Capacity Planning - The Practice of Cloud System Administration - page 373

18.2.3 Identifying Your Core Drivers
Core drivers are factors that strongly drive demand for a primary resource. They typically
include values such as MAU, QPS, the size of the corpus, or other high-level metrics that
represent well the factors that generate traffic or load on the service. These metrics also of-
ten have meaningful business implications, with links to sources of revenue, for example.
Asite may have millions ofregistered users, butit will typically have many fewer users
who are active. For example, many people sign up for accounts and use the service a few
times, but never return. Counting these users in your planning can be misleading. Many
people register with social networking sites, but rarely use their accounts. Some people
who are registered with online shopping sites may use the service only before birthdays
and gift-buying holidays. A more accurate representation of users may be how many were
active in the last 7 or 30 days. The number of users in the last 7 days is often called 7-day
actives (7DA) , while the term weekly active users (WAU) is used to indicate how many
were active in a specific calendar week. Likewise, 30-day actives (30DA) measures the
numberofusersinthelast30days,withtheterm monthly active users (MAU) usedifthe
measurementwasboundedbyaspecificcalendarmonth.Thesemeasurementsoftenreflect
usage much more accurately than the number of registered users.
Formetricssuchasactiveusersthathaveatimecomponent,differentvaluesofthattime
component may be appropriate to use in capacity planning for different services. For ex-
ample, for some services monthly active users may be the appropriate core driver, whereas
for another service minutely active users may be a better indicator to use in capacity plan-
ning. For highly transactional systems that are driven by active users, smaller time scales
like minutely active users may be appropriate. For storage-bound services that are driven
by users, the total number of registered users (or total user corpus number) may be more
appropriate.
The capacity model depicts the relationship between the core driver and the primary
resource. For a given service, the capacity model expresses how changes in the core driver
affect that service's need for its core driver.
Once you have identified the core drivers, and have determined the effect that each one
has on each of the primary and secondary resources, you can quantify the effect that each
will have on your requirements for ancillary resources, such as servers. You can also ana-
lyze whether your ancillary resources are well balanced. If servers run out of CPU cycles
long before they run out of RAM, then it may be more cost-effective to order servers with
less RAM or more or faster CPUs, for example. Or it may be possible to rework the ser-
vice by taking advantage of the extra RAM or making it less CPU-intensive. Similarly, if
yourswitches runoutofportslongbeforebackplane bandwidth,perhapsadifferentswitch
model would be more appropriate.