Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Introduction of Skyline

Introduction of Skyline

Avatar for Takumi Sakamoto

Takumi Sakamoto

July 12, 2013
Tweet

More Decks by Takumi Sakamoto

Other Decks in Programming

Transcript

  1. My LT Talk at Velocity • 3 Popular Ops Tools

    in Japan • serverspec, growthforecast, fluentd • Uploaded on youtube :-( • http://www.youtube.com/watch? v=bRYuBQyG5Sw
  2. The Most Interesting Talk • Avoiding Performance Regression at Twitter

    • fight against perf regression in an automated fashion • http://ameblo.jp/principia-ca/ entry-11561132297.html
  3. Background • Etsy deploys their app 30+ times per day

    • Optimize for quick recovery by anticipating problems, instead of fearing human error • Can’t fix what you don’t measure! • If it moves, graph it
  4. Too Many Graphs • 250,000+ dashboards • If a graph

    spikes and no one is watching, does it really spike? • There are things we do not know we don’t know.
  5. Kale • Skyline • Detect unknown anomalies • Oculus •

    Detect unknown correlations http://codeascraft.com/2013/06/11/introducing-kale/
  6. Architecture Horizon Horizon Horizon Redis Analyzer Manager Analyzer Analyzer Analyzer

    Assign Metrics Keys Get Metrics Keys Fetch Timeseries Metrics
  7. Horizon • Listeners • Receiving metrics & store them to

    queue • Workers • Inserting metrics to redis with Messagepack • Roombas • Purging metrics in Redis at a regular interval
  8. Analyzer • Assigning Redis keys to each process process •

    decode from Messagepack • run the detection algorithm
  9. How to Detect Anomalies? • Consensus model • If the

    majority of algorithms agree, the metric will be classified as anomalous • Use your own algorithm for each application • as long as you return a boolean, you can add any sort of algorithm you like
  10. Basic Algorithm • A metric is anomalous if its latest

    datapoint is over three standard deviations above its moving average
  11. Conclusion • Etsy monitors 250,000+ graphs • If a graph

    spikes and no one is watching, does it really spike? • Skyline • detects unknown anomalies • consensus with any algorithms you like