Data Center Overload(nytimes.com)
nytimes.com
Data Center Overload
http://www.nytimes.com/2009/06/14/magazine/14search-t.html?_r=1&hp
12 comments
If responses times are as low as 50 ms, it usually isn't a problem when it become 150 ms: people don't notice Similarly, if response times are 2 seconds, it doesn't matter that it becomes 2.1 seconds. However, when the response time is 400 ms, then it is a noticable difference when it becomes 500 ms. People notice that: the site feels more sluggish.
Now whenever I shop at Amazon, I browse around for a bit. It's not that a more sluggish site would cause me to abandon a purchase midway in the purchasing process, but it's rather that I might get annoyed at the sluggishness while browsing and postpone my purchase or take it elsewhere, where I can browse the available products at my leisure.
I think that is what they mean, but if the number is just based on assumptions about who intends to make a purchase, that also means the number is highly disputable. However, I think Amazon has enough data available to simply correlate response times with sales rates over a statistically signifcant period and statistically signifcant amount of sales. In that case, the number is pretty undisputable: if anyone has this kind of data, it's Amazon.
Now whenever I shop at Amazon, I browse around for a bit. It's not that a more sluggish site would cause me to abandon a purchase midway in the purchasing process, but it's rather that I might get annoyed at the sluggishness while browsing and postpone my purchase or take it elsewhere, where I can browse the available products at my leisure.
I think that is what they mean, but if the number is just based on assumptions about who intends to make a purchase, that also means the number is highly disputable. However, I think Amazon has enough data available to simply correlate response times with sales rates over a statistically signifcant period and statistically signifcant amount of sales. In that case, the number is pretty undisputable: if anyone has this kind of data, it's Amazon.
Basically, you just run the time to sale and how much cash you're processing and correlate that to time. True, it's not exactly causal, but Google found the same with I believe the quote is "1 second delay in search means $1M lost in revenue"
:)
Thanks (all) for responding.
I guess I never realised how much of amazons purchases must be spur-of-the-moment. (NOTE: I order stuff from amazon - but usually in groups as I'm shipping it to australia - so all of my stuff is planned - i'm going there to buy something)
But - if anyone has the metrics to prove a figure like that - it would be amazon.
But - if anyone has the metrics to prove a figure like that - it would be amazon.
Edit: whoops. this is supposed to be a response to the first comment.
I would guess that the 1% figure is the average loss for each 100ms of latency. So if it took 5 seconds longer to load each page, 50% less people would ultimately make a purchase.
Which is still a very high number, and makes me wonder if most of Amazon's customers are impulse shoppers, or whether these numbers were just fabricated to help sell AWS.
Either way, it's not a very meaningful without knowing the context and overall probability distribution of lost sales vs. latency.
I would guess that the 1% figure is the average loss for each 100ms of latency. So if it took 5 seconds longer to load each page, 50% less people would ultimately make a purchase.
Which is still a very high number, and makes me wonder if most of Amazon's customers are impulse shoppers, or whether these numbers were just fabricated to help sell AWS.
Either way, it's not a very meaningful without knowing the context and overall probability distribution of lost sales vs. latency.
Printable Link: http://www.nytimes.com/2009/06/14/magazine/14search-t.html?_...
EDIT: Doesn't work unless you have the correct Referer...
EDIT: Doesn't work unless you have the correct Referer...
I think the last point was the most interesting.. data substations. Every suburb/town could have a data substation that all connect to form the cloud. computing as a service, exactly the same as electricity. shit is about is get interesting.
Isn't the logical conclusion that every one's PC inside their own home forms part of the cloud?
It reminds me of a joke business plan developed over some beers: Provide free electric heaters for old people to run in winter, which were actually PCs connected to a phoneline doing data-intensive grid computing :-)
It reminds me of a joke business plan developed over some beers: Provide free electric heaters for old people to run in winter, which were actually PCs connected to a phoneline doing data-intensive grid computing :-)
And you thought the NIMBYness of cell towers was bad.
"Google search is not without environmental consequence — 0.2 grams of CO2 per search, the company claims"
I wonder how that compares to the CO2 cost of driving to the library to look up the same bit of information for the average person?
I wonder how that compares to the CO2 cost of driving to the library to look up the same bit of information for the average person?
It's a fallacious analogy. People didn't use to drive to the library to look up tidbits of info - search engines create new demand. Also, use of libraries has gone up greatly since they began installing computers for patrons. So driving to libraries is increasing even as Net use is increasing.
Can any of the bright minds here shed some light on this? 100ms doesn't seem at all enough to have that big an impact. (especially for something like amazon - where there is likely some level of intent and definate trust that the site is valuable). I just can't see how a .1sec delay could affect so many sales?