The Origins and Future of Distributed Computing and Clouds - The Practice of Cloud System Administration - page 453

way was not well understood—the traditional model of increased sales yielding a lower
cost per unit was assumed to still hold true. In addition, startup valuations were made in a
rather strange manner in those days. If a startup was achieving a small profit, it got a low
valuation. However, if it was showing a loss, it got a high valuation. Moreover, the larger
the loss, the higher the valuation. When these startups were floated on the stock market,
their stock prices went through the roof, and the investors made substantial profits, even if
the business model made no sense.
The saying at the time was “We may lose money on every sale, but we'll make it up in
volume.” Behind the scenes was the thought that if the company could corner the market,
itcouldraisepriceslater.However,thatassumptionfailedtotakeintoaccountthatthenext
startupwouldimmediately undercutthehigherpriceswithitsownloss-making corner-the-
market scheme.
This gold rush mentality kept the economy buzzing for quite a while. Then in 2000 the
bubble burst. The house of cards collapsed. This approach to scaling web-based services
was unsustainable. If the bubble had not burst due to issues of investment and cash flow, it
would have failed due to the bad economics of the technology being used.
High Availability
With the advent of the web, the user base changed from known internal company employ-
ees with predictable, cyclic availability (the 9-to-5 business day) requirements and access
schedules, to unknown external Internet users who required constant access. This change
createdtheneedforhigherreliability.Thefirstapproachtomeetingtheseavailabilitygoals
wastobuymoreexpensive hardware withbuilt-in higheravailability—for example, RAID
and multiple CPUs.
But even the most reliable system fails occasionally. Traditional options for critical sys-
tems were to have a service contract with a four-hour turnaround for replacement parts, or
to purchase spare parts to be stored near the machine. Either way, some downtime would
be required in the event of a hardware failure.
There was also the issue of downtime to perform software upgrades. Applying the in-
ternal corporate approach of notifying the user base, it became common practice for web
sites to pre-announce such downtime. There would be a warning that “This site will be
down Saturday from noon to 5 PM PST for an upgrade.” Regular users could plan around
this, but new or occasional customers would be caught unaware, as there was no other way
to notify them. Advertising upgrades in advance could also lead to adverse headlines if the
upgrade went poorly, as some people would watch to see how the upgrade went and to re-
port on the new service.