Information Technology Reference
In-Depth Information
• Splitting by type of user
Case Study: Separating Traffic Types at ServerFault.com
ServerFault.com is a question-and-answer web site for system administrators.
When displaying a question (and its answers) to a logged-in user, the page is aug-
mentedandcustomizedfortheparticularperson.Anonymoususers(userswhoare
not logged in) all see the same generic page.
To scale the system on the y -axis, the two ways to generate the same page were
split out. The anonymous pages are now handled by a different system that gener-
ates the HTML once and caches it for future requests. Since the vast majority of
queries are from anonymous users, this division greatly improved performance.
Anonymous page views are subdivided one additional way. Search engines
suchasGoogleandBingcrawleverypageatServerfault.comlookingfornewcon-
tent. Since this crawling hits every page, it might potentially overload the service,
due to both the volume of requests and the fact that hitting every page in order ex-
hausts the cache. Both factors make performance suffer for other users. Therefore
requests from web crawlers are sent to a dedicated pool of replicas. These replicas
are configured not to cache the HTML pages that are generated. Because the pool
is separate, if the crawlers overload it, regular users will not be affected.
5.3.3 z : Lookup-Oriented Split
A lookup-oriented split scales a system by splitting the data into identifiable segments,
eachofwhichisgivendedicatedresources. z -axisscalingissimilarto y -axisscalingexcept
that it divides the data instead of the processing.
A simple example of this is to divide, or segment , a database by date. If the database is
an accumulation of data, such as log data, one can start a new database server every time
thecurrentonefillsup.Theremaybeadatabasefor2013data,2014data,andsoon.Quer-
ies that involve a single year go to the appropriate database server. Queries that span years
are sent to all the appropriate database servers and the responses are combined. If a partic-
ular year's database is accessed so often that it becomes overloaded, it can be scaled using
the x -axistechniqueofreplication.Sincenonewdataiswrittentopastyears'servers,most
servers can be simple read-only replicas.
Search WWH ::




Custom Search