100% Uptime Anybody?

If you look at any web hosting service provider's site now you're bound to see a promise for “99.9% uptime”, “99.99” uptime", “99.999% uptime”. The hosting provider I used to use once proudly proclaimed “100% Infrastructure Availability” and “Zero Downtime Hosting™”.

Or course, given an unreliable network between the client and the server which is beyond the control of any individual company, all these claims should be treated with some degree of scepticism. But they do raise the interesting question, how much uptime do you actually need? Or put another way, how much will downtime cost you?

Let's look at some example companies and consider what their uptime needs are.

The companies with the highest requirements are the search engines and directories such as Google and Yahoo. What happens if a potential user types their address in a browser and gets no response? Do they wait for the service to come back? Unlikely. They just go to another search engine or directory to perform their search. Both sites are financed by pay-per-click or cost-per-impression advertising. If the site isn't there, the clicks or impressions are permanently lost.

Google received advertising revenue of about $796 million for the quarter ended 30 September 2004. Let's assume that Google traffic is uniform over a 24 hour period. There are 92 days in the quarter, giving ad revenue of $8.65 million per day, or about $100 per second. A one hour outage costs $360,416. A 99.99% uptime for Google corresponds to a 0.01% downtime, or a revenue loss of about $79,600 per quarter.

Assuming “Marketing Services” is how Yahoo describes advertising revenue on its financial statements, it made about $765 million in a similar period. Making the same uniform traffic assumption, ad revenue is about $96 per second. A one hour outage costs $346,647. A 99.9% uptime (or 0.01% downtime) is a revenue loss of $76,500 per quarter. (Note that Yahoo also made $37.4 million in listing fees and $104 million in "fees". Yahoo would continue to make these revenues unless their site was down sufficiently to damage their reputation.)

Financial services companies dealing with markets are probably in the next tier. Imagine that you are watching the price of a company tumble (or rise) while you are trying to sell (or buy) its shares. You probably won't do this too often (more than two or three times) before you are looking for a new company to deal with. For convenience, let's assume that everybody who tries to use the site while it is down immediately switches companies, never to return. A one hour outage therefore costs as many subscribers as try to use the system in that period.

E-Trade made about $180 million from about 2.9 million retail brokerage accounts during the same period, so each customer is worth about $60/quarter in terms of revenue. Rashly assuming that each account tries to log in once every three days, giving about 1 million logins per day. Using a 99.99% uptime, 0.01% of these logins (or 100) will fail. Assume that 50% of these people switch companies. Then our attrition due to uptime will be 50 people per day, or 4,600 per quarter. (Sanity check - E-Trade reported 66,501 accounts closed in the same period, so the turnover of customers is not implausible).

With a current prime rate of 5%, $60 paid quarterly is worth about $4800, so losing 4,600 customers is the same as losing about $22 million in revenue. (This assumes that there is no similar reduction in costs, that customers live forever(!) and never leave for any other reason). We could instead assume we have to regain these 4,600 customers and look instead at acquisition costs. Clearly with these assumptions permanent loss of customers due to downtime could be a serious cost for E-Trade.

Now look at an e-Commerce site such as Amazon. Question: if Amazon is down for an hour, do you go and buy from someone else? I suspect the answer for most people is no. You have a trust relationship with Amazon. Providing their service isn't always terrible, Amazon probably doesn't lose many of the people from one hour's downtime.

What does Amazon lose? Impulse purchases. As much as 70% of today's retail purchases can be classified as "impulse buys". An impulse buy is something that someone buys which they intend to buy before visiting a store. Amazon won't lose impulse buys from the people who tried again later while looking for a specific book (they will still visit Amazon and Amazon has the same possibility to cross-sell to them when they visit). What they will lose is the traffic from affiliates and other referrals during the period from people who weren't strongly motivated to buy, and will therefore continue elsewhere. How much might this be worth?

A sizeable chunk of Amazon's costs are related to cost of sales. If an order isn't placed, these costs don't exist. Therefore let's assume that 70% of Amazon's visitors are impulse buyers, and spread Amazon's gross profit of $355.6 million for the quarter amongst its buyers. That would put $249 million in the hand's of impulse buyers. Assume a 99.99% uptime and $24,900 would be lost over the quarter due to downtime (impulse buyer gross profit is $31 per second).

Let's now look at Intel? How much does an hour's downtime cost Intel? Do you immediately re-design your PC to use AMD chips because you can't get the specs off the Intel website? Hardly. Conceivably you might visit somewhere else while waiting for the Intel site to return and find a replacement you hadn't thought of. But it's unlikely. A well-known vendor such as Intel which I would argue doesn't produce commodity products, has little to lose from the occasional one hour's downtime.

Smaller companies with less brand loyalty will likely lose revenue in direct proportion to the amount of time they are down. If the web site of &ldqut;Fred's Book Emporium” isn't available, the chances of you revisiting it later are slim (unless you know Fred). So any profits Fred would have made in that time are gone.

These calculations are all based upon some very rash assumptions and some gross approximations, but they give some idea of the relative costs of web downtime for different types of companies and some of the figures that can be used in estimating these costs. I've ignored all the secondary effects of downtime (such as adverse publicity), and freely interpreted financial statements. I've also assumed that down means 100% down, something that is unlikely in a well designed system unless it is subject to a similarly well-designed distributed denial-of-service attack.

My website? I'm still trying to work out what my hosting provider's “100% Infrastructure Availability” actually means. Somehow I felt a little happier when they only promised me 99.7% uptime with a “100% satisfaction guarantee”. It seemed more plausible, even if it actually meant up to 26.28 hours of downtime a year.

They may be good, but nobody's 100% good.