このページは http://www.slideshare.net/kuma0177/velocity-ny-2014v5-39160794 の内容を掲載しています。
Frequent deployments, large set of in-flight A/B tests, new product launches etc. directly impact...
Frequent deployments, large set of in-flight A/B tests, new product launches etc. directly impact the profile of application metrics as well as system metrics. Specifically, the above can induce sudden breakouts – which manifest themselves as a mean-shift or a rampup (these are different from an anomaly) – in the time series of a given metric. Further, the profile on the incoming traffic may also experience a breakout due to a variety of reasons such as, but not limited to, roll out of a new feature or roll out for a new platform; this in turn results in breakouts in application and/or system metrics.
Breakouts can potentially impact performance of the corresponding service and consequently impact the end user experience. To alleviate the impact of breakouts – in other words, preventing user experience from ‘Breaking Bad’ – we developed statistically rigorous techniques to automatically detect breakouts in a timely fashion. The breakouts detected are used to guide capacity planning. In particular, there are two scenarios:
Positive breakout: Depending on the magnitude, deploy additionally capacity
Negative breakout: Depending on the magnitude, scale down the current capacity
We shall walk the audience through how the techniques are being at Twitter using REAL data.