intro.tex

Having system traces is useful in order to get high level explanations of system outcomes. 

Not all systems have traces - due to instrumentation and performance overloads. 

All systems generate logs of varying levels of details. 

Can we infer potential causal relationships from logs?

Decompose problem into:
\begin{itemize}
\item Identifying events 
\item Establishing dependencies amongst events
\end{itemize}

If we assume globally unique identifiers at the request level across nodes, the problem of identifying events becomes trivial, but the second problem remains.

If we make the assumption that we are able to run a given workload on a system repeatedly, the key intuition is that the true dependencies of any given event will be present in addition to others. 

We show that we can approach the true set of relationships with a combination of running the system repeatedly and randomly injecting delay on one or more nodes.