Stated that the only way to analyze production performance is using live production systems. In this paper, Google looks at Gmail performance
Users are partitioned randomly, one test and one control
Must make assumptions clear when using statistics and scrutinize the experiment to enure that assumptions (like event independence) are not violated
Before determining how to make an app faster, we must understand why it's slow or expensive
temporal context of a request can be used to find out what else was running at the time
operation context tells you how a request ran (what parameters, maybe what other calls are madE)
We need a method for tracing the context of these events. it seems temporal context may be as easy as timestamps when estimating? More accurately, we might require coordination between layers