhanced with L3 HA network,
+which only provides North-South routing. The East-West routing effectiveness of
+edge networks may be improved by using DVR or highly available Open Virtual
+Network (OVN). There is also BGPVPN and its backend specific choices.
+
+Network recommendations
+^^^^^^^^^^^^^^^^^^^^^^^
+
+Traditional provider networks with backbone routing at the edge may fulfill or
+complement a custom distributed routing solution, like L3 Spine-Leaf topology.
+
+.. note:: Neutron SDN backends that involve tunnelling may be sub-optimal for
+   Edge DCN cases because of the known issues 1808594_ and 1808062_.
+
+   .. _1808594: https://bugs.launchpad.net/tripleo/+bug/1808594
+   .. _1808062: https://bugs.launchpad.net/tripleo/+bug/1808062
+
+That said, when there is a network failure that disconnects the edge off the
+central site, there is no SLA for recovery time but only what the provider
+networks or a particular SDN choice can guarantee. For switched/routed/MPLS
+provider networks, that may span from 10's of ms to a few seconds. With
+the outage thresholds are typically considered to be a 15 seconds. These trace
+back on various standards that are relevant here.
+
+Config-drive/cloud-init details
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The simplest solution we recommend for DCN would involve only provider networks
+at the edge. For that case, it is also recommended to use config-drive or
+another configuration mechanism other than cloud-init. Otherwise, the latter
+requires a `169.254.169.254/32` route for the provider routers to forward data
+to the metadata service.
+
+IPv6 details
+^^^^^^^^^^^^
+
+IPv6 for tenants' workloads and infrastructure tunnels interconnecting
+the central site and the edge is a viable option as well. IPv6 cannot be used for
+provisioning networks though. Key benefits IPv6 may provide for DCN are:
+
+* SLAAC, which is a EUI-64 form of autoconfig that makes IPv6 addresses
+  calculated based on MAC addresses and requires no DHCP services placed on the
+  provider networks.
+* Improved mobility for endpoints, like NFV APIs, to roam around different links
+  and edge sites without losing its connections and IP addresses.
+* End-to-end IPv6 has been shown to have better performance by large content
+  networks. This is largely due to the presence of NAT in most end-to-end IPv4
+  connections that slows them down.
+
+Storage recommendations
+^^^^^^^^^^^^^^^^^^^^^^^
+
+DCN with only ephemeral storage is available for Nova Compute services.
+That is up to the edge cloud applications to be designed to provide enhanced
+data availability, locality awareness and/or replication mechanisms.
 
 Deploying from a centralized undercloud
 ---------------------------------------