Lies, Damn Lies, and Metrics
summary
Averages routinely lie: dashboards that show only means obscure bimodal traffic and hide one slow server among fifteen; New Relic defaults to averages even though raw data storage exists. Useful, specific practices include graphing median and percentiles (95th, 99th, 99.9th), breaking out per-server series to spot and replace a single bad node, and estimating runtime pauses by running a measure sleep(1) loop to detect GC or VM lag. Benchmarked VMs (Netflix style) expose up to 5x instance variance and Amazon reports 100 ms extra latency costing 1% revenue; use Pingdom Runscope ThousandEyes for end-to-end timing.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:21Metrics
- 2:31Metrics Everywhere
- 3:27Averages
- 5:57The Bell Curve
- 10:25The Median
- 15:50Aggregation
- 17:52Visualizing Data
- 20:05Aggregate Alerts
- 20:59Servers
- 25:56Virtual Machines
- 30:13Routing
