Monitoring Production Services at Amazon

Practical tips on how to make sure your alarms don't cry wolf, change your instrument to measure different "cost" of queries separately, how to measure client errors when they're actually your fault, do distributed tracing, and avoid being misled by single percentile aggregates.