Complex workflows lead to high tail latencies
Today, techniques used in an ad-hoc per-stage manner
The main work is
Framework to find per-stage parameters -〉
Minimize end-to-end tail latency of workflows with constraint on costs
1.Differest stages show dif gains from the same latency reduction technique.
2.Composing costs across stages is non-trivial
3.Decomposition of end-to-end latency percentiles into individual components is hard
The goal of Kwiken is to improve the latency of requestresponse workﬂows, especially on the higher percentiles. We pick the variance of latency as the metric to minimize because doing so will speed-up all of the tail requests; in that sense, it is more robust than minimizing a particular quantile.
solutions to the challenges:
1.Minimize variance in workflow latency
2.E2E cost decomposed into per-stage costs
3.Construct per-stage variance-response curves
1.Benefits of reissues
2.Combining reissues with early termination
More details in: