Key concepts

Redundancy and failover

High availability is implemented with redundant hardware running redundant instances of each service. If one piece of hardware running one instance of a service fails, the system can then failover to use another instance of a service that is running on hardware that did not fail.

A crucial aspect of high availability is the elimination of single points of failure (SPOFs). A SPOF is an individual piece of equipment or software that causes system downtime or data loss if it fails. In order to eliminate SPOFs, check that mechanisms exist for redundancy of:

  • Network components, such as switches and routers

  • Applications and automatic service migration

  • Storage components

  • Facility services such as power, air conditioning, and fire protection

In the event that a component fails and a back-up system must take on its load, most high availability systems will replace the failed component as quickly as possible to maintain necessary redundancy. This way time spent in a degraded protection state is minimized.

Most high availability systems fail in the event of multiple independent (non-consequential) failures. In this case, most implementations favor protecting data over maintaining availability.

High availability systems typically achieve an uptime percentage of 99.99% or more, which roughly equates to less than an hour of cumulative downtime per year. In order to achieve this, high availability systems should keep recovery times after a failure to about one to two minutes, sometimes significantly less.

OpenStack currently meets such availability requirements for its own infrastructure services, meaning that an uptime of 99.99% is feasible for the OpenStack infrastructure proper. However, OpenStack does not guarantee 99.99% availability for individual guest instances.

This document discusses some common methods of implementing highly available systems, with an emphasis on the core OpenStack services and other open source services that are closely aligned with OpenStack.

You will need to address high availability concerns for any applications software that you run on your OpenStack environment. The important thing is to make sure that your services are redundant and available. How you achieve that is up to you.

Active/passive versus active/active

Stateful services can be configured as active/passive or active/active, which are defined as follows:

active/passive configuration

Maintains a redundant instance that can be brought online when the active service fails. For example, OpenStack writes to the main database while maintaining a disaster recovery database that can be brought online if the main database fails.

A typical active/passive installation for a stateful service maintains a replacement resource that can be brought online when required. Requests are handled using a virtual IP address (VIP) that facilitates returning to service with minimal reconfiguration. A separate application (such as Pacemaker or Corosync) monitors these services, bringing the backup online as necessary.

active/active configuration

Each service also has a backup but manages both the main and redundant systems concurrently. This way, if there is a failure, the user is unlikely to notice. The backup system is already online and takes on increased load while the main system is fixed and brought back online.

Typically, an active/active installation for a stateless service maintains a redundant instance, and requests are load balanced using a virtual IP address and a load balancer such as HAProxy.

A typical active/active installation for a stateful service includes redundant services, with all instances having an identical state. In other words, updates to one instance of a database update all other instances. This way a request to one instance is the same as a request to any other. A load balancer manages the traffic to these systems, ensuring that operational systems always handle the request.

Clusters and quorums

The quorum specifies the minimal number of nodes that must be functional in a cluster of redundant nodes in order for the cluster to remain functional. When one node fails and failover transfers control to other nodes, the system must ensure that data and processes remain sane. To determine this, the contents of the remaining nodes are compared and, if there are discrepancies, a majority rules algorithm is implemented.

For this reason, each cluster in a high availability environment should have an odd number of nodes and the quorum is defined as more than a half of the nodes. If multiple nodes fail so that the cluster size falls below the quorum value, the cluster itself fails.

For example, in a seven-node cluster, the quorum should be set to floor(7/2) + 1 == 4. If quorum is four and four nodes fail simultaneously, the cluster itself would fail, whereas it would continue to function, if no more than three nodes fail. If split to partitions of three and four nodes respectively, the quorum of four nodes would continue to operate the majority partition and stop or fence the minority one (depending on the no-quorum-policy cluster configuration).

And the quorum could also have been set to three, just as a configuration example.

Note

We do not recommend setting the quorum to a value less than floor(n/2) + 1 as it would likely cause a split-brain in a face of network partitions.

When four nodes fail simultaneously, the cluster would continue to function as well. But if split to partitions of three and four nodes respectively, the quorum of three would have made both sides to attempt to fence the other and host resources. Without fencing enabled, it would go straight to running two copies of each resource.

This is why setting the quorum to a value less than floor(n/2) + 1 is dangerous. However it may be required for some specific cases, such as a temporary measure at a point it is known with 100% certainty that the other nodes are down.

When configuring an OpenStack environment for study or demonstration purposes, it is possible to turn off the quorum checking. Production systems should always run with quorum enabled.

Load balancing