gistfile1.txt

gotascii # question about terminology/definitions, does throughput impact 
availability? i.e. if a system is at maximum throughput and some users 
aren't able to submit requests is it still "available" as far as 
"highly-available" is concerned? possibly that is something an SLA 
would clarify? "available" would need to be defined in terms of 
throughput...just trying to wrap my head around some of this heh

gotascii # basically, i think this talk here 
http://www.royans.net/arch/talk-on-database-scalability/ 
and this blog post 
http://www.royans.net/arch/what-is-scalability/ 
confused me a bit as far as scalability goes specifically statements 
like "Scalability is not improving latency, but increasing throughput" 
it seems to me like throughput is just one  aspect of a system you might 
want to scale

benblack # true

gotascii # availability, storage capacity...a long list of other 
system properties... something like "If you can’t figure out how
to improve performance while scaling out, its okay." doesn't seem true 
if you are trying to scale out your performance

benblack # i think you'll find you can't scale out latency
only throughput

gotascii # does that make sense? can you scale out performance? 
something like, i want to be able to add another machine and make my 
some operation execute faster

benblack # scalability and scale out are not congruent

gotascii # you can scale up latency though, is that right?


benblack # right

gotascii # like theres no way to add another node to the cluster 
and make a process execute faster?

benblack # if you are overloaded, maybe

gotascii # what about something like a map reduce query?

benblack # what about it? latency _per operation_ unchanged
latency for entire mr job reduced...because you have more throughput

gotascii # ok makes sense, i wasnt understanding the granularity of "operation"
throughput of each operation is increased so the total job time,
job meaning a batch of operations, decreases
that is tricky because people talk about throughput in terms of
how many users can access the system at once

benblack # that is not throughput

gotascii # when a user might be issuing a job that has many operations...i 
do see now that those things are equivalent

benblack # throughput is operations per unit time
latency is time per operation

gotascii # awesome. ok that helps A LOT

gotascii # so going back to this statement "Scalability is not improving latency, 
but increasing throughput" is correct in the first half, and too narrow in the second half?
scalability is not improving latency, but about increasing _some scalable system property_
one of which is throughput thanks for explaining than ben, much appreciated