Eric Novikoff's Blog

IT managers and pundits speak of the reliability of a system in "nines." Two nines is the same as 99%, which comes to (100%-99%)*365 or 3.65 days of downtime per year, which is typical for non-redundant hardware if you include the time to reload the operating system and restore backups (if you have them) after a failure. Three nines is about 8 hours of downtime, four nines is about 52 minutes and the holy grail of 5 nines is 7 minutes.

From a users' point of view, downtime is downtime, but for a provider/vendor/web site manager, downtime is divided into planned and unplanned. Cloud computing can offer some benefits for planned downtime, but the place that it can have the largest effect on a business is in reducing unplanned downtime.

Planned downtime is usually the result of having to do some sort of software maintenance or release process, which is usually outside the domain of the cloud vendor, unless that vendor also offers IT operations services. Other sources of planned downtime are upgrades or scheduled equipment repairs. Most cloud vendors have some planned downtime, but because their business is based on providing high uptime, scheduled downtimes are kept to a minimum.

Unplanned downtime is where cloud vendors have the most to offer, and also the most to lose. Recent large outages at Amazon and Google have shown that even the largest cloud vendors can still have glitches that take considerable time to repair and give potential cloud customers a scare (perhaps it is because they didn't take some planned downtime??) On the other hand, cloud vendors have the experienced staff and proven processes that should produce overall hardware and network reliability that meets or exceeds that of the average corporate data center, and far exceeds anything you can achieve with colocated or self-managed servers.

However, despite claims of reliablity, few cloud vendors have tight SLAs (service level agreements) that promise controlled downtime or offer rebates for excess downtime. Amazon goes the opposite direction and doesn't offer any uptime guarantees, even cautioning users that their instance (or server) can disappear at any time and that they should plan accordingly. AppLogic-based clouds, provided by companies such as ENKI, are capable of offering better guarantees of uptime because of its inherent self-healing capabilities that can enable 3-4 nines of uptime. (The exact number depends on how the AppLogic system is set up and administered, which affects the time needed for the system to heal itself.) However, any cloud computing system, even even those based on AppLogic or similar technologies, can experience unplanned downtime for a variety of reasons, including the common culprit of human error. While I believe it is possible to produce a cloud computing service that exceeds 4-9s of uptime, the costs would be so high that few would buy it when they compared the price to the average cloud offering.

When you're purchasing cloud computing, it makes sense to look at the SLA of the vendor as well as the reliability of the underlying technology. But if your needs for uptime exceed that which the vendors and their technology can offer, there are time-honored techniques for improving it, most of which involve doubling the amount of computing nodes in your application. There's an old adage that each additional "9" of uptime you get doubles your cost, and that's because you need backup systems that are in place to take over if the primaries fail. This involves creating a system architecture for your application that allows for either active/passive failover (meaning that the backup nodes are running but not doing anything) or active/active failover (meaning that the backup nodes are normally providing application computing capability).

These solutions can be implemented in any cloud technology but they always require extra design and configuration effort for your application, and they should be tested rigorously to make sure they will work when the chips are down. Failover solutions are generally less expensive to implement in the Cloud because of the on-demand or pay-as-you go nature of cloud services, which means that you can easily size the backup server nodes to meet your needs and save on computing resources.

An important component of reliability is a good backup strategy. With cloud computing systems like AppLogic offering highly reliable storage as part of the package, many customers are tempted to skip backup. But data loss and the resulting unplanned downtime can result not just from failures in the cloud platform, but also software bugs, human error, or malfeasance such as hacking. If you don't have a backup, you'll be down a long time - and this applies equally to cloud and non-cloud solutions. The advantages of cloud solutions is that there is usually an inexpensive and large storage facility coupled with the cloud computing offering which gives you a convenient place to store your backups.

For the truly fanatical, backing up your data from one cloud vendor to another provides that extra measure of security. It pays to think through your backup strategy because most of today's backup software packages or remote backup services were designed for physical servers and not virtual environments having many virtual servers such as you might find in the cloud. This can mean very high software costs for doing backup if your backup software charges on a "per server" basis and your application is spread across many instances. If your cloud vendor has a backup offering, usually they have found a way to make backup affordable even if your application consists of many compute instances.

Another aspect of reliability that often escapes cloud computing customers new to the world of computing services is monitoring. It's very hard to react to unplanned downtime if you don't know your system is down. It's also hard to avoid unplanned downtime if you don't know you're about to run out of disk space or memory, or perhaps your application is complaining about data corruption. A remote monitoring service can scan your servers in the cloud on a regular basis for faults, application problems, or even measure the performance of your application (like how long it takes to buy a widget in your web store) and report to you if anything is out of the ordinary. I say "service" because if you were to install your own monitoring server into your cloud and the cloud went down, so would your monitoring! At ENKI, we solve this problem by having our monitoring service hosted in a separate data center and under a different software environment than our primary cloud hosting service.

The last aspect of reliability is security. However, that would require another entire article to cover, since security in the cloud is a complex and relatively new topic.

To sum up, the Cloud offers some enticing advantages with respect to reliability, perhaps the largest of which is that you can give your data center operations responsibility to someone who theoretically can do a much better job at a lower cost than you can. However, to get very good reliability, you must still apply traditional approaches of redundancy and observability that have been used in physical data centers for decades - or, you have to find a cloud computing services provider that can implement them for you.